I have run a multinomial logistic regression and am interested in reporting the results in a scientific journal. Would it be alright to include a model summary table with the coefficients, standard errors, and then the p-value from a likelihood ratio test determining overall statistical significance of that variable?
This makes more sense to me than reporting the p-value for each coefficient from each logistic regression in the model.
Rachael Mady, there will always be different opinions about reporting results from a statistical analysis. I think the best thing to do would be to see how multinomial logreg is reported in articles published in the outlets/journals you are planning to submit your work to, and follow the reporting style of the chosen journal.
You don't need to report model coefficients and standard errors. You can give ORs with confidence intervals and p values. Also some model fit statistics may be added.
I agree with Sir Mehmet, you can check published journals and check the format on how the results of Multilogistic regression is presented.
Thank you all for your responses, they are really helpful. What are "ORs" though?
@ Rachael Mady The coefficients reported in the output are most likely the log-odds-ratios so might need to anti-log (exponentiate) to obtain the odds ratios.
Said it simply, the Odds Ratio is a proportion that gives you how many times the outcome (response) could be, if the factor is present, and it must be over 1.0 (including the lower and upper limits of the confidence interval).
If you are using SPSS, the Odds Ratio is (like Ana Nora Alix said) the exponential function (antilogaritm) of the coefficients of the model, and you can find it (with it's 95% confidence interval) in the last three columns of the table named "Parameter Estimates" with the titles of "Exp(B) and "95% Confidence Interval for Exp(B)".
Best regards
Since it is a multinomial test so much can be learned from the individual coefficients. The overall p-value and coefficients tell you that somewhere in the complex model you have an interesting relationship. A table listing the overall coefficient and their SEs and p-value first and then the individual coefficients (including interactions) and their SEs and p-values will provide the information I would expect from such an analysis. This will describe the full model that you used to describe data, the significance of the individual effects and their interactions. You will then convey the structure and behavior of the data completely. Whether or not you report odds ratios depends on your application. Any reader of the coefficients will understand the relationships across the data without specifically calculating the odds ratios from the coefficients.
Patrice Showers Corneli thanks for your response! When you say "overall coefficient" what do you mean?
Let me rephrase. If you show all of the coefficients are you should, then you should show the associated SEs and p-values for each coefficient. Some may well be non significant but should be included anyway as the p=0.05 is arbitrary and too low for leaving a variable out of the regression equation.
Answering this we be much easier if I knew the full model or your output.
Interesting conversation on this question. I seem to need similar clarifications: Patrice Showers Corneli does it mean higher p-values can be included? What is the threshold of inclusion of a variable into the equation?
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics