Welcome to AnalyZ Solutions
AnalyZ Solutions covers the full research workflow: from loading and cleaning data, through statistical analysis and visualisation, to qualitative thematic coding. This guide walks you through every feature in detail.
What's in the platform
Typical workflows
Quantitative workflow
Qualitative workflow
Mixed-methods workflow
Account & Security
Creating an account
Creating an account is free and needs no payment details. You can sign up with an email address and password, or continue with a Google or GitHub account.
! or #). A checklist on the screen ticks off each rule as you meet it. Type the password a second time to confirm it.Signing in
On the sign-in screen you can use email and password, or the Google or GitHub buttons.
- Human verification. After a failed sign-in attempt, a human-verification check appears under the form. Complete it before trying again.
- Two-step verification code. If you have turned on two-step verification (next section), a second screen asks for the 6-digit code from your authenticator app after your password or Google or GitHub sign-in. Enter it and click Verify. Cancel and sign out returns you to the start.
- Cookie Settings. The link at the bottom of the sign-in screen reopens the cookie preferences banner at any time.
Two-step verification
Two-step verification adds a second lock to your account. After your password, AnalyZ asks for a 6-digit code that changes every 30 seconds and appears only in an app on your phone. Even if someone learns your password, they cannot sign in without the code. It is free.
Turning it on
Good to know
- First sign-in prompt. If you have not turned it on, a one-time Protect your account prompt appears after you sign in, offering Set up now or Skip for now. If you skip, the prompt returns after about 30 days.
- Add another device. Once it is on, Add another device (recommended as a backup) lets you register a second authenticator, so losing one phone does not lock you out.
- Turn it off. In the same tab, click Turn off beside the device and confirm.
Forgotten password
Click Forgot password? on the sign-in screen, enter your email and click Send reset link. For privacy the screen says the same thing whether or not the address is registered. Open the link in the email, choose a new password that meets the same rules as at sign-up, and sign in again. Reset links expire after a short time, so use them soon after they arrive.
Signing out
Click Sign out at the far right of the top bar (next to your email address). You are signed out on this device and returned to the landing page. Save a session file first if you want to keep your work (see Saving Sessions).
Account Settings
Open it with the Settings button in the top bar of the home screen. It has four tabs.
| Tab | What you can do there |
|---|---|
| Subscription | See your current plan (Free, Starter, Pro or Enterprise) and its renewal date, or the end date if you have cancelled. Usage this month shows bars for Sense Making interpretations, Dashboard interpretations, Data Canvas story generations and Data Canvas slide refinements. On the Free plan you can subscribe to Starter or Pro. On Starter you can upgrade to Pro. On a paid plan, Cancel plan stops renewal. |
| Profile | Your email is shown for reference and cannot be changed here. You can set your display name, organisation or institution, role (researcher, evaluator, academic, student, NGO staff, consultant or other) and country. Click Save profile. |
| Activity | This month's usage as simple counts, for the same four items. Usage resets on the 1st of each month. |
| Authentication | Turn two-step verification on or off and manage your authenticator devices. See the section above. |
Plans and monthly allowances
| Allowance per month | Free | Starter | Pro |
|---|---|---|---|
| Sense Making interpretations | 1 | 10 | 50 |
| Dashboard interpretations | 1 | 10 | 50 |
| Data Canvas story generations | 2 | 5 | 15 |
| Data Canvas slide refinements | 5 | 15 | 50 |
Starter is ₹200 a month and Pro is ₹400 a month. Allowances reset on the 1st of each month. Publishing public links for dashboards and Data Canvas stories is not available on the Free plan. The analysis modules themselves are available to every registered account.
Cancelling a plan
Click Cancel plan on the Subscription tab and confirm. The plan will not renew. It stays active until the end of the period you have already paid for, then returns to Free. Partial refunds are not issued. For billing questions, email support@analyz.solutions.
Data security
| What | How it works |
|---|---|
| Data in transit | All communication between your browser and AnalyZ is encrypted with TLS (HTTPS). Data is never sent over plain HTTP. |
| Data in session | Uploaded files and analysis results are held in browser memory only for the duration of your session. By default they're discarded the moment you close the tab or log out. |
| Local device save (opt-in) | You can turn on Save on this device in the sidebar to keep a working copy in this browser's local storage across closes. This is off by default, stored only on your device (never our servers), and expires automatically after 24 hours. |
| Server storage | AnalyZ does not store your uploaded data on any server. No copies are kept after your session ends. |
| Saved sessions (.analyz.json files) | When you save a session, the file is downloaded directly to your own device. AnalyZ does not keep a copy server-side. |
| Account credentials | Passwords are never stored in plain text: only a secure hash is kept. AnalyZ staff cannot see your password. |
| Authentication tokens | Login sessions use short-lived JWT tokens that expire automatically. Logging out invalidates the token immediately. |
Guest mode vs. registered account
| Feature | Guest | Registered |
|---|---|---|
| Data limit | 50 rows / 1 MB | Unlimited |
| Qualitative files | 1 file / 1 MB | No limit |
| Session saving (.analyz.json) | ❌ | ✅ |
| Download data | ❌ | ✅ (Excel, JSON, Stata, SPSS) |
| All modules | ✅ | ✅ |
| Account required | ❌ | ✅ (free) |
Loading Data
Everything starts in the sidebar on the left. At the top it asks What type of data do you want to analyze? with two buttons, Quantitative (spreadsheets and survey files) and Qualitative (interview and focus group transcripts). Choose the one that matches your data, then follow the matching steps below.
Load a quantitative file
What each file type brings with it
| File type | Variable labels and value labels | Notes |
|---|---|---|
| CSV | Not included | Add labels afterwards in Data View, using Edit variable labels, or by right-clicking a column header and choosing Variable & Value Labels. |
| Excel (.xlsx, .xls) | Optional, from a second sheet | An Import Excel data window opens first (see below), where you choose the sheet, the row with variable names and, optionally, a sheet of labels. |
| Stata (.dta) | Imported | A Select variables to load window opens first (see below). |
| SPSS (.sav) | Imported | Same variable picker as Stata. The file is read by AnalyZ's server, so see the privacy note below. |
| JSON | Not included | Should be a list of records, one per row. |
| .analyz.json | Included | Restores only the dataset half of the file here. To restore everything use Restore Session instead. |
Excel files: the Import Excel data window
Every Excel file opens a window titled Import Excel data, even when it has only one sheet. It shows the file name and, for workbooks with several sheets, how many there are.
Stata and SPSS files: choosing variables
Survey files often hold hundreds of variables. After you choose a .dta or .sav file, a window titled Select variables to load opens.
Connect to a survey platform (Connect API)
Use this to pull data straight from a survey tool instead of exporting a file. It needs an account (it is greyed out for guests with the text Connect API, sign up to unlock).
KoBo Toolbox in detail
Try a demo dataset
Under Or try a demo dataset: in the sidebar there are two buttons, Test data without labels and Test data with labels. Both load a 300-row example survey with age, income, gender, education, score, region and employment. The second one already has labels, so it is the better choice for following this guide.
Load interview transcripts
No transcripts yet? Under Or try a demo transcript: click Try dummy transcript. It loads a sample interview so you can practise coding.
To save transcripts, click Download Transcripts in the sidebar, then choose Word document (.docx), Plain text (.txt) or Markdown (.md). Several transcripts download together as a .zip.
Clear, undo and redo
| Button (sidebar) | What it does |
|---|---|
| Clear Dataset | Removes the loaded dataset and its results from the app. |
| Clear Transcripts | Removes all loaded transcripts. |
| ✕ Clear all | Removes the demo data or everything loaded, depending on where it shows. |
| Undo and Redo | Step backwards or forwards through changes you made to the dataset in Data Management or Data View, such as recodes, deletions or sorting. |
Saving & Restoring Sessions
Save a session file
The file holds: your dataset, variable and value labels, variable types, survey design, transcripts, codebook, coded segments, notes, connected API sources (without credentials) and saved dashboard charts.
Restore a session
Keep your work on this device (optional)
By default everything is lost when you close the tab. You can ask AnalyZ to keep a working copy in your browser.
This copy is stored only in your browser and is not sent to AnalyZ's servers.
Download your data
Download Data in the sidebar saves the current dataset, including any recodes and new variables. Choose a format: Excel (.xlsx), JSON (.json), Stata (.dta) or SPSS (.sav). There is no CSV option here. Stata and SPSS files include variable and value labels. These two are prepared by AnalyZ's server, which handles the data only for the duration of the request.
Choosing an Analysis
The right method depends on two things: what you are asking, and what kind of variables you have (see Statistics Basics if the types are unfamiliar). Use the tables below, then follow the link to the full explanation. You can also open Guided Analysis from the top bar and answer a few questions to be pointed to a method.
I want to describe my data
| My question | Use | Where |
|---|---|---|
| Is my data clean enough to analyse? | Data Quality | Descriptive Analysis |
| How many people gave each answer? | Frequency | Descriptive Analysis |
| How do two categorical questions line up? | Cross-Tabulate | Descriptive Analysis |
| I need a standard 'Table 1' for a report | Table | Descriptive Analysis |
| What is the typical value and how spread out is it? | Summary Statistics | Descriptive Analysis |
I want to compare groups or time points
Work from left to right: how many groups, whether they are the same people measured repeatedly, and whether the numbers are roughly bell-shaped. If you are unsure about the shape, the rank-based test is the safer option.
| Groups being compared | Numeric outcome, roughly bell-shaped | Ordinal, skewed or small sample |
|---|---|---|
| One group against a fixed value | One-sample t-test | No direct equivalent here. Report the median and use Frequency. |
| Two separate groups | Independent t-test | Mann-Whitney U |
| Same people measured twice | Paired t-test | Wilcoxon Signed-Rank |
| Three or more separate groups | One-way ANOVA | Kruskal-Wallis |
| Two factors at once (for example sex and region) | Two-way ANOVA | No direct equivalent. Use regression with both variables. |
| Same people measured three or more times | Repeated-Measures ANOVA | Friedman Test |
For percentages instead of averages: use the Chi-square test to compare a category breakdown between groups, the Z-test for one proportion to compare a percentage with a target, and the Z-test for two proportions to compare two percentages.
I want to know whether two things are related
| The two variables are | Use |
|---|---|
| Both categorical | Chi-square test (or Cross-Tabulate for a quick look) |
| Both numeric, roughly straight-line, no extreme outliers | Pearson Correlation |
| Ordinal, skewed, or with outliers | Spearman Correlation |
| One numeric, one categorical | Compare groups (see above) |
I want to predict or explain an outcome
Regression lets you look at several predictors at once, so each one's effect is shown while the others are held steady. The method is chosen by the type of outcome.
| My outcome is | Example | Use |
|---|---|---|
| A number | Income, test score, weight | Linear Regression |
| Yes or no | Attended, employed, recovered | Logistic Regression (or Probit) |
| Three or more unordered categories | Which programme was chosen | Multinomial Regression |
| Ordered categories | Satisfaction from 1 to 5, severity | Ordinal Regression |
| A count of events | Number of visits or incidents | Poisson or Negative Binomial Regression |
| Any of the above, but people are grouped in schools, clinics or villages | Scores of pupils in schools | Multilevel Regression |
| A number, and you want to see how much each set of predictors adds | Demographics first, then programme variables | Hierarchical Regression |
I want to know whether a programme caused a change
| Your situation | Use |
|---|---|
| One round of data, no random assignment, rich background information on everyone | Propensity Score Matching |
| Programme and comparison groups measured before and after | Difference-in-Differences |
| Eligibility decided by a cut-off score, age or income | Regression Discontinuity Design |
| Take-up is tied up with unmeasured factors and you have a credible instrument | Instrumental Variables |
| You want to know how an effect happens, through which middle step | Mediation or Counterfactual Mediation |
I want to check a questionnaire scale
| My question | Use |
|---|---|
| Do these items hang together as one scale? | Cronbach's Alpha |
| How many underlying factors do my items form? (I have no theory yet) | Exploratory Factor Analysis |
| I just want fewer variables or a combined score | Principal Component Analysis |
| Does my planned structure of items and factors fit the data? | Confirmatory Factor Analysis |
| Does the scale work the same way in every group? | Compare Groups (measurement invariance) |
I am planning a study
| My question | Use |
|---|---|
| How many people do I need to survey? | Sample Size Estimation |
| Will my planned sample be big enough to detect a difference? | Power Analysis |
| I have a list of units and need to pick the sample | Draw a Sample |
Statistics Basics
You do not need a statistics background to use AnalyZ. Every result comes with a written interpretation. This page explains the ideas those interpretations rely on, so you can check them against your own judgement and explain them to others. Each method in the guide links back here when it uses one of these ideas.
Sample and population
The population is everyone you want to learn about, for example all households in a district. The sample is the smaller group you actually measured, for example 400 of those households. Almost every test in AnalyZ asks the same underlying question: could what I see in my sample just be luck of who happened to be picked, or does it probably hold for the whole population?
This is why a bigger sample makes results more trustworthy, and why a result from 12 people deserves more caution than the same result from 1,200.
Types of variables
Choosing the right method starts with knowing what kind of variable you have. AnalyZ guesses the type for you (a number column with more than 10 different values is treated as continuous, otherwise as categorical) and you can change it in Data Management, under Variable Types.
| Type | What it is | Examples | Usually analysed with |
|---|---|---|---|
| Continuous (numeric) | A measured number where differences are meaningful | Age, income, test score, weight, time in minutes | Means, correlation, linear regression, t-tests, ANOVA |
| Binary | Two possible answers | Yes/No, male/female, passed/failed, treated/not treated | Percentages, chi-square, logistic regression |
| Nominal | Three or more categories with no natural order | Region, religion, type of programme, occupation | Frequencies, cross-tabulation, chi-square, multinomial regression |
| Ordinal | Categories with a natural order but unequal gaps | Satisfaction from 1 to 5, education level, severity (mild, moderate, severe) | Medians, rank-based tests, Spearman, ordinal regression |
| Count | Whole numbers of events, never negative | Number of clinic visits, number of incidents, children in a household | Poisson or negative binomial regression |
Significance and p-values
Most tests begin by assuming that nothing is going on: the two groups are really the same, the variables are really unrelated, the programme really did nothing. This starting assumption is called the null hypothesis. The test then asks how surprising your data would be if that assumption were true.
The p-value is the answer to that question, as a number between 0 and 1. A small p-value means your data would be very unlikely if nothing were going on, so you take the result seriously. A large p-value means the pattern you see is the kind of thing chance produces easily.
| Marker in the tables | Meaning | How AnalyZ words it |
|---|---|---|
*** | p-value below .001 | significant at the 0.1% significance level |
** | p-value below .01 | significant at the 1% significance level |
* | p-value below .05 | significant at the 5% significance level |
ns | p-value .05 or above | no significant difference or relationship |
The significance level (5% is the usual choice) is the line you draw in advance. Choosing 5% means you accept being fooled by chance about 1 time in 20.
Confidence intervals
A single number from a sample, such as an average of 72, is only an estimate. A 95% confidence interval gives a range, for example 69 to 75, that is built so that it would capture the true population value in 95 out of 100 repeats of the study. Wide intervals mean you know little. Narrow intervals mean the estimate is precise.
Intervals also give you a quick significance check. If the interval for a difference does not include 0, or the interval for a ratio (odds ratio, rate ratio) does not include 1, the result is significant at the 5% level.
Effect sizes
An effect size answers how big, while a p-value answers whether it is distinguishable from chance. AnalyZ reports an effect size alongside most tests. These are the rules of thumb it uses to describe them in words:
| Effect size | Used with | Small | Medium | Large |
|---|---|---|---|---|
| Cohen's d | t-tests (difference in means) | 0.2 | 0.5 | 0.8 |
| Eta-squared and partial eta-squared | ANOVA, Kruskal-Wallis | 0.01 | 0.06 | 0.14 |
| Correlation r or rho | Correlation | 0.1 | 0.3 | 0.5 |
| Effect size r | Mann-Whitney U, Wilcoxon | 0.1 | 0.3 | 0.5 |
| Cramér's V | Chi-square (2 by 2 table; larger tables need smaller cut-offs) | 0.1 | 0.3 | 0.5 |
| R-squared | Regression (share of variation explained) | 0.02 | 0.13 | 0.26 |
| Odds ratio | Logistic regression | 1.5 (or 0.67) | 2.5 (or 0.40) | 4.3 (or 0.23) |
These are conventions, not laws. In some fields a correlation of 0.2 is valuable, and in others 0.7 would be disappointing. Judge size against what is normal in your topic.
Assumptions in plain terms
Every test works properly only when your data behave in certain ways. You do not need to memorise the mathematics, only the four ideas below.
| Assumption | In plain words | Check it by | If it fails |
|---|---|---|---|
| Roughly bell-shaped data (normality) | Most values sit near the middle, with fewer and fewer further away. Matters most for small samples (under about 30 per group). | A histogram in Data Visualization, and the skewness value in Summary Statistics | Use the rank-based alternative under Non-parametric Tests |
| Similar spread in each group (equal variance) | The groups are equally scattered around their averages. | Compare the SDs in the group table, and the Levene p-value on the independent t-test | Interpret with caution, or use Mann-Whitney U |
| Independent observations | One person's answer does not influence another's. Broken by repeated measures on the same person or by people grouped in the same village or school. | Think about how the data were collected | Use a paired or repeated-measures test, or Multilevel Regression |
| Straight-line relationship (linearity) | Each extra unit of X changes Y by about the same amount. | A scatter plot in Data Visualization | Transform the variable, or use Spearman for monotone patterns |
Correlation and causation
A relationship between two variables does not show that one causes the other. Children with bigger shoe sizes read better, but both depend on age. Regression can adjust for other variables you measured, which helps, but only designs that build a fair comparison group can support a claim about cause. That is the purpose of the methods in Effect Analysis (matching, difference-in-differences, regression discontinuity and instrumental variables) and of Counterfactual Mediation in Structural Equation Modelling.
Reading a results page
Every analysis page in AnalyZ follows the same layout, so once you know one you know them all.
| Part | What it gives you |
|---|---|
| Analysis set up | Where you choose variables and options. The Run or Generate button sits at the bottom, next to the Apply survey weight tick box when a survey weight has been set. |
| Tables | The numbers: group statistics, the test statistic, the p-value with its stars, and effect sizes. Each table can be exported. |
| Charts | Where a picture helps (means with error bars, forest plots, scree plots). Charts can be edited and downloaded. |
| Interpretation | A short written reading of the main result in plain language, suitable as a starting point for your report. It names the groups and variables using your labels. |
| Export and share buttons | CSV, Excel and Word downloads, Add to Canvas (a slide for your data story) and Add to Sense Making (sends the result for AI synthesis). |
Glossary
Type in the box to narrow the list. Terms are written the way they appear in AnalyZ tables and interpretations.
| Term | In plain words | Where you meet it |
|---|---|---|
| Alpha (Cronbach's) | A number from 0 to 1 showing how well a set of questionnaire items hang together as one scale. | Scale & Reliability |
| Alpha (significance level) | The cut-off you set for calling a result significant. 0.05 means 5%. | All tests |
| ANOVA | A test that compares averages across three or more groups at once. | Inferential |
| Assumption | A condition your data should roughly meet for a test to be trustworthy. | All tests |
| ATT, ATC, ATE | The programme effect on those treated, on those not treated, or on everyone. | Effect Analysis |
| Average variance extracted (AVE) | How much of an item set's variation is captured by its factor rather than by noise. 0.50 or more is good. | SEM |
| Balance (covariate balance) | Whether two groups look alike on background characteristics. | Propensity Score Matching |
| Bandwidth | How close to a cut-off a case must be to be used in the comparison. | Regression Discontinuity |
| Beta (B, coefficient) | In regression, how much the outcome changes for a one-unit increase in a predictor. | Regression |
| Bonferroni correction | A way of making p-values stricter when many comparisons are made at once. | Non-parametric, RM ANOVA |
| Bootstrap | Re-running an analysis on many random re-draws of your data to see how much the answer wobbles. | Effect Analysis, SEM |
| Categorical variable | A variable whose values are groups or labels. | All |
| CFI and TLI | Fit measures for a factor or structural model. 0.95 or more is good. | SEM |
| Chi-square (χ²) | A measure of how far observed counts are from what would be expected if nothing were going on. | Chi-square, SEM fit |
| Cohen's d | The size of a difference between two means, in standard deviations. | t-tests |
| Confidence interval | A range that likely contains the true population value. | All |
| Confounder | A third factor that influences both the thing you are studying and the outcome, and can fake or hide an effect. | Effect Analysis |
| Correlation (r) | A number from -1 to +1 showing how closely two variables move together. | Correlation |
| Covariate | An extra variable you adjust for in an analysis. | Regression, Effect Analysis |
| Cramér's V | The strength of a link between two categorical variables, from 0 to 1. | Chi-square, Cross-Tabulate |
| Degrees of freedom (df) | A technical count that depends on sample size and number of groups. It is needed to look up the p-value, and you rarely need to interpret it. | All tests |
| Dependent variable | The outcome you are trying to explain or predict. | Regression, ANOVA |
| Design effect (DEFF) | How much larger a sample must be because people were chosen in clusters rather than at random. | Sampling |
| Dummy variable | A 0/1 variable standing for one category. AnalyZ creates these for you. | Regression |
| E-value | How strong an unmeasured confounder would need to be to explain a result away. | Effect Analysis, SEM |
| Effect size | How big a difference or relationship is, whether or not it is significant. | All |
| Eigenvalue | How much of the total variation in your items a component or factor captures. Above 1 is the usual retention rule. | PCA, EFA |
| Endogenous | A variable that is itself influenced by other things in the model, or by unmeasured factors. | Instrumental Variables |
| F statistic | The ratio of differences between groups to differences within groups. Larger means stronger evidence of a group difference. | ANOVA, regression |
| Factor (statistical) | A hidden trait, such as 'anxiety', that is assumed to drive answers to several questions. | EFA, CFA |
| Factor loading | How strongly an item reflects its factor, on a scale where higher is stronger. | EFA, CFA, PCA |
| Heaping | Respondents rounding to favourite numbers, such as ages ending in 0 or 5. | Data Quality |
| Hypothesis (null) | The starting assumption that nothing is going on. | All tests |
| ICC (intraclass correlation) | The share of variation that sits between groups rather than within them. | Multilevel, Sampling |
| Independent variable (predictor) | A variable you use to explain or predict the outcome. | Regression |
| Instrument | A variable that affects take-up of a treatment but affects the outcome only through that treatment. | Instrumental Variables |
| Interaction | When the effect of one variable depends on the level of another. | Two-way ANOVA, regression |
| Intercept | The predicted outcome when all predictors are zero. | Regression |
| KMO | A check that your items share enough common variation for factor analysis. Above 0.60 is acceptable. | PCA, EFA |
| Latent variable | A trait that cannot be measured directly and is inferred from several questions. | SEM |
| Likert scale | A set of ordered answer options such as strongly disagree to strongly agree. | Surveys |
| Margin of error | The plus-or-minus range around a survey estimate. | Sampling |
| Mean, median, mode | The average; the middle value when sorted; the most common value. | Descriptive |
| Mediator | A middle step through which one variable affects another. | SEM |
| Multicollinearity | Predictors that are so closely related that their separate effects cannot be told apart. | Regression |
| Non-parametric test | A test that works on ranks and so does not require bell-shaped data. | Non-parametric Tests |
| Odds ratio (OR) | How many times higher (or lower) the odds of an outcome are for each one-unit increase in a predictor. 1 means no effect. | Logistic regression |
| Outlier | A value far from the rest, which may be an error or a genuine extreme case. | Data Quality |
| p-value | How surprising your data would be if nothing were going on. Small means surprising. | All tests |
| Parallel trends | The assumption that without the programme both groups would have changed by the same amount. | Difference-in-Differences |
| Power (statistical) | The chance that a study detects a real effect of a given size. 80% is the usual target. | Sampling |
| Propensity score | A person's estimated chance of being in the programme given their characteristics. | Propensity Score Matching |
| R-squared (R²) | The share of the outcome's variation that your predictors explain. | Regression |
| Random effect | A group-specific difference, such as a school's own baseline, that is allowed to vary. | Multilevel Regression |
| Reference category | The group everything else is compared against. | Regression |
| Reliability | How consistently a measure gives the same answer. | Scale & Reliability |
| RMSEA and SRMR | Misfit measures for a model. Lower is better. RMSEA 0.06 or less and SRMR 0.08 or less are good. | SEM |
| Running variable | The score or measure whose cut-off decides who gets a programme. | Regression Discontinuity |
| Sampling frame | The list from which a sample is drawn. | Sampling |
| Scree plot | A line chart of eigenvalues used to decide how many factors to keep. | PCA, EFA |
| Skewness | Whether a distribution leans to one side. Positive skew has a long tail of high values. | Summary Statistics |
| Standard deviation (SD) | The typical distance of values from the average. | Descriptive |
| Standard error (SE) | How much an estimate would vary from sample to sample. | All tests |
| Standardised effect | An effect expressed in standard deviations so that different variables can be compared. | SEM, regression |
| Straight-lining | Giving the same answer to every item in a grid. | Data Quality |
| Treatment and comparison group | The people who received a programme and those who did not. | Effect Analysis |
| Variance | The SD squared. Used inside many tests. | Descriptive |
| Weak instrument | An instrument that barely predicts take-up, making results unreliable. | Instrumental Variables |
| z-score | How many SDs a value lies from the mean. | Tests of proportions |
Data View
Data View is your first stop after loading a file. Use it to confirm everything loaded correctly and to get familiar with your variables before running any analysis.
