How to impute a two-level model with repeated measures

[TiagoVPereira]

Dear all,

My question is how to organize data in order to impute a two-level model with 5 repeated measures.

To sum up: We have been using Stata and RealComImpute. Level 1 = patient id, Level 2 = month.

Month is a repeated measure taken from the same patient at month 0 (baseline), 1, ..., 4.

Letting i denote the i-th month (i = 0,...,4) and j = 1,...,n participants, our model is like:

y_ij = β_0 + β_1month_ij + β_2gender_j + ....+ u_0j + u_1j month_ ij + e_ ij


We have missing data for i = 3 and i = 4 only (~10% and 90% respectively). Data for the remaining months are complete.

y from month 2 is highly correlated with y from month 3 (r>0.90), y from month 2 is also highly correlated with y from month 4 (apparently - with r>0.70).

So, we would like to use y from month 2 as one of the predictors for y month 3 and 4.

In order to use y from month 2 as predictor, I created (in Stata) an additional column called pred_2, which contains y from month 2 for every subject.

In Stata, the final dataset would look like this:

*/------------------- start -----------------------------
clear
set seed 12345
set obs 1000
gene id = _n
gene age = round(runiform()*30)+20
gene gender = round(runiform())
gene covariate1 = runiform()
gene covariate2 = runiform()
forvalues i = 0/4 {
gene y`i' = round(rnormal(100,20))
}
replace y3 = . if runiform()<0.10
replace y4 = . if runiform()<0.90
gene pred_2 = y2
reshape long y, i(id) j(month)
gene cons = 1
sort month id
order id month y pred_2
realcomImpute y age gender covariate1 covariate2 pred_2 using mydata , numresponses(1) cons(cons) level2id(month)
*/------------------- end -----------------------------

Is this the correct set up for the two-level imputation?

Look forward to hearing from you.

Tiago

[Harvey Goldstein]

As I understand it, your model is a repeated measures model where you have some responses missing for some individuals at some occasions (months). If you just set this up as a 2-level model then there is typically no problem since the level 2 units (individuals) just have different numbers of level 1 measures. Any multilevel software that can handlerandom coefficients will fit this. Imputation only arises if one or more predictors has missing data, or if you happen to be using a standard MANOVA or SEM design that requires balance.
Harvey Goldstein

[TiagoVPereira]

Thanks, Dr. Goldstein.

So, apparently there is no point in imputing responses in that repeated measure analysis? No gain in terms of standard errors?

Tiago

[TiagoVPereira]

It took me some minutes only: Journal of Clinical Epidemiology 66 (2013) 1022e1028
Multiple imputation of missing values was not necessary before
performing a longitudinal mixed-model analysis
Jos Twisk a,b, *, Michiel de Boer b , Wieke de Vente c , Martijn Heymans a,b

[TiagoVPereira]

Dear all,

After some days of reading, I've decided that I would like to see if MI will provide different results for my data. Assuming that I want to impute response variables (repeated measures), is the approach I pointed out above OK?

[rajveervyas]

Thanks for sharing this information.