1
$\begingroup$

I'm conducting a study on how pregnancy weight gain affects risk of breast cancer and decided to go with a logistic regression model (outcome is yes/no for breast cancer) and my primary independent variable is categorical (<10lbs, 10-19 lbs, 30-39 lbs and >40 lbs each compared to the referent 20-29lbs). I've recently been told that mixed modeling may be a better alternative to account for random effects (which as I understand is basically variation between subjects if I treat that as a random effect for example).

My question is: are there any major drawbacks to using mixed-effects logistic regression? Is it a more complex model by any chance that I may not necessarily need? Could it inflate odds ratios?

In other words, how do I defend my use of logistic regression over mixed-effects logistic regression, if I can?

$\endgroup$
  • $\begingroup$ Are there known causes of breast cancer that you have no data for? (eg, smoking & alcohol consumption) $\endgroup$ – probabilityislogic Jan 13 '17 at 0:12
  • $\begingroup$ I have data for most breast cancer risk factors and do plan on including them as covariates in my model. What are your thoughts? $\endgroup$ – Moony Jan 13 '17 at 0:18
  • $\begingroup$ What would your 'groups' of observations be in a mixed model? E.g. Do you have multiple measurements per subject? $\endgroup$ – Michael Oberst Jan 13 '17 at 3:30
  • $\begingroup$ Yes, pretty much. The way the data is set up is that the subjects are asked to fill out a questionnaire every 2 years and they're asked if they've been diagnosed with breast cancer, so I suppose those count as repeated measurements? $\endgroup$ – Moony Jan 13 '17 at 4:10
  • 1
    $\begingroup$ You started off on the wrong foot by classifying a continuous variable and pretending that subjects are homogeneous in the <10lbs and >40lbs groups. This will do more damage to the model than any benefit you will achieve by better modeling. $\endgroup$ – Frank Harrell Jan 14 '17 at 14:49
0
$\begingroup$

There are different approaches of including random effects in your model (random slope model, random slope + intercept model, random intercept model). You can check for the magnitude of random effects using for example "induced correlation" (Zuur et al. 2008).

Based on its magnitude you can then justify whether to include the random effect in your model or not.

While there is no threshold (at least that I know of) at which to include or exclude a random effect in your model (talking about "induced correlation"), you can most often easily justify the exclusion of random effects with a value close to 0. Furthermore, you have a generalized comparison of the influences of your random effects.

In general, simples models are preferred over complex ones in statistical modelling.

$\endgroup$
  • $\begingroup$ @Moony If you are satisfied with this answer please accept it :) By doing so you give credit to somebody's answer on stackexchange sites. $\endgroup$ – pat-s Jan 14 '17 at 17:42

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.