- This event has passed.
Michael KNAUS (University of St Gallen) – “Implied Weights of Common Estimators for (Local) Average Treatment Effects: Review and New Results “
Time: 12:15 pm – 13:30 pm
Date: 8th of February 2022
Room : 3001 et visio
Abstract: Many estimators for causal inference can be written as a weighted sum of observed outcomes. This paper considers the implied weights of estimators for three research designs (experiments, selection-on-observables, instrumental variables) and three assumed outcome models (nonparametric, partially linear, linear). This includes, e.g. the Wald estimator, regression adjustment, (augmented) inverse probability weighting, partially linear estimators, and OLS. We show that these estimators can be unified as special cases of two-stage least squares and under which condition this provides a functional form of the implied weights. This includes weights for augmented inverse probability weighting and partially linear estimators that have not been derived in the literature. We analyze basic properties of the weights and demonstrate how they permit finite sample diagnostics to (i) check covariate balancing of the methods (internal validity), (ii) understand the targeted population of estimated effects (external validity).
Organizers:
Benoît SCHMUTZ (Pôle d’économie du CREST)
Anthony STRITTMATTER (Pôle d’économie du CREST)
Sponsors:
CREST
Time: 12:15 pm – 13:30 pm
Date: 8th of February 2022
Room : 3001 et visio
Michael KNAUS (University of St Gallen) – “Implied Weights of Common Estimators for (Local) Average Treatment Effects: Review and New Results”
Abstract: Many estimators for causal inference can be written as a weighted sum of observed outcomes. This paper considers the implied weights of estimators for three research designs (experiments, selection-on-observables, instrumental variables) and three assumed outcome models (nonparametric, partially linear, linear). This includes, e.g. the Wald estimator, regression adjustment, (augmented) inverse probability weighting, partially linear estimators, and OLS. We show that these estimators can be unified as special cases of two-stage least squares and under which condition this provides a functional form of the implied weights. This includes weights for augmented inverse probability weighting and partially linear estimators that have not been derived in the literature. We analyze basic properties of the weights and demonstrate how they permit finite sample diagnostics to (i) check covariate balancing of the methods (internal validity), (ii) understand the targeted population of estimated effects (external validity).
Organizers:
Benoît SCHMUTZ (Pôle d’économie du CREST)
Anthony STRITTMATTER (Pôle d’économie du CREST)
Sponsors:
CREST