In cluster-randomized trials, individuals within a cluster are assigned the same treatment condition, but the treatment uptake status may vary across individuals due to noncompliance. We propose a semiparametric framework to evaluate the individual compliance effect and network assignment effect within principal strata exhibiting different patterns of noncompliance. The individual compliance effect captures the portion of the treatment effect attributable to changes in treatment receipt, while the network assignment effect reflects the pure impact of treatment assignment and spillover among individuals within the same cluster. We characterize new structural assumptions for nonparametric point identification, and we develop semiparametrically efficient estimators that combine data-adaptive machine learning methods with efficient influence functions.
In the presence of interference, IPW estimators often suffer from high variance. Under low-rank assumptions on the potential outcomes and in the presence of interference, we design optimal covariate balancing estimators. The framework encompasses commonly-invoked assumptions such as stratified or additive interference.
We study design-based causal inference when there are two distinct sets of units, one on which interventions are applied, and one on which the outcome is measured. We introduce various causal estimands and propose weighting estimators from a design-based perspective. We derive the estimators' variance and prove consistency for a growing bipartite graph. In our analysis, we illustrate complications that arise from the positivity assumption, the experimental design, and the structure of the graph.
When the treated units are spatial areas, their relationship with the control units is expected to exhibit a spatial relationship. Under the vertical regression framework, we propose a Bayesian approach for estimation of causal effects with spatial panel data.
Estimating the parameters of a temporal, spatio-temporal, or mutually-exciting Hawkes process based on data that are available in aggregated form by time, space, or both.
Investigating treatment effect heterogeneity with spatio-temporal data, point pattern treatment and outcome, and spatial or spatio-temporal potential moderators
In cluster randomized experiments with selection bias due to recruitment, data are often only available on those that were recruited. Based on a principal stratification framework, we show that causal effects on the overall population are identifiable based on the recruited sample only.
We investigate the complications and opportunities when drawing causal inference from spatial observational data. We introduce causal diagrams that allow us to investigate the impact of spatial confounders, interference, and the inherent spatial structure in the exposure variable, and we illustrate that causal inference with spatial data has crucial differences to counterparts with independent observations. We then propose an approach that mitigates bias from unmeasured spatial confounding and incorporates interference within one framework.