Assessing Balance After Matching
What you'll learn: How to verify that matching successfully removed covariate imbalances between treatment and control groups.
Why Balance Assessment Matters
After performing matching (exact, propensity score, or CEM), you can't just assume it worked. You need to check that the treated and control groups now look similar on the covariates that matter. Think of it like balancing a scale—matching is your attempt to balance it, but you must verify the scale is actually level before trusting your causal estimates.
Standardized Mean Differences (SMD)
The primary diagnostic tool is the standardized mean difference. For each covariate, you calculate:
- The difference in means between treated and control groups
- Divide by the pooled standard deviation
A rule of thumb: SMD values below 0.10 (sometimes 0.20) indicate good balance. Values above suggest persistent imbalance that could bias your treatment effect estimate.
Visual Diagnostics
Love plots (or balance plots) display SMDs before and after matching, making it easy to see which covariates improved and which remain problematic. Each covariate gets a dot showing its SMD, with a vertical reference line at zero (perfect balance) and threshold lines at ±0.10.
You might also use side-by-side density plots or histograms to visually compare covariate distributions between groups after matching.
What to Do If Balance Fails
If balance is poor after matching, you need to refine your approach: try different matching methods, change caliper widths (in propensity score matching), include additional covariates, or consider alternative methods entirely.
Key Takeaway: Balance assessment using standardized mean differences and visual diagnostics is essential to verify matching eliminated covariate imbalances—never skip this diagnostic step before estimating treatment effects.