I have a large dataset from hospitals that follows a hierarchical structure with over 6 million observations. I am trying to run a multilevel logistic regression using Stata 14.2. The outcome variable is a binary variable indicating whether patients were admitted to the hospital for certain conditions. The predictor variables include both categorical and continuous data. I removed the unused variables and now I'm working with a set of 13 variables. The following is the command I used:
melogit HOSPITALIZATIONS RACE AGE SEX INCOME REGIONS INSURANCE COMORBIDITY_SCORES || HOSPITAL:, or
However, I keep receiving the error message "initial values not feasible" r(1400); whenever I try to use sample weights. For weighting, I was using the following command:
melogit HOSPITALIZATIONS RACE AGE SEX INCOME REGIONS INSURANCE COMORBIDITY_SCORES [pw=SAMPWT] || HOSPITAL:, or
I have tried using one predictor variable at a time to identify when the issue occurs. I received the same error message even with a single predictor variable. My command looked like this:
melogit HOSPITALIZATIONS RACE [pw=SAMPWT] || HOSPITAL:, or
Then I tried to limit the number of observations to a smaller size initially and gradually increase it. However, I received the same error message when the number of observations reached 950k. The following command was used for this:
melogit HOSPITALIZATIONS RACE AGE SEX INCOME REGIONS INSURANCE COMORBIDITY_SCORES if _n<900000 [pw=SAMPWT] || HOSPITAL:, or
I have been struggling to understand the potential reasons behind this issue and how to resolve it. Could it be due to the size of my dataset?
This is the Stata output:
[Fitting fixed-effects model:
Iteration 0: log likelihood = -9661061
Iteration 1: log likelihood = -9395233.1
Iteration 2: log likelihood = -9393493.5
Iteration 3: log likelihood = -9393493.3
Refining starting values:
Grid node 0: log likelihood = -9178954.3
Fitting full model:
initial values not feasible
r(1400);]
Thanks so much,
Munira
