Combining mcmc(cc) logit results using mi estimate
Posted: Fri Apr 23, 2021 9:03 pm
Thank you in advance! I am relatively new to runmlwin and Bayesian models.
I am running a longitudinal cross-classified logit model (10 rounds of individual observations, and individuals can move across states over the 10 years). Due to missingness, I used "mi impute chained" to impute missing data. Based on my reading of previous posts, I know I should be able to use "mi estimate, cmdok" to combine the results across imputed datasets. Here are my codes:
STEP 1:
gen cons=1
mi estimate, cmdok noisily post imputations (1/5): ///
runmlwin DV cons IV1 IV2 IV3, ///
level3 (state_fips: cons) ///
level2 (study_id: cons) ///
level1 (year:) ///
discrete(dist(binomial) link(logit) denom(cons)) nopause forcesort
estimates store m_prior
STEP 2:
mi estimate, cmdok noisily imputations (1/5): ///
runmlwin DV cons IV1 IV2 IV3, ///
level3 (state_fips: cons) ///
level2 (study_id: cons) ///
level1 (year:) ///
discrete(dist(binomial) link(logit) denom(cons)) mcmc(cc) initsmodel(m_prior) nopause forcesort
1) Do my codes look correct? (I have been running them successfully for my analysis, but want to double-check with experts here)
2) More specifically, in both step 1 and step 2, when showing the combined results, I observed "Average RVI = .", "Largest FMI=. ", "DF: avg =.", and "DF: max =.". Are these normal to see?
3) Most importantly, in step 2 with MCMC(CC), how can we interpret "Coef. Std. Err. t P>|t| [95% Conf. Interval]" in the combined results? In each iteration (i.e., m=1, 2, etc.), we get "Mean Std. Dev. ESS P [95% Cred. Interval]". Can we interpret [95% Conf. Interval] in the combined results the same as [95% Cred. Int] derived in each iteration (i.e., m=1, 2, etc.)? How about the p-values in the combined results? Are these p-values in the combined results one-sided or two-sided? Any other things I should be aware of when interpreting the combined results?
Thank you again!
I am running a longitudinal cross-classified logit model (10 rounds of individual observations, and individuals can move across states over the 10 years). Due to missingness, I used "mi impute chained" to impute missing data. Based on my reading of previous posts, I know I should be able to use "mi estimate, cmdok" to combine the results across imputed datasets. Here are my codes:
STEP 1:
gen cons=1
mi estimate, cmdok noisily post imputations (1/5): ///
runmlwin DV cons IV1 IV2 IV3, ///
level3 (state_fips: cons) ///
level2 (study_id: cons) ///
level1 (year:) ///
discrete(dist(binomial) link(logit) denom(cons)) nopause forcesort
estimates store m_prior
STEP 2:
mi estimate, cmdok noisily imputations (1/5): ///
runmlwin DV cons IV1 IV2 IV3, ///
level3 (state_fips: cons) ///
level2 (study_id: cons) ///
level1 (year:) ///
discrete(dist(binomial) link(logit) denom(cons)) mcmc(cc) initsmodel(m_prior) nopause forcesort
1) Do my codes look correct? (I have been running them successfully for my analysis, but want to double-check with experts here)
2) More specifically, in both step 1 and step 2, when showing the combined results, I observed "Average RVI = .", "Largest FMI=. ", "DF: avg =.", and "DF: max =.". Are these normal to see?
3) Most importantly, in step 2 with MCMC(CC), how can we interpret "Coef. Std. Err. t P>|t| [95% Conf. Interval]" in the combined results? In each iteration (i.e., m=1, 2, etc.), we get "Mean Std. Dev. ESS P [95% Cred. Interval]". Can we interpret [95% Conf. Interval] in the combined results the same as [95% Cred. Int] derived in each iteration (i.e., m=1, 2, etc.)? How about the p-values in the combined results? Are these p-values in the combined results one-sided or two-sided? Any other things I should be aware of when interpreting the combined results?
Thank you again!