Hello,
I am currently looking at a scale (PDI) measured every third month over a year (baseline, 3-, 6-, 9- and 12 months followup).
I have two groups where the first group received treatment between baseline and 3-month followup, and the second received treatment between 3-month followup and 6 month followup.
The research question is whether there is a difference between the groups depending on when they were treated.
The scale has both item-level missingness (only few) as well as subject-level missingness ( dropout, many). If one or more individual items are not measured, the total score cannot be directly calculated. I was asked to multiple impute the missings on item-level when they are item-level missing, and use linear mixed models to handle the subject wise missings.
This I did, imputing all items missing in wide format, but only calculating the total score (using mi passive) in those cases where not all items were missing from the beginning.
After this first step I transformed the data to long format, and did a mixed linear model. Here is my code, I have simplified and deleted the names of items so that it is easier to read.
Output (of the regression only):
1) Does this approach seem reasonable?
2) How do I find out whether the total score in group A and B differ at timepoint 6, 9 and 12? I would use lincom in a normal setting, but MI doesn't support this. I have also considered margins, but am unsure how to interpret the output.
Thanks a lot in advance.
-Anna
I am currently looking at a scale (PDI) measured every third month over a year (baseline, 3-, 6-, 9- and 12 months followup).
I have two groups where the first group received treatment between baseline and 3-month followup, and the second received treatment between 3-month followup and 6 month followup.
The research question is whether there is a difference between the groups depending on when they were treated.
The scale has both item-level missingness (only few) as well as subject-level missingness ( dropout, many). If one or more individual items are not measured, the total score cannot be directly calculated. I was asked to multiple impute the missings on item-level when they are item-level missing, and use linear mixed models to handle the subject wise missings.
This I did, imputing all items missing in wide format, but only calculating the total score (using mi passive) in those cases where not all items were missing from the beginning.
After this first step I transformed the data to long format, and did a mixed linear model. Here is my code, I have simplified and deleted the names of items so that it is easier to read.
Code:
mi set flong mi misstable summarize mi register imputed /*all missing variables*/ mi impute chained (pmm, knn(5)) /*all missing items*/ = /*nonmissing items and auxiliiary variables*/, add(20) rseed(30042019) mi passive: gen PDI_base=/*sum of variables*/ if PDI_base_m!=0 mi passive: gen PDI1= /*sum of variables*/ if PDI_md3_m!=0 mi passive: gen PDI2=/*sum of variables*/ if PDI_md6_m!=0 mi passive: gen PDI3=/*sum of variables*/ if PDI_md9_m!=0 mi passive: gen PDI4= /*sum of variables*/ if PDI_md12_m!=0 mi reshape long PDI, i(id) j(timepoint) /*Longitudinal on imputed datasets, PDI*/ mi estimate: mixed PDI i.timepoint##i.Gruppe PDI_base||id:timepoint, covariance(unstructured) vce(robust)
PHP Code:
Multiple-imputation estimates Imputations = 20
Mixed-effects regression Number of obs = 228
Group variable: id Number of groups = 76
Obs per group:
min = 1
avg = 3.0
max = 4
Average RVI = 0.0000
Largest FMI = 0.0000
DF adjustment: Large sample DF: min = 1.11e+60
avg = 6.89e+62
max = .
Model F test: Equal FMI F( 8, 2.7e+64) = 9.83
Within VCE type: Robust Prob > F = 0.0000
(Within VCE adjusted for 76 clusters in id)
------------------------------------------------------------------------------------
PDI | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------------+----------------------------------------------------------------
timepoint |
6-month | .4600555 1.824221 0.25 0.801 -3.115351 4.035462
9-month | .6995459 1.648965 0.42 0.671 -2.532367 3.931459
12-month | -.683934 1.659344 -0.41 0.680 -3.936188 2.56832
|
Gruppe |
Gruppe B | 10.6341 3.043612 3.49 0.000 4.668732 16.59947
|
timepoint#Gruppe |
6-month#Gruppe B | -7.730362 2.433259 -3.18 0.001 -12.49946 -2.961261
9-month#Gruppe B | -6.499903 2.72415 -2.39 0.017 -11.83914 -1.160666
12-month#Gruppe B | -3.960838 2.62529 -1.51 0.131 -9.106312 1.184636
|
PDI_base | .7692552 .1270858 6.05 0.000 .5201715 1.018339
_cons | -6.166558 4.649119 -1.33 0.185 -15.27866 2.945548
------------------------------------------------------------------------------------
------------------------------------------------------------------------------
Random-effects Parameters | Estimate Std. Err. [95% Conf. Interval]
-----------------------------+------------------------------------------------
id: Unstructured |
sd(timepo~t) | 1.732742 .8726095 .6457561 4.649426
sd(_cons) | 11.30863 1.440699 8.80985 14.51616
corr(timepo~t,_cons) | .0651884 .4009863 -.6193758 .6934409
-----------------------------+------------------------------------------------
sd(Residual) | 5.736988 .6589014 4.580598 7.185313
------------------------------------------------------------------------------
2) How do I find out whether the total score in group A and B differ at timepoint 6, 9 and 12? I would use lincom in a normal setting, but MI doesn't support this. I have also considered margins, but am unsure how to interpret the output.
Thanks a lot in advance.
-Anna
