Dear colleagues,
How to interpret adjusted p-value>0.05 of a predictor, which has entered into the model (stepwise multiple logistic regression - SAS)?
The predictor is associated with the model, but does not enhance the model fit parameters? Should we report both adjusted and unadjusted p-value, with OR and 95%CL?
Best regards,
Paulo
The following link may have a helpful answer
Adjust any set of p-values by PROC MULTTEST in a data set,
https://communities.sas.com/t5/SAS-Statistical-Procedures/Adjusting-for-multiple-comparisons-in-Logistic-Regression-models/td-p/16743
Are you saying that between the univariable and multivariable models the predictor has lost significance? If so, many researchers opt to drop non-significant predictors. Stepwise methods should automatically do this to create the most parsimonious (best fit for variables included) model unless you forced these into the model. If you choose to report it, the 95% CI will straddle 1, the p-value > 0.05, and it will be there mostly for reference.
Some researchers will force in certain model variables such as age, race, gender to show whether these terms are significant. If they aren't significant you can say that "controlling for age, race, gender, your model shows that variables x, y, z impact the outcome, etc". If the variables you are concerned about lose significance when you leave in variables like these you can say "Once age, race, and gender are controlled for, variable x is no longer a significant predictor for outcome." Alternatively, if these are significant, you can say "In models that control for age, race, and gender, variable x independently impacts outcome to Beta_x amount per unit change." The interpretation changes when talking about interaction terms.
Dear Philips,
Thank you for your amazing answer!
The predictor was significant on the univariable model. In the multivariable stepwise, this predictor was the last to enter into the model, but it has been taken as a significant predictor according to the wald criteria (p
This would really depend on whether you have a sample size large enough to assure that the Wald test is giving you a "good" confidence interval. Generally, one hopes that the sample size is large enough that the Wald and the ML-estimates are close to one another even if their p-values fall on either side of 0.05. As these tests are asymptotically equivalent, the larger the sample size the better they should approximate each other. In small sample situations, the ML estimate is generally preferred.
Can you post the results of your model with and without the variable? If not, you might want to read up on the difference between the two tests.
Here are a few resources:
- http://wps.aw.com/wps/media/objects/2387/2445250/Web_Extensions/4885_2fMURR_EX11W.pdf
- https://web.as.uky.edu/statistics/users/pbreheny/760/S13/notes/3-19.pdf
- http://people.math.umass.edu/~daeyoung/Stat608/readingpaper2.pdf
- Badges
- Science topic
- Similar topics
- Geoscience
- Asia
- India