Code:
mixed htagesd i.curchew2##i._1_wmAge4 i.male c.cagecenter c._1_mheightcenter, || CS: curchew2, cov(un)
CS=cluster (village)
htagesd=continuous outcome
curchew2=binary exposure of interest
wmAge4=woman's age categories (1 to 4, with 4 oldest)
Stata output pasted below.
I am analyzing cross-sectional data and the code above is my reduced model; I removed covariates and interaction terms that did not improve model fit. Is it appropriate to put curchew2, a binary variable, in the random effects with unstructured covariance? My main question is whether or not there is an association between curchew2 and htagesd at the individual level. I want to allow the the effect of curchew2 on htagesd to vary depending on the cluster and I also have reason to believe that the effect of curchew2 could vary depending on the level of the outcome (htagesd), which is why I've included curchew2 in the random effects and selected unstructured. Putting the other covariates into random effects doesn't improve the model and I'm not really interested in them - is it appropriate to only include one variable in random effects?
If there is not something wrong with my syntax, it appears that curchew2 is negatively associated with htagesd among older women. It also seems that there is significant covariance, and the negative sign on the covariance indicates that as htagessd increases, there is less effect from curchew2 (reduced slope). Is this the correct interpretation of covariance when dealing with a binaryXcontinous?
My next step is to re-run the model with MI estimate because I have a few missing values for covariates (the results here are from the first imputed dataset). I tried this already and had trouble with the unstructured covariance because all 40 of the imputed databases will not converge. Any suggestions on how to keep my covariance random effect, but not use unstructured? I was able to converge all 40 databases with id. I see that STATA has other structures, such as Toeplitz, etc, but it seems like those are only meant for longitudinal data. Is it correct that in STATA the correct options for my data and research question are limited to unstructured, ex, ind, and id?
I feel like I'm close, but not quite there. Can someone give me a nudge? Thanks, Joe
Iteration 5: log likelihood | = | -825.78781 | ||
Computing standard errors: | ||||
Mixed-effects ML regression | Number of obs | = | 441 | |
Group variable: CS | Number of groups | = | 222 | |
Obs per group: min | = | 1 | ||
avg | = | 2.0 | ||
max | = | 8 | ||
Wald chi2(10) | = | 63.62 | ||
Log likelihood = -825.78781 | Prob > chi2 | = | 0.0000 |
htagesd | Coef. | Std. Err. | z | P>z | [95% Conf. | Interval] |
1.curchew2 | .3209291 | .456974 | 0.70 | 0.482 | -.5747235 | 1.216582 |
_1_wmAge4 | ||||||
2 | .7935068 | .3380072 | 2.35 | 0.019 | .131025 | 1.455989 |
3 | 1.013153 | .3802861 | 2.66 | 0.008 | .2678063 | 1.758501 |
4 | .8584673 | .5064598 | 1.70 | 0.090 | -.1341758 | 1.85111 |
curchew2#_1_wmAge4 | ||||||
1 2 | -.6724435 | .4933902 | -1.36 | 0.173 | -1.63947 | .2945835 |
1 3 | -.8854803 | .5481432 | -1.62 | 0.106 | -1.959821 | .1888605 |
1 4 | -1.512746 | .7949464 | -1.90 | 0.057 | -3.070812 | .0453203 |
1.male | -.2779354 | .1490301 | -1.86 | 0.062 | -.570029 | .0141583 |
cagecenter | -.1308908 | .0320338 | -4.09 | 0.000 | -.1936758 | -.0681058 |
_1_mheightcenter | .0821842 | .0145201 | 5.66 | 0.000 | .0537254 | .110643 |
_cons | -2.06025 | .3608977 | -5.71 | 0.000 | -2.767597 | -1.352904 |
Random-effects Parameters | Estimate Std. Err. | [95% Conf. Interval] |
CS: Unstructured | ||
var(curchew2) | .6217895 .3436879 | .2104507 1.837116 |
var(_cons) | .8191444 .2542355 | .4458364 1.505031 |
cov(curchew2,_cons) | -.7136772 .2799136 | -1.262298 -.1650567 |
var(Residual) | 2.09134 .1704813 | 1.782529 2.45365 |
LR test vs. linear regression: | chi2(3) = 16.20 | Prob > chi2 = 0.0010 |