Marginal structural models: estimating the controlled direct effect WITHIN imputations

[Nicole Hammond]

Hello,

I am using marginal structural models to estimate the controlled direct effect of parenting practices on adolescent self-harm in the presence of time-dependent confounding.

Variable list:
propensity scores = IPTCW_withdrawal_mi_trunc
dichotomous outcome = COPY_any_harm
dichotomous exposure = COPY_mean_withdrawal_dich_MI
mediator (quartiles) = COPY_quart_SDQ_comp_MI

Using: Stata/SE 18.0

I conducted multiple imputation using MICE and generated my propensity scores in the MI set data using “mi xeq:”. The problem I am having is with estimating effects within each imputation and then pooling the results.
To date, I have been doing the following to svyset the data and obtain an estimate of the controlled direct effect:

mi svyset SPTN00 [pweight=IPTCW_withdrawal_mi_trunc], strata(PTTYPE2) fpc(NH2)
mi estimate: svy: glm COPY_any_harm i.COPY_mean_withdrawal_dich_MI i.COPY_quart_SDQ_comp_MI, fam(poisson) link(log) nolog eform

However, if I understand correctly, the svyset command acts globally, so even though the propensity scores vary across imputation, Stata is only using one set (from one imputation) of propensity scores to estimate the controlled direct effect when using the above code. The other svyset variables (SPTN00, PTTYPE2, and NH2) do not vary across imputations (i.e., complete data).

I have tried several workarounds, but none seem to work. Example of one attempt:

.     scalar M = r(M)

.  forvalues m = 1/75 {
  2.     mi xeq `m': svy: glm COPY_any_harm i.COPY_mean_withdrawal_dich_MI i.COPY_quart_SDQ_comp_MI [pweight=IPTCW_withdrawal_mi_trunc], ///
> fam(poisson) link(log) nolog eform
  3. }

m=1 data:
-> svy: glm COPY_any_harm i.COPY_mean_withdrawal_dich_MI i.COPY_quart_SDQ_comp_MI [pweight=IPTCW_withdrawal_mi_trunc], fam(poisson) link(log) nolog eform
weights not allowed with the svy prefix
    The svy prefix assumes survey weights were already specified using svyset.
r(101);

end of do-file

r(101);

I cannot seem to combine “mi xeq” with svyset or use “pweight” without svyset. Is it possible to svyset and estimate a model within each imputation and then pool (using Rubin’s rules) to obtain an overall measure of association?

Thank you kindly,
Nicole G. Hammond

[Andrew Musau]

I cannot seem to combine “mi xeq” with svyset or use “pweight” without svyset.

Just a comment on your code in relation to your question:

. forvalues m = 1/75 {
2. mi xeq `m': svy: glm COPY_any_harm i.COPY_mean_withdrawal_dich_MI i.COPY_quart_SDQ_comp_MI [pweight=IPTCW_withdrawal_mi_trunc], ///
fam(poisson) link(log) nolog eform
3. }

m=1 data:
svy: glm COPY_any_harm i.COPY_mean_withdrawal_dich_MI i.COPY_quart_SDQ_comp_MI [pweight=IPTCW_withdrawal_mi_trunc], fam(poisson) link(log) nolog eform
weights not allowed with the svy prefix
The svy prefix assumes survey weights were already specified using svyset.
r(101);

end of do-file

Here, you are using both the -svy- prefix and specifying -pweight- (highlighted), which causes Stata to throw an error. You can use either svy (after setting survey design with svyset) or specify weights directly, but not both at the same time.

Also, note that -pweights- are only equivalent to using the svy prefix if there is no stratification, which is not the case in your setup.