The aim of my project is to find the association between a group of independent variables and a categorical dependent variable (with three levels). For variable selection, my supervisor suggested employing the ANOVA test for each variable separately among the levels of the categorical variable. (he also provided me this link as a reference https://padhokshaja.medium.com/using-anova-and-multinomial-logistic-regression-to-predict-human-activity-cd2101a5e8bf) But I was wondering if the ANOVA test could be employed before running the logistic regression. As I remember, for the Anova test, the response variable needs to be continuous. So, before conducting a multiple logistic regression, should I perform a univariate logistic regression instead? Thanks so much for your help.
-
2$\begingroup$ can you help us understand why you feel the need to do an ANOVA before just diving into a logistic regression? $\endgroup$– John MaddenCommented May 16, 2023 at 20:39
-
$\begingroup$ @ John Madden, thanks for the comment. Usually before applying the regerssion modeling, we present the results of univariate analysis in medical research. So, my supervisor suggested me to apply the anova test to see wether there is a significant different among the means between groups or not and then to run the logistic regression. But my question is that when the dependent variable has categorical format, does it make sure to compare the means among its level? $\endgroup$– ElRCommented May 16, 2023 at 21:09
-
$\begingroup$ I took the liberty to modify the title on the basis of the contents of your query. $\endgroup$– utobiCommented May 17, 2023 at 7:50
-
$\begingroup$ @ utobi , thank you so much $\endgroup$– ElRCommented May 17, 2023 at 7:56
-
$\begingroup$ an analog to ANOVAs would be to do individual logistic regressions with only 1 indep variable at a time. $\endgroup$– John MaddenCommented May 17, 2023 at 17:34
1 Answer
This response is based on my experience working with colleagues who have made similar analysis suggestions. The rationale behind their suggestions (and what I am assuming is the same for your advisor) is the idea of reducing the potential number of predictor variables in the model. And this is coupled with the idea that if variable A is related to B, then you can view the relationship as B being related to A.
While this is not the best practice in doing a multiple regression analysis, the idea of trimming out variables that don't show a group difference at the bivariate level (using an ANOVA with the continuous predictor as the dependent variable in the ANOVA analysis and the grouping variable as the independent variable) is one that might be used to reduce the list of potential predictors (at least at the bivariate level).
Note: I'm not necessarily supporting this approach, but I'm hoping to shed insight into why the suggestion may have been offered.
