mkuzma96 / dta

Deconfounding Temporal Autoencoder: Estimating Treatment Effects over Time Using Noisy Proxies

Abstract: Estimating individualized treatment effects (ITEs) from observational data is crucial for decision-making. In order to obtain unbiased ITE estimates, a common assumption is that all confounders are observed. However, in practice, it is unlikely that we observe these confounders directly. Instead, we often observe noisy measurements of true confounders, which can serve as valid proxies. In this paper, we address the problem of estimating ITE in the longitudinal setting where we observe noisy proxies instead of true confounders. To this end, we develop the Deconfounding Temporal Autoencoder (DTA), a novel method that leverages observed noisy proxies to learn a hidden embedding that reflects the true hidden confounders. In particular, the DTA combines a long short-term memory autoencoder with a causal regularization penalty that renders the potential outcomes and treatment assignment conditionally independent given the learned hidden embedding. Once the hidden embedding is learned via DTA, state-of-the-art outcome models can be used to control for it and obtain unbiased estimates of ITE. Using synthetic and real-world medical data, we demonstrate the effectiveness of our DTA by improving over state-of-the-art benchmarks by a substantial margin.

Paper available at: https://proceedings.mlr.press/v158/kuzmanovic21a/kuzmanovic21a.pdf

dta's Issues

Reproducing Figure 3

I have read your paper with great interest and was looking into running the code in this repository on my own machine. Unfortunately, when running the code in sim_exp.py (or main.py) out-of-the-box I am unable to do so. Instead of a replicate of Figure 3a (marginal structural model), I get the following results:

Two main differences immediately jump out:

1. There is little difference between the three models i) with proxy confounders X only (base), ii) DTA deconfounded proxies (dta), and iii) an oracle with access to the true confounders L (oracle).
2. The absolute value of MSE is similar to that reported in the paper for no confounding (gamma=0) but is much lower than reported for gamma=0.4 or gamma=0.8

Due to computation time, I've only done full 10 seeds 50 hyperparameter search for MSM with the above three levels of gamma and rank=5 but I get comparable results on individual tries with different gammas, ranks and with RMSN.

Do you maybe have any idea as to what might be going wrong?