Propensity Score Analysis
Learn how propensity scores work, when to use them, and run your own analysis — upload a test/control dataset and instantly see how balanced your groups are.
Learn
What are propensity scores?
A propensity score is the probability that a person was assigned to the treatment (test) group, given their observed characteristics. In plain English: it's a single number that summarises everything you know about a respondent — their age, attitudes, behaviours — and tells you how likely they were to end up in the test group. Introduced by Rosenbaum & Rubin (1983), propensity scores are a cornerstone of causal inference. They let you compare apples to apples: instead of directly comparing test and control groups (which may differ systematically), you compare people with similar propensity scores.
When should you use them?
Propensity scores are most valuable when you cannot randomise perfectly — which is most of the time in marketing and consumer research. Use them when: • Your test and control groups weren't randomly assigned (observational data) • You suspect the groups differ on key background characteristics • You want to check balance after an experiment ran • You're weighting or matching respondents to create a fairer comparison They're commonly applied in media mix modelling, brand lift studies, A/B test diagnostics, and causal attribution.
How are they calculated?
The standard approach is logistic regression: 1. Combine your test and control groups into one dataset 2. Create a binary outcome: 1 = Test, 0 = Control 3. Use your pre-treatment variables (demographics, attitudes, behaviours) as predictors 4. Train the model — the predicted probability for each row is their propensity score Once you have scores, you assess balance by comparing the score distributions across test and control. Good overlap = well-balanced groups. Poor overlap = systematic differences that could bias your results. The Standardized Mean Difference (SMD) is the standard metric: SMD < 0.10 is considered well balanced; SMD > 0.25 signals concern.
Try it
Upload a CSV with test and control groups, or use the sample dataset. The tool will calculate propensity scores and show you how well your groups are balanced.
Upload your data
Drag & drop a .csv file here, or click to browse
Needs a group column (test / control) and numeric feature columns.