🤖 AI

When Prediction Error Is Not Enough: Evaluating Nuisance-Function Prediction for Causal Estimation

This paper demonstrates through Monte Carlo simulations that while prediction error is a useful metric for assessing nuisance-function estimation, it does not consistently correlate with causal estimator performance (such as bias or confidence interval coverage), indicating that it should not be relied upon as a direct proxy for the quality of causal inference.

Original authors: Cong Cao

Published 2026-09-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Cong Cao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of data science, researchers often face a puzzle: how to tell if one thing causes another when they cannot run a controlled experiment. Imagine trying to understand if a new medicine works, but you only have records of people who chose to take it on their own, mixed with those who did not. To untangle this, scientists use a method called causal inference. This approach relies on building mathematical models to describe the background factors—like age, income, or health history—that influence both the decision to take the medicine and the health outcome. These background models are known as "nuisance functions." They are called "nuisance" not because they are annoying, but because they are necessary distractions; the researcher must estimate them accurately to isolate the true effect of the treatment. For years, the standard way to check if these background models were any good was to see how well they predicted the data, much like a weather forecaster checking if their temperature predictions matched the actual weather. The assumption was simple: if the model predicts the background data well, it should also help find the correct cause-and-effect answer.

A recent study by Cong Cao at Yale University challenges this long-held assumption. The researcher set out to test whether a model that is excellent at predicting background data is necessarily excellent at finding the true cause. To do this, Cao did not look at real-world medical records but instead created thousands of simulated worlds on a computer. In these simulations, the true cause-and-effect relationship was known by design, allowing the researcher to see exactly how well different computer programs performed. The study compared several popular methods for building these background models, ranging from traditional statistical tools to modern, flexible machine learning algorithms. The goal was to see if the score a model gets for predicting the background data matched the score it gets for finding the correct causal answer.

The results revealed a surprising disconnect. The study found that the method which was best at predicting the background data was not always the best at estimating the causal effect. In the simulations, a machine learning tool called XGBoost was the most accurate at predicting the background variables, producing the smallest errors in its guesses. However, when it came to the final goal of estimating the causal effect, another method called Double Machine Learning, which used XGBoost inside a specific statistical framework, performed better at a different crucial task: providing reliable confidence intervals. A confidence interval is a range of values that researchers use to express how sure they are about their answer. In these simulations, the method with the best prediction accuracy gave a range that was too narrow, meaning it was overconfident and often missed the true answer. The method with slightly worse prediction accuracy, however, produced wider, more honest ranges that captured the true answer about 93 percent of the time, which is very close to the ideal 95 percent.

This suggests that being a good predictor is not the same as being a good detective for cause and effect. The study showed that a model can be very precise at guessing the background numbers but still fail to give a trustworthy range for the final answer. The researchers also tested a new idea: whether looking at the combined errors of two different background models could predict the final result. They found that this combined measure was weak and did not reliably tell them which method would work best. The findings indicate that while checking how well a model predicts data is useful, it is not enough on its own to guarantee a good causal conclusion. Researchers cannot simply pick the model with the lowest prediction error and assume the causal answer will be correct. Instead, they must look at the specific properties of the causal method itself, such as how it handles uncertainty, to ensure the final conclusion is robust.

The study also explored what happens when data points are not independent, such as when patients are grouped together in the same hospital or family, which creates a hidden link between them. Even in these more complex, clustered situations, the pattern held true. The method that predicted the background data best still did not provide the most reliable confidence intervals. The researchers noted that their work was based on computer simulations with specific conditions, so the results might look different in other real-world scenarios with different types of data. However, the core message remains clear: in the quest to find cause and effect, a model's ability to predict the past does not automatically translate into an ability to explain the future. The tools used to clean up the background noise must be judged by how well they protect the final answer, not just by how well they guess the noise itself.

