Universität zu Köln

Bolwin, Lennart (2026). Empirical Methods for Causal Inference and Forecasting in Panel Data Settings. PhD thesis, Universität zu Köln.

PDF Bolwin_Dissertationsschrift.pdf - Published Version Download (9MB)

Abstract

Economic forecasting must cope with short samples, model uncertainty and heterogeneity across units. This dissertation develops three econometric approaches to these obstacles. Chapter 2 builds on the synthetic control method, whose restriction of the donor weights to a convex combination regularizes the estimate and keeps it interpretable. Preserving this idea in more flexible form, the developed REGSC-estimator shrinks the weights toward zero and their sum toward unity. It is available in closed form, admits a Bayesian representation yielding credibility intervals and, with lagged donors, covers non-stationary and cointegrated data. Chapter 3 addresses model uncertainty, forecasting German GDP growth by dynamic model averaging enriched with search-based indicators. A parsimonious modification lets coefficient variability differ across regressors, and controlled experiments give the first systematic evidence on when they help. In a second step, DMA serves as a diagnostic tool for assessing expert forecasts. Chapter 4 turns to dynamic panels with many units and few periods. Common parameters are precise here, so accuracy hinges on the individual effects, whose estimation error dominates the forecast error. Initial conditions govern this trade-off: conditioning on all observations is efficient once the process has converged while quasi-differencing is safer otherwise. MSE-optimal fixed- and random-effects predictors are derived, extended to time-varying covariates via a Mundlak specification, and benchmarked against an empirical Bayes alternative. Simulations and a design-based application to firm-level productivity data favor random-effects prediction under non-stationary initialization.