The data table
Use the row range slider at the top of the controls bar to set which rows to display. Drag the left handle to set the start row and the right handle to set the end row, for example, drag to rows 100–300 to inspect a specific section of a large dataset. The label beside the slider shows the active range (e.g. 1–50). A missing value indicator appears automatically, a green badge if the dataset is complete, or an orange warning with a count if there are missing cells.
Showing value labels
Click Show value labels above the table. Any cell with a label mapping will display as 1 (Male) instead of just 1. Click again to return to raw values. When a dataset has value labels (e.g. imported from Stata or SPSS), this toggle is switched on automatically: so you see decoded labels by default and can turn them off to inspect raw codes.
Column details
Expand the Column details section below the table for a full inventory of every variable:
| Field | What it shows |
|---|---|
| Label | Human-readable variable description (if one has been added) |
| Value labels | Mappings e.g. 1 = Male, 2 = Female |
| Type | Data type (int64, float64, object, etc.) |
| Non-null | Count of rows with a value |
| Nulls | Count of missing values |
| Unique | Number of distinct values |
| Sample | The first non-missing value in the column |
Editing variable labels
Expand Edit variable labels at the bottom of the page. Select a variable, type a descriptive label (e.g. "Age of respondent in years"), and click Save label. Labels appear throughout the app wherever that variable is used. Leave the field blank and click Save to clear a label.
Column menu
Right-click a column header to open a menu of quick actions for that column, without leaving Data View.
| Menu item | What it does |
|---|---|
| Variable Type | Switches the column between continuous and categorical. The change applies in every analysis module. |
| Variable & Value Labels | Adds or edits the column's label and the labels for its coded values. |
| Define Missing Values | Marks coded values, such as 99 for refused or 999 for don't know, as missing, so analyses treat them like blanks. |
| Survey Design | Opens the survey design form. See the Survey weight section below. |
| Rename Variable, Recode Variable, Delete Variable | The same operations as in Data Management, applied to this column. |
| Duplicate Column, Split Column, Encode Variable | The same operations as in Data Management, applied to this column. |
| Sort Ascending, Sort Descending | Sorts the rows by this column. The row order changes in the dataset, and the Undo button in the controls bar restores it. |
| View Frequency | Shows a frequency table of the values in this column. |
Column Inspector
Click any column header in the data table to open the Column Inspector: a slide-in panel on the right side of the screen. It shows the full distribution of that column alongside key statistics, so you can explore your data without running a formal analysis.
Click a different column header to switch instantly. Close the panel with the ✕ at its top. The Inspect tab on the right edge of the screen is always visible and lets you open or hide the drawer at any time.
Categorical columns
The inspector shows a horizontal bar chart with one bar per distinct value, sorted by frequency. Each bar displays the count and percentage inline. Click any bar to filter the data grid to rows containing that value, or exclude it. Reset removes the filter; Exclude all hides every value at once. Filters applied here are view-only and do not modify your dataset.
Numeric columns
The inspector shows a binned histogram. Use the Bins slider to adjust the number of bins (5 to 40) and reshape the distribution in real time. A yellow dashed line marks the mean and a blue dashed line marks the median, when they diverge noticeably, the distribution is skewed. Key statistics are shown above the chart: valid count, missing count, unique values, range, mean, median, and standard deviation.
Changing variable type from the inspector
The inspector includes a Variable type row showing the current type, continuous or categorical: with an (auto) suffix when the type has been inferred rather than set explicitly. Click Change to expand the type editor inline. Once saved, the type badge updates immediately and the distribution chart adjusts to reflect the new classification. The change propagates to all analytical modules, Descriptive Analysis, Inferential, Visualization, for the remainder of the session.
Column menu
Right-click any column header in the data table for quick actions on that variable, without leaving Data View:
- Variable Type, Variable & Value Labels, Rename Variable, Recode Variable, Delete Variable, Duplicate Column, Split Column and Encode Variable. Each opens a short version of the matching Data Management tool. Split Column and Encode Variable are greyed out for columns where at least 80% of the values are numbers.
- Sort Ascending / Sort Descending reorders the rows by that column.
- Hide Column takes the column out of the table view only. It stays in the dataset and in every analysis. It is greyed out when only one column is left visible. See Moving, hiding and reordering columns below.
- View Frequency jumps to a frequency table for the column.
Moving, hiding and reordering columns
These change only what you see in Data View. Your data and analyses are not affected.
- Move a column. Point at a column header. Small arrows appear on either side of the name. Click the left arrow to move it left, the right arrow to move it right.
- Hide a column. Right-click the header and choose Hide Column. A Hidden columns (n) button appears in the controls bar. Click it, then click a column name to show it again, or choose Show all columns.
- Restore the original order. Click Reset order in the controls bar.
- Undo. The Undo (n) button in the controls bar reverses your last change here: a sort, a duplicate column, a recode, or a move or hide. It is greyed out when there is nothing to undo.
Filtering the table
Each column header has a small funnel icon. Click it to open a value filter: type in Search values…, tick or untick the values you want to see, and use Select all or Clear. Remove this filter clears that column, and Clear filters (n) in the controls bar clears all of them. Blank cells appear as (Blank). Filters only change what the table shows. They never alter or delete data, and the missing-value count always reflects the full dataset.
Survey weight
If your data come from a survey with sampling weights, declare them once and the analysis modules will use them. Click Survey design in the Data View controls bar, or right-click a column header and choose Survey Design. The same form opens from the Set survey weight button beside Generate on the analysis pages.
| Field | What to enter |
|---|---|
| Weight variable | The numeric column that holds each respondent's sampling weight. This is the only field you need. |
| Divide weight by (optional) | A number that every weight is divided by, without changing your data. DHS weights are stored multiplied by 1,000,000, so enter 1000000 for them. A link in the form fills it in. |
| Strata (optional) | The variable that identifies the strata of the sample design. |
| Cluster or PSU (optional) | The variable that identifies the primary sampling units, for example villages or enumeration areas. |
Before you apply, the form shows how many rows can be used, the weighted sample size, the design effect from the weights, the weight range, and the number of strata and clusters. Rows with a missing or zero weight are left out of weighted results. Once a design is set, the controls bar button shows the weight variable, and the variables in the design carry a small W, S or C badge in their column headers.
On the analysis pages, an Apply survey weight option beside Generate is ticked by default. Untick it to see the unweighted result. How each analysis changes is described under Descriptive Analysis and Inferential Analysis. Data Quality does not use survey weights.
Data Management
Data Management gives you a full set of data preparation tools without writing any code. A live row / column / missing-value count is shown at the top of each action screen. Every action is undoable (up to 20 steps).
How to use any Data Management tool
| Tool card | Button to click | Typical use |
|---|---|---|
| Variable & Value Labels | Download label template (.csv), then Upload filled template (.csv); or the manual editor and Apply labels | Give variables readable names and show 1 as Male. |
| Encode Text Variable | Encode selected column | Turn Yes/No or Male/Female text into numeric codes. |
| Variable Types | Pick an Override for a variable; Reset all overrides undoes them | Treat a 1 to 5 rating as categorical or continuous. |
| Sort Data | Choose Sort by, add Then by rules, apply | Order rows by region, then by age. |
| Duplicate Column | Duplicate Column | Keep a safe copy before recoding. |
| Create Variable | Create Variable | Total score, ratios, pass/fail flags. |
| Recode Variable | Apply Recode | Group ages, reverse-score items, collapse categories. |
| Rename Variables | Rename, Add Prefix or Add Suffix | Tidy names. Spaces are not allowed. |
| Delete Variables | Delete Selected Variables | Drop columns you will not use. |
| Delete Observations | Delete Rows | Remove rows by number (for example 3, 10-15) or by a condition such as age < 18. |
| Manage Duplicates | Remove Duplicates | Drop repeated records, keeping the first or the last. |
| Merge Datasets | Upload second dataset, add Key variables, choose Join type, then Merge Datasets | Add household data to individual data. |
| Append Dataset | Upload a second dataset, then Append Datasets | Stack survey waves. |
| Split Column | Preview split, then Split Column | Break 'First Last' into two columns, or split multi-select answers into 1/0 columns. |
| Reshape Data | Choose Wide to Long or Long to Wide, set the variables, apply | Switch between one row per person and one row per measurement. |
Variable Types
AnalyZ auto-detects variable types. Numeric columns with more than 10 unique values are treated as continuous; those with 10 or fewer unique values are treated as categorical.
Use Variable Types to override this for individual columns, for example, a Likert scale coded 1–5 has more than 10 possible values if recoded but represents categories. Getting types right is important because they control which tests and chart options appear throughout the app.
You can also change the type directly from the Column Inspector in Data View, click a column header, then click Change in the Variable type row. Both routes update the same setting.
Variable & Value Labels
Define variable labels (column descriptions) and value labels (code to text) for any file type. Only numeric-coded columns are listed, so convert text columns with Encode Text Variable first. There are two routes:
- Template: download a CSV pre-filled with every numeric-coded column and its distinct values, fill in the labels, and upload it. For each column in the file, value labels are fully replaced by what the file contains.
- Manual editor: search for a column, type its variable label, and add code and label pairs directly.
Encode Text Variable
Converts a text column (for example Male/Female or Yes/No) into numeric codes. Codes are assigned alphabetically (1, 2, 3 and so on) and the original text is saved as value labels, so the meaning is kept. Review the mapping before applying, because codes cannot be reordered afterwards.
Create Variable
Add a new column to the dataset using one of four methods:
- Formula: arithmetic using existing columns (e.g.
score / age) - Constant: a fixed value for all rows
- Conditional (if/else): e.g. if score > 50 then "Pass" else "Fail"
- Sum / Mean of items: aggregate multiple scale items into a composite score
Give the new variable a name, choose its data type, and optionally add a label and value labels before clicking Create.
Recode Variable
Map old values to new values, for example to collapse age groups or reverse-score Likert items. Select the source variable, add rules row by row, and choose the output: Create new variable from recode (the original is kept) or Recode into same variable (the original values are overwritten, so use Undo if you change your mind). Rules can be single values, ranges (1-5 or 18 thru 65), or thresholds (<18, >65). Two catch-all rules, All other non-missing and All missing, let you set a value for everything not matched.
Rename Variable
Rename variables in one of three ways: Assign new names to chosen variables one by one, Add prefix, or Add suffix to several at once. Variable and value labels follow the new names, and the change is reflected everywhere in the app immediately.
Delete Variables / Delete Observations
Delete Variables removes one or more columns. A preview shows what the dataset will look like before you confirm.
Delete Observations removes rows either by row number(s) (for example 3, 10-15) or by condition, a filter you define (e.g. age < 18).
Duplicate Column
Copies a column as-is and inserts it right after the original as name_copy. The label, value labels and variable type carry over. Use it as a backup before transforming or recoding a variable.
Sort Data
Sort rows by one or more variables, each ascending or descending. Earlier rules take priority and later ones break ties; reorder the rules with the arrows. Blank values always sort to the end. Sorting changes row order for the session, so use Undo to revert.
Manage Duplicates
Identifies fully duplicated rows, or duplicates based on a key column you choose. Preview the duplicates, then remove them keeping either the first or last occurrence.
🔗 Merge Datasets
Join a second dataset to the current one using one or more key variables. The key columns can have different names in the two datasets, and you can add several key pairs. Choose a join type:
| Join type | Behaviour |
|---|---|
| Inner | Keep only rows that match in both datasets |
| Left | Keep all rows from the current dataset; add columns from the second where they match |
| Right | Keep all rows from the second dataset; add columns from the current where they match |
| Outer | Keep all rows from both datasets; fill gaps with missing values |
📥 Append Dataset
Stack a second dataset on top of the current one (row-wise). Columns present in one dataset but not the other are filled with missing values. Useful for combining data collected in batches (e.g. survey wave 1 and wave 2).
🔀 Split Column
Splits a text column into multiple new columns using a delimiter you specify (comma, space, semicolon, pipe, tab, or any custom string). Each part becomes a separate column named colname_A, colname_B, colname_C, and so on, inserted immediately after the source column.
Use Preview split before applying to verify how the delimiter will divide the values. Useful for columns such as "A B C", "Yes, No, Maybe", or multi-response fields encoded as delimited strings.
Reshape Data
Converts your dataset between wide and long formats, the two fundamental data structures in analysis:
| Direction | When to use | How it works |
|---|---|---|
| Wide → Long | Repeated-measures or longitudinal analysis; many value columns represent the same measurement at different times or conditions | Select ID variables (columns to keep) and value variables (columns to unpivot). Each value variable becomes a row. Uses pd.melt(). |
| Long → Wide | Creating a summary matrix; pivoting category values into columns | Select an index column (row identifier), a columns variable (whose unique values become column headers), and a values column. Uses pd.pivot_table() with an aggregation function you choose. |
Variable & Value Labels
Adds variable labels (a description for each column) and value labels (for example 1 = Male, 2 = Female) to your dataset. Click Download template to get a CSV pre-filled with your variable names, fill in the labels in Excel or Google Sheets, then upload the file. The labels are applied at once and appear in all analysis outputs and charts. You can also add or edit single labels in the manual editor below the upload area.
Encode Text Variable
Turns a text column, such as Male/Female, Yes/No or region names, into numeric codes. Select the column, review the codes AnalyZ assigns (1, 2, 3 and so on), and click Encode. The text column is replaced by the numeric codes, and value labels are created automatically so the meaning is kept.
Duplicate Column
Copies a column. The copy is inserted straight after the original and named with _copy added (or _copy_2 if that name is taken). It keeps the original's variable label, value labels and variable type. This is useful as a safety copy before you recode or transform a column.
Sort Data
Reorders the rows by one or more variables. Choose the first variable and the direction, then click Add variable to add tie-breakers, for example region first, then score descending within each region. Earlier variables in the list take priority, and you can move them up or down. Blank or missing values always sort to the end.
↩ Undo
The ↩ Undo button in the sidebar rolls back the last data management action, and ↪ Redo reapplies it. Up to 20 steps of history are maintained. The dataset, variable labels, and value labels are all restored together.
Worked examples
Short, concrete examples of the most-used tools. Your own variable names will differ.
| Task | Example | How |
|---|---|---|
| Recode Variable | Age in years has values from 15 to 80. You want groups: under 25, 25 to 44, 45 and over. | Recode Variable, choose age, define each range and give it a new label, and save into a new variable so the original stays intact. |
| Recode Variable | A Likert item runs 1 (strongly disagree) to 5 (strongly agree) and you want agree versus not agree. | Recode 4 and 5 to 1, and 1 to 3 to 0, into a new variable. Add value labels such as 1 = Agree. |
| Create Variable | You want a total stress score from five items. | Create Variable, choose a formula that adds the five columns (or their mean). Check any reverse-coded item first, as in Cronbach's Alpha. |
| Merge Datasets | Household data in one file and individual data in another, sharing a household ID. | Merge on the ID column. Check the row count afterwards, since a merge on a non-unique ID can multiply rows. |
| Append Dataset | Two survey rounds with the same columns. | Append the second file below the first. Add a round variable so you can tell them apart, which Difference-in-Differences needs. |
| Reshape Data | One row per person per time point (long) but a method wants one row per person (wide), or the reverse. | Reshape, choose the ID column and the columns that repeat. |
| Manage Duplicates | The same respondent was entered twice. | Run Manage Duplicates on the ID column, review the flagged rows, then keep the first or the most complete. |
Descriptive Analysis
Descriptive Analysis answers the first questions of any project: Is my data clean? What does a typical respondent look like? How are the answers spread out? It has five analysis cards, arranged in the order most projects use them. Click a card to open it, and use Back to return to the cards. The optional Filter data panel restricts any analysis to a sub-group, for example women aged 18 to 35.
| Card | In one line | Typical question |
|---|---|---|
| Data Quality | A health check of the dataset | Are there missing, duplicated or impossible values? |
| Frequency | Counts and percentages for each answer | What share of respondents chose each option? |
| Cross-Tabulate | How two categorical questions line up | Do men and women answer differently? |
| Table | A ready-made 'Table 1' for reports | How do programme and comparison groups compare at baseline? |
| Summary Statistics | Average, spread and shape of numeric variables | What is the typical income, and how much does it vary? |
Data Quality
Data Quality
Use it when
- You have just received or exported a dataset
- Data collection is still under way and you want to monitor fieldwork
- A result looks too good, too odd or too uniform
- You are about to run any other analysis
Your data needs
- Any loaded dataset
- For the Enumerator tab: a column that identifies the interviewer
- For interview duration: start-time and end-time columns
- For straight-lining: the grid of rating questions to check
Data Quality: how to run it
Data Quality: reading the results
| You see | It means |
|---|---|
| % Missing, coloured red | More than 30% of people did not answer. Ask why before using the variable. |
| Unique values = 1 (constant) | Every respondent has the same value, so the variable cannot explain anything. |
| Outliers (beyond 3 SD) | Values far from the average. They may be typing errors or genuine extremes. Look at them before deciding. |
| Duplicate groups | Sets of identical records. Often the same form submitted twice. |
| Range violations | Values outside the range you set, such as an age of 214. |
| Heaping detected | Too many answers end in 0 or 5 (or another favoured digit). Respondents are rounding, which blunts precision. |
| High straight-lining rate | Many respondents gave the same answer to every item in a grid. |
| Short median duration | An interviewer finishes much faster than colleagues. A prompt to review their work. |
The situation. A household survey of 600 interviews has just come in from five interviewers.
What AnalyZ shows. Overview shows monthly income is 34% missing (red) and age has 5 outliers. Duplicate Rows finds 6 pairs of identical records. Value Issues shows 5 ages above 110 and reports that 41% of ages end in 0 or 5. Enumerator Patterns shows interviewer E07 with a median interview of 9 minutes against 31 minutes for the others, and 72% straight-lining.
How to read it. The 6 duplicate pairs are probably double submissions, so keep one of each. The five impossible ages are probably typing errors, so check the paper forms. The heaping is typical when people do not know their exact age, and you should avoid analysing age in single years. The missing income needs a decision before analysis (see below). E07's interviews deserve a spot check, because a short duration and heavy straight-lining together suggest the questions were not asked in full. These are leads to follow up, not proof of anything.
Data Quality: things to watch for
- Missing is not always a mistake. A question about pregnancy may be skipped by design for men. Check the questionnaire before treating gaps as errors.
- Outliers are flagged by a statistical rule, not by common sense. A very large household may be real.
- Many missing values do not simply disappear. Tests use only the cases with an answer on every variable involved, so heavy missingness quietly shrinks your sample.
- Data Quality does not use survey weights.
Data Quality: key terms
- Missing value
- A blank where an answer should be.
- Outlier
- A value unusually far from the rest.
- Heaping
- Rounding of answers to favourite digits.
- Straight-lining
- Answering every item in a grid identically, often a sign of low effort.
- Enumerator
- The person who conducts an interview.
