Dear community,

I have an unbalanced panel dataset of 165 countries from 1991-2019, in which my dependent variable is dichotomous (1 = civil war onset, 0 = no civil war onset). There are 4722 possible country-year observations. I want to run a simple logistic regression, testing whether a variable measuring state capacity (l1_factor1) can predict civil war onset, controlling for GDP/capita and population. The code I have used is this:

Code:
. logit  onset l1_factor1 l1_loggdp_capita l1_logtotpop, vce(cluster country_id)
´

However, intuitively, I'm assuming I should use fixed effects in order to control for county effects (country_id). When using the code below, 2,577 observations are dropped because they have no variation in y (i.e., most countries never experience conflict onset). I am new to STATA and statistics in general, and I am wondering whether random effects should be used instead? If that's the case: why?

Code:
xtlogit onset l1_factor1 l1_loggdp_capita l1_logtotpop, fe


Thank you, and my apologies for the perhaps very basic question.
Charlotte