0
$\begingroup$

I am conducting a project where I have to conduct regression on a data set with rather large amount of variables and am in need of some help.

To give some context, we are looking at whether certain prognostic variables make individuals more likely to undergo surgical or non-surgical treatment. There are a variety of variables, which are grouped into different prognostic categories (each category has 3 to 20 different variables). Categories include past medical history, which may include the presence of hypertension, obese, etc. These prognostic variables are usually binary, but can also be categorical with >2 categories or continuous.

From my understanding, I would conduct some sort of multinomial logistic regression, where the outcome is surgical or non-surgical treatment (categorical). However, I am uncertain whether I should include all variables in this model or conduct several logistic regressions for the prognostic categories. I also am wondering what other considerations I should have when conducting this analysis.

I would like to use either R or SPSS to conduct this analysis.

Thank you!

$\endgroup$
2
$\begingroup$

I'd recommend that you have a read of the third edition of Hosmer, Lemeshow, and Sturdivant's "Applied Logistic Regression." Specifically, Chapter 4, section 4.2 provides a step by step approach to what they call the "Purposeful Section of Covariates" approach to logistic regression modelling. It essentially involves several steps:

  1. Univariate analysis of each independent variables
  2. Fitting a multivariable logistic model to all variables that passed the first round of univariate screning and reducing the model to include only those variables that produced statistically meaningful results
  3. Comparing the coefficients from the reduced model in step 2 with the full model in that step to verify that you haven't excluded certain variables that might have a dramatic effect on the other variables (and hence should be left in the model).
  4. Obtaining a preliminary main effects model (described in detail in the book)
  5. Assumption checking and verification of the main effects model
  6. Checking for interactions in the main effects model and developing a preliminary final model
  7. Performing adequacy checks of the final model.

You can read Chapter Four (or at least parts of it online).

Of course there are other approaches to building statistical models (including automated approaches -- but I'm not a fan of these) and each approach depends on exactly what you are trying to accomplish.

So, I'd recommend that you start out by reading this material and some of the works referenced in the book. It should help you get a sense for popular model building and variable selection strategies.

$\endgroup$
  • $\begingroup$ Holmer's approach is some morbid distortion of stepwise selection, unique to (and popular in) the medical field. Why would one prefer univariate tests over subject knowledge, i.e. using DAGs to decide? Univariate tests are entirely meaningless in typical heavily confounded epidemiology settings. And with that many stages of correcting and refitting the model, the end result will overestimate everything even worse than standard stepwise selection. $\endgroup$ – juod May 17 '17 at 10:24
  • $\begingroup$ What would you propose @juod? We're all ears. I agree subjective knowledge is important, but that' incorporated in the approach outlined he Chapter 4 of Applied Logistic Regression. They don't advocate a completely blind approach. $\endgroup$ – StatsStudent May 29 '17 at 20:17

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.