Frequency
Frequency
Use it when
- You want to report the breakdown of a categorical question
- You want to check that value labels and codes look right
- You need percentages for a report or chart
Your data needs
- One or more variables, usually categorical
- Numeric variables with only a few distinct values also work well
Frequency: how to run it
Frequency: reading the results
| You see | It means |
|---|---|
| Value | The answer. If a value label exists it is shown instead of the code. |
| Frequency | How many respondents gave that answer. |
| % | Share of all respondents, including those who did not answer. |
| Valid % | Share of respondents who did answer. This is the figure most reports use. |
| Cumulative % | Running total of the valid percentages, top to bottom. Useful for ordered answers such as education. |
| Weighted count and 95% CI | Appear when a survey weight is applied. See Survey weights below. |
The situation. Highest education level was asked of 500 respondents. 20 skipped the question.
| Education | Frequency | % | Valid % | Cumulative % |
|---|---|---|---|---|
| No schooling | 72 | 14.4 | 15.0 | 15.0 |
| Primary | 168 | 33.6 | 35.0 | 50.0 |
| Secondary | 150 | 30.0 | 31.3 | 81.3 |
| Higher | 90 | 18.0 | 18.8 | 100.0 |
| Missing | 20 | 4.0 |
How to read it. Report the Valid % column: 35.0% finished primary school only, and half of respondents (50.0%) have primary education or less. The plain % column is smaller (33.6%) only because the 20 non-responders are counted in the base. Say which one you used.
Illustrative figures.
Frequency: things to watch for
- Choose the base deliberately. Valid % and % answer different questions.
- Pie charts are readable with up to about five slices. For many categories use the horizontal bar chart.
- A categorical variable with hundreds of values (a village name, say) makes a very long table. Group it first with Recode.
Frequency: key terms
- Valid percent
- Percent among people who gave an answer.
- Cumulative percent
- Percent at or below each category.
Cross-Tabulate
Cross-Tabulate
Use it when
- You want to compare the breakdown of one category across the groups of another
- You want a quick test of whether two categorical questions are linked
Your data needs
- Two categorical variables, such as sex and phone ownership
- A reasonable number of cases in each cell
Cross-Tabulate: how to run it
Cross-Tabulate: reading the results
| You see | It means |
|---|---|
| Cell counts and percentages | Read the percentages along the direction that adds to 100. With groups in rows, read across each row. |
| Chi-square, df | The size of the gap between your counts and the counts expected if the two questions were unrelated, and the degrees of freedom. |
| p-value | Below 0.05 means the pattern is unlikely to be chance. |
| Cramér's V | Strength of the link from 0 to 1, described in words: very weak (under 0.1), weak (under 0.3), moderate (under 0.5) or strong. |
The situation. Phone ownership by sex among 400 respondents (180 women and 220 men).
| Owns a phone | Does not | Total | |
|---|---|---|---|
| Women | 126 (70.0%) | 54 (30.0%) | 180 |
| Men | 187 (85.0%) | 33 (15.0%) | 220 |
What AnalyZ shows. Row percentages are shown, so each row adds to 100%. Chi-square = 12.22, df = 1, p-value below 0.001, Cramér's V = 0.17 (weak).
How to read it. Phone ownership is 70% among women and 85% among men, a gap of 15 percentage points. The gap is significant, so it is unlikely to be chance. Cramér's V of 0.17 describes a weak but real link: sex matters, but it is far from the whole story. Reading the column percentages instead would answer a different question (of all phone owners, what share are women?).
Illustrative figures.
Cross-Tabulate: things to watch for
- A significant chi-square shows that a link exists, not that one question causes the other.
- Cells with very few people (under 5 expected) make the chi-square unreliable. Merge sparse categories with Recode first.
- Choose the percentage direction before you look at the result, and say which you used when you report it.
Cross-Tabulate: key terms
- Cell
- One box in the grid.
- Row percent
- Percent within each row.
- Column percent
- Percent within each column.
Table (Table 1)
Table
Use it when
- You need a profile of respondents by programme group, region or sex
- You want to show that two groups were similar at the start (baseline balance)
Your data needs
- One grouping variable for the columns
- One or more variables for the rows, numeric or categorical
Table: how to run it
Table: reading the results
| You see | It means |
|---|---|
| Continuous row, for example Age | Shown as mean (standard deviation) for each group. |
| Categorical row, for example Sex | Shown as count and percentage, usually percentage within the group column. |
| Total column | The whole sample, for reference. |
| p-value column | A chi-square test for categorical rows and a t-test or ANOVA for numeric rows. With survey weights these become design-based tests. |
The situation. A training programme compares 200 participants with 200 non-participants before it starts.
| Programme (n=200) | Comparison (n=200) | p-value | |
|---|---|---|---|
| Age, mean (SD) | 31.4 (8.2) | 32.0 (8.6) | 0.46 |
| Female, n (%) | 104 (52.0%) | 96 (48.0%) | 0.44 |
| Years of school, mean (SD) | 8.9 (3.1) | 8.7 (3.3) | 0.55 |
| Employed at baseline, n (%) | 58 (29.0%) | 81 (40.5%) | 0.02 |
How to read it. The two groups look alike on age, sex and schooling, since all the p-values are well above 0.05. They differ on employment: 29.0% of participants were employed against 40.5% of the comparison group. That matters, because a later difference in employment could reflect where each group started. This is the kind of imbalance that matching or a difference-in-differences design is meant to handle.
Illustrative figures.
Table: things to watch for
- With many rows, a few p-values will fall below 0.05 by chance alone. Look at the size of the gap, not only the stars.
- In a randomised study, differences at baseline are chance by definition, and many journals prefer you report the numbers without p-values.
- Categorical rows with very small cells give unreliable p-values.
Table: key terms
- Baseline
- The situation before a programme starts.
- Header variable
- The grouping that forms the columns.
Summary Statistics
Summary Statistics
Use it when
- You need the typical value and the variation for age, income, scores, weight
- You want to compare the numbers across groups (for example by region)
- You are checking whether a variable is skewed before choosing a test
Your data needs
- One or more numeric variables
- An optional categorical variable to split the results by
Summary Statistics: how to run it
Summary Statistics: reading the results
| You see | It means |
|---|---|
| N and Missing | How many people have a value, and how many do not. |
| Mean | The average. Sensitive to extreme values. |
| Median | The middle value when sorted. Steady even with extremes. |
| Mode | The most frequent value. |
| SD and Variance | How far values typically sit from the mean. SD is in the same units as the data. Variance is SD squared. |
| Min, Max, Range | The smallest and largest values, and the gap between them. |
| Skewness | Near 0 is symmetric. Above 1 means a long tail of high values. Below -1 means a long tail of low values. |
| Kurtosis | Tail weight. Near 0 looks like a bell. Above 1 means heavy tails and probable outliers. Below -1 means values bunch in the middle. |
The situation. Monthly household income for 450 households.
| N | Missing | Mean | Median | SD | Min | Max | Skewness |
|---|---|---|---|---|---|---|---|
| 450 | 12 | 412 | 310 | 380 | 40 | 3,200 | 2.4 |
How to read it. The mean (412) is well above the median (310) and skewness is 2.4, meaning a small number of high-income households pull the average up. For a typical household, quote the median. An SD of 380 is almost as large as the mean, which also shows how unequal the incomes are. Because the data are strongly skewed, use rank-based tests such as Mann-Whitney U or Kruskal-Wallis for comparing income across groups, or log the variable first.
Illustrative figures.
Summary Statistics: things to watch for
- Do not report a mean alone. Pair it with the SD, or use the median when the data are skewed.
- The mode is mainly useful for categories. For a continuous variable every value can be unique.
Summary Statistics: key terms
- Standard deviation
- The typical distance of values from the mean.
- Skewness
- Which way the distribution leans.
- Kurtosis
- How heavy the tails are.
Survey weights
When a survey weight is set, Frequency, Cross-Tabulate, Table and Summary Statistics use it by default. Untick Apply survey weight beside Generate to compare with the unweighted result.
| Module | What changes |
|---|---|
| Frequency | Percentages are weighted, a weighted count is added, and each percentage comes with a 95% confidence interval that reflects the weights, strata and clusters. The Frequency column stays the number of respondents. |
| Summary Statistics | The mean, median and measures of spread are weighted. Weighted N, the standard error of the mean and its 95% confidence interval can be added as statistics. N stays the number of respondents. |
| Cross-Tabulate | Percentages are weighted. The counts shown can be the number of respondents or the weighted count. The chi-square test is replaced by the Rao-Scott corrected F test, which adjusts for the weights, strata and clusters. |
| Table | Percentages, means and standard deviations are weighted, with the same choice of counts. The p-values come from the Rao-Scott F test for categorical rows and from a design-based test of equal means for continuous rows. |
Inferential Analysis
Descriptive results tell you what happened in the people you measured. Inferential tests ask the next question: is this pattern big enough, given how many people I measured, that I should believe it is real, rather than a quirk of who happened to be in my sample? If you are new to this, read Statistics Basics first, in particular the parts on p-values and effect sizes.
Inferential Analysis is organised into five groups. Choose a group, then choose the test, set up your variables and click Run test. Every result ends with a written interpretation, and results can be exported to Word, Excel or CSV, added to your Data Canvas, or sent to Sense Making.
| Group | Tests | Question it answers |
|---|---|---|
| Compare Means | Independent t-test, One-sample t-test, Paired t-test, One-way ANOVA, Two-way ANOVA, Repeated-Measures ANOVA | Do the averages differ? |
| Compare Percentages | Chi-square test, Z-test for one proportion, Z-test for two proportions | Do the percentages differ? |
| Correlation | Pearson, Spearman | Do two things move together? |
| Regression | Linear, Logistic, Probit, Ordinal, Multinomial, Poisson, Negative Binomial, Hierarchical, Multilevel | What predicts the outcome, once other things are held steady? |
| Non-parametric Tests | Mann-Whitney U, Wilcoxon Signed-Rank, Kruskal-Wallis, Friedman | The same as the means tests, for ordinal, skewed or small-sample data |
Before you run any test
- Check the variable types. A numeric outcome and a grouping variable need to be recognised correctly. If a test does not offer your variable, check its type in Data Management.
- Run Data Quality first. Outliers and impossible values can quietly dominate a test.
- Know your sample size. Tests on fewer than about 30 cases per group are fragile. If you are planning, use Sampling to work out how many you need.
- Remember who is included. A test uses only the cases that have an answer on every variable you selected. The N shown in the results tells you how many that was.
- Decide on the question before looking at results. Running many tests and reporting only the significant ones will produce false findings.
Compare Means
These tests compare averages. Pick by how many groups you have and whether they are separate people or the same people measured again.
Independent t-test
Use it when
- You are comparing the average of a number between exactly two groups
- The groups are made of different people (men and women, programme and comparison)
Your data needs
- A numeric outcome such as score, income or age
- A grouping variable with exactly 2 values
- Roughly bell-shaped data in each group, or about 30 or more people per group
Independent t-test: how to run it
Independent t-test: reading the results
| You see | It means |
|---|---|
| N, Mean, SD per group | How many people, their average and their typical spread. Start here: how far apart are the means, in real units? |
| t | How many standard errors apart the two means are. Further from 0 means stronger evidence. |
| df | Degrees of freedom: N minus 2. Needed for the p-value, rarely worth interpreting. |
| p-value and Sig | Below 0.05 (one or more stars) means the gap is unlikely to be chance. |
| Cohen's d | Size of the gap in standard deviations: 0.2 small, 0.5 medium, 0.8 large. |
| Levene p | A check that both groups are equally spread out. If it is below 0.05, the spreads differ and the result deserves caution. In that case also run Mann-Whitney U as a cross-check. |
The situation. 80 pupils took a reading test after being taught by Method A (40 pupils) or Method B (40 pupils).
| Group | N | Mean | SD |
|---|---|---|---|
| Method A | 40 | 74.2 | 9.6 |
| Method B | 40 | 68.9 | 10.4 |
What AnalyZ shows. t = 2.37, df = 78, p-value 0.020 (*), Cohen's d = 0.53. Levene p is above 0.05. Interpretation: Method A has a significantly higher mean reading score than Method B at the 5% significance level. The difference is medium in practical terms.
How to read it. Method A pupils scored 5.3 points higher on average. The p-value says this gap is unlikely to be chance, and d = 0.53 says it is a medium-sized gap, about half a standard deviation. Both matter: significance tells you the gap is probably real, and d tells you whether it is worth caring about. A significant result with d = 0.1 would be real but small.
Illustrative figures.
Independent t-test: things to watch for
- Three or more groups need One-way ANOVA. Running several t-tests instead raises the chance of a false positive.
- The same people measured twice need the Paired t-test, not this one.
- One or two extreme values can distort a mean. Check Summary Statistics or a boxplot first.
- With survey weights applied, the test does not assume equal spread and shows the mean difference with its 95% interval instead of Levene's test.
Independent t-test: key terms
- t statistic
- The gap between means divided by its uncertainty.
- Cohen's d
- The gap measured in standard deviations.
- Levene's test
- A check that the groups are equally scattered.
One-sample t-test
Use it when
- You have a benchmark to compare against
- You want to test whether an average is different from zero, for example average change
Your data needs
- One numeric variable
- A test value you decided before looking at the data
One-sample t-test: how to run it
One-sample t-test: reading the results
| You see | It means |
|---|---|
| N, Mean, SD, SE | Sample size, average, spread and the standard error of the mean (how much the average would wobble between samples). |
| Test value | The benchmark you entered. |
| t, df, p-value, Sig | How far the mean is from the benchmark in standard errors, and whether that is beyond chance. |
| 95% CI lower and upper | A range for the true average. If it does not contain the test value, the result is significant. |
The situation. 80 respondents rated their satisfaction from 1 to 5. Is the average different from the neutral value of 3?
| N | Mean | SD | Test value | t | df | p-value | 95% CI |
|---|---|---|---|---|---|---|---|
| 80 | 3.46 | 0.92 | 3 | 4.47 | 79 | < 0.001 | 3.26 to 3.67 |
How to read it. The average of 3.46 is 0.46 points above neutral. The interval 3.26 to 3.67 stays clear of 3, so the result is significant. Satisfaction is modestly positive on average. This says nothing about how many people are satisfied. For that, use Frequency or the Z-test for a proportion.
Illustrative figures.
One-sample t-test: things to watch for
- Choose the benchmark first. Picking it after seeing the mean is circular.
- For a single 5-point question, the median or the share who agree is often more informative than a mean.
One-sample t-test: key terms
- Standard error
- How much the sample average is likely to wobble from sample to sample.
- Test value
- The fixed number you compare against.
Paired t-test
Use it when
- Pre and post scores on the same participants
- Two measurements per person under two conditions
- Matched pairs, for example twins or couples
Your data needs
- Two numeric columns, one row per person, for example pre_score and post_score
- Rows with a missing value in either column are left out
- If your data are stacked in rows, reshape to wide format first in Data Management
Paired t-test: how to run it
Paired t-test: reading the results
| You see | It means |
|---|---|
| Mean and SD of each variable | The two averages. Compare them in real units. |
| Mean diff and SD diff | The average change per person, and how much that change varies between people. |
| t, df, p-value, Sig | Whether the average change is beyond what chance would produce. |
| 95% CI lower and upper | A range for the true average change. If it excludes 0, the change is significant. |
| Interpretation | Names which measurement is higher and gives a practical size, from the mean change relative to its spread. |
The situation. 36 participants took a skills test before and after a training course. Post-test is entered first, so improvement is positive.
| Variable | N | Mean | SD |
|---|---|---|---|
| Post-test | 36 | 61.3 | 10.1 |
| Pre-test | 36 | 56.5 | 9.7 |
What AnalyZ shows. Mean diff = 4.8, SD diff = 8.2, t = 3.51, df = 35, p-value 0.001 (**), 95% CI 2.0 to 7.6. Interpretation: Post-test has a significantly higher mean than Pre-test at the 1% significance level. The difference is medium in practical terms.
How to read it. On average people gained 4.8 points, and the interval 2.0 to 7.6 does not include 0. The SD of the change (8.2) shows that individuals varied widely, and some may have gone down. Without a comparison group, you cannot say the course caused the gain, as people might have improved anyway. For that question see Effect Analysis.
Illustrative figures.
Paired t-test: things to watch for
- A before-and-after gain is not proof that a programme worked. Anything else that changed between the two times is mixed in.
- Three or more time points need Repeated-Measures ANOVA.
- For ordinal scores or very skewed changes, use Wilcoxon Signed-Rank.
Paired t-test: key terms
- Paired data
- Two measurements linked to the same person or matched unit.
- Mean difference
- The average of each person's change.
One-way ANOVA
Use it when
- Three or more separate groups
- You want one overall answer before looking at pairs
Your data needs
- A numeric outcome
- A grouping variable with 3 or more values
- Reasonably similar spread in each group, and at least a handful of people per group
One-way ANOVA: how to run it
One-way ANOVA: reading the results
| You see | It means |
|---|---|
| Group, N, Mean, SD | The averages you are comparing. |
| F | The ratio of between-group to within-group variation. Larger means stronger evidence of a difference somewhere. |
| p-value and Sig | Whether the overall differences are beyond chance. It does not say which groups differ. |
| η² (eta-squared) | The share of all variation explained by group: 0.01 small, 0.06 medium, 0.14 large. |
| Post-hoc comparisons | One row per pair of groups, with the difference in means and a p-value. Use this table to see where the difference is. |
The situation. 105 adults from three districts completed a knowledge test (35 per district).
| District | N | Mean | SD |
|---|---|---|---|
| North | 35 | 58.2 | 11.0 |
| Central | 35 | 64.5 | 10.4 |
| South | 35 | 67.1 | 11.6 |
What AnalyZ shows. F(2, 102) = 6.05, p-value 0.003 (**), η² = 0.11. Post-hoc comparisons: North vs Central, mean diff -6.3, p 0.016 (*). North vs South, mean diff -8.9, p 0.002 (**). Central vs South, mean diff -2.6, p 0.327 (ns).
How to read it. The overall test is significant, and η² = 0.11 says district explains about 11% of the variation in scores, a medium-to-large effect. The pair table shows where the difference sits. North scores lower than both other districts, while Central and South do not differ from each other. Report it as 'North was lower than Central and South', not as 'all three districts differ'.
Illustrative figures.
One-way ANOVA: things to watch for
- A significant F tells you at least one group is different, not which one. Use the pair table.
- In the standard analysis the pair p-values are not adjusted for the number of comparisons. With many groups, treat a borderline pair with caution. When a survey weight is applied, the pairwise p-values are Bonferroni adjusted.
- Strongly skewed outcomes or tiny groups call for Kruskal-Wallis instead.
One-way ANOVA: key terms
- Between-group variation
- How far group averages sit from the overall average.
- Within-group variation
- How much people differ inside their own group.
- Post-hoc test
- A follow-up comparing pairs after the overall test is significant.
Two-way ANOVA
Use it when
- You have two categorical factors and one numeric outcome
- You suspect an effect works differently for different groups
Your data needs
- A numeric outcome
- Two grouping variables (Factor A and Factor B)
- Reasonable numbers in every combination of the two, at least 5 per cell and ideally many more
Two-way ANOVA: how to run it
Two-way ANOVA: reading the results
| You see | It means |
|---|---|
| Factor A and Factor B (main effects) | Does the average differ across the categories of that factor, ignoring the other? |
| A × B (interaction) | Does the effect of one factor change across the levels of the other? A significant interaction means it depends. |
| F, p-value, Sig | For each row: the strength of evidence and whether it is beyond chance. |
| Error | Variation left over, inside the groups. |
The situation. 200 pupils took a test, 50 in each combination of teaching method (A or B) and sex.
| Female | Male | All | |
|---|---|---|---|
| Method A | 80.0 | 70.0 | 75.0 |
| Method B | 69.0 | 70.0 | 69.5 |
| All | 74.5 | 70.0 |
What AnalyZ shows. Method: F(1, 196) = 18.67, p-value < 0.001. Sex: F(1, 196) = 12.50, p-value 0.001. Method × Sex: F(1, 196) = 18.67, p-value < 0.001.
How to read it. All three rows are significant, but the interaction changes how the other two should be told. The cell means show that Method A helps girls (80 against 69 under Method B) but makes no difference for boys (70 against 70). Saying 'Method A is better' would be misleading. The accurate statement is that Method A works for girls only. When the interaction is significant, describe the cell means rather than the overall method or sex averages.
Illustrative figures.
Two-way ANOVA: things to watch for
- Read the interaction first. If it is significant, main effects can mislead.
- Cells with very few people give unreliable results.
- Plot the cell means in Data Visualization to see an interaction. Lines that are not parallel signal one.
Two-way ANOVA: key terms
- Main effect
- The effect of one factor averaged over the other.
- Interaction
- When the effect of one factor depends on the level of the other.
Repeated-Measures ANOVA
Use it when
- The same people are measured at three or more times or under three or more conditions
- The outcome is numeric and roughly bell-shaped
Your data needs
- Three or more numeric columns, one per time point, such as score_baseline, score_midline, score_endline
- One row per person, with a value in every column
- Optionally a subject ID column
Repeated-Measures ANOVA: how to run it
Repeated-Measures ANOVA: reading the results
| You see | It means |
|---|---|
| Condition means | The average at each time point. Look at the pattern first. |
| F and p-value | Whether the averages differ across time points. |
| Partial η² | How much of the variation is explained by time: 0.01 small, 0.06 medium, 0.14 large. |
| Mauchly's W | A check that the changes between time points vary evenly (sphericity). Close to 1 is fine. Below 0.75, treat the p-value with caution. |
| Pairwise comparisons (Bonferroni) | Each pair of time points compared, with p-values made stricter for the number of comparisons. Shown when the overall result is significant. |
The situation. 30 participants were tested at baseline, midline and endline.
| Time point | Mean |
|---|---|
| Baseline | 53.6 |
| Midline | 56.2 |
| Endline | 61.8 |
What AnalyZ shows. F(2, 58) = 55.96, p-value < 0.001, partial η² = 0.66. Mauchly's W is above 0.75, so sphericity looks acceptable. All three pairs differ after Bonferroni adjustment.
How to read it. Scores rose steadily, and time explains a large share (66%) of the variation. The pairwise table says each step is a real change, not just the overall baseline-to-endline gain. As with the paired t-test, you cannot tell from this alone whether a programme caused the rise.
