# Dependent count variable with negative values

In a regression model my dependent variable is a count variable that can have negative numbers.

What econometric method can I use? Poisson regression and negative binomial only accept positive dependent variable

In particular, the dependent variable it's the difference in the number of managers of a company after and before an investment round. So a value of 2 indicates that there are 2 more managers, while a value of -2 indicates that there 2 less managers. In my dataset, the dependent variable it's the subset of integer numbers from -6 to 12. There is no literature on how far the dependent variable can go. In theory after the investment the new shareholder can fire all of the existing managers and replace them all. This is what I am trying to discover

• Any Poisson or negative binomial routine that rejects data with zeros is incompetent! Restriction to zero or positive values is common, but not universal, as arguably the key assumption is that means are strictly positive, not the data. The problem with negative values is knowing how low they can go. Is the range from -5 to 5 empirical or a matter of principle? If the latter, a binomial model might make sense. Nov 6, 2017 at 13:43
• Why not track the two components of the difference, both of which are counts? This retains information that otherwise is lost when keeping just the differences.
– whuber
Nov 6, 2017 at 14:22
• I updated the question with the exact range Nov 7, 2017 at 8:11
• I'd proceed with linear regression and caution, in the absence of an obvious specific model for the process. Nov 7, 2017 at 9:15

While the number of managers before and after are count variables, your dependent variable no longer is: Counts can't be negative, after all. So you don't need to use either poisson or negative binomial.

• Happily, one thing you don't have to watch out for is zero-inflation: All your firms will have managers.
– Mox
Aug 17, 2018 at 21:52
• Look into the literature on 'change scores' or 'difference scores' as outcome variables.
– Mox
Sep 17, 2018 at 21:42

For this kind of analysis I would suggest that you follow whuber's advice in his comment, and track the individual counts of the number of managers before and after each change. It should not be too hard to create a GLM that uses the non-negative count variable managers as the response variable and uses a binary indicator investment to indicate whether or not an investment round occurred. This kind of model would have a coefficient for the effect of the investment round on the count of managers, and this coefficient could be positive or negative.

Since your underlying count values for this kind of model would be non-negative integers, a good place to start would be a standard negative binomial GLM. You could program this in R as shown below. If you are able to add some of your data, and give some additional information on other covariates, we could have a look at what you get.

library(MASS)

MODEL <- glm.nb(managers ~ 1 + investment + covariates, data = DATA_FRAME);


Y is not actually a count - counts can only have positive values. It's a difference, which is, well, different.

One option would be to treat Y as a continuous variable and use ordinary least squares regression. Another would be to add 5 to Y and use a count regression model. You could try both and see if the results are similar.

• If there's symmetry in so far as changes can be positive or zero or negative, that symmetry wouldn't be respected by an upwards shift. That is, a model is wanted that respects zeros as a natural reference; no standard model will, I imagine, think that 5 is special. Nov 6, 2017 at 13:59
• True enough, but is there a better model for this? Nov 6, 2017 at 14:00
• As in my comment on the question, I think we need to know more about the stated limits. It's hard for me to imagine that there is no literature for this kind of data; I just don't know what it is. Nov 6, 2017 at 14:03