I use simple regressions. My independent variable is a count data (# of a drug used per year), and it has too many zeroes. Depending upon a dependent variable, for some simple regressions, I have 13 observations, and there are only 3 non-zero values out of 13. For other simple regressions, I have 150 observations and only 12 are non-zeros out of 150. Is it appropriate to use simple regressions for these data? Is there a rule about how many zeros should be in independent variable to be able to model the relationship? My dependent variable is continuous (weight).

  • 1
    $\begingroup$ If you think there's a qualitative difference between drug users & abstainers, then heed @Nick's point & consider using a dummy variable as a predictor in addition to the count. $\endgroup$ Commented Nov 26, 2013 at 18:33
  • $\begingroup$ @Scortchi-ReinstateMonica could you explain why you would include the dummy along with the count please. I am curious to know. Thank you. $\endgroup$ Commented Jun 23, 2023 at 12:11
  • $\begingroup$ @Seanosapien: There's little context here, but often enough 'not at all' bucks the trend from 'a lot' to 'a little'. (Suppose it's a diet drug, & those who take it more often are lighter - except for those who don't take it at all, because they're not overweight.) See stats.stackexchange.com/a/103823/17230 for a bit more detail. $\endgroup$ Commented Jun 23, 2023 at 15:53

2 Answers 2


Regression methods do not make assumptions about the distribution of your independent variable. Strictly speaking, you would have too many zeros for linear regression when all of your data are zeros. Instead the issue here is lower statistical power and reduced ability to check your assumptions. Although it is discussed in terms of the t-test, you might be able to get the main idea from my answer here: How should one interpret the comparison of means from different sample sizes?


(1) by simple regression I assume you mean linear regression.
(2) zero values are not to my mind an issue for independent variables (in the same way they are for DV count data).
(3) what is an issue is that you're unlikely to have a linear relationship between IV and DV. There is a whole section in most regression textbook on how to approach this issue. I usually categorize the IV into categories that seem reasonable. Many statisticians dislike this approach - it is data-driven, arbitrary and wastes statistical power - and would recommend:
(a) transforming variable, (b) adding cubic/quadratic terms, (c) using restricted splines, (d) or something even more complex. Depends out what relationship between ID/DV looks like.

  • 2
    $\begingroup$ +1 to this clear advice. In addition, the scatter plot is a simple key. If a linear model seems applicable then zeros in the predictor are not intrinsically invalid. It may be that zeros and non-zeros are better handled via an indicator (dummy) variable. $\endgroup$
    – Nick Cox
    Commented Nov 26, 2013 at 16:04
  • $\begingroup$ I agree. I re-read my answer and noticed that I am answering a different problem - I have been plagued by the opposite problem recently (too many 0's in the dependent variable) and now I seem to see it everywhere. Answer removed. Thanks @charles! $\endgroup$ Commented Nov 26, 2013 at 16:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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