I am running a binary logistic regression with compositional predictors that sum to 100% (demographic categories). I've looked at several postings about this, but can't find a good solution to my problem. Would dropping a single predictor be useful in cases where 0% of the data comes from that category? I.e., if my predictors are race, and I drop "Hispanic/Latino", the hispanic/latino rate in my data ranges from 0% to 6% in each of my cases, so in many/most cases the data is still correlated.

Would a transformation be appropriate here?

I do have the ability to calculate a (rough) number for each category, since I do have the total number of individuals in each case, but I am more interested in effect of the proportion of the racial categories on my independent variable.

I've found these, but they don't present a solution.

What regression model to use when independent variables are percentages to predict % outcome?

Proportions (compositions) in logistic regression

  • $\begingroup$ I see explicit solutions in the first thread you reference, so could you please elaborate on what you might be looking for in addition to them? $\endgroup$
    – whuber
    Jan 10 '17 at 17:02
  • 1
    $\begingroup$ I'm sorry if it wasn't clear: for a majority of the cases, 5/7 of my categories are 0%. So I'm not certain if dropping a category with low explanatory power, or several categories even, would help: they would still sum to 100%, and be correlated due to the racial population of the city I'm researching. In the first case referenced, it could be expected that none of those categories (bone/muscle/fat) would be 0, which is not the case in my data. $\endgroup$
    – SLAstats
    Jan 10 '17 at 17:09
  • $\begingroup$ I'm afraid I don't follow: could you explain the distinctions between a "case," a "category," and a "predictor"? $\endgroup$
    – whuber
    Jan 10 '17 at 17:12
  • 1
    $\begingroup$ Each of my cases represents a group that is broken down by demographic data that I'm using as predictors for my binary outcome. In this case, mutually exclusive race categories that sum to 100%. $\endgroup$
    – SLAstats
    Jan 10 '17 at 17:57

You can use the additive logistic (or multivariate logit) transformation to transform your predictor variables.

$z = log(\frac{x_i}{x_1}), i=2,..,D$ (number of categories) and then perform logistic regression on z.

  • 3
    $\begingroup$ This was your fifth answer which was very short. Please extend your answer and read our tour stats.stackexchange.com/tour $\endgroup$
    – Ferdi
    Feb 28 '17 at 8:11
  • $\begingroup$ Actually this is the answer. I cannot do anything about it. $\endgroup$
    – Michail
    Mar 1 '17 at 12:14

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.