Technical Summary: When Prediction Error Is Not Enough

Problem Statement
In causal inference using observational data, flexible machine learning methods are increasingly employed to estimate nuisance functions (e.g., the relationship between covariates and exposure, or covariates and outcome). These nuisance functions are typically evaluated using standard supervised learning metrics, such as prediction error (e.g., mean squared error) or cross-validation performance. However, it remains unclear whether superior prediction accuracy of these nuisance functions translates directly to superior performance of the resulting causal estimator. Specifically, the paper investigates whether minimizing nuisance-function prediction error is a sufficient proxy for minimizing causal bias, root mean squared error (RMSE), and ensuring valid confidence interval coverage. This question is particularly complex in settings like Double Machine Learning (DML), where errors in multiple nuisance functions interact to affect the final causal estimate.

Methodology
The author conducted a Monte Carlo simulation study within a partially linear causal model framework with a continuous exposure. The study compared four estimation approaches:

Ordinary Least Squares (OLS): A parametric benchmark.
Generalized Additive Models (GAMs): Allowing for nonlinear additive relationships.
XGBoost: A flexible non-parametric learner allowing for nonlinearities and interactions.
Double Machine Learning with XGBoost (DML-XGBoost): Combining XGBoost with Neyman-orthogonal scores and cross-fitting to reduce the impact of nuisance estimation error.

The simulation evaluated performance across two settings:

Scenario 1: Independent observations.
Scenario 2: Clustered data with within-cluster dependence, utilizing cluster-level cross-fitting and cluster-robust variance estimation.

Performance Metrics
The study assessed two distinct categories of metrics:

Nuisance-Function Metrics:
- Mean squared prediction errors for the exposure ( $E_m$ ) and outcome ( $E_h$ ) nuisance functions.
- A joint-error measure, $D_{joint} = |E\{\delta_m(C)\delta_h(C)\}|$ , representing the absolute magnitude of the cross-product of estimation errors between the two nuisance functions.
Causal Estimator Metrics:
- Bias (difference between the mean estimate and the true causal effect $\tau$ ).
- RMSE.
- 95% Confidence Interval (CI) coverage.

Key Results

Divergence of Metrics: Prediction error did not consistently track causal performance. In both scenarios, XGBoost (without cross-fitting) achieved the lowest RMSE among non-oracle methods, indicating superior point-estimation accuracy. However, DML-XGBoost provided significantly better confidence interval coverage (0.928 and 0.908 in Scenarios 1 and 2, respectively) compared to standard XGBoost (0.856 and 0.838), which suffered from under-coverage.
Prediction Error vs. Inference: While XGBoost had lower nuisance-function prediction errors than DML-XGBoost in Scenario 1, this lower error did not correspond to better inferential calibration. Thus, the method with the best point-estimation performance was not the method with the best confidence interval coverage.
Joint-Error Measure: The proposed joint-error measure ( $D_{joint}$ ) showed only weak associations with absolute causal bias (correlations ranging from approximately -0.11 to 0.11). The author concludes that $D_{joint}$ is not a reliable standalone predictor of finite-sample causal bias or a useful criterion for selecting causal estimators.
Clustered Data: The main patterns held in the clustered setting, with DML-XGBoost maintaining better coverage than standard XGBoost, though the analysis was limited to a single cluster size and variance component configuration.

Significance and Claims
The paper argues that while prediction error is a valid and useful tool for assessing the quality of nuisance-function estimation, it should not be treated as a direct measure of the quality of the resulting causal estimator.

The author emphasizes that:

Point estimation and inference are distinct: A method can minimize prediction error and achieve low bias (good point estimation) while failing to provide valid uncertainty quantification (poor CI coverage).
Nuisance assessment is multi-faceted: Prediction error measures accuracy relative to the true function, but it does not capture sample-to-sample variation or the specific way errors in multiple nuisance functions interact to affect the sampling distribution of the causal estimator.
Limitations of Joint Measures: Simple summaries of joint nuisance errors, such as the cross-product of errors, do not sufficiently characterize finite-sample causal bias and should be viewed as descriptive diagnostics rather than selection criteria.

Ultimately, the study suggests that researchers should not rely solely on prediction metrics to select learners for causal inference, as better prediction does not necessarily guarantee better causal estimation performance, particularly regarding inferential validity.

Technical Summary: When Prediction Error Is Not Enough

More like this