Binary logistic regression with no significant predictors

[micdhack]

Hi everyone,

I have a question for logistic regression but i guess it applies to all kinds of regression analysis.

I have a model with many predictors that satisfies all assumptions and is a significant improvement as opposed to having just the outcome majority result.

My issue is that from all the predictors(10), only 2 are significant. Is it worth publishing the model? Or should i try to revise it with just the 2 predictors?

And what if some of the predictors are categorical variables that achieve significance for certain levels while for others they don't?

[noetsi]

There are two answers to your question. First, does the model with just two predictors generate anything of substantive or theoretical interest? And, related to the first, will anyone be interested in the results. The answer to that has little to do with statistics, it has to do with your audience and you research.

[micdhack]

There are two answers to your question. First, does the model with just two predictors generate anything of substantive or theoretical interest? And, related to the first, will anyone be interested in the results. The answer to that has little to do with statistics, it has to do with your audience and you research.

Hi noetsi, and thank you for your answer. The model definitely is of theoretical interest and the results are good but only two of the predictors are significant(however the classification accuracy is high, 88,9%). The predictors were significantly correlated with the outcome variable and that's why they were included in the model, it's just for me confusing when i don't achieve significance if they should be included in the final formula or not.

[SiBorg]

AFAIK, you should delete all the non significant predictors from the model and only include the significant ones. As you delete predictors, you will reduce the degrees of freedom in the model and hence the model should become more powerful. You should pick the model with the fewest predictors and the highest predictive ability - getting this balance right can be difficult and there is no single way to correctly prune the model down, but deleting non-significant predictors stepwise from the model (and re-running it each time and then deleting the next least significant predictor) is one way of going about it, but there are other ways of deciding which predictors should stay and which should go.

You'll probably find that deleting the non-significant predictors does not harm your model, and even if it does, a model with more predictors needs to be 'penalised' for having a greater number of predictors, and this penalised difference should favour the model with the fewer number of predictors.

Finally, although you chose your predictors because they were significantly correlated with the outcome variable (I guess in a univariate analysis), what your multivariate analysis is telling you is that when you take the other predictors into account, only 2 remain significantly associated and the others lose their predictive ability. Whether a predictor is significant or not also depends on the other predictors in the model and how many there are, so by deleting the least significant ones, you might find that other predictors become significant.

An excellent introduction to modelling and how to prune models down to the best model with the fewest significant predictors is given in the R Book by Michael Crawley:

http://www.amazon.co.uk/The-Book-Mic...7427711&sr=8-1

Hope this helps.