5
$\begingroup$

I am trying to adjust my covariates in a model. My dependent variable is count data variance is greater than mean which means I have to apply negative binomial regression, main independent variable (bullying) is binary (0/1), and I have few covariates such as age, gender, family status and so on which I want to adjust in the model. One options for the adjustment is that I can adjust the main independent variable and age (model I), bullying, age and gender (model II), bullying, age, gender and family status (model III) and add other covariates one by one and have more models. However, some of the reviewers in my previous commented of not applying forward or backward method for covariates adjustment. So my main question is can we apply stepwise forward or backward selection for negative binomial regression in SPSS?

$\endgroup$

1 Answer 1

18
$\begingroup$

I don't use SPSS, so I don't know if you can do this, but I do know that you shouldn't.

When you use stepwise, forward, or backward, the output is all wrong. Standard errors are too small, p values too low, and parameter estimates biased away from 0.

In addition, automated methods stop the investigator (that's you!) from thinking, and there are reasons to include covariates that won't show up in any automated method, e.g. a variable could be an important covariate, and a low parameter estimate might be interesting.

I wrote a short and somewhat informal paper on this called "Stopping Stepwise". If you search my name and that title, you can find it at various spots on the web.

For a more extended discussion, with proofs, see Regression Modeling Strategies by Frank Harrell.

$\endgroup$
3
  • 8
    $\begingroup$ +1 ($+\infty$, really :) There are many similar answers on CrossValidated providing the same insight to Peter Flom's answer. $\endgroup$
    – Alexis
    Jun 7, 2023 at 15:15
  • 4
    $\begingroup$ Very ex cathedra view @peter-flom:) I disagree in general, i.e. that significance levels are a required condition that prevents overfitting. I agree in particular: the problem is the misguided attempts such as in step-wise regression to automate inference rather than get rid of if altogether. Regularized models such as LASSO (still in "pre-inference" state, despite being obsoleted by boosted trees and then DNNs) trained properly (i.e. with cross-validation) on tiny data sets with noisy targets can have entirely insignificant features, yet still achieve decent generalization. $\endgroup$
    – mirekphd
    Jun 8, 2023 at 7:12
  • 1
    $\begingroup$ @peter thank you so much. $\endgroup$ Jun 8, 2023 at 10:20

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.