Suppose that in a linear regression, some of the continuous variables (such as number of pills taken each day) have small discrete values. For e.g. the number of pills taken each day can take values 0,1,2,3.
Would it be better to treat these continuous variables as categorical or ordinal variables? What should be the appropriate approach?
Thanks.
Your question suggests that the variable(s) at hand are your independent variables. If not (or if one of them is the dependent variable), you might want to take up Clark's advice and do Poisson regression.
For the independent variables: assuming you're looking at association, you can simply test whether it is necessary to treat the variable as categorical. If you compare a model: $$ out = \beta_0 + \beta_1 I(x=1) + \beta_2 I(x=2) + \epsilon $$ (have as many dummies as you need for the variable at hand) it's straightforward to see that this model: $$ out = \beta_0 + \beta_1 x + \epsilon $$ Is a submodel (just assume $\beta_2==2\beta_1$). This allows for a likelihood ratio test that will either indicate that you need the more complex model (categorical interpretation) or (given your sample size makes the power of the test trustworthy) warrant using the simpler numerical interpretation.
This still holds if the model contains other covariates (just keep those the same in the two models you're comparing.
Note: besides other details left out in the above, at the least you should be aware that adding more tests like this theoretically requires some form of multiple testing correction. Fortunately (?) no one ever does this, so you can get by without - just be aware.
