Categorical predictor

Hi all

I have one question concerning the presentation of results from a regression model. If you have a predictor with more than 2 categories what do you usually present in the publication: The overall p-value of the effect or the individual p-values compared to the reference category (which then would correspond to the confidence intervals).

Best regards and thanks


Omega Contributor
What type of regression are we talking about (e.g., linear, logistic, multinomial, survival, ridge, robust, etc.). Also when you clarify, provide the beta coefficient values or proxies, so we can provide an example of presentation.
Hi hlsmith

it is a general question and I think it applies to linear as well as logistic regression. As an example: If you have as a predictor a variable race with values "white", "African american", "Asian". Lets say it is a logistic regression. Then you will get beta values for the categories compared to the reference. You will have for each category expect for the reference a p-value. Lets say reference is white, then you have a beta for African American and one for Asian both with p-values. In addition you will have an overall p-value for the test that all the beta-coefficients for that variable are zero. What do you present in the publication: the overall test or the individual test for each category compared to the reference. In general when it is a logistic regression, I guess you also would present the OR compared to the reference, correct? For me it is the question whether to present the overall p-value or the p-values for the individual categories, which will depend on the reference category that you have chosen.
Thanks for your help.
Hi Lisa,
It might be useful to know a bit more about what it is you are doing- what is the hypothesis for the research? Does it related to the development of a model, is it about identifying predictors? If you are identifying predictors only for say a binary outcome using logistic regression then ordinarily people quite the odds ratio and the 95% confidence interval (you can include the p-value in there too if for completeness). For categorical data, you would ordinarily have a reference as you mention.

If, for example, you are designing a model then it would be usual to include the beta coefficients too.

What do you mean by the "overall test"? Are you trying to assess the performance of the score that you have developed? Can you do as hlsmith suggests and give more specifics, as there may be important nuances which means the above is not optimal and I may be leading you astray!