Illustrative figures.
Repeated-Measures ANOVA: things to watch for
- AnalyZ reports Mauchly's W but does not apply a sphericity correction. If W is low, treat borderline results with care or use Friedman.
- Every person needs a value at every time point. People with a gap are dropped.
- If the data are skewed or ordinal, use the Friedman test.
Repeated-Measures ANOVA: key terms
- Sphericity
- The assumption that differences between time points vary equally.
- Partial eta-squared
- The share of variation explained by time once person-to-person differences are set aside.
Compare Percentages
These tests compare shares of people rather than averages. Use them when your variables are categories.
Chi-square test
Use it when
- Both variables are categories (education by employment, region by whether a service was used)
- You want to know whether groups differ in how they answered
Your data needs
- Two categorical variables
- Enough people in each cell. About 5 expected per cell is the minimum.
Chi-square test: how to run it
Chi-square test: reading the results
| You see | It means |
|---|---|
| Observed counts | The real number of people in each combination. |
| Expected counts | How many you would see in each cell if the two questions were unrelated. Compare with observed to see where the pattern is. |
| Chi-square (χ²) and df | The overall size of the gap between observed and expected, with the degrees of freedom. |
| p-value and Sig | Whether the gap is beyond chance. Significant means the questions are linked. |
| Cramér's V | Strength of the link from 0 to 1. Roughly: under 0.1 negligible, 0.1 to 0.3 small, 0.3 to 0.5 moderate, above 0.5 strong. |
The situation. 200 people took part in a job programme (100) or a comparison group (100). Six months later each was recorded as employed full-time, employed part-time, or not employed.
| Full-time | Part-time | Not employed | |
|---|---|---|---|
| Programme (observed) | 48 | 30 | 22 |
| Comparison (observed) | 27 | 41 | 32 |
| Expected in each row | 37.5 | 35.5 | 27.0 |
What AnalyZ shows. Chi-square = 9.44, df = 2, p-value 0.009 (**), Cramér's V = 0.22.
How to read it. Full-time employment is 48% in the programme group and 27% in the comparison group. The expected count for each group was 37.5 people, so the programme group has 10.5 more full-timers than 'nothing is going on' would give. The result is significant, and V = 0.22 is a small to moderate link. Chi-square says the two variables are linked overall. It does not tell you which cell is responsible, so compare observed and expected counts or the row percentages from Cross-Tabulate.
Illustrative figures.
Chi-square test: things to watch for
- Cells with an expected count under 5 make the p-value unreliable. Merge sparse categories.
- A large sample can make a trivial link significant. Look at Cramér's V.
- It shows a link, not a cause, and not the direction.
- To compare just two percentages, the Z-test for two proportions is more direct.
Chi-square test: key terms
- Contingency table
- A grid of counts for two categorical variables.
- Expected count
- The count you would see if the variables were unrelated.
- Cramér's V
- A 0 to 1 measure of how strongly two categories are linked.
Z-test for one proportion
Use it when
- You have a target or benchmark percentage
- You want to know whether your observed percentage is meaningfully above or below it
Your data needs
- One categorical variable, usually yes/no
- A hypothesised proportion between 0 and 1, decided before looking at the data
Z-test for one proportion: how to run it
Z-test for one proportion: reading the results
| You see | It means |
|---|---|
| Observed p | The share in your sample with the success value. |
| Hypothesised p | The target you entered. |
| z | How many standard errors the observed share is from the target. |
| p-value and Sig | Whether the gap from the target is beyond chance. |
The situation. 150 participants were assessed and 93 improved. The programme aimed for more than half to improve.
| N | Observed p | Hypothesised p | z | p-value |
|---|---|---|---|---|
| 150 | 0.62 | 0.50 | 2.94 | 0.003 |
How to read it. 62% improved, which is 12 percentage points above the 50% target, and the gap is unlikely to be chance. Interpretation: The proportion of Improved (62.0%) is significantly different from the expected 50.0% at the 1% significance level. For a statement about how much above 50% the true figure is, use a confidence interval from Frequency.
Illustrative figures.
Z-test for one proportion: things to watch for
- Fix the target before you look at the results.
- Works best when both the expected successes and failures are at least 10 or so.
- With survey weights the test uses a t statistic with design degrees of freedom.
Z-test for one proportion: key terms
- Proportion
- A share expressed as a number from 0 to 1.
- Success value
- The answer you are counting as a hit.
Z-test for two proportions
Use it when
- You want to compare the share with a characteristic across two measures, such as baseline and endline columns
- You want a direct test of a gap between two percentages
Your data needs
- Two categorical variables with a success value each
- Both measured on comparable people
Z-test for two proportions: how to run it
Z-test for two proportions: reading the results
| You see | It means |
|---|---|
| N, Successes, Proportion | Each variable's count and share. |
| Diff | The first proportion minus the second. |
| z, p-value, Sig | Whether the difference is beyond chance. |
The situation. 54 of 120 households used a service in the first year (45.0%), and 72 of 130 in the second year (55.4%).
| N | Successes | Proportion | |
|---|---|---|---|
| Year 1 | 120 | 54 | 0.450 |
| Year 2 | 130 | 72 | 0.554 |
What AnalyZ shows. Diff = -0.104, z = -1.64, p-value 0.101 (ns).
How to read it. Usage was about 10 percentage points higher in year 2, but with these sample sizes the gap is within what chance can produce, so you cannot claim a real rise. A bigger sample might settle it. Say 'no significant difference' rather than 'no difference', because the test cannot prove there is none.
Illustrative figures. The test treats the two groups as independent samples.
Z-test for two proportions: things to watch for
- The test assumes the two groups are independent. If they are the same people measured twice, the result is only approximate.
- For more than two groups, or a full category breakdown, use the Chi-square test.
Z-test for two proportions: key terms
- Percentage-point difference
- The simple gap between two percentages.
Correlation
Correlation measures how closely two numeric variables move together, on a scale from -1 to +1. It does not say that one causes the other.
| Value of r | Reading |
|---|---|
| +1 | Perfect upward line: as one rises, the other rises by a steady amount |
| +0.5 | Strong positive link |
| +0.3 | Moderate positive link |
| +0.1 | Weak positive link |
| 0 | No linear link |
| -0.3 | Moderate negative link: as one rises, the other tends to fall |
| -1 | Perfect downward line |
Pearson Correlation
Use it when
- Two numeric variables that you expect to rise or fall together
- A first look at many variables at once
Your data needs
- Two or more numeric variables
- A roughly straight-line relationship and no extreme outliers
- Complete answers on all selected variables. Only people with a value on every selected variable are used.
Pearson Correlation: how to run it
Pearson Correlation: reading the results
| You see | It means |
|---|---|
| r | The correlation for each pair, from -1 to +1. The sign gives direction and the size gives strength. |
| Stars | * p-value below .05, ** below .01, *** below .001. No stars means the link is not distinguishable from chance. |
| Heatmap colours | Darker cells are stronger. One colour is positive and the other negative. |
| r squared (mentally) | Square r to get the share of variation shared. r = 0.5 means about 25% is shared. |
The situation. 80 learners recorded their weekly study hours and their exam score.
| Study hours | Exam score | |
|---|---|---|
| Study hours | 1.000 | 0.639 *** |
| Exam score | 0.639 *** | 1.000 |
How to read it. r = 0.64 is a strong positive link: learners who studied more tended to score higher, and study hours account for about 41% of the variation in scores (0.64 squared). The three stars say this is very unlikely to be chance. It does not prove that studying caused the higher scores, since motivation or prior ability could drive both.
Illustrative figures.
Pearson Correlation: things to watch for
- Pearson only detects straight-line links. A U-shaped pattern can give r near 0. Check a scatter plot in Data Visualization.
- One outlier can create or hide a correlation. If you see outliers or skew, use Spearman.
- With a large sample even a tiny r is significant. Read the size, not just the stars.
- A correlation can be produced by a third variable that drives both.
Pearson Correlation: key terms
- r
- The Pearson correlation coefficient.
- Linear
- Following a straight line.
Spearman Correlation
Use it when
- Ordinal variables such as ratings or education level
- Skewed data, or data with outliers
- The pattern rises or falls steadily but not along a perfectly straight line
Your data needs
- Two or more variables, numeric or ordered categories
- Only cases with a value on every selected variable are used
Spearman Correlation: how to run it
Spearman Correlation: reading the results
| You see | It means |
|---|---|
| ρ (rho) | The Spearman correlation, read exactly like r. |
| Stars | Significance, as with Pearson. |
The situation. The same 80 learners also rated their motivation from 1 to 5.
What AnalyZ shows. Spearman ρ between motivation and exam score = 0.43 (***).
How to read it. Higher motivation goes with higher scores, a moderate link. Because motivation is a 1-to-5 rating, Spearman is the right choice. Pearson would treat the gap between 1 and 2 as equal to the gap between 4 and 5, which ratings do not guarantee.
Illustrative figures.
Spearman Correlation: things to watch for
- Spearman looks for a steadily rising or falling pattern. It does not detect U-shapes.
- Many ties (lots of people with the same rating) reduce precision slightly.
Spearman Correlation: key terms
- Rank
- A value's position when all values are sorted.
- Monotone
- Always rising, or always falling, but not necessarily in a straight line.
Regression
Regression answers the question what goes with the outcome? when several things could matter at once. It looks for the pattern that best links your outcome to a set of predictors, and reports how much the outcome changes for each predictor while the other predictors are held steady. That last part is the strength of regression: it separates the effect of schooling from the effect of age, for instance, even when older people tend to have less schooling.
Four ideas that apply to every regression
| Idea | In plain words |
|---|---|
| Outcome and predictors | The outcome (dependent variable) is what you want to explain. The predictors (independent variables) are the things you think are related to it. |
| The coefficient | One number per predictor: how much the outcome changes when that predictor goes up by one unit, with everything else held constant. It may be shown as B, an odds ratio or a rate ratio, depending on the model. |
| Categorical predictors and the reference category | A category variable is turned into a set of yes/no comparisons against one reference category, which AnalyZ chooses as the first category by default. Use Change reference category if another group makes a more natural comparison (for example the largest group, or the untreated group). Every coefficient then reads 'compared with the reference'. |
| Fit and significance | Each predictor has its own p-value and 95% confidence interval. The model as a whole has a fit measure (R-squared for a numeric outcome, pseudo R-squared for others) saying how much of the outcome the predictors explain. |
Linear Regression
Use it when
- Your outcome is a number: income, score, weight, days to recovery
- You want the effect of each predictor with the others held constant
Your data needs
- A numeric outcome
- One or more predictors, numeric or categorical
- A reasonable sample: at least 10 to 20 people for every predictor
- A broadly straight-line relationship and no extreme outliers
Linear Regression: how to run it
Linear Regression: reading the results
| You see | It means |
|---|---|
| N | People used. Anyone missing on any variable is left out. |
| R² and Adj R² | Share of the variation in the outcome explained by all the predictors together. Adjusted R² penalises predictors that add nothing. Prefer it when comparing models of different sizes. |
| F and its p-value | Whether the predictors as a group explain anything. |
| B | The coefficient: change in the outcome for one more unit of the predictor, others held constant. For a category it is the gap from the reference category. |
| SE | How uncertain B is. |
| p-value and Sig | Whether that predictor's effect is distinguishable from zero. |
| 95% CI | The plausible range for B. If it includes 0, the predictor is not significant. |
| AIC | A fit score used to compare models on the same data. Lower is better. It has no meaning alone. |
The situation. 200 adults. Outcome: monthly income. Predictors: age in years, years of schooling, and sex (male is the reference).
| Predictor | B | SE | p-value | 95% CI |
|---|---|---|---|---|
| (Intercept) | 373.6 | 108.1 | 0.001 | 160.4 to 586.9 |
| Age | 19.8 | 2.2 | < 0.001 | 15.5 to 24.1 |
| Years of schooling | 51.3 | 5.7 | < 0.001 | 40.0 to 62.6 |
| Female (vs Male) | -95.0 | 37.0 | 0.011 | -167.9 to -22.1 |
What AnalyZ shows. R² = 0.44, Adj R² = 0.43, F(3, 196) = 51.79, p-value < 0.001.
How to read it. Each extra year of schooling goes with about 51 more in monthly income, and each extra year of age with about 20 more, holding the other predictors constant. Women earn about 95 less than men of the same age and schooling. All three intervals exclude zero. The model explains 44% of the variation in income, so more than half is down to things not in the model. The intercept is the predicted income for a man aged 0 with no schooling. It is a mathematical anchor, not a real person, so ignore it.
Illustrative figures.
Linear Regression: things to watch for
- Predictors that are very closely related to each other (for example height in cm and height in inches) make individual effects impossible to separate.
- A large coefficient is not a strong effect unless you consider the unit. An effect of 51 per year of schooling is bigger in practice than 20 per year of age, but compare like with like by looking at the spread of each predictor.
- Do not extend predictions beyond the range of your data.
- If the outcome is skewed (income often is), results can be driven by a few high values. Consider logging the outcome with Create Variable.
Linear Regression: key terms
- B
- The unstandardised coefficient, in the outcome's own units.
- R²
- The proportion of variation explained.
- Reference category
- The comparison group for a categorical predictor.
- Holding constant
- Comparing people who are alike on the other predictors.
Logistic Regression
Use it when
- The outcome has exactly two values: yes/no, 0/1, pass/fail
- You want to know which factors are linked to the outcome and how strongly
Your data needs
- A binary outcome with a decent number of cases in each category. A rough guide is at least 10 of the rarer outcome per predictor.
- One or more predictors, numeric or categorical
Logistic Regression: how to run it
Logistic Regression: reading the results
| You see | It means |
|---|---|
| OR (odds ratio) | 1 means no effect. Above 1: the odds of the event are higher. Below 1: lower. An OR of 1.5 is 50% higher odds, and 0.80 is 20% lower odds, for each one-unit increase. |
| 95% CI on the OR | If it includes 1, the predictor is not significant. |
| p-value, Sig, Wald z | Whether that predictor's effect is beyond chance. |
| LR chi2 and Prob > chi2 | Whether the predictors as a group improve on a model with none. |
| Pseudo R² | A rough fit measure. Values look small compared with linear R², and 0.2 to 0.4 is already considered a good fit. |
| AIC and Log-likelihood | Model-comparison scores. Lower AIC is better. |
| Accuracy % | The share of cases the model classifies correctly at a 50% cut-off. Compare it with simply guessing the commoner outcome. |
| Converged | If No, the estimates are not reliable. See the warnings below. |
The situation. 400 people were asked whether they attended a clinic in the last year (55% did). Predictors: distance to the clinic in km, and whether the household owns a means of transport.
| Predictor | OR | 95% CI | p-value |
|---|---|---|---|
| Distance to clinic (per km) | 0.80 | 0.73 to 0.88 | < 0.001 |
| Owns transport (vs does not) | 1.49 | 0.99 to 2.25 | 0.055 |
What AnalyZ shows. LR chi2 = 29.78, df = 2, p-value < 0.001. Pseudo R² = 0.05. Accuracy = 60%.
How to read it. Each extra kilometre lowers the odds of attending by 20%, a clear and significant effect. Owning transport raises the odds by 49%, but the interval (0.99 to 2.25) just touches 1, so there is no firm evidence of an effect in this sample. The pseudo R² is small, which is normal when behaviour has many causes. The accuracy of 60% is only modestly better than the 55% you would get by always predicting attendance, so use the model to understand the factors rather than to predict individuals.
Illustrative figures.
Logistic Regression: things to watch for
- Odds are not probabilities. Odds ratios exaggerate when the outcome is common, so do not say 'twice as likely' for an OR of 2 unless the outcome is rare. Say 'twice the odds'.
- A warning about separation or multicollinearity means a predictor perfectly splits the outcome or predictors overlap heavily. Remove or merge variables.
- Very few cases in one outcome category make estimates unstable.
Logistic Regression: key terms
- Odds
- Chance of the event divided by chance of no event. A 75% chance is odds of 3 to 1.
- Odds ratio
- How the odds change per unit of the predictor.
- Log-odds
- The raw scale the model works on. The B column.
Probit Regression
Use it when
- Your field or supervisor expects probit, as is common in economics and some health research
- You want to check that a logistic result does not depend on the model shape
Your data needs
- The same as logistic: a yes/no outcome and one or more predictors
Probit Regression: how to run it
Probit Regression: reading the results
| You see | It means |
|---|---|
| B | The change in the underlying 'probit score' per unit of the predictor. The sign tells you direction. The size has no everyday meaning. |
| p-value, Sig, 95% CI | Read exactly as in other models. |
| Model fit rows | Log-likelihood, LR chi2, pseudo R² and AIC, as in logistic regression. |
The situation. The clinic attendance data from the logistic example, run as a probit model.
What AnalyZ shows. Distance: B = -0.13, p-value < 0.001. Owns transport: B = 0.24, p-value 0.06.
How to read it. The signs and significance agree with the logistic result: distance lowers attendance and transport may raise it. You cannot say 'by how much' in plain terms from a probit coefficient. If you need a size that readers will understand, report the logistic odds ratios alongside or instead.
Illustrative figures.
Probit Regression: things to watch for
- Probit coefficients are on a different scale from logistic ones (roughly 1.6 to 1.8 times smaller). Do not compare the numbers across the two models.
Probit Regression: key terms
- Probit
- A model for yes/no outcomes using the normal curve.
Ordinal Regression
Use it when
- The outcome is a Likert item or a severity or grade rating
- Treating the scale as a plain number (linear regression) seems doubtful
Your data needs
- An ordered outcome with 3 or more categories
- One or more predictors
- Enough people in each category
Ordinal Regression: how to run it
Ordinal Regression: reading the results
| You see | It means |
|---|---|
| OR for each predictor | The change in the odds of being in a higher category rather than a lower one, per unit of the predictor. Above 1 pushes people up the scale. Below 1 pushes them down. |
| Threshold rows (/cut1, /cut2 …) | Where the scale divides between categories. They are mechanical and rarely reported. |
| p-value, 95% CI, Sig | As in other models. |
| LR chi2, pseudo R², AIC | Overall fit. |
The situation. 500 patients rated service as Low, Medium or High. Predictors: waiting time in 10-minute blocks, and whether the patient has insurance.
| Predictor | OR | 95% CI | p-value |
|---|---|---|---|
| Waiting time (per 10 min) | 0.78 | 0.70 to 0.87 | < 0.001 |
| Insured (vs not insured) | 1.90 | 1.30 to 2.80 | 0.001 |
How to read it. Each extra 10 minutes of waiting cuts the odds of a higher rating by 22%. Insured patients have 90% higher odds of rating the service higher. The word 'higher' means any step up, whether Low to Medium or Medium to High. One odds ratio summarises all the steps.
Illustrative figures.
Ordinal Regression: things to watch for
- The model assumes each predictor has the same effect at every step (the proportional odds assumption). If a predictor matters a lot for Low versus Medium but not for Medium versus High, a single odds ratio hides this.
- Do not use it for categories with no natural order. Use Multinomial.
Ordinal Regression: key terms
- Proportional odds
- The assumption that a predictor's effect is the same across every step of the scale.
- Threshold
- The cut-points between adjacent categories.
Multinomial Regression
Use it when
- The outcome is a choice or type: programme chosen, main occupation, source of water
- Categories cannot be ranked
Your data needs
- An outcome with 3 or more unordered categories
- One or more predictors
- A decent number of people in every category
Multinomial Regression: how to run it
Multinomial Regression: reading the results
| You see | It means |
|---|---|
| Outcome blocks | One block of rows per non-reference category. A block answers: what predicts choosing this category over the reference? |
| RRR (relative risk ratio) | Above 1: the predictor raises the chance of choosing this category relative to the reference. Below 1: lowers it. Use the switch to see the raw coefficient B instead. |
| p-value, 95% CI, Sig | As in other models. |
| LR chi2, pseudo R², AIC | Overall fit. |
The situation. 600 trainees chose Farming, Vocational or Business training. Farming is the reference. Predictors: age and sex.
| Outcome vs Farming | Predictor | RRR | p-value |
|---|---|---|---|
| Vocational | Age (per year) | 0.94 | < 0.001 |
| Vocational | Female (vs male) | 1.10 | 0.55 |
| Business | Age (per year) | 1.01 | 0.62 |
| Business | Female (vs male) | 2.10 | 0.002 |
How to read it. Older trainees are less likely to choose Vocational over Farming (6% lower per year of age). Women are more than twice as likely as men to choose Business over Farming. Sex makes no clear difference to choosing Vocational over Farming, and age makes no clear difference to choosing Business. Notice that every statement is a comparison with Farming. Nothing here says anything about Vocational against Business.
Illustrative figures.
Multinomial Regression: things to watch for
- The results change meaning if you change the reference category, so state the reference when you report.
- With many categories and few cases per category, estimates are unstable. Merge rare categories.
Multinomial Regression: key terms
- Reference category
- The outcome everything else is compared with.
- Relative risk ratio
- How the chance of one outcome relative to the reference changes per unit of the predictor.
Poisson Regression
Use it when
- The outcome counts events
- Counts are not wildly spread out (the variance is close to the average)
Your data needs
- A count outcome (0, 1, 2, …) with no negative or decimal values
- One or more predictors
Poisson Regression: how to run it
Poisson Regression: reading the results
| You see | It means |
|---|---|
| IRR | 1 means no effect. 1.30 means the count is 30% higher for each extra unit (or for that category compared with the reference). 0.80 means 20% lower. |
| 95% CI, p-value, Sig | As in other models. |
| Dispersion (Pearson chi-square / N) | Close to 1 is fine for Poisson. Above about 1.5 the counts are more spread than Poisson allows, and the p-values will look too good. Switch to Negative Binomial. |
The situation. 300 households reported clinic visits in a year (average 2.9, but a variance of 8.3). Predictors: an asset score (per standard deviation) and whether the household is in a health scheme.
| Predictor | IRR | 95% CI | p-value |
|---|---|---|---|
| Asset score (per SD) | 1.33 | 1.24 to 1.44 | < 0.001 |
| In health scheme (vs not) | 1.30 | 1.14 to 1.49 | < 0.001 |
What AnalyZ shows. Dispersion = 2.56.
