2
$\begingroup$

I am conducting an ordinal logistic regression. I have an ordinal variable, let's call it Change, that expresses the change in a biological parameter between two time points 5 years apart. Its values are 0 (no change), 1 (small change), 2 (large change).

I have several other variables (VarA, VarB, VarC, VarD) measured between the two time points. My intention is to perform an ordinal logistic regression to assess whether the entity of Change is more strongly associated with VarA or VarB. I'm really interested only in VarA and VarB, and I'm not trying to create a model. VarC and VarD are variables that I know may affect Change, but probably not very much, and in any case I'm not interested in them. I just want to know if the association in the period of observation (5 years) was stroger for VarA or for VarB.

Would it be wrong to not include VarC and VarD in the regression?

$\endgroup$
3
  • 6
    $\begingroup$ You should include them. At the very least, not including them makes your research vulnerable to criticism as to whether you did not control for C and D, and if you did not, why you didn't. Interpretation of regression models and coefficient changes depending on what's in the model, and to include stuff that you may not need in the end is a more conservative approach than to obnoxiously state that you know the true relation, and only include A and B. If you knew the true relation, you would not do any data analysis in first place, and shout at every corner that you have this divine knowledge. $\endgroup$
    – StasK
    Aug 11, 2014 at 17:54
  • 1
    $\begingroup$ Why not report both? Give an odds ratio and 95% CI for VarA (and VarB) alone (i.e., univariate/unadjusted ORs) and then report another set of ORs that have been adjusted for C and D (the potential confounders). $\endgroup$
    – Alexander
    Aug 12, 2014 at 4:47
  • $\begingroup$ @StasK but my model will be always virtually incomplete, because I cannot exclude that there is a number of unknown variables that influence Change and that I did not include in my model simply because I did not thought of measuring them. $\endgroup$ Aug 12, 2014 at 12:55

2 Answers 2

3
$\begingroup$

This depends on the relationships between the predictor variables (how are VarC and VarD related to VarA and VarB?) and also what question you are trying to answer.

Consider the possible case where VarA causes VarC which causes the response. If your only interest is the relationship between VarA and the response then including VarC would hide the indirect relationship. But if we are interested in if VarA has a direct effect on the response above and beyond the indirect effect through VarC then including VarC is important.

Sometimes it is helpful to draw a diagram with all the different variables and then draw lines/arrows showing the potential and/or interesting relationships between all the variables. Then use that along with the question of interest to decide on the model.

$\endgroup$
1
$\begingroup$

I suppose it is best to show one model,the original model, that includes them. Including any possible interactions.Than you may reason, based on your findings, whether they can be removed due to their statistical properties (no significance, causing non normal errors, variance inflation, etc.) or domain specific factors (e.g., HP may be significant if you try to forecast how much a car is desired but may not be adequate in a study on color preferences of car vendors). The result would be a model that is suitable to the domain and statistically relevant.

$\endgroup$

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.