# Dummy's, significant and not significant

I have in my model two dummy's, for a categorical variable with 3 categories (native, EU-immigrant and non-EU-immigrant). The reference category is 'native'. These are merely control variables.

Now I find a significant coefficient for one dummy, and not for the other. See last 2 rows here. Note these are log odds.

    L1PRED  S.E.
Response    R

Fixed Part
cons    -1.775  0.080
female  0.431   0.042
age 0.016   0.001
FA  1.053   0.052
TV  1.778   0.112
second  -0.277  0.052
third   -0.269  0.070
fourth  -0.298  0.093
fifth   -0.163  0.102
EU_imm  -0.011  0.092
nEU_imm 0.953   0.439


I was wondering what the right thing to do here is? Because any control variable that has no effect is usually left out of the model. But I cannot do that here, I think?

Would it be most common to recode 'native' and 'EU-immigrants' into one category, since there is no significant effect on the outcome?

• You seem to be taking for granted that "any control variable that has no effect is usually left out of the model" - why is that? Unless you have a specific reason for doing so, it seems that any variables that must be "controlled for" (based on substantive knowledge) should be left in the model regardless of their significance. – Macro Jan 4 '13 at 13:34
• Oh, ok, thats easy. And it does make sense, because it's not about the effect on the outcome, but rather the correlation with the predictors. Is that right? – Marloes Jan 4 '13 at 13:41
• I'm not sure what you mean. What I meant was that, if you included the variable in the model because you believe, for substantive reasons, that it should be controlled for, an insignificant $p$-value shouldn't change that. – Macro Jan 4 '13 at 14:25
• +1 to @Macro for this reminder and for fighting against the cult of significance. – Peter Flom Jan 4 '13 at 15:33
• @Marlo, usually when one talks about "controlling" for a variable, they mean that, if that variable is left out of the model (or not entered into the model correctly), then the effects of interest will not be estimated correctly. Thus, controlling is done to prevent confounding. A typical toy example of confounding is a drug efficacy trial where baseline characteristics (e.g. what if young people tend to slightly improve regardless of treatment?) are not accounted for. In that case you wouldn't want to delete age based only on its $p$-value. – Macro Jan 4 '13 at 16:07