How to read it. The IRRs suggest each standard deviation of assets goes with 33% more visits and scheme members have 30% more. But the dispersion of 2.56 is well above 1.5, so the variance is far larger than the average and these intervals are too narrow. Move to Negative Binomial and trust that result instead.
Illustrative figures.
Poisson Regression: things to watch for
- A count outcome measured over different time spans or population sizes needs those taken into account. Compare like with like, such as visits per year.
- Many more zeros than expected can also cause trouble. Check with Frequency.
Poisson Regression: key terms
- IRR
- Incidence rate ratio: how the average count changes per unit of a predictor.
- Overdispersion
- More variation in counts than Poisson allows.
Negative Binomial Regression
Use it when
- A Poisson run flagged overdispersion
- Your counts are very uneven, with many small values and a few big ones
Your data needs
- The same as Poisson
Negative Binomial Regression: how to run it
Negative Binomial Regression: reading the results
| You see | It means |
|---|---|
| IRR, 95% CI, p-value | Read as in Poisson, but with intervals that honestly reflect the extra variation. |
| Alpha | Close to 0 means Poisson would have done. Larger values mean more overdispersion. |
The situation. The same 300 households.
| Predictor | IRR | 95% CI | p-value |
|---|---|---|---|
| Asset score (per SD) | 1.36 | 1.20 to 1.53 | < 0.001 |
| In health scheme (vs not) | 1.33 | 1.07 to 1.64 | 0.010 |
What AnalyZ shows. Alpha = 0.53.
How to read it. The IRRs are similar to Poisson, but the intervals are wider. The health scheme effect (1.07 to 1.64) is now much less certain, and its p-value rose from below 0.001 to 0.010. This is the honest picture. The Poisson result overstated how sure we were. Alpha of 0.53 confirms substantial overdispersion.
Illustrative figures.
Negative Binomial Regression: things to watch for
- AnalyZ uses the common NB2 form, in which the variance equals the mean plus Alpha times the mean squared.
Negative Binomial Regression: key terms
- Alpha
- The extra spread in counts beyond Poisson.
Hierarchical Regression
Use it when
- You want to know whether a set of variables adds anything beyond what you already account for
- You have a theory about the order in which predictors matter
Your data needs
- A numeric outcome
- Two or more blocks of predictors
Hierarchical Regression: how to run it
Hierarchical Regression: reading the results
| You see | It means |
|---|---|
| R² for each block | The explained share at each stage. |
| ΔR² | How much explanatory power the new block added. |
| ΔF and its p-value | Whether the added power is more than chance would give. |
| Block coefficients | The predictor effects in each stage, which show how effects shift as variables are added. |
The situation. The income model, built in two blocks: Block 1 = age and sex, Block 2 adds years of schooling.
| Model | R² | ΔR² | ΔF | p-value |
|---|---|---|---|---|
| Block 1: age, sex | 0.213 | |||
| Block 2: + schooling | 0.442 | 0.229 | 80.4 | < 0.001 |
How to read it. Age and sex alone explain 21% of the variation in income. Adding schooling lifts this to 44%, so schooling adds 23 percentage points, a substantial and significant improvement. If schooling had added 1% the verdict would be that it does not matter much beyond age and sex.
Illustrative figures.
Hierarchical Regression: things to watch for
- The order of blocks changes what each block appears to add. Put controls first and the variables of interest later, and justify the order.
Hierarchical Regression: key terms
- Block
- A group of predictors entered together.
- ΔR²
- The extra variation explained by a block.
Multilevel Regression
Use it when
- People are nested in schools, clinics, villages or districts
- You have repeated measurements on the same people
- You want group-level and individual-level predictors in one model
Your data needs
- An outcome that is numeric, binary or a count
- A variable that says which group each case belongs to, for each level
- Several groups. Fewer than 5 groups give unreliable group-level estimates, and 20 or more are comfortable.
Multilevel Regression: how to run it
Multilevel Regression: reading the results
| You see | It means |
|---|---|
| Fixed effects | The average effect of each predictor across all groups: B for continuous outcomes, odds ratio for binary, rate ratio for counts. Read them like ordinary regression. |
| Variance components | How the unexplained variation splits between groups and individuals. |
| ICC | The share of variation that lies between groups. Above about 0.05 means grouping matters and an ordinary regression would be misleading. |
| AIC, BIC, log-likelihood | Fit scores for comparing models. |
| Random intercept BLUPs | Each group's own baseline relative to the average, after accounting for the predictors. |
The situation. 1,200 pupils in 40 schools took a maths test. The predictor is weekly study hours.
| Part | Estimate |
|---|---|
| Fixed effect: study hours | B = 2.3, SE 0.4, p-value < 0.001 |
| Variance between schools | 24.6 |
| Variance between pupils within schools | 112.1 |
| ICC | 0.18 |
How to read it. Each extra weekly study hour goes with 2.3 more marks on average. The ICC of 0.18 says 18% of the variation in marks lies between schools and 82% between pupils in the same school, so schools matter and a model that ignored them would give intervals that are too narrow. Without the multilevel model you would be treating 1,200 pupils as 1,200 independent pieces of evidence, when they behave partly like 40 clusters.
Illustrative figures.
Multilevel Regression: things to watch for
- With very few groups the group variances are very uncertain and are often estimated at zero.
- Binary and count models use an approximation (penalised quasi-likelihood), which can understate group variation in small samples or when the ICC is high. For high-stakes results, check against R or Stata.
- A group-level predictor, such as school size, can only be estimated from the number of groups, not the number of pupils.
Multilevel Regression: key terms
- ICC
- Intraclass correlation: the share of variation between groups.
- Random intercept
- A different starting level for each group.
- Fixed effect
- The average effect of a predictor across all groups.
- Nested
- Each lower-level unit belongs to exactly one higher-level unit.
Non-parametric Tests
Non-parametric tests replace the raw values with their ranks (1st, 2nd, 3rd and so on) before comparing groups. That makes them steady against skewed data, outliers and ordinal answers. Each one is the counterpart of a test in Compare Means, and the cost is slightly less power when the data really are bell-shaped.
| Instead of | Use | When |
|---|---|---|
| Independent t-test | Mann-Whitney U | Two separate groups |
| Paired t-test | Wilcoxon Signed-Rank | Two measurements on the same people |
| One-way ANOVA | Kruskal-Wallis | Three or more separate groups |
| Repeated-Measures ANOVA | Friedman Test | Three or more measurements on the same people |
Mann-Whitney U
Use it when
- You are comparing two separate groups on an ordinal rating or a skewed number
- Samples are small and normality is doubtful
Your data needs
- A numeric or ordinal variable
- A grouping variable with exactly two groups
Mann-Whitney U: how to run it
Mann-Whitney U: reading the results
| You see | It means |
|---|---|
| Median and mean rank per group | The median is the typical value. The mean rank shows which group sits higher overall. Higher mean rank means higher scores. |
| U | The test statistic. Smaller means the groups overlap less. |
| p-value and Sig | Whether the difference in rank positions is beyond chance. |
| Effect size r | 0.1 small, 0.3 medium, 0.5 large. |
The situation. 60 people rated their satisfaction from 1 to 5 after a service delivered either online (30) or in person (30).
| Group | N | Median | Mean rank |
|---|---|---|---|
| Online | 30 | 3 | 25.8 |
| In person | 30 | 4 | 35.2 |
What AnalyZ shows. U = 308, p-value 0.030 (*), effect size r = 0.28.
How to read it. People served in person gave higher ratings, with a median of 4 against 3 and a higher mean rank (35.2 against 25.8). The difference is significant, and r = 0.28 is a small-to-medium effect. A t-test would have averaged the 1-to-5 codes, which assumes the steps are equal. Mann-Whitney does not.
Illustrative figures.
Mann-Whitney U: things to watch for
- It compares whole distributions. If the groups have very different shapes, a significant result does not strictly mean their medians differ.
- Many tied values, such as a 5-point scale with 60 people, are handled but reduce precision slightly.
Mann-Whitney U: key terms
- Rank
- A value's position in the sorted list of all values.
- Mean rank
- The average rank of a group's members.
Wilcoxon Signed-Rank
Use it when
- Paired measurements that are ordinal, skewed, or from a small sample
- The paired t-test's bell-shape assumption looks doubtful
Your data needs
- Two numeric or ordinal columns for the same people
- Pairs with a missing value are dropped automatically
Wilcoxon Signed-Rank: how to run it
Wilcoxon Signed-Rank: reading the results
| You see | It means |
|---|---|
| Medians | The typical score at each time. |
| W | The test statistic: the smaller of the sums of ranks for increases and decreases. |
| p-value and Sig | Whether the typical change is beyond chance. |
| Effect size r | 0.1 small, 0.3 medium, 0.5 large. |
The situation. 28 participants were scored before and after a short course.
| Time | Median |
|---|---|
| Before | 50 |
| After | 55 |
What AnalyZ shows. W = 19, p-value < 0.001 (***), effect size r about 0.8.
How to read it. The median rose from 50 to 55, and nearly everyone improved: a small W means the 'down' changes were few and small. The effect is large. Unlike a t-test, a single huge change by one person would not have swung the result.
Illustrative figures.
Wilcoxon Signed-Rank: things to watch for
- It uses the size ranking of changes, so it needs differences that can be ranked. Many people with exactly no change reduce its power.
Wilcoxon Signed-Rank: key terms
- Signed rank
- The rank of a change's size with its direction (up or down) attached.
Kruskal-Wallis
Use it when
- Three or more separate groups on an ordinal or skewed outcome
- Small or uneven groups where ANOVA's assumptions are doubtful
Your data needs
- A numeric or ordinal variable
- A grouping variable with 3 or more groups
Kruskal-Wallis: how to run it
Kruskal-Wallis: reading the results
| You see | It means |
|---|---|
| Median and mean rank per group | The typical value and relative position of each group. |
| H | The test statistic. Larger means the groups are further apart. |
| p-value and η² | Whether the groups differ, and how much of the ranking is explained: 0.01 small, 0.06 medium, 0.14 large. |
| Post-hoc comparisons (Dunn, Bonferroni-adjusted) | Each pair of groups compared, with p-values made stricter for the number of pairs. Use this table to find which groups differ. |
The situation. 75 workers in three age bands rated their job stress (25 per band).
| Band | N | Median stress score |
|---|---|---|
| Under 30 | 25 | 49 |
| 30 to 45 | 25 | 54 |
| Over 45 | 25 | 67 |
What AnalyZ shows. H = 29.2, p-value < 0.001 (***), η² about 0.38. In the Dunn table, Over 45 differs from both other bands. Under 30 and 30 to 45 do not differ from each other.
How to read it. Stress is clearly higher among workers over 45, and this stands out against the other two groups, which are similar. The effect is large. A significant H alone would only say 'somewhere there is a difference', so the pair table is where the finding is.
Illustrative figures.
Kruskal-Wallis: things to watch for
- A significant overall result needs the pair table to say where the difference is.
- The Bonferroni adjustment is cautious, which guards against false positives at the cost of some sensitivity.
Kruskal-Wallis: key terms
- Dunn test
- Pairwise comparisons that follow a Kruskal-Wallis test.
- Bonferroni
- A way of making p-values stricter when many pairs are compared.
Friedman Test
Use it when
- Three or more measurements on the same people, with ordinal or skewed outcomes
- Ratings of several items by the same respondents
Your data needs
- Three or more numeric or ordinal columns
- One row per person, with every measurement present
Friedman Test: how to run it
Friedman Test: reading the results
| You see | It means |
|---|---|
| Median per condition | The typical value at each time or condition. |
| Chi-square and p-value | Whether the conditions differ. |
| Kendall's W | Agreement across people, from 0 (none) to 1 (everyone ranks the conditions the same way). It is the effect size. |
| Pairwise Wilcoxon (Bonferroni) | Each pair of conditions compared. Pairs marked ns still show their adjusted p-value, worth checking when borderline. |
The situation. 30 people were scored at baseline, midline and endline.
What AnalyZ shows. Chi-square = 41.3, p-value < 0.001 (***), Kendall's W = 0.69.
How to read it. W = 0.69 means most people showed the same ordering, with baseline lowest and endline highest. The pairwise table then tells you which steps are real. Use this instead of repeated-measures ANOVA when the scores are ordinal or skewed.
Illustrative figures.
Friedman Test: things to watch for
- Each row must be the same person in every column, or the test is not valid.
Friedman Test: key terms
- Kendall's W
- The degree of agreement between people in how they rank the conditions.
Survey weights
Every test in Inferential Analysis can use the survey weight set in Data View. The Apply survey weight option beside Generate is ticked by default. The tests take the survey design into account (weights, strata and clusters), so p-values and confidence intervals reflect how the sample was drawn. A filter is treated as a subgroup of the full sample.
| Tests | What changes |
|---|---|
| t-tests | Means are weighted, and the standard error and degrees of freedom come from the survey design. The independent t-test also shows the mean difference with its 95% interval and no longer shows the Levene test, because it does not assume equal variances. |
| ANOVA | The F test is design-based. One-way and two-way ANOVA show weighted sums of squares and eta squared, and pairwise comparisons are Bonferroni adjusted. Repeated-measures ANOVA uses a joint test that needs no sphericity assumption. |
| Chi-square test | Replaced by the Rao-Scott corrected F test, with weighted frequencies. |
| Z-test for Proportion | The proportion is weighted. The test uses a t statistic with the design degrees of freedom in place of z. |
| Pearson and Spearman correlation | Correlations are weighted and the test of no correlation uses the survey design. Spearman is worked out on population ranks estimated from the weights. |
| Mann-Whitney U and Kruskal-Wallis | Values are ranked by their estimated position in the population. The tests compare the groups' average rank scores with the survey design and report t or F instead of U or H. This follows the approach of Lumley and Scott (2013). |
| Wilcoxon Signed-Rank and Friedman | Built on the same principle. Wilcoxon scores each difference by its sign and the estimated population rank of its size, and Friedman compares weighted average ranks within persons. Both use the survey design and report t or F. |
| Regression models | Coefficients are estimated with the weights, and the standard errors, confidence intervals and p-values come from the survey design. A design-based F test replaces the likelihood-ratio test, adjusted R-squared, AIC and BIC are not shown, and Weighted N appears next to N. |
| Multilevel Regression | The survey weight is used within groups, and groups are treated as equally likely to be selected, as in Stata with a single weight. With two levels you can also enter the weight of the group itself and choose how the within-group weights are rescaled. Two to four levels are supported, and four-level models with random slopes are limited to continuous outcomes. With few groups the group variances are very uncertain and are often estimated at zero. |
📖 Reading the output
All tests use standard significance thresholds: *** p < .001, ** p < .01, * p < .05, ns p ≥ .05. Each result section ends with a plain-language interpretation paragraph suitable for direct use in a report.
Effect Analysis
Inferential tests ask is X related to Y? Effect Analysis asks something harder: did our programme cause this change? That is difficult because we can never see the same people both with and without the programme. The people who joined a programme are usually different from those who did not (more motivated, closer to the road, poorer), so simply comparing the two groups mixes the effect of the programme with the effect of those differences.
The four methods here are different ways of building a fair comparison group, which statisticians call the counterfactual: what would have happened to the participants if the programme had not existed. Each method rests on a different assumption, and the guide spells out what that assumption is, how to check it, and what to do when it is doubtful.
| Method | The idea in a sentence | Use when |
|---|---|---|
| Propensity Score Matching | Find non-participants who look just like participants, and compare those. | One round of data, no random assignment, and good information on how participants and non-participants differ. |
| Difference-in-Differences | Compare how much each group changed, not where each ended up. | A programme and a comparison group, each measured before and after. |
| Regression Discontinuity Design | Compare people just above and just below an eligibility cut-off. | Eligibility is decided by a score, age or income threshold. |
| Instrumental Variables (2SLS) | Use something that nudges participation but has no effect on the outcome by any other route. | Take-up is tied up with unmeasured factors and you have a credible instrument. |
Propensity Score Matching
Use it when
- Participation was not randomly assigned
- You have only one round of data but good information on who joined and who did not
- You can name the characteristics that drive both joining and the outcome
Your data needs
- A binary treatment variable (1 = participated, 0 = did not)
- A numeric or binary outcome
- Covariates measured before the programme, for example age, sex, schooling. Covariates affected by the programme will bias the answer.
Propensity Score Matching: how to run it
Propensity Score Matching: reading the results
| You see | It means |
|---|---|
| Treatment effect estimate(s) | The headline number: the average difference in the outcome between participants and their matched counterparts, with standard error, p-value and 95% CI. |
| Matching summary | How many participants were matched, and how many controls were used or left out. |
| Covariate balance (SMD before and after) | For each covariate, how different the two groups were before matching and after. An absolute standardised mean difference below 0.1 is good balance. |
| Matching quality | How many covariates reached good balance, and the percentage of bias removed by matching. |
| Rubin's B and R | A second check on the propensity scores themselves. B (standardised bias) should be under 25, R (variance ratio) between 0.5 and 2. |
| Propensity score model fit | Pseudo R² for the model that predicts participation. About 0.2 to 0.4 suggests your covariates separate participants from others reasonably well. This is not the quality of the effect estimate. |
| Sensitivity analysis (E-value) | How strong an unmeasured factor would have to be, in its link to both participation and the outcome, to explain the whole effect away. Bigger is more reassuring. |
| Bias grid | How the estimate would shift under a range of hypothetical hidden confounding strengths. |
Choosing an estimation technique
| Technique | How it works | When to prefer it |
|---|---|---|
| Nearest neighbor matching | Pairs each participant with the single most similar non-participant, as long as they are close enough (within the caliper). Participants with no close partner are dropped. | You want a transparent, easy-to-explain method and have plenty of comparison people. |
| Kernel matching | Builds each participant's counterfactual from all non-participants, giving more weight to the most similar. Nobody is dropped. | You have few comparison people, or want to use all the data. |
| Inverse probability weighting (IPW) | Re-weights the whole sample so that the two groups resemble each other. People who looked unlikely to join but did are given more weight. | You want to use everyone and avoid explicit pairs. |
| Doubly robust (AIPW) | Combines propensity weighting with an outcome regression. It gives a correct answer if either of the two models is right. | You want protection against a mistake in one of the two models. A good default when sample size allows. |
For nearest neighbor matching, the caliper is the largest allowed difference between a participant and their match. It is set as a share of the standard deviation of the propensity score on the logit scale, with 0.2 as the usual default. A tighter caliper gives closer matches but drops more people. For kernel matching you can set the bandwidth. IPW and doubly robust results use a bootstrap for their standard errors, so you set the number of Bootstrap replications (more means slower and steadier). For IPW you can also choose stabilised or unstabilised weights, and stabilised weights are less erratic.
Choosing an estimand
| Estimand | The question it answers | Typical choice when |
|---|---|---|
| ATT | What was the effect on those who actually took part? | Most evaluations. You want to know whether the programme worked for its participants. |
| ATC | What would the effect have been on those who did not take part? | You are considering extending the programme to non-participants. |
| ATE | What is the effect across the whole population? | The programme could be applied to anyone. |
The situation. A job-training programme enrolled 300 people. 700 others did not enrol. The outcome is monthly earnings six months later. Covariates: age, years of schooling, prior earnings and household size. Method: nearest neighbor matching, estimand ATT.
| Covariate | SMD before | SMD after |
|---|---|---|
| Age | 0.35 | 0.04 |
| Years of schooling | 0.41 | 0.06 |
| Prior earnings | 0.52 | 0.05 |
| Household size | -0.12 | -0.03 |
What AnalyZ shows. ATT = +42 per month (95% CI 12 to 72, p-value 0.006). All four covariates have an absolute SMD below 0.1 after matching, so balance is good. Rubin's B = 12 and R = 1.1, both within the target. Propensity model pseudo R² = 0.12. E-value = 1.8.
How to read it. Before matching, participants were younger, better educated and had higher earnings before the programme than non-participants, so a raw comparison would have overstated the benefit. After matching the groups look alike. Participants earn about 42 more per month than comparable non-participants, and the interval stays above zero. The E-value of 1.8 says an unmeasured factor would need to be associated with both participation and earnings by a factor of about 1.8 to explain the whole effect away. That is moderate protection, not strong, because motivation, which you did not measure, could plausibly be that strong. Report the finding with that caveat.
Illustrative figures.
Propensity Score Matching: things to watch for
- Matching can only correct for characteristics you measured. Unmeasured ones, such as motivation, can still distort the result. That is what the E-value is for.
- Include only covariates measured before the programme started.
- If many participants have no close match, you are comparing only the overlapping part of the population. Say so.
- Do not read the propensity model's fit as the quality of your result. A low pseudo R² only says that background does not predict participation strongly.
Propensity Score Matching: key terms
- Propensity score
- A person's estimated probability of being in the programme.
- SMD
- Standardised mean difference: a gap between groups in standard deviations.
- Caliper
- The largest allowed distance between a person and their match.
- Counterfactual
- What would have happened without the programme.
Difference-in-Differences
Use it when
- You have baseline and endline data on a programme group and a comparison group
- You worry that the groups started at different levels (which DiD allows) but think they would have moved in parallel without the programme
Your data needs
- A group variable (programme = 1, comparison = 0)
- A round variable with two values (before and after)
- A numeric or binary outcome
- Optional covariates. Rounds can be in one long file or a second file you upload.
Difference-in-Differences: how to run it
Difference-in-Differences: reading the results
| You see | It means |
|---|---|
| Group, Round and Group × Round coefficients (robust SE) | Group is the baseline gap between the groups. Round is the change over time in the comparison group. Group × Round is the programme effect: the extra change in the programme group. |
| Raw 2 × 2 difference | The same calculation done directly from the four averages. It matches the regression exactly when there are no covariates, which makes it a useful cross-check. |
| Model summary | R², adjusted R² and the F statistic. |
| For a binary outcome | A logistic model is also fitted, and the effect is given as an average marginal effect, a change in probability, not the raw interaction term, which would not be the right DiD estimate. |
The situation. A savings-group programme ran in some villages. Household income was measured before and after, in 150 programme and 150 comparison households.
| Group | Before | After | Change |
|---|---|---|---|
| Programme | 100 | 130 | +30 |
| Comparison | 98 | 112 | +14 |
| Difference-in-differences | +16 |
What AnalyZ shows. Group × Round = 16.0 (robust SE 5.2, p-value 0.002, 95% CI 5.8 to 26.2). The raw 2 × 2 estimate is also 16.
