0
$\begingroup$

I've come accross the following paragraph in the To Explain or To Predict? paper by Galit Shmueli.

In explanatory modeling, data partitioning is less common [than in predictive modeling] because of the reduction in statistical power. When used, it is usually done for the retrospective purpose of assessing the robustness of ˆf. A rarer yet important use of data partitioning in explanatory modeling is for strengthening model validity, by demonstrating some predictive power. Although one would not expect an explanatory model to be optimal in terms of predictive power, it should show some degree of accuracy.

I understand why data partitioning is useful in the case of a predictive model, which is to assess the generalization capacity of a model. However, in the case of an explanatory model, I don't understand why it should show some degree of accuracy in terms of predictive power since it's not the objective of the model.

Here comes my question: is data partitioning necessary for an explanatory model and why?

$\endgroup$
0
$\begingroup$

An explanatory model is used to identify and explain what causes some particular outcome (i.e. identify/quantify the drivers of effect).

Even though the model will not be used for prediction, it needs to be accurate enough to adequately predict the response so that you can be sure the conclusions you draw from the model are valid. As an extreme example, if your explanatory model does not predict the outcome at all, then all the inferences you make from this model are useless.

So then, how does one validate model accuracy? In most cases, data partitioning and validation on the hold-out sample is used; in your case the validation is more of a sense-check for accuracy.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.