# What to do with statistically insignificant dummy/categorical variables? [duplicate]

From the research I've done the common answer is that you can not remove insignificant dummy variables from a regression. I'm having a hard time finding academic papers or books that back up this thought. If anyone knows of any papers dealing with this issue or a good place to search through statistics papers that would be great!

• This is contradictory. What "research" is this that doesn't involve reading? You seem to be saying "I've found some [unexplained] references arguing a point of view. Please suggest some more references arguing likewise." Commented Dec 3, 2019 at 15:19
• Pedantically or otherwise, the issue is not whether you can do this, but whether you should. In essence one or more indicator variables coding two or more categories are a team effort. The usual argument is that you accept the whole team, or not. But I can't see that some aggregation of infrequent categories might not make some sense, but you need to be open about why you do it. Commented Dec 3, 2019 at 15:21
• What I was trying to ask is if anyone knows of any stats papers that deal with this issue. The research I've done has been mostly web based forums like this one, but I was hoping to find a more concrete source. [relevant links: stats.stackexchange.com/questions/94441/…, stats.stackexchange.com/questions/154758/… Commented Dec 3, 2019 at 15:27
• See stats.stackexchange.com/questions/24298/… and search this site. Commented Dec 3, 2019 at 18:14

I fear that your question might be considered ambiguous. Let us say you have a regression of the form: $$yY_i = a . X_i + b2 . D2_i + b3 . D3_i + \epsilon_i$$ Where D2 and D3 are dummies variable respectively equal to 1 if the underlying variable $$D_i==2$$ (respectively $$D_i==3$$). Reference category is $$D_i ==1$$.
Case B: why not? But this will change the interpretation of coefficient b3, as reference category is now: $$D_i==$$ 1 or 2.