|
|
MLM and complex survey data |
|
Message/Author |
|
|
Hi Bengt, I read your article "MLM of Complex Survey Data" and I am trying to get a handle on how it applies to my data. Briefly, here is the sampling strategy that was used: "In Botswana, Lesotho, Swaziland, Malawi, Mozambique, Namibia, Zambia and Zimbabwe we stratified the most recent available census into rural, urban (not within the capital region), and urban capital sites. In each country, we drew a last stage random selection of enumeration areas, with probability proportional to the national population (Table 1). Interviewers tried to cover all households in each enumeration area, without sub-sampling. In each household, they interviewed all adults aged 16-60 years present at the time of the visit." It is my understanding that I have 3 levels - enumeration areas, households, and individuals. I have weights for country and EA type and a stratification variable for EA type. How would you recommend that I model this in Mplus? Can I use MLM as opposed to TYPE=COMPLEX, or would it be a combination? Also, can I use IMPUTATION with TYPE=TWO LEVEL? Thanks for your assistance, as always. Cheers, Alison |
|
|
I would suggest TYPE=COMPLEX TWOLEVEL; with CLUSTERS=enumeration household; STRATIFICTION= rural/urban/urbancap; WEIGHT= product of two weight variables; You can use IMPUTATION with COMPLEX TWOLEVEL. |
|
|
Thanks so much, Linda. |
|
|
Drs Muthen, I just wanted to follow up on your previous recommendation. Do I understand correctly that the method you are suggesting is a combination of multilevel modelling and an adjusted chi-square approach? If so, is there a particular reason you have adopted this approach versus a 3 level model (or is it simply because Mplus cannot yet perform an analysis with more than 2 levels)? Also, can I use TYPE=TWOLEVEL for an EFA? Cheers, Alison |
|
|
Yes, see the description in the introduction to Chapter 9. Mplus currently cannot do three levels with cross-sectional data only with repeated measures. This is another approach that can be considered because often the highest level does not have enough variability to model. |
|
Back to top |
|
|