Multivariable Logistic Regression Interpretation

[#1]

Hello,

I have used a multiple regression to explore the important features in running.  I am finding it very hard to interpret these results.  Is there anyone who could please offer some advice?  I have read the statistical knowledge portal all day and things are still no clearer.

Thanks,

Scouse.

[#2]

Could you supply some of those results? Are there specific questions that you want answered? A bit more information is needed in order to try and help you.

[#3]

Hello Dan,

Thanks for your quick reply.

Here is what I am struggling to understand:-

1.  What is the difference between Actual by Predicted Plot ( AbPP) and Leverage Plot (LP) - why does the AbPP have a P, Rsq and RMSE value, but the LP only has a P value?

2.  Effect summary - do you report LogWorth or only the P values?

3.  Summray of Fit - Is there any guide for rating your Root Mean Square Error?  (mine is 0.0977 - is this good or bad)

4.  Analysis of Variance - what should be reported from here?

Thanks,

Scouse

[#4]

1) Multiple logistic regression does not have leverage plots.

    A regression model for a continuous response will have both the observed vs. predicted plot and leverage plots. Obs. vs. pred plot gives you information regarding the entire model. The entire model has statistics such as R-square, an overall p-value, and RMSE. Leverage plots are giving you information about each individual term in the model.

2) You can report either value, but what you report will depend on who you are reporting those results to. If you take the negative of the log base 10 of the p-value, you will obtain the LogWorth. It is the same information, just reported on a different scale.

3) Root Mean Square Error is the differences are between the response and p (the fitted probability for the event that actually occurred). This typically means that a lower value is better because that indicates the response and p are closer together. For example, an observation may have a result of "True". If the model predicts the probability of the True event is 0.6, then the RMSE would have a value of 1 - 0.6 = 0.4^2 = 0.16 for that observation. Add up those "deviations" and divide by n to get the RMSE (the formula is presented with the output). There is no way for me to know what a good number is since it will depend on the number of observations.

4) I cannot tell you what to report. That will depend on the information you wish to communicate and who you are communicating to at a minimum. In some cases maybe you won't need to report the Analysis of Variance results. Other times you will.This is not a statistical question.

[#5]

Thanks Dan,

Some really great information there broken down for a 'lay-man'.  Really appricate your time and effort.