for my thesis I need to build a logistic regression model and test the significance of several indicators on a certain outcome, that is testing if the independent variables has a significant effect on the dependent variable.

For the first try, I entered all the IV simultaneously, resulting in a significant LLR p-Value, but some of the IV dont have a significant p-Value or dont have the effect on the AV that I was expecting. Obviously some Variables are some sort of confounders. Therefore I came up with trying to build seperate models for each IV. Each model had the effect that I was expecting of the IV and were also significant but I'm not sure if this would be the correct proceeding for testing the indicators, as I think any valid Information that is entered to the blank constant regression model is somehow significant.

I'm not so much interested in what indicator is the best for predicting the outcome (which would yield in a forced entry of the IV to the model) but as I mentioned I just want to know if the indicators are signifcant concerning the AV.

What would your suggestion be in this particular case?

  • $\begingroup$ Are you looking for a joint test of significance for all your indicators? $\endgroup$ Apr 12 '16 at 11:29
  • $\begingroup$ Regarding the prior formulated hypotheses, I want to check for the significance for every indicator independently, so I assume that I'm not looking for a joint test (actually not sure what this is). $\endgroup$
    – TheDude
    Apr 12 '16 at 11:46
  • $\begingroup$ In which model(s)? The bivariate regressions with one indicator each are restricted versions of the multiple regression model with all indicator variables included, under the maintained hypothesis that all but the included indicator are insignificant. If the results of the multiple regression do not meet your expectations, you should check whether your regression is otherwise misspecified. $\endgroup$ Apr 12 '16 at 11:57
  • $\begingroup$ Not just "insignificant' but conditional on the effect being exactly zero. $\endgroup$ Apr 12 '16 at 12:52

In the field of medical research it is quite common to present the coefficients from the full model (with a confidence interval) alongside the coefficients (and confidence interval) from the series of bivariate models. This is sometimes extended to a set of intermediate models using theoretically justified subsets of the predictor variables (so one might first add the demographic variables). If your predictor variables fall into several theoretically justified subsets you could consider fitting the full model and then checking the effect of dropping each subset using a LRT. What is best depends strongly on (a) the underlying science (b) publication practice in your discipline. In any event you want to avoid automatic variable selection methods.

  • 1
    $\begingroup$ In my humble opinion this represents not so good statistical practice and results in very hard to interpret estimates. Just stick with the full model. $\endgroup$ Apr 12 '16 at 12:53
  • $\begingroup$ @Frank I agree the full model is and should be the primary interest but sometimes it is hard to understand and the other suggestions help to disentangle. $\endgroup$
    – mdewey
    Apr 12 '16 at 14:05
  • 1
    $\begingroup$ How does an unadjusted estimate with unknown confounding give you any useful information? Unadjusted estimates are very hard to interpret and may be misleading. They also sidetrack the reader from the important issues and create confusion in the minds of readers. $\endgroup$ Apr 12 '16 at 16:34

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.