Research highlight: Maurice Bun (UvA-Econometrics)
In economics typically observational or non-experimental data are used for policy intervention analysis or program evaluation. A major issue is selection bias or omitted variables bias. We should be confident that unobserved determinants of the outcome variable of interest are uncorrelated with the policy intervention. The use of longitudinal or panel data instead of cross-section analysis can be of considerable advantage in controlling for these unobserved systematic differences between entities (individuals, firms, sectors, regions) affected and not affected by the policy change or treatment.
In my NWO-VIDI project we aim at developing accurate statistical inference methods for panel data regression models used for policy intervention analysis. Although selection bias can be mitigated by the use of panel data, there still remain other important threats to the internal validity of such research results. In the research project we analyse causal inference with panel data models allowing for (1) endogenous policy interventions; (2) dynamic adjustment processes; (3) extensive modelling of unobserved heterogeneity; (4) heterogeneous causal effects.
In applied economic research standard practice nowadays is to use Instrumental Variables (IV) methods or the Generalized Method of Moments (GMM) to overcome these various endogeneity and parametrisation issues. At the same time, however, application of IV or GMM inference methods has proved to be notoriously difficult due to, among other things, weak identification issues and lack of invariance with respect to crucial nuisance parameters.
Regarding the use of IV and GMM methods for panel data models we aim at developing: (1) IV/GMM based statistical inference methods robust for number and weakness of instruments; (2) IV/GMM testing procedures for detecting weak instruments; (3) alternative inference procedures with null distributions largely invariant to important nuisance parameters; (4) alternative models and methods allowing for time varying unobserved heterogeneity and heterogeneous causal effects.