Machine learning models can often predict outcomes accurately, but prediction alone does not tell us why those outcomes occur.
In economics, researchers usually want more than prediction. They want to understand which variables matter, how they matter, and whether relationships reflect genuine economic mechanisms.
This chapter examines feature importance and explains why prediction should not be confused with causation.
Which product characteristics contribute most to predicting milk prices?
Machine learning models can identify variables that improve prediction. However, variables that are important for prediction are not necessarily causes of the outcome.
Prediction importance ≠ Economic importance ≠ Causal importanceA variable can be useful for prediction without causing the outcome.
Feature importance measures the contribution of a variable to predictive performance. When a model uses several predictors, some variables may help prediction more than others.
For example, a Random Forest predicting milk prices may use volume, brand, fat content, package type, and freshness. Feature importance attempts to quantify how much each variable helps improve prediction.
| Variable | Importance |
|---|---|
| Volume | 0.48 |
| Brand | 0.27 |
| Fat | 0.15 |
| Package | 0.07 |
| Freshness | 0.03 |
Higher importance values indicate that a variable contributes more to prediction. They do not indicate economic significance or causality.
Random Forests naturally generate importance measures. The algorithm evaluates how much each variable improves prediction across many trees.
import pandas as pd
importance = pd.DataFrame({
"Variable": X_train.columns,
"Importance": forest_model.feature_importances_
})
importance = importance.sort_values(
"Importance",
ascending=False
)
print(importance) Variable Importance
0 Volume 1.0import matplotlib.pyplot as plt
importance.sort_values("Importance").plot(
x="Variable",
y="Importance",
kind="barh",
legend=False
)
plt.title("Feature Importance")
plt.show()
The graph ranks predictors according to their contribution to model performance. Researchers can use this ranking as a starting point for further investigation.
Standard importance measures can sometimes be misleading. Permutation importance provides an alternative.
For each variable:
If prediction accuracy deteriorates sharply, the variable was important.
from sklearn.inspection import permutation_importance
perm = permutation_importance(
forest_model,
X_test,
y_test,
n_repeats=10,
random_state=4107
)
perm_importance = pd.DataFrame({
"Variable": X_test.columns,
"Importance": perm.importances_mean
}).sort_values("Importance", ascending=False)
print(perm_importance) Variable Importance
0 Volume 1.650983Permutation importance asks what happens when the model loses useful information from one variable. If accuracy falls sharply, the variable was important for prediction.
Feature importance tells us which variables matter for prediction. Partial dependence plots help us understand how predictions change as a variable changes.
For example:
from sklearn.inspection import PartialDependenceDisplay
PartialDependenceDisplay.from_estimator(
forest_model,
X_train,
["Volume"]
)
Partial dependence plots improve transparency. They help researchers understand how a machine learning model behaves. However, they still do not establish causal relationships.
Suppose a model finds that premium brands are important predictors of price. Can we conclude that becoming a premium brand causes higher prices?
Not necessarily.
Premium brands may differ in quality, marketing, consumer preferences, distribution, and product positioning. The model shows predictive value, not a clean causal effect.
A classic example is ice cream sales and drowning incidents. Both may increase during hot weather. Ice cream sales may help predict drowning incidents, but ice cream sales do not cause drowning.
Machine learning is useful when the objective is prediction. Examples include forecasting food prices, predicting crop yields, demand forecasting, credit scoring, and consumer behavior prediction.
Its strengths include handling large datasets, capturing nonlinear relationships, detecting complex interactions, and often improving prediction.
Machine learning does not automatically solve fundamental econometric problems, such as omitted variable bias, endogeneity, reverse causality, measurement error, and selection bias.
A model may predict agricultural exports accurately, but this does not mean it identifies the causal effect of trade agreements on exports. Causal analysis requires research design, theory, and appropriate econometric methods.
Machine learning should be viewed as a complement to econometrics, not a replacement.
A practical workflow is:
This combines the strengths of both traditions.
| Chapter | Main theme |
|---|---|
| 22 | Train-test split and prediction |
| 23 | Regression versus machine learning |
| 24 | Decision Trees and Random Forests |
| 25 | XGBoost and model comparison |
| 26 | Feature importance and prediction limits |
Together, these chapters introduce predictive analytics while maintaining the econometric focus of the course.
Do not interpret feature importance as evidence of causality. Feature importance identifies variables that improve prediction. Causal inference requires theory, research design, and econometric methods.
The next part of the course moves from prediction to research practice. Students will learn how to formulate research questions, organize data, write methodology sections, present empirical results, and prepare a complete empirical article.