mjhjmtu

I am running a linear regression model and I have a binary predictor that has a highly skewed distribution. For example, one category represents 96% of the data. In terms of frequency, the other 4% represents 26 observations.

Should I keep/remove this binary predictor variable? And, what is the rationale for doing so? Thank you in advance!

In general, it's not an issue; you should keep it if it makes sense to be in the model, which presumably it does or it wouldn't be there to begin with.

Consider, for example, a model for weekly sales of chayote squash in the New Orleans area (see https://en.wikipedia.org/wiki/Chayote, down in the "Americas" section.) Such a model would likely need a dummy variable for Thanksgiving week in order to capture the very large increase in chayote sales at Thanksgiving (> 5x "regular" sales.) This dummy variable would take on the value "1" once every 52 weeks and "0" the rest of the time, so the "not Thanksgiving week" category represents roughly 98% of the data. If we take the dummy variable out, our Thanksgiving forecasts will be terrible and likely all the rest of our forecasts will be a lot worse, because they would be affected by the Thanksgiving data point in various ways (e.g., trends look much steeper if Thanksgiving is near the end of the modeling horizon, ...).

It's important, however, to note the following caveat. @Henry's comment in response to the OP is of course correct; if you only have one observation for one of the two categories, including the dummy variable will, in effect, simply remove that observation from the data set, and all your (other) parameter estimates would be the same as if you had just deleted that observation.