Causal Inference in Observational Studies and Experiments: Theory and Applications

Sammanfattning: This thesis consists of six papers that study the design of observational studies and experiments.Paper I proposes strategies to consistently estimate the average treatment effect of the treated using information derived from a large number of pre-treatment measurements of the outcome. The key to this strategy is to use two-level time-series model estimates to summarize the inter-unit heterogeneity in the sample. It is illustrated how this approach is in line with the conventional identifying assumptions, and how sensitivity analyses of several key assumptions can be performed.Paper II contains an empirical application of the identification strategy proposed in Paper I. This study provides the first causal analysis of the demand response effects of a billing demand charge involuntarily introduced to small and medium sized electricity users.Paper III proposes strategies for rerandomization. First, we propose a two-stage allocation sample scheme for randomization inference to the units in balanced experiments that guarantees that the difference-in-means estimator is an unbiased estimator of the sample average treatment effect for any experiment, conserves the exactness of randomization inference, and halves the time consumption of the rerandomization design. Second, we propose a rank-based covariate-balance measure which can take into account the estimated relative weight of each covariate.Paper IV discusses the concept of optimal rerandomization. It is shown that depending on whether inference is to be drawn to the units of the sample or the population, the notion of optimal differs. We show that it is often advisable to aim for a design that is optimal for inference to the units of the sample, as such a design is often near-optimal also for inference to the units of the population.Paper V summarizes the current knowledge on asymptotic inference for rerandomization designs and proposes some simplifications for practical applications. Drawing on previous work, we show that the non-normal sampling distribution of the difference-in-means test statistic approaches normal as the rerandomization criterion approaches zero. Furthermore, the difference between the correct non-normal distribution and the proposed approximation based on a normal distribution is in many situations negligible even for near optimal rerandomization criteria.Paper VI investigates and clarifies the relation between the traditional blocked designs and rerandomization. We show that blocking and rerandomization is very similar, and in some special cases identical. Moreover, it is shown that combining blocking and rerandomization is always at least as efficient as using only rerandomization, but the difference is in many cases small.