Can Predictive Models Uncover Causal Effects? A Truck Risk Case Study

Abstract

This article uses a road freight risk prediction case to illustrate that treating the interpretability of predictive models as causal inference can be deceptive; it outlines situations where a predictive model can legitimately address causal questions and advises the use of causal inference packages such as ecoml for trustworthy effect estimation.

Introduction

Algorithm engineers often face business requests like “Which risk factors are most critical?” or “What drives abnormal claim rates?” in the context of road freight accident forecasting. While predictive metrics (AUC, accuracy, recall) and tools like SHAP enhance model transparency, assuming that SHAP explanations can directly prescribe actions leads to confusion between correlation and causation.

Truck Risk Control Case

We built a model to predict whether a truck will be involved in an accident. After feature engineering we obtained 28 important variables (total mileage, highway mileage, night‑time mileage, trajectory span, etc.) and trained a basic XGBoost classifier. model = fit_xgboost(X, y) With the trained model we applied SHAP to obtain global feature importance.

explainer = shap.Explainer(model)
shap_values = explainer(X)
clust = shap.utils.hclust(X, y, linkage="single")
shap.plots.bar(shap_values, clustering=clust, clustering_cutoff=1)

The bar chart highlights night‑time travel proportion, village‑road proportion, and highway proportion as the top three drivers of accident predictions.

Deeper inspection with SHAP scatter plots reveals less intuitive patterns. For example, higher night‑time travel share correlates with higher predicted accident probability, while a larger east‑west to north‑south travel ratio (wh_ratio) appears to reduce risk.

shap.plots.scatter(shap_values)

Prediction vs. Causal Tasks

A prediction task aims to estimate model(X) that approximates Y under the same data distribution, relying on statistical correlation between X and Y. A causal task, however, asks whether intervening on a feature X would change the outcome Y, which requires stability of the relationship across environments and independence from unobserved confounders.

Causal Effect Estimation Challenges

Unobserved confounders such as pandemic impacts or policy changes can bias causal estimates. Open‑source libraries like ecoml provide frameworks for rigorous causal effect estimation.

When Predictive Models Can Answer Causal Questions

In the truck risk example, the feature wh_ratio shows strong independence: its predictive power is not redundant with any measured or unmeasured variables, as evidenced by the feature‑redundancy clustering tree where it merges only at the top level.

This independence suggests that, in limited scenarios, a predictive model can provide causal insight for features that are not entangled with other variables. Nevertheless, true causal inference still demands explicit causal modeling.

Conclusion

Flexible models like XGBoost or LightGBM excel at prediction, and SHAP enhances interpretability. However, correlation does not imply causation; using predictive explanations for causal decisions requires caution. For reliable causal effect estimation, dedicated causal tools such as Microsoft’s ecoml should be employed.