Carrying Out Interventions Based on ML "Feature Importances"

Question

Recently, I have been studying causal inference and have come to a bit of a crossroads with respect to making decisions based on the analysis of data (especially in a business/industry setting). Specifically, I am referring to common problems like "churn modelling", segmentation, and lifetime value problems where the goal is to figure out specific demographics to "target" to increase revenue or to decrease churn, etc.

Often, I see these problems solved in the following way (whether good or bad). Take a bunch of predictors that are plausibly associated in some way with the outcome variable (whether that is churn, lifetime value, or some other profitability metric) and then fit a machine learning model to the problem (using the standard test sets/data splitting, etc.). Then, look at the feature importances of the best predictive model (perhaps using a method that corrects for multicollinearity, like SHAP scores) and determine the most impactful features, from which we understand as the most predictive variables. We can make then decisions on who to target, market to, etc. based on these influential variables.

Now, we know that none of this is causal in any way since we are just exploiting correlations. We didn't consider the actual causal structure of the problem, draw out DAG's like Pearl suggests, and condition on sufficient adjustment sets to derive causal effects (and ultimately see the impact of "treatments"). Through careful handling of causality, we can deal with issues that may arise from the above approach like Simpson's paradox, for example.

My question is as follows: is the first method of modelling, and ultimately, the business decisions made from the first method, incorrect or dangerous? Equivalently, is absolute causality needed in this setting? I can see why this may be the case - but in say a huge dataset with many predictors and proper regularization, I have a tough time believing that the ML approach would lead to outright bad decisions (though perhaps not quite as strong conclusions). In addition, I think many would agree that the first method is less time-consuming. Writing out a causal model is difficult, especially when there is a lack of expertise.

Answer 1

Considering this came up in a work context, perhaps there was some confusion on what the goal of the analysis was.

If the goal is to identify who to market to people likely to click -- as in the case of uplift analysis or similar-- then a predictive model should be fine. In uplift, the goal is to target only those people who are likely to open the email/click the ad/whatever. The mechanism of why they clicked is irrelevant. You just want to know who is most likely to click and that is a prediction problem.

If, however, the goal is to take a customer who is unlikely to click intervene on them in such a way to cause them to click the ad, then a causal approach is needed. "Data-driven recommendations to improve customer experience " seems causal to me, at least in the way you're written it, so I'm willing to think this is the context we find ourselves in.

OK, but that doesn't answer the question. Why should we draw dags and do our causal analysis this way rather than just throw everything in a regression model? Richard McElreath gives some pretty compelling examples of why "Causal Salad" -- his pejorative name for throwing everything in a linear or machine learning model -- doesn't work. In chapters 5 & 6 of Statistical Rethinking, Richard gives several examples through simulation in which the true causal mechanism is poorly estimated when you don't draw the dag. I won't take the time to regurgitate those examples here as I wouldn't do them justice.

Suffice to say, you can very easily think your intervention is helpful when in reality it is hurtful if you don't take the time to draw your assumptions before your conclusions. So your approach is technically wrong, but the danger is presently unknown. For example, assume you estimated a positive treatment effect but in reality the effect was null. Nothing gained, nothing lost -- except money.