Why Causal Inference Matters More Than Prediction in Development Research
March 15, 2026
Predictive accuracy is a seductive metric. A model that predicts child stunting with 90% AUC feels like progress. But in development research, prediction is rarely the goal — intervention is.
The Core Problem
When a ministry of health asks “which children should we target with nutritional support?”, they are not asking for a prediction model. They are asking a causal question: which children will benefit most from the intervention?
These are different questions. A model trained to predict outcomes from observational data conflates correlation with causal effect. High socioeconomic status correlates with good outcomes — but targeting rich children for nutrition programs would be absurd.
Potential Outcomes Framework
The counterfactual framework makes this precise. For each child $i$, define:
- $Y_i(1)$: outcome if treated
- $Y_i(0)$: outcome if untreated
- $\tau_i = Y_i(1) - Y_i(0)$: individual treatment effect
We can never observe both $Y_i(1)$ and $Y_i(0)$ for the same child — this is the fundamental problem of causal inference. But we can estimate the average treatment effect $\mathbb{E}[\tau_i]$ or the conditional average treatment effect $\mathbb{E}[\tau_i \mid X_i = x]$ under assumptions.
What This Means for Our Work
In the Green-NAS paper, we were careful to frame our contributions as predictive — we are not claiming that wider transformers cause better performance in all settings. In the child development encoder work, we explicitly model the deployment setting as a transfer learning problem, not a causal one.
The field needs more researchers who can draw this line clearly.