95%Credible intervals are smaller than OR
Posted: Sat Mar 20, 2021 6:10 pm
Hello,
I am running the following multilevel model with a binary outcome with mcmc:
My outcome is whether urine dipstick are available (0=no; 1=yes). I am using mcmc because the response proportion is extreme (0=42/504; 1=462/504).
I checked the diagnostics and following the Raftery-Lewis and Brooks-Drappers diagnostics I re run the model with 550,00 iterations.
Some of the odds ratio and Standard deviations are huge for 2 of the variables (OR for gov:1156.948 , and VHIOOP:13730.1), and the 95%Credible interval is smaller than the parameter for VHIOOP: 95%CI = 0.050643 - 164.6105
Burnin = 1000
Chain = 550000
Thinning = 1
Run time (seconds) = 156
Deviance (dbar) = 214.32
Deviance (thetabar) = 195.41
Effective no. of pars (pd) = 18.91
Bayesian DIC = 233.23
Disptick_available OR Std. Dev ESS P [95% Cred.Int]
gov 1156.948 24956.07 8874 0.003 2.891191 2926.618
VHIOOP 13730.1 1120034 14575 0.307 0.050643 164.6105
Can anyone explain why is this happening, and what should I do to get correct OR and 95%CI?
Thank you!
Manuel
I am running the following multilevel model with a binary outcome with mcmc:
runmlwin unique_dipstick_avai2cats cons private secondaryhosp paed_mother_child centered_unique_turn19_time centered_exp_capita gov VHIOOP, level2(country2: cons) level1(id:) discrete(distribution(binomial) link(logit) denom(cons) pql2) nopause
I have 504 observations (level 1) and 29 second level units.runmlwin unique_dipstick_avai2cats cons private secondaryhosp paed_mother_child centered_unique_turn19_time centered_exp_capita gov VHIOOP, level2(country2: cons, residuals(u)) level1(id:) discrete(distribution(binomial) link(logit) denom(cons) pql2) mcmc(burnin(1000) chain(550000)) initsprevious nopause nogroup
My outcome is whether urine dipstick are available (0=no; 1=yes). I am using mcmc because the response proportion is extreme (0=42/504; 1=462/504).
I checked the diagnostics and following the Raftery-Lewis and Brooks-Drappers diagnostics I re run the model with 550,00 iterations.
Some of the odds ratio and Standard deviations are huge for 2 of the variables (OR for gov:1156.948 , and VHIOOP:13730.1), and the 95%Credible interval is smaller than the parameter for VHIOOP: 95%CI = 0.050643 - 164.6105
Burnin = 1000
Chain = 550000
Thinning = 1
Run time (seconds) = 156
Deviance (dbar) = 214.32
Deviance (thetabar) = 195.41
Effective no. of pars (pd) = 18.91
Bayesian DIC = 233.23
Disptick_available OR Std. Dev ESS P [95% Cred.Int]
gov 1156.948 24956.07 8874 0.003 2.891191 2926.618
VHIOOP 13730.1 1120034 14575 0.307 0.050643 164.6105
Can anyone explain why is this happening, and what should I do to get correct OR and 95%CI?
Thank you!
Manuel