I am relatively new to statistical analysis and so please forgive me for any cumbersome language/explanations.

I'm interested in how sex and race interact to affect the experience of discrimination in low income health care settings. I have two focal independent variables (sex and race/ethnicity) and I would like to create interaction terms for use in binary logistic regression in SPSS. Sex is a 0/1 dummy variable and race/ethnicity has 5 categories: non-Hispanic white, non-Hispanic black, non-Hispanic Asian, non-Hispanic "other" and Hispanic. So I would want interaction terms (in dummy variable form?) for each possible combination: white men, white women, Black men, Black women, Asian men....and so on. How should I best go about doing this? I've already constructed a single categorical variable "sex_race" but am not sure how to properly use it in logistic regression.

  • $\begingroup$ I don't understand, why don't you estimate the interaction between sex and race/ethnicity directly through your model, rather than creating dummy variables yourself? What are you trying to achieve here? Also, what is your dependent variable and what type is it? $\endgroup$ – user2974951 Jul 12 '19 at 6:04

You won't want to use a binary logistic regression. This type of model would be used to find the probability of an event given a value of your dependent variable(s). Use multiple linear regression to test these hypotheses. I am assuming your dependent variable is health care costs or something of that form? The "interaction terms" in this sense are just a set of terms, for which only one will be a nonzero value for any given observation. Because all possible values of these variables are {0,1}, they basically just act as conditional intercept terms. You will want 5 (races) x 2 (genders) = 10 - 2 terms, with one (say black and woman) of each category as the control variable. Be wary that having this many parameters will penalize your Adj R^2. The interpretation will be as follows:

Say that you have a term δ(wh*m), where wh = white, m = male, and δ is the coefficient. If δ=-250, this means that, if the observation is white AND male, ceteris peribus, the dependent variable (cost of hc?) will be 250 less than that of someone that is black OR a woman.

Depending on the data, you will likely also have non-conditional terms such as δ(wh) and δ(m) that are significant. If the coefficient of δ(wh) is -150, that means that regardless of gender, the y-hat is 150 less if you are white, than if you are black. The difference between δ(wh) + δ(m), and δ(wh*m) is due to the fact that the affect of (wh ∩ m) is different than (wh ∪ m). In other words, there is an extra benefit of being white AND male.

One last tip; make the control variable something that, for your research and motive, you want to be compared to. If you are looking at how each race/sex compare to a white male, use white and male as the controls to make the coefficients more useful and applicable.

If you want a simpler model that preserves some DFs, then you could make your races white and nonwhite. This would, however, limit analysis between groups inside "non-white".


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.