How to read it. Income rose in both groups, by 14 in the comparison villages simply from the passage of time. Programme households gained 30, which is 16 more. The extra 16 is the estimated programme effect. A simple before-and-after look at the programme group alone would have claimed 30, nearly double the credible figure. A simple end-of-programme comparison would have claimed 18 (130 vs 112), which wrongly ignores the 2-point head start the programme group had.
Illustrative figures.
Difference-in-Differences: things to watch for
- Anything that changes for one group between the two times, such as a new road or a second programme, is mixed into the estimate.
- People lost between rounds (attrition) can bias the result if the dropouts differ between groups.
- The second-round file needs the same column names as round 1 for the ID, outcome and covariates.
Difference-in-Differences: key terms
- Parallel trends
- The assumption that both groups would have followed the same path without the programme.
- Interaction
- Here, the extra change belonging to the programme group in the later round.
- Baseline and endline
- The first and last measurements.
Regression Discontinuity Design
Use it when
- Eligibility is decided by a threshold on a score, index, income level or age
- You have data on people near the threshold
Your data needs
- A running variable: the continuous score that defines the cut-off
- A numeric or binary outcome
- Enough people close to the cut-off
- For a fuzzy design, a variable showing who actually received the programme
Regression Discontinuity Design: how to run it
Regression Discontinuity Design: reading the results
| You see | It means |
|---|---|
| Treatment effect at the cutoff | The size of the jump in the outcome at the threshold: the effect for people right at the line. |
| Manipulation check (density at the cutoff) | Tests whether unusually many people sit just above or just below the cut-off, a sign that scores were manipulated or people sorted themselves. A significant result is a warning. |
| Sorting check (counts within the bandwidth) | Compares how many observations are just below and just above. |
| Covariate balance at the cutoff | Background variables should not jump at the cut-off. If they do, something besides the programme changes there too. |
| Bandwidth sensitivity | The estimate re-run with windows from half to double the chosen width. A stable estimate is reassuring. Big swings mean the result is fragile. |
| Placebo cutoff test | The same test at fake cut-offs where nothing should happen. A significant placebo is a warning. |
| First stage (fuzzy only) | How strongly the cut-off predicts who actually got the programme. If the jump is small, the estimate is unstable. |
| Model fit | R², adjusted R² and F for the local regression within the bandwidth. |
The situation. A scholarship goes to students with an entrance score of 60 or more. The outcome is a test score two years later. A sharp design, automatic bandwidth (8 points either side of 60).
What AnalyZ shows. Effect at the cutoff = +6.2 points (SE 2.1, p-value 0.003). Manipulation test: not significant (p-value 0.47). Covariate balance for family income and age: no jump (p-values 0.61 and 0.38). Bandwidth sensitivity: estimates from 4.9 to 7.0 across half to double the bandwidth. Placebo cutoffs at 50 and 70: not significant.
How to read it. Students who just scraped in did about 6 points better two years later than those who just missed out. The checks all support this. The scores do not bunch around 60, so nobody seems to have gained access by manipulation. Family income and age do not jump at the line, so the two sides really are alike. The estimate holds from 4.9 to 7.0 when the window is changed, and fake cut-offs produce nothing. The estimate applies to students near the cut-off, not to high or low scorers.
Illustrative figures.
Regression Discontinuity Design: things to watch for
- The result is local. It tells you about people near the threshold, not everyone.
- A fuzzy design with a weak first stage gives a shaky estimate. AnalyZ warns you if the first-stage jump is small.
- If people can influence their own score to land just above the line, the design fails. The manipulation check looks for this.
Regression Discontinuity Design: key terms
- Running variable
- The score that decides eligibility.
- Cutoff
- The threshold value.
- Bandwidth
- How close to the cutoff someone must be to be used.
- LATE
- Local average treatment effect: the effect for those whose treatment was changed by crossing the cutoff.
Instrumental Variables (2SLS)
Use it when
- Participation is tied up with unmeasured factors such as motivation, ability or reverse causation
- You have a credible instrument you can defend on substantive grounds
Your data needs
- A numeric or binary outcome
- A numeric or binary endogenous treatment variable
- At least one instrument
- Optional exogenous covariates
Instrumental Variables: how to run it
Instrumental Variables: reading the results
| You see | It means |
|---|---|
| First-stage strength (F statistic) | How strongly the instrument predicts the treatment. Above 10 is the usual 'not weak' threshold. Below that the 2SLS estimate becomes unreliable. |
| 2SLS estimate | The headline effect of the treatment on the outcome, using only the variation the instrument creates. |
| OLS versus 2SLS | The naive estimate beside the corrected one. A big gap shows how much the hidden factors were distorting the simple comparison. |
| Reduced form | The outcome regressed directly on the instrument, which is a building block of the final estimate. |
| Endogeneity test (Durbin-Wu-Hausman) | Whether the treatment really is tied up with hidden factors. A significant result supports using 2SLS rather than ordinary regression. |
| Overidentification test (Sargan) | Only shown with more than one instrument. Tests whether the instruments agree with each other. A significant result suggests at least one may be invalid. |
| Model fit | First-stage R² and adjusted R², with a 2SLS R² for reference only, because it is not a reliable quality measure for 2SLS. |
The situation. A financial-skills workshop was open to all, but only some chose to attend. The programme also mailed invitation letters to a randomly chosen half of households. Instrument: received an invitation. Outcome: monthly income. Treatment: attended the workshop.
What AnalyZ shows. First-stage F = 38. OLS effect = +85 per month. 2SLS effect = +52 per month (SE 18, p-value 0.004). Endogeneity test p-value 0.03.
How to read it. The invitation strongly predicted attendance (F = 38, well above 10), so it is not a weak instrument. The naive OLS estimate of +85 is much larger than 2SLS (+52), and the endogeneity test confirms that attendance was tied up with hidden factors. Plausibly, motivated people both attend and earn more. Correcting for this brings the effect down to +52. This applies to people whose attendance was influenced by the invitation.
Illustrative figures.
Instrumental Variables: things to watch for
- A weak instrument (F below 10) gives estimates that are unreliable and standard errors that are too small.
- A valid instrument is hard to find. If there is any route from the instrument to the outcome other than through the treatment, the result is biased.
- 2SLS estimates are noisier than OLS. Wide intervals are normal.
Instrumental Variables: key terms
- Instrument
- A variable that affects take-up but not the outcome directly.
- Endogenous
- Tied up with hidden factors that also affect the outcome.
- Exclusion restriction
- The assumption that the instrument affects the outcome only through the treatment.
- First stage
- The step that predicts treatment from the instrument.
Survey weights
When a survey weight is set in Data View, all four methods use it by default. Untick Apply survey weight beside Run to compare with the unweighted result. Estimates are weighted, and uncertainty comes from the survey design (weights, strata and clusters) in place of the usual robust or bootstrap errors.
| Method | What changes |
|---|---|
| Propensity Score Matching | The propensity model is fitted with the survey weights, and matching, the balance checks (standardised differences and Rubin's B and R) and the effect estimates carry the weights. Nearest-neighbour matching uses the design variance of the weighted mean of matched differences. Kernel matching, inverse probability weighting and the doubly robust estimate use a bootstrap that resamples the survey's clusters within strata and re-estimates the propensity model each time. Weighted N for the treated and control groups appears in the model table. |
| Difference-in-Differences | The effect comes from weighted regression with design-based standard errors, and the cell means in the 2x2 table are weighted. For a binary outcome the weighted logit and the marginal DiD are shown, with the marginal effect's uncertainty from a bootstrap of the survey's clusters. The weight cannot be used when the second round is uploaded as a separate file, so put both rounds in one dataset. |
| Regression Discontinuity | A weighted local linear regression within the bandwidth, with design-based standard errors. The cross-validated bandwidth, density check, covariate balance, bandwidth sensitivity and placebo cutoffs all use the weights. |
| Instrumental Variables (2SLS) | The first stage, the 2SLS and naive OLS estimates and the reduced form are weighted, with design-based standard errors. The first-stage F statistic is a design-based Wald F, and the endogeneity test uses the same errors. The over-identification (Sargan) test is not shown with a survey weight, because it has no accepted design-based version. |
📖 Reading the output
All four methods report standard significance thresholds, a plain-language interpretation, and a model fit table appropriate to that method. For Propensity Score Matching, always read the balance diagnostics and sensitivity analysis alongside the point estimate. For Regression Discontinuity Design, treat the headline estimate as provisional until the manipulation test, covariate balance, and bandwidth sensitivity checks all come back clean. For Instrumental Variables, check the first-stage F-statistic before reading anything else, a weak instrument undermines every downstream number.
Structural Equation Modelling
Many things we care about, such as trust, wellbeing or food security, cannot be measured with one question. We ask several questions and treat them as signs of an underlying idea, called a factor or construct. Structural Equation Modelling (SEM) does two jobs together: it checks that your questions really measure the idea you intended (the measurement part), and it tests how those ideas affect each other (the structural part). Because it allows for measurement error, it gives cleaner effect estimates than averaging items and running a regression.
| Analysis | Use it when | Start here if you are new |
|---|---|---|
| Confirmatory Factor Analysis | You want to check that a set of questions measures the factors you expect. | Yes. Always check your scales here first. |
| Structural Equation Model | You want to test paths between factors (or between factors and measured variables). | After your CFA is sound. |
| Compare Groups | You want to know whether a scale means the same thing in every group, or whether effects differ by group. | Before comparing groups on a scale. |
| Path Analysis | You have measured variables only and want direct and indirect effects. | When you have no multi-item scales. |
| Mediation Analysis | You want to test whether X affects Y through one mediator M. | Simple, all-numeric chains. |
| Counterfactual Mediation | A mediator or outcome is binary or ordered, or you need covariates, several mediators or group results. | Anything beyond the simple chain. |
Things that apply to every analysis
| Topic | What to know |
|---|---|
| Numeric variables | All variables must be numeric. Likert items stored as numbers work well. |
| Variable types | Choose All continuous to treat items as numeric scores, or Mix of binary, ordinal and continuous to treat each by its number of different values: 2 is binary, 3 to 10 is ordinal, more than 10 is continuous. A table under the model shows how each variable will be treated. |
| Categorical predictors | Choose Categorical to code the variable into one 0/1 variable per category against a reference category you pick, or Continuous to use its values as numbers. |
| Survey weights | If you set a survey weight or design in Data View it is used and the weighted N is reported. Bootstrap intervals then resample whole clusters. |
| Subsample and missing data | The subsample filter applies. Only complete cases on the variables in the model are used, and the number used is shown with the results. |
| Sharing | Add to Canvas lets you pick which chart to send. Add to Sense Making sends the tables and interpretation. Word and Excel exports include all tables and charts. |
| Sample size | Aim for at least 10 cases per estimated parameter, never fewer than 100 cases, and at least 200 for structural equation models. |
Reading model fit
A model is a simplified story about how variables relate. Fit measures how closely that story reproduces the relationships actually seen in your data. A poor fit means the story leaves out something important.
| Measure | Good fit | In plain words |
|---|---|---|
| CFI and TLI | 0.95 or higher | How much better your model is than a model that assumes nothing is related. Closer to 1 is better. |
| RMSEA | 0.06 or lower | The average size of the mismatch between model and data, allowing for model simplicity. Smaller is better. |
| SRMR | 0.08 or lower | The average gap between the correlations in your data and those the model implies. Smaller is better. |
Confirmatory Factor Analysis
Use it when
- You use a multi-question scale and want proof it works
- You want reliability and validity evidence for a report
- You need to confirm two factors are really distinct
Your data needs
- Numeric items, such as Likert answers
- At least two items per factor (three for a single factor)
- Each item belongs to one factor
- At least 100 cases, and 10 per estimated parameter
Confirmatory Factor Analysis: how to run it
Confirmatory Factor Analysis: reading the results
| You see | It means |
|---|---|
| Fit summary (rated good, acceptable or poor) | Whether the structure you specified matches the data. See the fit table above. |
| Standardised loadings | How strongly each item reflects its factor, like a correlation. 0.50 or higher is reasonable. Below 0.40 the item is a weak signal. |
| Composite reliability | How consistently the items measure the factor. 0.70 or higher is acceptable. |
| Average variance extracted (AVE) | How much of the items' variation is captured by the factor rather than noise. 0.50 or higher is good. |
| Discriminant validity (HTMT) | Whether two factors are really different. A ratio above 0.85 means respondents are probably not telling them apart. |
| Suggested changes | Places where letting two items' errors relate would improve fit. Add them only if the question wording makes sense of it, and report every change. |
The situation. A 6-question scale: three questions on trust in health workers and three on trust in local government, answered by 420 people on a 1 to 5 scale. Two factors, three items each.
| Item | Factor | Standardised loading |
|---|---|---|
| Trust1 | Health workers | 0.78 |
| Trust2 | Health workers | 0.81 |
| Trust3 | Health workers | 0.69 |
| Trust4 | Government | 0.74 |
| Trust5 | Government | 0.80 |
| Trust6 | Government | 0.43 |
What AnalyZ shows. CFI 0.97, TLI 0.95, RMSEA 0.05, SRMR 0.04. Composite reliability 0.83 (health) and 0.76 (government). AVE 0.58 (health) and 0.50 (government). HTMT 0.62.
How to read it. Fit is good on all four measures. Five items load at 0.69 or more. Trust6 loads at 0.43, below the 0.50 guide, so it is a weak signal of government trust, though the factor is still acceptably reliable and its AVE just reaches 0.50. HTMT of 0.62 is well under 0.85, so the two kinds of trust are distinguishable. Next step: look at the wording of Trust6 and consider dropping it. Re-run to confirm fit and reliability hold.
Illustrative figures.
Confirmatory Factor Analysis: things to watch for
- Do not fit the structure to the data and then confirm it on the same data. Exploratory methods (EFA) are for discovering structure.
- Adding many suggested changes can make a model fit well for the wrong reasons. Each needs a substantive reason.
- Items with very uneven answers (almost everyone picks the same option) behave badly. Check them with Frequency first.
Confirmatory Factor Analysis: key terms
- Factor
- An underlying idea measured through several questions.
- Loading
- How strongly an item reflects its factor.
- AVE
- Average variance extracted: the share of item variation the factor explains.
- HTMT
- A ratio showing how distinct two factors are.
Structural Equation Model
Use it when
- You have a theory with several linked constructs
- You want effects that are not shrunk by measurement error
Your data needs
- Everything CFA needs
- At least 200 cases
- A clear theory about which factor affects which
Structural Equation Model: how to run it
Structural Equation Model: reading the results
| You see | It means |
|---|---|
| Structural paths | Each arrow's estimate with p-value. The standardised version tells you relative strength: about 0.1 small, 0.3 moderate, 0.5 large. |
| R² for each outcome | How much of that factor the model explains. |
| Effects route by route | Each way one factor reaches another: direct, through each middle factor, and total. |
| Residual diagnostics and suggested changes | Relationships the model leaves unexplained. Large ones suggest a missing path. |
| Diagram | Drag boxes to tidy it. Download any chart as PNG or SVG. |
The situation. Using the two trust factors from the CFA example, test whether trust in health workers predicts a service use factor (4 items), which in turn predicts wellbeing (5 items). N = 420.
| Path | Standardised effect | p-value |
|---|---|---|
| Health-worker trust → Service use | 0.46 | < 0.001 |
| Service use → Wellbeing | 0.34 | < 0.001 |
| Health-worker trust → Wellbeing (direct) | 0.12 | 0.09 |
What AnalyZ shows. CFI 0.95, RMSEA 0.05, SRMR 0.05. R² = 0.21 for service use and 0.18 for wellbeing.
How to read it. Fit is acceptable to good. Trust has a fairly strong effect on service use (0.46), and service use has a moderate effect on wellbeing (0.34). The direct path from trust to wellbeing is small and not significant, which suggests that trust works through service use rather than directly. The model explains about a fifth of each outcome, so much remains unexplained. To formally test the indirect route, use Mediation.
Illustrative figures.
Structural Equation Model: things to watch for
- Paths imply a direction, but the data cannot prove it. A good fit does not prove the story, only that it is not contradicted.
- Several different models can fit equally well. Prefer the one backed by theory.
- A model with very good fit and every path included is not impressive. See the step above.
Structural Equation Model: key terms
- Path
- An arrow from one variable or factor to another.
- Structural model
- The part of the model that links factors.
- Direct and indirect effect
- The part that goes straight, and the part that passes through a middle variable.
Compare Groups
Use it when
- You want to compare scale scores across countries, sexes or programme groups
- You want to know whether a path is stronger in one group than another
Your data needs
- A group variable with 2 to 8 groups
- At least 30 cases per group
- Items that already pass CFA overall
Compare Groups: how to run it
Compare Groups: reading the results
| You see | It means |
|---|---|
| Configural invariance | Same structure in every group. If supported, groups understand the scale in the same overall way. |
| Metric invariance | Loadings are equal. If supported, you can compare correlations and paths across groups. |
| Scalar invariance | Loadings and item intercepts are equal. If supported, you can also compare group averages on the factor. |
| Change in CFI, RMSEA, SRMR | A level is supported when fit barely changes from the one before. This matters more than the chi-square difference, which is nearly always significant in big samples. |
| Partial invariance table | Shows which loadings or intercepts differ most. You may free a few and re-run. |
The situation. The six-item trust scale was compared between women (n = 230) and men (n = 190).
| Level | CFI | Change in CFI | Verdict |
|---|---|---|---|
| Configural | 0.970 | Supported | |
| Metric | 0.968 | -0.002 | Supported |
| Scalar | 0.951 | -0.017 | Not supported |
How to read it. The scale has the same structure and the same loadings in both sexes, so you can compare how trust relates to other variables. But the scalar level fails, because CFI dropped by 0.017 (a drop above about 0.01 is the usual warning). That means that you cannot yet compare average trust between women and men. The partial invariance table points to Trust6 as the main culprit. Freeing its intercept and re-running may restore scalar invariance partially, which is usually enough. If many items needed freeing, the scale would not be measuring the same thing in both groups and no comparison would be safe.
Illustrative figures.
Compare Groups: things to watch for
- Do not compare group means of a factor unless scalar invariance is supported, in full or in part.
- With ordered categories, paths are on the probit scale, so read direction and relative size, not percentage points.
Compare Groups: key terms
- Measurement invariance
- The scale measures the same thing in every group.
- Intercept
- An item's baseline level when the factor is zero.
Path Analysis
Use it when
- You have several measured outcomes linked in a chain
- You want direct, indirect and total effects
- Some outcomes are binary or ordered
Your data needs
- Numeric variables, treated as measured without error
- A clear theory of which variable predicts which
Path Analysis: how to run it
Path Analysis: reading the results
| You see | It means |
|---|---|
| Estimates | Each path's strength. |
| Variance explained | R² for each outcome. |
| Direct, indirect and total effects | An indirect effect is significant when its 95% interval does not include zero. |
The situation. Do years of schooling affect income directly, or through the type of job a person gets? Variables: schooling, job skill level (1 to 5), income. N = 300.
| Effect of schooling on income | Standardised | 95% CI |
|---|---|---|
| Direct | 0.21 | 0.09 to 0.33 |
| Indirect (via job skill level) | 0.28 | 0.18 to 0.38 |
| Total | 0.49 | 0.39 to 0.59 |
How to read it. Schooling raises income by two routes. About 43% of the total (0.21 of 0.49) is direct. The rest, 0.28, runs through getting a higher-skilled job. Both intervals exclude zero, so both routes matter. Because schooling has a sizeable effect on income even after taking job skill into account, job type alone does not explain the benefit.
Illustrative figures.
Path Analysis: things to watch for
- Every variable is treated as measured without error. For a scale built from several items, check it with CFA first.
- Effects on a binary or ordered outcome are on an underlying continuous scale. Read direction and relative size, not percentage points.
- Including every possible path makes the model fit perfectly by design.
Path Analysis: key terms
- Indirect effect
- An effect that passes through a middle variable.
- Total effect
- Direct plus indirect.
Mediation Analysis
Use it when
- You want to open the black box of how a programme works
- You have one mediator and everything is numeric
Your data needs
- X, M and Y, all numeric and measured on the same cases
- A sensible time or causal order: X before M before Y
Mediation Analysis: how to run it
Mediation Analysis: reading the results
| You see | It means |
|---|---|
| Total, direct and indirect effects | Total = direct + indirect. |
| Bootstrap 95% CI for the indirect effect | Significant if the interval does not include zero. |
| Sobel test | An older test of the same thing. Less reliable than the bootstrap. |
| Proportion mediated | The share of the total effect that goes through M. Unstable when the total effect is close to zero, so report it only when the total effect is clearly non-zero. |
| Type of mediation | Full: indirect significant, direct not. Partial: both significant. Suppression: opposite signs, so interpret with care. None: indirect not significant. |
The situation. Does a training programme (hours attended) raise income through skill score? N = 250.
| Effect | Estimate | 95% bootstrap CI |
|---|---|---|
| Total | 4.8 | 3.1 to 6.5 |
| Direct | 1.9 | 0.2 to 3.6 |
| Indirect (via skills) | 2.9 | 1.8 to 4.2 |
What AnalyZ shows. Sobel p-value < 0.001. Proportion mediated = 60%. Partial mediation.
How to read it. Each extra training hour is linked to 4.8 more in income. About 2.9 of that (60%) travels through improved skills, and 1.9 does not. Both parts are significant, so this is partial mediation: skills explain a large part of the benefit but not all of it. Something else the training does, such as networking, may account for the rest.
Illustrative figures.
Mediation Analysis: things to watch for
- Mediation needs a causal story to be causal. With data from one round of a survey, a significant indirect effect shows consistency with the story, not proof.
- Unmeasured factors affecting both M and Y bias the result. See the sensitivity tools under Counterfactual Mediation.
- For binary mediators or outcomes, or for covariates, use Counterfactual Mediation.
Mediation Analysis: key terms
- Mediator
- The middle step on the route from X to Y.
- Bootstrap
- Re-running the analysis on many resampled versions of your data to see how much the answer varies.
