If needed, more relevant papers

[dvirzg]

Can someone take a look at these papers and methods?

• Continuous-Time Modeling of Counterfactual Outcomes Using Neural Controlled Differential Equations
• Estimating Counterfactual Treatment Outcomes Over Time Through Adversarially Balanced Representations
• Bayesian Neural Controlled Differential Equations for Treatment Effect Estimation

[dvirzg]

Code for TE-CDE (Continuous-Time Modeling of Counterfactual Outcomes Using Neural Controlled Differential Equations)

[dvirzg]

@sneha3799 and @safiyamak here's what I mentioned

[sneha3799]

I will take a look at Continuous-Time Modeling of Counterfactual Outcomes Using Neural Controlled Differential Equations if it's fine and try working on GNet implementation. I think all the papers have made their code available online.

[dvirzg]

Bayesian Neural Controlled Differential Equations for Treatment Effect Estimation seems to good for personalized modelling, good re uncertainty quantification, but missing in correction for time-varying confounders:

As with any other method for treatment effect estimation, ours relies upon mathematical assumptions. Importantly, the above implementation of our BNCDE does not account for confounding bias.

The other two articles do look very promising, I skimmed the Adversarially Balanced Representations paper and it addresses time deconfounding, autocorrelation (since the model is based on an LSTM), and personalized ITEs rather than ATEs. The dependencies are a bit old so when I tried to get it to run on my machine I had to change some of the code, but worth a deeper dive before next week's session.

@sneha3799

[dvirzg]

@jogong2718 did you take a look at the TE-CDE paper?

[jogong2718]

Sorry just saw this. Yes I took a look at the paper and I think it has some real promise. The main point is how the model is best at tackling irregular sampling (could be gaps from imperfect glucose monitoring?) using their encoder and controlled diff eq model. Main tradeoffs are if the data are pretty well scheduled (5-mins between each sample) the model is not really needed since the paper says that other more lightweight models do well on that data already. They also highlight continuous counterfactual paths when comparing with rnns, but most causal models do that already. Overall, their application of their model is on irregular lung cancer tumor data, not sure how irregular and patchy our data is for this model to prove useful but we can certainly try.