2
$\begingroup$

I have implemented machine learning (RF) for analysis of a dataset, evaluating observations' survival.
After analysis (including variable importance) one of my collegues asked me that does ML method consider independent effect of a variable, like what we see in adjusted multivariate coxph analysis ?
I did not know the exact answer.
As I know in ML methods, for example RF, split rule is log rank test, which does not account and does not adjust the confounding effect of the other variables.
So for example if another variable has remarkable effect on both target and our variable of interest, based on logrank splitrule,RF would not know the confounding effect of another variable, and thinks that this is our variable of interest which causes the target.
So I don't know that if my conclusion is correct or not? I did not find the answer in SE pointing out machine learning. I really appreciate your help

$\endgroup$

2 Answers 2

3
$\begingroup$

Many machine learning models do not care about the effect of confounding variables, because they only care about the accuracy of the model, but not casual relationship between independent variables and dependent variable.

We can easily make an example using simulation: let variable $A$ have prior distribution of $P(A)$, and generate two other variables $B$ and $C$ with conditional distribution $P(B|A)$ and $P(C|A)$ given. Then generate the prediction target $Y$ from all three variables. Finally we can use random forest to predict and check the performance.

You can see, random forest will work "fine" on all variable included, although there are confounding variables exists.

$\endgroup$
1
  • 1
    $\begingroup$ Thanks, but it is so disappointing, because we can not conclude with confidence about our variables effect on the outcome. $\endgroup$ Apr 20, 2021 at 7:57
2
$\begingroup$

Haitao Du is correct that this is the most common approach. Adam Kelleher references Sendhil Mullainathan in talking about machine learning vs social science as follows:

In machine learning, the goal is often just to reduce prediction error, and not to estimate the effects of interventions in a system. Social science is much more concerned with the effects of interventions, and how those might inform policy. At IC2S2 recently, in Sendhil Mullainathan’s keynote, he called this the “beta-problem” (the focus on the regression coefficients, more common in social science) and the “y-hat problem” (the focus on the actual prediction, more common in machine learning).

That said, as Adam also points out, people working in causal inference ML, building on work by Judea Pearl and others, bridge this gap between ML and social science as they are interested in the beta-problem.

The DoWhy package for causal inference in Python by Amit Sharma and Emre Kiciman is a good example. They're also working on a book on this. Another good example is the “Elements of Causal Inference: Foundations and Learning Algorithms” book by Jonas Peters and colleagues.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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