#propensity-score

Design strong quasi-experiments when randomization impossible. Design types: 1. Non-equivalent groups: compare treatment and comparison groups without random assignment. 2. Interrupted time series: multiple observations before and after intervention. 3. Regression discontinuity: treatment assigned based on cutoff score. 4. Instrumental variables: use natural randomization to estimate causal effects. Strengthening designs: 1. Propensity score matching: match treatment and control on likelihood of receiving treatment. 2. Difference-in-differences: compare changes over time between treatment and control areas. 3. Multiple baselines: stagger intervention across participants or sites. Threats to validity: 1. Selection bias: groups differ systematically. 2. History: events coinciding with intervention. 3. Maturation: natural change over time. Analysis considerations: intention-to-treat vs. per-protocol analysis, sensitivity analysis for unobserved confounders, report effect sizes and confidence intervals.

Control for confounding variables in observational studies. Design-based controls: 1. Randomization: Random assignment eliminates selection bias. 2. Restriction: limit study to homogeneous group (e.g., only males, specific age range). 3. Matching: match cases and controls on potential confounders (age, gender, education). Analysis-based controls: 1. Stratification: analyze results within strata of confounder levels. 2. Multiple regression: include confounders as covariates in regression model. 3. Propensity score matching: calculate probability of exposure, match on propensity scores. 4. Instrumental variables: use natural randomization when available. Assessment: Create directed acyclic graphs (DAGs) to identify confounders vs. mediators vs. colliders. Use causal inference framework to determine which variables to control. Report all controlled variables and rationale for inclusion.