Counterfactual Mediation
Use it when
- The mediator or outcome is binary or ordered
- You need to control for covariates
- You have 2 or 3 mediators
- You want effects for each group or for each category of exposure
Your data needs
- An exposure, one or more mediators, an outcome, and optional covariates
- Include covariates that could affect both the mediator and the outcome
Counterfactual Mediation: how to run it
Counterfactual Mediation: reading the results
| You see | It means |
|---|---|
| Total, direct and indirect effects | For a binary outcome the effects are in percentage points. An ordered outcome is shown as the probability of reaching a category or higher, or the expected score. |
| Several mediators | Parallel mediators do not affect each other. A serial model has exactly two, the first affecting the second, and splits the route through the first, the second and both. |
| Categorical exposure | Each other category is compared with your reference, with its own total, direct and indirect effect. Comparisons are not adjusted for how many you make. |
| Each group separately | The whole model is fitted within each group. Each group is compared with a reference group, with intervals from paired resamples. |
| E-values | The smallest strength of link a hidden factor would need with both sides of the relationship to explain the effect away. Bigger is more reassuring. For direct and indirect effects they are a guide only. |
| Hidden-correlation curve | For one continuous mediator and outcome, shows the correlation between hidden factors at which the indirect effect would fall to zero. |
The situation. Does a cash transfer (binary: received or not) raise the chance a child is enrolled in school (binary) through household food security (binary mediator)? Covariates: household size and head's schooling. N = 800.
| Effect on school enrolment | Percentage points | 95% bootstrap CI |
|---|---|---|
| Total | +9.0 | +4.1 to +13.8 |
| Direct | +5.4 | +0.3 to +10.4 |
| Indirect (via food security) | +3.6 | +1.5 to +5.9 |
What AnalyZ shows. E-value for the total effect = 1.7.
How to read it. Receiving the transfer raises enrolment by about 9 percentage points. Roughly 3.6 of those points (40%) work through improved food security, while 5.4 do not. An E-value of 1.7 suggests only a moderately strong hidden factor, linked to both receiving the transfer and enrolment, would explain the effect away, so keep this in mind in your write-up and do not claim certainty.
Illustrative figures.
Counterfactual Mediation: things to watch for
- The causal reading needs that no unmeasured factor affects the exposure, the mediator and the outcome together. Sensitivity analysis shows how fragile the result is but does not remove the problem.
- With a binary outcome or interactions, how the effect splits between mediators can depend on their order.
- Include covariates that could affect both the mediator and the outcome.
Counterfactual Mediation: key terms
- Counterfactual
- What would have happened under a different exposure.
- E-value
- How strong an unmeasured factor must be to explain an effect away.
- Serial mediation
- Two mediators in a chain, the first affecting the second.
Data Visualization
Open Modules in the top bar, then Data Visualization. The screen has a Chart type panel on the left, the settings for the chart you choose in the middle, and your chart below the Generate button. You need a dataset loaded first.
Build your first chart, step by step
Buttons above the chart
| Button | What it does |
|---|---|
| HTML | Downloads this chart as an interactive web page you can open in any browser. |
| Add to Sense Making | Saves the chart and a summary of its data into a Sense Making project, where the AI can help you interpret it. Choose an existing project or create one, and give the item a title. |
| Add to Canvas | Adds the chart as a slide in a Data Canvas story. Write the slide headline and choose a canvas, or create one. |
| Add to dashboard | Puts the chart on your dashboard so it can be shown with others. See below. |
Available chart types
The groups below are the ones shown in the Chart type panel.
Compare groups
How do categories differ from each other?
| Chart | Best for | Variables needed |
|---|---|---|
| Bar | Frequency or value of a category | 1 category (+ optional split) |
| Stacked Bar | Part-of-whole comparison across groups | 1 category + 1 group variable |
| Scatter | Relationship between two numeric variables | 2 numeric |
| Treemap | Relative size of groups (tile area is the number of rows per category) | 1 category (+ optional second grouping) |
| Bubble | Three-variable relationships (size encodes a third variable) | 3 numeric |
| Quadrant | Two-axis positioning with centre reference lines | 2 numeric + 1 label |
| Range chart | Min–max or CI ranges across categories | 1 label + 2 numeric (lo/hi) |
Trends
How does a value change over time or sequence?
| Chart | Best for | Variables needed |
|---|---|---|
| Line | Change over time or ordered categories | 1 X column + 1 numeric |
| Area | Cumulative trends over time | 1 X column + 1 numeric |
| Streamgraph | Stacked proportional flow over time by category | X + numeric + 1 category |
| Sparkline | Minimal trend line with min, max and latest value marked; one per group if you split | 1 numeric (+ optional order variable, + optional split) |
Proportions
How does each part contribute to the whole?
| Chart | Best for | Variables needed |
|---|---|---|
| Pie | Proportions of a whole | 1 category variable |
| Donut | Same as pie, modern hollow style | 1 category variable |
| Sunburst | Hierarchical part-of-whole with nested rings | 1 label + optional parent column |
Relationships
How are variables related or correlated?
| Chart | Best for | Variables needed |
|---|---|---|
| Heatmap | Correlation matrix across multiple variables | 2+ numeric |
| Contour | Density surface over two numeric variables | 2 numeric |
| Radar | Multivariate comparison on radial axes | 1 category + 1 numeric |
| Radial Bar | Circular bar chart for part-of-whole comparisons | 1 category + 1 numeric |
Distributions
How is a variable spread across its range?
| Chart | Best for | Variables needed |
|---|---|---|
| Histogram | Distribution of a single numeric variable | 1 numeric |
| Boxplot | Spread and outliers across groups | 1 numeric (+ optional group) |
| Violin | Full distribution shape across groups | 1 numeric (+ optional group) |
| Density plot | Smooth probability density curve | 2 numeric |
| Radial histogram | Circular frequency distribution (e.g. directional data) | 1 numeric |
Specialised
Niche chart types for specific use cases.
| Chart | Best for | Variables needed |
|---|---|---|
| Circular gauge | Single KPI against a target value | 1 numeric |
| Forest plot | Effect sizes with confidence intervals (meta-analysis) | Label + effect + CI lo/hi columns |
| Sankey | Flow between two sets of categories | 2 category columns |
| Table | A formatted table of selected columns, with optional sorting | Any columns, with optional sort column and maximum rows shown (50 by default) |
| Multiple Response | Frequency of each option in a "select all that apply" question, optionally by group | Either one column with answers separated by a delimiter (such as ;) or one Yes/No (1/0) column per option |
Survey weights
When a survey weight is set in Data View, charts use it by default. Untick Apply survey weight beside Generate to see the unweighted chart. The weight is saved with each chart, so charts you add to a dashboard, filter or share keep the setting they were built with.
| Charts | What changes |
|---|---|
| Bar, Stacked Bar, Pie, Donut, Treemap, Sunburst, Sankey, Multiple Response, Radial histogram | Each respondent counts with their weight, so counts are weighted counts (estimates for the population) and percentages are weighted. Axes and tooltips say "Weighted count". |
| Histogram, Density plot, Contour | Each value adds its weight to its bin, so the heights show weighted frequency. |
| Line, Area, Streamgraph, Radar, Radial Bar, Circular gauge, and the mean overlay on a Bar chart | Means, medians and totals are weighted. Axis titles read, for example, "Weighted mean of income". |
| Boxplot, Violin | Drawn from the weighted distribution, so the quartiles, median and mean are weighted. The violin smoothing uses the effective sample size implied by the weights. |
| Heatmap, Scatter trend line | Correlations are weighted, and Spearman uses weighted ranks. The trend line is a weighted least squares line. |
| Scatter points, Bubble, Quadrant, Range chart, Forest plot, Table, Sparkline line | Not changed, because they show one point or row per respondent and nothing is counted or averaged. The Sparkline reference line is the weighted mean. |
Build and share a dashboard
Save & Share
My Dashboards lists dashboards you saved. Click Load beside one to reopen it for editing. If the dataset is still loaded, every chart can be rebuilt. Without the dataset, the charts show their saved pictures only.
Which chart should I use?
| Your question | Good chart | Why |
|---|---|---|
| How do categories compare (for example % by district)? | Bar chart | Length is the easiest thing to judge by eye. Sort the bars from largest to smallest. |
| How does something change over time? | Line chart | The slope shows direction and speed. Use one line per group. |
| What share of the whole does each part make up? | Stacked bar, or a pie or donut with few parts | Keep to 5 parts or fewer, or the slices become hard to compare. |
| Are two numbers related? | Scatter plot | Each dot is one person. A tilt shows a relationship, and a cloud shows none. |
| What does the spread of a number look like? | Histogram or box plot | Shows the shape, skew and outliers that an average hides. |
| How do groups differ on a numeric variable? | Box plot by group, or grouped bars of the mean | Box plots show spread as well as the typical value. |
Scale & Reliability
A scale is a set of questions that together measure one idea, such as stress, trust or satisfaction. Adding the answers into one score is more dependable than relying on a single question, but only if the questions really do belong together. This module checks that. Choose a card, open the analysis, and use Back to return to the card selection. Results can be exported to Excel or Word.
| Analysis | The question it answers | Use it when |
|---|---|---|
| Cronbach's Alpha | Do my items hang together as one scale? | You already have a scale and want to report its reliability. |
| PCA | Can I boil many variables down to a few summary scores? | You want to simplify data or combine overlapping variables. |
| EFA | What hidden ideas are behind my items? | You are developing or exploring a questionnaire and do not yet know its structure. |
Cronbach's Alpha
Use it when
- You use a set of questions that you will add or average into one score
- A funder, journal or ethics board wants evidence your scale is reliable
Your data needs
- Numeric items (Likert answers work well)
- Items that all point the same way. Reverse-worded items must be flagged first.
- At least 3 items, and ideally 100 or more respondents
Cronbach's Alpha: how to run it
Cronbach's Alpha: reading the results
| You see | It means |
|---|---|
| Cronbach's alpha | Overall reliability. See the table below. |
| Standardised alpha | The same, if all items were put on the same scale first. Useful when items use different response ranges. |
| Inter-item correlation matrix | How strongly each pair of items agrees. The average is ideally 0.2 to 0.5. Much higher means items are near-duplicates, and negative values point to a reverse-coded item. |
| Corrected item-total correlation | How strongly each item agrees with the rest of the scale. Aim for 0.30 or higher. A low value means the item does not belong. |
| Alpha if item deleted | What alpha would be if that item were removed. If removing one item raises alpha substantially, consider dropping it. |
| Alpha | Interpretation |
|---|---|
| 0.90 or higher | Excellent (above about 0.95 can mean items are repeating each other) |
| 0.80 to 0.89 | Good |
| 0.70 to 0.79 | Acceptable |
| 0.60 to 0.69 | Questionable |
| Below 0.60 | Poor. Consider revising the scale |
The situation. A 5-item stress scale, answered on a 1 to 5 scale by 150 people.
| Item | Item-total r | Alpha if deleted |
|---|---|---|
| Stress1 | 0.56 | 0.742 |
| Stress2 | 0.51 | 0.759 |
| Stress3 | 0.58 | 0.734 |
| Stress4 | 0.60 | 0.729 |
| Stress5 | 0.55 | 0.747 |
What AnalyZ shows. Cronbach's alpha = 0.78. Average inter-item correlation about 0.54.
How to read it. An alpha of 0.78 falls in the acceptable-to-good band. Every item correlates 0.30 or more with the rest, and removing any item would lower alpha (every alpha-if-deleted figure is below 0.78), so there is no reason to drop any. The scale is reliable enough to add the five answers into one stress score.
Illustrative figures.
Cronbach's Alpha: things to watch for
- Alpha rises with the number of items, so a long scale can look reliable even when it is not well focused.
- A high alpha shows items agree, not that they measure what you intend (validity).
- Alpha assumes one idea. A scale that actually measures two ideas can still return a decent alpha. Check with EFA or CFA.
- Always reverse-code before running, or alpha will be understated.
Cronbach's Alpha: key terms
- Internal consistency
- How well items agree with each other.
- Reverse-coded item
- A question worded in the opposite direction, so its scores must be flipped.
- Item-total correlation
- How closely an item tracks the rest of the scale.
Principal Component Analysis (PCA)
Use it when
- You want fewer variables for a regression, or want to remove overlap between predictors
- You want a single score from several related variables such as an asset index
Your data needs
- Several numeric variables that are correlated
- At least 5 cases per variable, ideally many more
Principal Component Analysis: how to run it
Principal Component Analysis: reading the results
| You see | It means |
|---|---|
| KMO measure | Whether your variables are correlated enough to summarise. 0.60 or higher is acceptable. Below that, PCA may not be worthwhile. |
| Eigenvalues and variance explained | How much of the total variation each component captures. An eigenvalue above 1 means the component explains more than one original variable would on its own. |
| Scree plot | A line chart of eigenvalues. Keep components up to the 'elbow', where the line flattens. |
| Component loadings | How strongly each variable relates to each component. Above 0.4 (or below -0.4) is meaningful. Look at which variables load highly to name the component. |
| Communalities | How much of each variable the kept components capture. Below 0.40 is flagged as poorly represented. |
The situation. Eight household asset questions (radio, phone, bicycle, and so on) from 500 households.
What AnalyZ shows. KMO = 0.82. Component 1: eigenvalue 3.4, 42% of variance. Component 2: eigenvalue 1.3, 16%. Component 3: eigenvalue 0.8.
How to read it. KMO of 0.82 means the variables are well suited to summarising. Two components pass the eigenvalue rule and together explain 58% of the variation. If every asset loads positively on Component 1, it can serve as a general wealth index. Component 2 might separate productive assets (such as a plough) from consumer goods. Any variable with communality below 0.40 is not well captured, and you might drop it and re-run.
Illustrative figures.
Principal Component Analysis: things to watch for
- PCA is a data-reduction tool, not a theory of hidden causes. For that, use EFA.
- Eigenvalue above 1 often keeps too many components. Cross-check with the scree plot.
- Components are not automatically meaningful. Give them a name only when the loadings support one.
Principal Component Analysis: key terms
- Component
- A weighted combination of your original variables.
- Eigenvalue
- The amount of variation a component captures.
- Communality
- How much of a variable the retained components explain.
- KMO
- A check that your data suit factor-style methods.
Exploratory Factor Analysis (EFA)
Use it when
- You are developing or shortening a questionnaire
- You do not know how many dimensions your questions contain
Your data needs
- Numeric items that are correlated
- At least 100 cases, and about 5 to 10 cases per item
Exploratory Factor Analysis: how to run it
Exploratory Factor Analysis: reading the results
| You see | It means |
|---|---|
| KMO | 0.60 or higher is acceptable. |
| Eigenvalues, variance explained and scree plot | As in PCA. |
| Factor loadings | How strongly each item reflects each factor. Above 0.4 suggests a good indicator. An item that loads strongly on two factors (a cross-loading) is ambiguous and may need rewording or removal. |
| Factor correlation matrix (oblique rotation) | How related the factors are. If all correlations are below 0.32, Varimax would have given about the same answer. |
| Communalities | Below 0.30 is flagged as poorly explained by the factors. |
The situation. Ten wellbeing questions from 300 people, two factors extracted with Principal Axis Factoring and Oblimin rotation.
What AnalyZ shows. KMO = 0.86. Two factors explain 54% of the variation. Factor correlation = 0.41.
How to read it. Items 1 to 5 (all about mood) load 0.60 to 0.80 on Factor 1 and almost nothing on Factor 2. Items 6 to 10 (all about sleep and energy) do the opposite. That reveals two ideas, emotional wellbeing and physical wellbeing. One item loading 0.45 on both would be a candidate for rewording. The factors correlate at 0.41, above 0.32, so Oblimin was the right choice. Once the structure is found, confirm it on a fresh sample with CFA.
Illustrative figures.
Exploratory Factor Analysis: things to watch for
- EFA is for discovering structure. Do not test it on the same data afterwards and call it confirmed.
- Different choices of method, rotation and number of factors can give different answers. Choose with a reason and report them.
- If unsure about rotation, run Oblimin and check the factor correlations.
Exploratory Factor Analysis: key terms
- Latent factor
- An unseen idea that causes related answers.
- Rotation
- A step that makes the factor pattern easier to read.
- Cross-loading
- An item that relates strongly to more than one factor.
Survey weights
When a survey weight is set in Data View, all three analyses use it by default. Untick Apply survey weight beside Run to compare with the unweighted result. Each analysis works from the survey-weighted covariance or correlation matrix of your items, so respondents count in proportion to their weights, and Weighted N appears next to N.
| Analysis | What changes |
|---|---|
| Cronbach's Alpha | Alpha, standardised alpha and the inter-item correlations come from the weighted matrix. Item means and standard deviations are weighted, and the corrected item-total correlations and alpha-if-item-deleted use the weighted covariances. Reverse-coded items are detected from the weighted correlations. |
| PCA | Eigenvalues, variance explained, loadings, communalities and KMO come from the weighted correlation matrix, or from the weighted covariance matrix if you do not standardise. |
| EFA | Extraction and rotation work on the weighted correlation matrix, so loadings, communalities, variance explained, factor correlations and KMO reflect the weights. Bartlett's test uses the effective sample size implied by the weights, which is smaller than N, so it is more cautious. |
Sampling
Sampling brings together three planning tools that you use before collecting data (or independently of any loaded dataset): Sample Size Estimation, Power Analysis and Draw a Sample. Start on the landing screen, choose the tool, then pick the calculator or design.
Five ideas to know before you start
| Idea | In plain words |
|---|---|
| Confidence level | How sure you want to be that the true value lies within your margin of error. 95% is standard. |
| Margin of error | How far your estimate may be from the truth. A margin of 5% means a survey result of 40% really means 35% to 45%. A smaller margin needs a larger sample. |
| Statistical power | The chance of detecting a real effect if it exists. 80% is the usual minimum, 90% for important studies. |
| Effect size | How big a difference you hope to detect. Small differences need much larger samples than big ones. Choose the smallest difference that would matter in practice. |
| Design effect (DEFF) | A multiplier for cluster designs. People in the same village or school resemble each other, so you learn less per person and need more people. Typical field surveys use 1.5 to 2.0. |
Sample Size Estimation
Use it when
- You are planning a survey or evaluation
- A protocol or funder requires a sample size justification
Your data needs
- Your target confidence level and margin of error (or power and effect size)
- A rough guess of the proportion or standard deviation you expect, from earlier studies or a pilot
Sample Size Estimation: how to run it
Sample Size Estimation: reading the results
| You see | It means |
|---|---|
| The main figure (large text) | The sample you need, for example per arm in an RCT or total in a cross-sectional survey. |
| Secondary figures | The minimum before non-response, number per stratum, or clusters needed. |
| Interpretation panel | Plain guidance and any design-specific caveats. |
| Formula | The exact calculation, for transparency and referencing. |
Cross-sectional: the flexible calculator
| Choice | Options |
|---|---|
| What are you measuring? | Proportion (%), a rate or share. Defaults to 50% if you have no estimate, which gives the largest and safest sample. Mean (numeric), an average such as income, which needs an expected standard deviation. |
| How will you select participants? | Simple random, equal chance for everyone. Stratified, subgroups sampled separately (add strata names and sizes, or tick I need this precision within each stratum if you will report by subgroup). Cluster / multi-stage, whole groups first, which also covers PPS designs. |
| Design effect | Defaults to 1 and is always editable. With cluster designs you may calculate it from the intra-class correlation: DEFF = 1 + (m - 1) × ρ, where m is the average cluster size and ρ the ICC. For example, with 20 per cluster and ρ = 0.05, DEFF is 1.95. |
| Population size and non-response | Optional. A population size applies a finite-population correction. Non-response inflates your recruitment target. |
Measuring change and other designs
| Calculator | Use when | Key inputs |
|---|---|---|
| Longitudinal Study | Following the same people over several rounds. | Confidence, margin of error, number of waves, dropout per wave, population size |
| Change in Proportion | Detecting a change in a share, between two groups, or in one group before and after. | Confidence, power, baseline and endline proportion. For before-and-after, the expected correlation between the two measurements, since the higher it is, the smaller the sample needed. |
| Randomized Controlled Trial | Random assignment to treatment and control. | Confidence, power, control and treatment outcome rates, expected attrition |
| Case-Control | Comparing people with an outcome (cases) against those without (controls). | Confidence, power, exposure rate among controls, odds ratio to detect, controls per case |
| LQAS | Classifying whether coverage meets a threshold, instead of estimating a precise rate. Common in humanitarian monitoring. | Coverage threshold, upper (acceptable) and lower (unacceptable) bounds |
LQAS example. With a sample of 19 and a decision rule of 13, a site with truly acceptable coverage has about a 7% chance of being wrongly failed and a site with truly unacceptable coverage has about an 8% chance of being wrongly passed. These are the two error risks the calculator displays so you can judge the rule.
The situation. Cross-sectional survey of the percentage of households vaccinated. No prior estimate (use 50%), 95% confidence, 5% margin of error.
| Step | Sample size |
|---|---|
| Simple random sample (DEFF = 1) | 385 |
| Clustered design, DEFF = 2 | 770 |
| Add 10% expected non-response | 855 |
| Simple random sample if the population is only 2,000 (finite-population correction) | about 323 |
How to read it. A simple random survey needs 385 households. Because you plan to visit villages (clusters), people within them resemble each other, so you must double the sample, to 770. Allowing for 10% of households not responding, plan to approach 855. If the whole population is only 2,000 households, you need fewer (about 323 in the simple random case), because you are sampling a large share of everyone. Other examples: to estimate an average with a standard deviation of 12 to within 2 units needs about 139 people. To detect a rise in a proportion from 40% to 55% with 80% power needs about 170 per group. A case-control study with 30% exposure among controls and an odds ratio of 2 needs about 141 cases (and as many controls at 1:1).
Calculated for illustration.
Sample Size Estimation: things to watch for
- The sample size is for the smallest group you plan to report on. If you will report by district, each district needs its own sample.
- Always add a margin for non-response and dropout.
- For high-stakes surveys with complex designs, consult a sampling statistician.
Sample Size Estimation: key terms
- DEFF
- Design effect: the multiplier for clustering.
- Finite-population correction
- A reduction when you sample a large share of a small population.
- Attrition
- People lost between rounds.
