Search

We combine two recently proposed nonparametric difference-in-differences methods, extending them to enable the examination of treatment effect heterogeneity in the staggered adoption setting using machine learning. The proposed method, machine learning difference-in-differences (MLDID), allows for estimation of time-varying conditional average treatment effects on the treated, which can be used to conduct detailed inference on drivers of treatment effect heterogeneity. We perform simulations to evaluate the performance of MLDID and find that it accurately identifies the true predictors of treatment effect heterogeneity. We then use MLDID to evaluate the heterogeneous impacts of Brazil's Family Health Program on infant mortality, and find those in poverty and urban locations experienced the impact of the policy more quickly than other subgroups.

Machine learning (ML) estimates of conditional average treatment effects (CATE) can guide policy decisions, either by allowing targeting of individuals with beneficial CATE estimates, or as inputs to decision trees that optimise overall outcomes. There is limited information available regarding how well these algorithms perform in real-world policy evaluation scenarios. Using synthetic data, we compare the finite sample performance of different policy learning algorithms, machine learning techniques employed during their learning phases, and methods for presenting estimated policy values. For each algorithm, we assess the resulting treatment allocation by measuring deviation from the ideal ("oracle") policy. Our main finding is that policy trees based on estimated CATEs outperform trees learned from doubly-robust scores. Across settings, Causal Forests and the Normalised Double-Robust Learner perform consistently well, while Bayesian Additive Regression Trees perform poorly. These methods are then applied to a case study targeting optimal allocation of subsidised health insurance, with the goal of reducing infant mortality in Indonesia.

While machine learning (ML) methods have received a lot of attention in recent years, these methods are primarily for prediction. Empirical researchers conducting policy evaluations are, on the other hand, pre-occupied with causal problems, trying to answer counterfactual questions: what would have happened in the absence of a policy? Because these counterfactuals can never be directly observed (described as the "fundamental problem of causal inference") prediction tools from the ML literature cannot be readily used for causal inference. In the last decade, major innovations have taken place incorporating supervised ML tools into estimators for causal parameters such as the average treatment effect (ATE). This holds the promise of attenuating model misspecification issues, and increasing of transparency in model selection. One particularly mature strand of the literature include approaches that incorporate supervised ML approaches in the estimation of the ATE of a binary treatment, under the \textit{unconfoundedness} and positivity assumptions (also known as exchangeability and overlap assumptions). This article reviews popular supervised machine learning algorithms, including the Super Learner. Then, some specific uses of machine learning for treatment effect estimation are introduced and illustrated, namely (1) to create balance among treated and control groups, (2) to estimate so-called nuisance models (e.g. the propensity score, or conditional expectations of the outcome) in semi-parametric estimators that target causal parameters (e.g. targeted maximum likelihood estimation or the double ML estimator), and (3) the use of machine learning for variable selection in situations with a high number of covariates., Comment: 40 pages 3 Figures

In studies based on electronic health records (EHR), the frequency of covariate monitoring can vary by covariate type, across patients, and over time. This can lead to major challenges: first, the difference in monitoring protocols may invalidate the extrapolation of study results obtained in one population to the other, and second, monitoring can act as a time-varying confounder of the causal effect of a time-varying treatment on the outcomes of interest. This paper demonstrates how to account for non-systematic covariate monitoring when evaluating dynamic treatment interventions, and how to evaluate joint dynamic treatment-censoring and static monitoring interventions, in a real world, EHR-based, comparative effectiveness research (CER) study of patients with type II diabetes mellitus. First, we show that the effects of dynamic treatment-censoring regimes can be identified by including indicators of monitoring events in the adjustment set. Second, we demonstrate the poor performance of the standard inverse probability weighting (IPW) estimator of the effects of joint treatment-censoring-monitoring interventions, due to a large decrease in data support resulting in a large increase in standard errors and concerns over finite-sample bias from near-violations of the positivity assumption for the monitoring process. Finally, we detail an alternate IPW estimator of the effects of these interventions using the No Direct Effect assumption. We demonstrate that this estimator can result in improved efficiency but at the cost of increased bias concerns over structural near-violations of the positivity assumption for the treatment process. To conclude, this paper develops and illustrates new tools that researchers can exploit to appropriately account for non-systematic covariate monitoring in CER, and to ask new causal questions about the joint effects of treatment and monitoring interventions.

We consider the estimation of the average treatment effect in the treated as a function of baseline covariates, where there is a valid (conditional) instrument. We describe two doubly robust (DR) estimators: a locally efficient g-estimator, and a targeted minimum loss-based estimator (TMLE). These two DR estimators can be viewed as generalisations of the two-stage least squares (TSLS) method to semi-parametric models that make weaker assumptions. We exploit recent theoretical results that extend to the g-estimator the use of data-adaptive fits for the nuisance parameters. A simulation study is used to compare standard TSLS with the two DR estimators' finite-sample performance, (1) when fitted using parametric nuisance models, and (2) using data-adaptive nuisance fits, obtained from the Super Learner, an ensemble machine learning method. Data-adaptive DR estimators have lower bias and improved coverage, when compared to incorrectly specified parametric DR estimators and TSLS. When the parametric model for the treatment effect curve is correctly specified, the g-estimator outperforms all others, but when this model is misspecified, TMLE performs best, while TSLS can result in large biases and zero coverage. Finally, we illustrate the methods by reanalysing the COPERS (COping with persistent Pain, Effectiveness Research in Self-management) trial to make inference about the causal effect of treatment actually received, and the extent to which this is modified by depression at baseline., Comment: 24 pages, 3 figues