Prior-data Fitted Networks for Causal Inference: a Simulation Study with Real-world Scenarios

Prior-data fitted networks (PFNs) represent a paradigm shift in tabular data prediction. We present the principles of this new paradigm and evaluate two PFNs for estimating the average treatment effect (ATE) of a binary treatment on a binary outcome, using simulated clinical scenarios based on real-world data. We assessed TabPFN, a predictive PFN, in combination with causal inference procedures such as g-computation and inverse probability of treatment weighting (IPTW), as well as CausalPFN, a PFN specifically designed for causal inference that directly provides an ATE estimate with its uncertainty interval. Confidence intervals for the TabPFN-based methods were derived using bootstrap resampling. We found that computation times for TabPFN were too long for causal inference applications relating to the need for bootstrap methods to compute uncertainty. Moreover, g-computation with TabPFN produced a biased estimator, which was partially corrected by fitting separate models for each treatment group (T-learner approach). CausalPFN, by contrast, was computationally efficient but exhibited poor coverage of the 95\% uncertainty interval for the ATE, due to both estimation bias and its uncertainty quantification procedure. Beyond automating model specification, some PFNs variants -- like CausalPFN -- attempt to automate causal modeling, but in the settings we evaluated its estimates were biased. However, their application in routine causal inference tasks needs further investigation.