Power Analysis
Use it when
- You have chosen your analysis and need to justify the sample
- You have a fixed sample and want to know what you can detect
- You want to see how power changes across sample sizes
Your data needs
- The test you plan to use
- An effect size, drawn from earlier studies or the smallest effect that would matter
Power Analysis: how to run it
Power Analysis: reading the results
| You see | It means |
|---|---|
| Required sample size or achieved power | The headline number. |
| Power curve | Where adding more people stops paying off. The steep part shows where each extra person helps most. |
| Test | Use when | Effect size measure |
|---|---|---|
| Two-sample t-test | Comparing means of two independent groups | Cohen's d (0.2 small, 0.5 medium, 0.8 large) |
| One-sample t-test | Comparing a group mean with a known value | Cohen's d |
| Paired t-test | Two measurements on the same people | Cohen's dz |
| Chi-square | Association between two categorical variables | Cohen's w (0.1, 0.3, 0.5) |
| McNemar's test | Change in a yes/no outcome in the same people | Discordant pair proportions |
| One-way ANOVA | Means across 3 or more groups | Cohen's f (0.1, 0.25, 0.4) |
| Linear regression | Detecting a given R² with several predictors | Cohen's f² (0.02, 0.15, 0.35) |
| Logistic regression | Binary outcome with several predictors | Odds ratio and events per variable (EPV) |
Use the paired t-test or McNemar's test when the same people are measured twice, and the two-sample t-test or chi-square when comparing separate groups. Paired designs generally need fewer people for the same effect.
The situation. Planning to compare two groups with a t-test, 80% power, 5% significance level.
| Test and effect size | Result |
|---|---|
| Two-sample t-test, d = 0.5 (medium) | 64 per group |
| Two-sample t-test, d = 0.3 (small to medium) | 176 per group |
| One-sample t-test, d = 0.5 | 34 in total |
| One-way ANOVA, f = 0.25, 3 groups | about 158 in total (about 53 per group) |
| Chi-square, w = 0.3, 1 degree of freedom | 88 in total |
| Linear regression, f² = 0.15, 5 predictors | 92 in total |
What AnalyZ shows. Achieved power with d = 0.5 and 30 per group: 0.48.
How to read it. Halving the effect size (0.5 to 0.3) nearly triples the sample needed, which is why choosing a realistic effect size is the most important step. If you can only recruit 30 per group, you have a 48% chance of detecting a medium effect, worse than a coin flip. Either recruit more, accept that you can only detect larger effects, or use a design with more power, such as a paired comparison.
Calculated for illustration.
Power Analysis: things to watch for
- Do not choose the effect size to hit a convenient sample size. Choose it from evidence or from what would matter in practice.
- Power calculated after a study from its own observed effect adds nothing. Use it before.
- Allow for non-response and dropout on top of the figure.
Power Analysis: key terms
- Power
- The chance of detecting a real effect.
- Effect size
- How big the difference is, in standardised units.
- Significance level
- The risk of a false alarm you accept, usually 5%.
Draw a Sample
Use it when
- You have a full list of the population and need a defensible random selection
- You need the same draw to be repeatable later
Your data needs
- A CSV, XLS or XLSX file with one row per unit
- Columns for strata, clusters or size, depending on the design
Draw a Sample: how to run it
_selected flag column.Draw a Sample: reading the results
| You see | It means |
|---|---|
| Preview table | The selected units. Check it looks sensible before downloading. |
| Download options | Selected rows only, or the whole frame flagged. |
| Design | Use when |
|---|---|
| Simple Random | Each unit has an equal chance of selection. The default when there is no further structure. |
| Systematic | Taking every k-th unit after a random start. The frame should be in a random order. |
| Stratified | You must guarantee representation of subgroups such as region or sex. |
| Cluster | You select whole groups (villages, schools) and survey everyone, or a sample, within them. |
| PPS (probability proportional to size) | Multi-stage designs where clusters differ in size. Larger clusters have proportionally higher selection chances. |
The situation. A list of 1,200 households across 3 districts. You need 300, with enough from each district.
How to read it. A simple random draw could, by bad luck, take very few from the smallest district. A stratified draw with district as the strata column guarantees each district gets its share. Setting a seed (say 2025) means a colleague can regenerate the exact same list from the same file later.
Draw a Sample: things to watch for
- Your frame is processed in your browser. It is never uploaded to or stored on AnalyZ servers.
- A seed is specific to AnalyZ and will not reproduce the same draw in R, Stata or other software. Keep the downloaded list as the record.
- A random sample is only as good as the frame. If the list leaves out some of the population (new settlements, for example), so will the sample.
Draw a Sample: key terms
- Sampling frame
- The complete list the sample is drawn from.
- Stratum
- A subgroup sampled separately.
- PPS
- Selection chance proportional to size.
Qualitative Analysis
The Qualitative module is organised into five sub-modules. Load your transcripts first (see Loading Data → Qualitative data), then open Qualitative Analysis from the home screen. Use the buttons below to jump directly to any sub-module's guide.
🔄 Recommended workflow
View Transcripts
Opens with a card selection screen. Four tabs are available: Transcripts, Word Cloud, Word Search, and Transcript Settings. Use the Back button to return to the card selection at any time.
Transcripts, overview and detail
The Transcripts tab opens in overview mode: all loaded transcripts displayed as cards showing word count, sentence count, average sentence length, and coded segments. Click any card to open that transcript's detail view.
In detail view: read the transcript in full, with coding stripes shown in a thin column on the right margin, coloured bars indicating which codes are applied at each point in the text. A "What the stripes mean" key below maps each colour to its code. Click Edit to make corrections, Save to commit, or Cancel to discard. Inline tag dropdowns let you assign participant attributes without leaving the transcript view. Click All transcripts to return to the overview.
Word Cloud
Shows word frequency across all transcripts, a single transcript, or coded segments only. Adjust the max words slider, exclude specific words using the exclusion list with autocomplete suggestions, and filter by a specific code when in coded segments mode. The frequency table below lets you exclude words individually. Export as CSV.
Word Search
Search for any word or phrase across all transcripts or a single transcript. Each result shows the matching sentence in context with the search term highlighted. Toggle case sensitivity as needed. Export all matches as CSV.
Transcript Settings, tagging transcripts
Define tag categories to classify each transcript by participant attributes. This enables grouped comparisons in Explore Themes.
Thematic Coding
What it does: Assigns meaning to sections of your transcripts by tagging them with codes, labels that capture what is happening in that passage. Over time, a pattern of codes builds into themes that describe recurring ideas, experiences, or perspectives in your data.
What you will see: A card selection screen with three cards: Manage Codes, Assign Codes, and Review Codes. Each card opens its own sub-view; the Back button in the header always returns you to the card selection.
Thematic Coding opens with a card selection screen, Manage Codes (your codebook), Assign Codes (the coding workspace), and Review Codes (audit all coded segments). Click a card to enter that section. Use the Back button to return to the card selection at any time.
Manage Codes, building your codebook
Click ➕ Add new code. Enter a code name (e.g. Housing Insecurity), choose a colour, and optionally add a description. Click Add code. Each code shows a usage count (how many segments it has been assigned to). Codes can be edited (✏️ button, opens an inline form) or deleted (🗑 button, also removes the code from all segments).
Manage Codes, building the codebook
Create, edit, colour, and delete codes. Each code can have a name, description, analytical memo (a working note on when and how to apply it), and keywords for auto-suggest. The colour picker shows 15 preset swatches; click + for a custom colour.
Assign Codes, the coding workflow
The screen is split: transcript on the left, Coding Panel on the right.
Review Codes, auditing and querying segments
Review Codes opens with three cards. Click any card to enter that view. Use the Back button in the top header to return to the card selection.
- Coded Segments: browse, filter, sort, edit, and export all coded segments. Two filter modes: Simple filter (transcript and code dropdowns) and Query builder (Boolean AND/OR/NOT conditions across codes). Export as CSV, Excel, or JSON (for inter-rater reliability).
- Code Frequency: frequency table, bar chart, and Code × Transcript heatmap showing how often each code appears across transcripts. Includes a plain-language interpretation panel.
- Coding Map: heatmap showing where each code appears within a single transcript, by paragraph or by interview quarter. Coverage bars show what proportion of the transcript each code covers. Use this to spot uncoded stretches and check coding coverage across the interview.
Compare Coding
A standalone module for inter-rater reliability analysis. Use it when two researchers have independently coded the same transcripts and want to measure agreement.
Explore Themes
Requires coded segments to be present. Opens with four cards in a 2×2 grid. Each card opens its own analysis view with a plain-language interpretation panel. Use the Back button to return to the card selection.
| Card | What it shows | Best for |
|---|---|---|
| Saturation | Bar and line chart showing new codes introduced per transcript (cumulative codebook growth) | Assessing whether coding has reached thematic saturation, a flat line means no new themes are emerging |
| Theme Comparison | Two modes: Per transcript (code share heatmap and raw counts table) and By participant group (normalised frequency by tag category, bar chart and heatmap) | Per transcript: seeing which themes dominated each interview. By group: comparing theme prevalence across subgroups (e.g. gender, location, domain) |
| Theme Connectedness | Two views toggled within the card: Co-occurrence (heatmap and ranked table of how often code pairs appear together) and Distinctiveness (exclusivity scores for code pairs that rarely or never appear together) | Co-occurrence: finding theme clusters and candidates for higher-level labels. Distinctiveness: identifying themes that occupy separate conceptual territory |
| Segment Length | Average word count per coded segment for each code, sorted by elaboration depth | Identifying which themes participants elaborate on at length, a proxy for depth of engagement or emotional significance |
Analyze Relationships
Explores structural and configurational relationships between codes and participants across your corpus. Requires at least 2 codes and 2 coded transcripts. Opens with seven cards. Each card includes an automated plain-language interpretation panel and Add to Sense Making / Add to Canvas buttons.
| Card | What it shows | How to read it |
|---|---|---|
| Network Map | Draggable SVG network where codes are nodes (sized by frequency) and edges represent co-occurrence counts. Toggle connection count labels and adjust edge colour. Export as image for Canvas. | Hub codes connected to many others are likely core themes. Isolated codes may be peripheral or under-coded. Drag nodes to explore the structure. |
| Mind Map | A branching diagram with your central topic in the middle. Each branch is a group of codes that tend to be applied to the same segments (or to the same transcripts), and the number beside a code is how many segments carry it. Controls: Centre (all codes, or one code to see everything linked to it), Central topic (the label in the middle), Group codes that share a Segment or Transcript, Branches (Auto or a number), Min segments per code, Show segment count, and Expand all / Collapse all. Click a branch to fold it, and hover over a code to read an example segment. Add to Canvas saves the map as an image slide. | Branches are suggestions from your data, not finished themes: read the codes on a branch together and name the theme yourself. Choose a single code under Centre to list the codes that share a segment with it, strongest link first. Codes that share nothing with any other appear on a grey Unlinked codes branch. If every code lands on one branch, group by Transcript, or add a second code to segments that cover more than one theme. |
| Thematic Clusters | Jaccard similarity matrix and presence heatmap, codes that appear in the same transcripts are grouped together | Tight clusters of co-occurring codes are candidates for higher-level theme labels. Codes that cluster together across transcripts form the backbone of your thematic structure. |
| Code Sequence | Sankey diagram of code-to-code transitions, the order themes follow one another within transcripts. Self-transitions (same code consecutive) are filtered out. | Dominant transitions reveal narrative structure and causal reasoning. A theme that consistently precedes another may be a precondition or trigger for it. |
| Configurational Analysis | Presence/absence matrix of codes per transcript, sorted so transcripts with identical thematic profiles appear together. Identifies participant types (configurations) and classifies codes as Universal, Differentiating, or Rare. | Transcripts sharing the same configuration form a participant type. Universal codes are present in every transcript. Differentiating codes are what distinguish one type from another. Rare codes appear in only one transcript and should be treated as individual cases rather than group patterns. |
| Qualitative Comparative Analysis | Crisp-set QCA in four steps: select conditions (codes), assign a binary outcome per transcript, build a truth table with consistency scores, and compute the minimised Boolean solution. Consistency threshold is adjustable (default 0.75). Add to Sense Making sends the full truth table and solution. Add to Canvas sends the truth table as a slide. | Use when you want to explain why some transcripts achieve an outcome and others do not. Identifies the minimal combinations of coded themes sufficient for the outcome. Requires at least 2 coded transcripts with outcomes assigned and at least 2 conditions selected. Outcome assignments are session-only and do not affect other analyses. |
| Persona Builder | Identifies thematic personas from the co-occurrence patterns in coded segments in three steps. Step 1: select which codes define the thematic space. Step 2: review the silhouette-scored suggestion, the analysis tests cluster solutions from 2 to 8 personas and recommends the most analytically distinct solution; click any row to select a different number. Step 3: persona cards appear in a grid with auto-generated descriptive names, thematic frequency profiles, participant characteristics from transcript tags, and sample segments. Expand a card to see frequency bars and click a theme to view all segments. Use Merge into to combine similar personas. Add to Sense Making sends all personas with frequency data and interpretation. Add to Canvas sends a frequency profile table. | Suitable for both IDIs and FGDs. Personas represent thematic positions, not individual participants or transcripts. A silhouette score above 0.7 indicates strong separation between personas; below 0.5 suggests the code selection should be reconsidered. Participant characteristics (Predominantly female, Mostly rural) are computed from transcript tags counted per unique source file. |
Technical note, Configurational Analysis: Each unique combination of present/absent codes across transcripts constitutes a configuration. Configurations shared by 2 or more transcripts are analytically robust; singleton configurations (unique cases) may represent outliers.
Technical note, Persona Builder: Personas are derived from hierarchical clustering of coded segments using Jaccard similarity on code co-occurrence patterns. The silhouette score measures how well-separated the clusters are (above 0.7 is strong, 0.5 to 0.7 is reasonable, below 0.5 suggests overlap). Personas represent thematic positions in the data, not individual participants, and are suitable for both IDI and FGD transcripts without requiring speaker attribution.
Technical note, QCA: AnalyZ implements crisp-set QCA (csQCA) using the Quine-McCluskey algorithm for Boolean minimisation. The conservative solution approach is used, logical remainders (unobserved configurations) are excluded from minimisation. Outcome assignments are stored in session state only and are not written to transcript data, so the same transcripts can be used for multiple QCA runs with different outcome variables.
Key ideas in plain language
| Idea | In plain words | Example |
|---|---|---|
| Code | A short label you attach to a piece of text to say what it is about. | 'Cost of transport' on a sentence about the bus fare to the clinic. |
| Codebook | The list of codes with a definition and when to use each, so everyone codes the same way. | A shared table: code name, meaning, include and exclude rules. |
| Theme | A bigger pattern formed by several related codes. | The codes bus fare, distance and time off work together form the theme barriers to access. |
| Saturation | The point where new interviews stop producing new codes or ideas. | If the last three interviews added no new codes, you may have enough data on that topic. |
| Inter-coder agreement (Cohen's kappa) | How often two coders agree, after removing the agreement expected by chance. Above 0.60 is substantial and above 0.80 is strong. | Two researchers code the same 20 segments. Kappa of 0.72 means good agreement. Discuss the disagreements and refine the codebook. |
Sense Making
Sense Making brings together the analyses you have run across any module, quantitative tests, qualitative codes, visualisations, and uses AI to synthesise them into a single coherent interpretation. Instead of describing each result individually, the AI identifies patterns, connections, and contradictions across your full body of evidence.
How it works
Projects
A project is a named collection of analyses grouped for a specific synthesis. Key facts:
- You can have multiple projects, one per research question, study, or deliverable.
- Each project holds up to 20 analyses. A warning appears at 15; the button is blocked at 20.
- Projects are private to your account: no other user can see them.
- Deleting a project permanently removes all its analyses and saved interpretations.
- You can copy an analysis from one project to another from inside the project workspace.
Interpretation modes
Choose a mode in the project workspace before interpreting. The mode persists and can be changed at any time, switching mode and re-interpreting is one click.
| Mode | Best for | Style |
|---|---|---|
| Plain language | Non-specialist audiences, internal reads | Simple words, short sentences, jargon-free |
| Academic | Research reports, papers, dissertations | Formal tone, statistical terminology, methodology references |
| Executive summary | Leadership briefings, decision-makers | Concise, action-oriented, implications-first |
| Narrative | Donor reports, public summaries | Flowing prose, story-driven, accessible |
Length
Each mode offers a length setting that sets the target word count of the interpretation:
| Mode | Short | Standard | Detailed |
|---|---|---|---|
| Plain language | about 300 words | about 600 | about 1,000 |
| Academic | about 600 | about 1,200 | about 2,000 |
| Executive summary | about 200 | about 400 | not offered |
| Narrative | about 400 | about 800 | about 1,400 |
Context field
The context field (in the project workspace) tells the AI what this project is about. It is optional but significantly improves interpretation quality. Good context includes:
- The research question or objective
- How the data was collected (e.g. "survey of 300 B2B clients")
- Who the findings are for
- Any hypotheses you want tested or confirmed
Context is saved to the project automatically when you click 💾 Save context.
Refining the interpretation
After the initial interpretation, a Refine interpretation panel appears below the synthesis. Type a plain-language instruction and the AI rewrites the interpretation incorporating it, the refined version replaces the original. Examples:
- "What might explain the contradiction between the t-test result and the qualitative themes?"
- "Focus more on the findings for the female sub-group."
- "Summarise the three most important findings in one paragraph."
- "What are the limitations of this analysis?"
The AI rewrites the full interpretation incorporating your instruction, drawing on the original analyses and any context you provided. The revised version replaces the previous one in the project.
Re-interpret and version history
Re-interpret generates a fresh interpretation from your current analyses, optionally steered by a short focus note (for example "lead with the qualitative findings" or "add recommendations"). It replaces the current interpretation, and counts as one interpretation from your monthly allowance. If the response is cut off, nothing is saved or counted and your previous version is kept.
Every interpretation and refinement is stored under Version history in the workspace. Open it to read earlier versions and click Restore to bring one back. Use Copy text to copy the current interpretation to the clipboard, and drag the grip handle on an analysis to change its order in the project.
Exporting the interpretation
Click ⬇ Export to Word in the synthesis view. The downloaded .docx file includes:
- Project title and interpretation mode
- Research context (if provided)
- Numbered list of all analyses included, with module type and user notes
- Full AI interpretation
Privacy and the AI
The AI receives no raw data. It only sees what you explicitly added to the project: aggregated statistics, auto-generated interpretation text from each module, and any notes you wrote. Your dataset stays in your browser session and is never transmitted to the AI.
Data Canvas
Data Canvas turns your analysis outputs into a structured, shareable narrative story, powered by AnalyZense. Unlike the dashboard (which lets people explore data interactively), Data Canvas tells the story of what the data means for a non-technical audience: donors, community members, board members, or anyone who needs to understand findings without doing the analysis themselves.
Open Data Canvas from the home page card or the sidebar navigation.
How it works
Data Canvas has two layers:
- Findings tab: your evidence repository. Add charts, tables, network diagrams, and participant quotes from any analysis module. Each finding is shown as a thumbnail card. You can delete individual findings at any time.
- Story tab: the AI-generated narrative. AnalyZense reads all your findings, groups related ones, writes headlines and narratives, and arranges them into slides. This tab appears after your first Generate Story.
Adding findings
Every analysis output has an Add to Canvas button. Clicking it opens a modal where you choose the destination canvas (or create one inline). Supported finding types:
| Type | Source |
|---|---|
| Chart | Any Plotly chart from any analysis module |
| Table | Frequency tables, crosstabs, regression outputs, test statistics |
| Infographic / Network | Network map from Analyze Relationships (qualitative) |
| Quote | Coded participant excerpts from Thematic Coding |
Deleting a finding from the Findings tab clears the generated story, you will need to regenerate after removing or adding findings.
Generating the story
Viewing and editing the story
Use the toggle in the canvas header to switch between Vertical scroll (all slides in one scrolling page) and Slide-by-slide (one slide at a time). The choice is saved with the canvas and applies to the published story.
AnalyZense picks one of eight layouts for each slide: single chart, single table, single infographic, chart and table, two charts, chart and quote, chart and text, or full text. To change the wording yourself without using a refinement, click Edit text on a slide and edit the headline and narrative directly.
AnalyZense AI features
| Action | What it does | Free | Starter | Pro |
|---|---|---|---|---|
| Generate Story | Reads all findings, decides grouping and layout, writes headlines and narratives for the full story | 2/month | 5/month | 15/month |
| Refine Slide | Updates a single slide based on a plain-language instruction, headline, narrative, layout, or chart type | 5/month | 15/month | 50/month |
Your remaining allowance is shown in the canvas header above the title. Regenerating the story (after adding new findings) counts as one Generate use.
Publishing
Anyone with the link can read the story in their browser, no AnalyZ account required, no software to install. The viewer shows the story title, each slide with its headline, narrative, and visual, and a navigation bar.
Privacy and the AI
AnalyZense receives only the finding summaries, headlines, context text, chart metadata, and table data. Your raw dataset is never sent to the AI. Participant quotes are sent as written text only, no metadata from your dataset accompanies them.
Exporting Results
Tables
Every analysis module that produces a results table includes export buttons below the output. The buttons are labelled Excel (.xlsx, several sheets when there are several tables), Word (.docx, ready to paste into a report) and, on some results, Image (a picture of the table).
Charts
In the Visualization module, the HTML button above a chart downloads it as an interactive web page. For a picture, use the camera icon in the chart's own toolbar (top right of the chart, visible when you hover), which saves a PNG. In Descriptive Analysis, the chart menu has Download PNG.
💾 Dataset
Use ⬇ Download Data in the sidebar to export your current (possibly modified) dataset. Available formats:
| Format | Best for |
|---|---|
| JSON (.json) | Passing data to other software; no labels embedded |
| Excel (.xlsx) | Easy to open and share; no labels embedded. Reading it back in loses labels. |
| Stata (.dta) | Variable labels and value labels are fully preserved |
| SPSS (.sav) | Variable labels and value labels are fully preserved |
💼 Sessions
Use 💾 Save Session in the sidebar to download a complete snapshot of your work as a .analyz.json file, dataset, labels, variable type overrides, connected API sources, dashboard charts, and (for qualitative sessions) all transcripts and coded segments. Click 📂 Restore Session at any time to load that file back in and resume your exact session.