Estimate treatment likelihood from observed covariates quickly today. Check overlap, weights, and model fit easily. Export results, improve balance, and document assumptions clearly now.
| T | X1 | X2 | X3 | X4 | X5 |
|---|---|---|---|---|---|
| 1 | 35 | 1 | 0.42 | 120 | 0 |
| 1 | 40 | 1 | 0.30 | 115 | 0 |
| 0 | 29 | 0 | 0.10 | 90 | 1 |
| 0 | 33 | 0 | 0.15 | 95 | 1 |
A propensity score is the probability of receiving treatment given observed covariates: e(x) = P(T=1 | X=x).
This calculator uses logistic regression: logit(e(x)) = b0 + b1·X1 + b2·X2 + b3·X3 + b4·X4 + b5·X5, where e(x) = 1 / (1 + exp(-logit(e(x)))).
For weighting, common choices are: ATE weights (treated 1/e, control 1/(1-e)) and ATT weights (treated 1, control e/(1-e)).
Propensity scores summarize the probability of receiving treatment given observed covariates. In this calculator, treatment is T=1 and the score e(x) is computed from selected X1–X5 inputs. Using a single number helps compare treated and control units on a common scale, supporting matching, stratification, or weighting. When e(x) is well estimated, downstream effect estimation can reduce bias from measured confounding, while keeping modeling assumptions explicit.
The tool fits a logistic model, logit(e)=b0+b1X1+b2X2+b3X3+b4X4+b5X5, using iterative reweighted least squares. After estimation, it reports log-likelihood, McFadden R², and AUC. Log-likelihood tracks overall fit, McFadden R² compares the fitted model to an intercept-only baseline, and AUC summarizes discrimination between treated and control observations. These diagnostics guide whether additional covariates, transformations, or interaction terms may be justified.
A key validity condition is positivity: for each covariate pattern, both treatment states should be plausible. The calculator reports propensity ranges separately for treated and control groups to highlight overlap. Limited overlap often produces extreme weights, unstable estimates, and sensitivity to minor model changes. Practical steps include trimming non-overlapping regions, restricting to a common support interval, or redefining the target population to where comparisons are credible.
For causal estimation, weights depend on the estimand. For ATE, treated units receive 1/e and controls receive 1/(1−e). For ATT, treated units receive 1 and controls receive e/(1−e). The calculator computes standardized mean differences (SMD) before and after weighting for each selected covariate. As a rule of thumb, absolute SMD below 0.10 indicates good balance, though tighter thresholds may be preferred in high-stakes analyses.
To support auditability, exportable CSV output includes each row’s T, covariates, propensity, and weight, while the PDF summary captures diagnostics, overlap, coefficients, and balance results. Record the covariate set, any trimming decisions, and the estimand used, because these choices define the causal question. When results change materially across reasonable specifications, treat conclusions as fragile and consider sensitivity analyses or alternative adjustment strategies in operational settings today.
Include variables measured before treatment that affect both treatment assignment and outcome. Avoid post-treatment variables. Start with key drivers, then refine using balance checks and domain expertise.
No. The score only reflects treatment likelihood given covariates. It is not a risk score or outcome prediction. Use it to improve comparability between treated and control groups.
Extreme weights usually indicate limited overlap or very small e(x) or 1−e(x). They can inflate variance and make estimates unstable. Consider trimming, stabilizing weights, or restricting to common support.
Many analyses target absolute SMD below 0.10 after weighting or matching. For sensitive decisions, aim lower and also inspect overlap and weight distributions, not just a single threshold.
ATE targets the average effect in the full population represented by your data. ATT targets the effect among treated units. Choose the estimand that matches the decision context and reporting audience.
Not necessarily. High AUC can even worsen overlap by separating groups too strongly. Focus on achieving covariate balance and adequate overlap, then report diagnostics transparently.
Important Note: All the Calculators listed in this site are for educational purpose only and we do not guarentee the accuracy of results. Please do consult with other sources as well.