IPW (GEE calibration)

Inverse probability weighting where the propensity score parameters solve a generalised estimating equation (calibration-style constraints) rather than the maximum-likelihood score. Both smoothing variants from the literature are shown — gee_h_fun = 1 (logarithmic) and gee_h_fun = 2 (exponential).

nonprobsvy version: 0.3.0  |  R: 4.6.0  |  run: 2026-05-24 08:07:58  |  commit: 4b2ba9a

gee_h_fun var_method num_boot alpha n_reps bias rmse mean_se coverage ci_width
1 analytic NA 0.05 500 0.008 0.085 0.078 0.920 0.307
1 analytic NA 0.10 500 0.008 0.085 0.078 0.872 0.258
1 bootstrap 200 0.05 500 0.008 0.085 0.080 0.928 0.312
1 bootstrap 200 0.10 500 0.008 0.085 0.080 0.872 0.262
2 analytic NA 0.05 500 0.010 0.099 0.099 0.946 0.386
2 analytic NA 0.10 500 0.010 0.099 0.099 0.904 0.324
2 bootstrap 200 0.05 500 0.010 0.099 0.104 0.956 0.409
2 bootstrap 200 0.10 500 0.010 0.099 0.104 0.916 0.343

Notes

  • DGP: the same simple default DGP used for the MLE table (linear outcome, logistic selection on x1 and x2).
  • Compared to MLE, GEE typically gives slightly wider CIs because it solves a constraint system rather than maximising likelihood; that often translates into over-coverage.