Stata dropping two dummy variables to avoid collinearity?

[Evan Carron-Kee]

Hi there,

I'm trying to run a pooled cross-sectional regression model by combining yearly datasets from the UK LFS. I have eight datasets, one for each year between 2015 and 2022 inclusive. I'm having some trouble however with getting dummy variables for each year, so that I can control for differences associated with each time period.

As I'm appending each of the datasets, I thought I could use the 'generate' option on the append command to create a variable, 'year', to distinguish between each year. I would then run the regression with i.year as a dependent variable, and introduce my other variables along with it.

However, when I try this, Stata drops two categories - 2022 and 2021. My understanding until now was that to avoid the dummy variable trap, one had to include n-1 dummy variables, where n is the number of categories. However, in this case Stata seems to be dropping two categories - this is a problem because the coefficients on the year dummies are important to my analysis. I've included my code below:

use 2015.dta
append using 2016 2017 2018 2019 2020 2021 2022, generate (year)
probit job_found i.year if INCAC051==5

[Richard Williams]

What effect does the if qualifier have? Does it, say, cause all the 2021 cases to drop out? Try running it without the if qualifier.

[Evan Carron-Kee]

Apologies I should have explained this better in the original post: the ultimate goal of the analysis is to see if the effect on job finding probability of focusing on informal methods of job searching changes depending on the year - specifically I want to test whether informal job searching methods were less effective during covid-19 lockdowns. The qualifier is ensuring that all observations were unemployed in the first period, and doesn't change the problem of it dropping two categories.

However, if I add the noconstant option to the regression and specify year as ibn.year, it only drops one category, but then I've no constant. My main result will be from a regression which interacts an 'informal search' dummy with a 'lockdown' dummy (the 2020 and 2021 periods), so this isn't of major importance - but I would love to understand what's going on here!

probit job_found ibn.year, noconstant

Thanks