I have some pilot data from a social survey of people's belief in the social cohesion of their neighbourhoods. The survey was developed from a theoretical stance and has four subscales built into it which should be part of one higher order scale.
People were sampled by neighbourhood so there is also a multilevel element to the survey.
I can fit a single level second-order factor model which fits the data well using CFA in Stata, but can I extend this to account for the nested structure of the data. All items (24 in total) are Likert scored 1-5. There is also the need to add sample weights to take into account differential probability of selection in different neighbourhoods according to the sampling design.
Thanks!
James
Hi James, I've never used Stata for Multilevel CFA. I mainly use R, MlWin, Lisrel, SuperMix, M-plus or SPSS for dealing with multilevel datasets.
Best,
Fabio
Dear James, same for me...I use Lisrel or SPSS (amos) or EQS...
Sorry can't help,
Sabina
Dear James,
I have not yet found convenient ways to conduct multilevel analyses including latent factors in AMOS or EQS and I am unfamiliar with STATA.
Even though it is not what you want to do, a first exploration of your ideas might be based on invariance tests of multi-group models. For example, you might apply mean / median splits (or clusters) on neighborhood characteristics and use invariance tests to see whether your model is invariant across the resulting neighborhood groups.
James,
It is unclear whether the the nesting is to be represented as strata or primary sampling units/PSUs (which is my guess given there are sample/probability weights) or as clusters/levels in the multilevel modeling sense.
If the sampling was conducted by neighborhood, it is likely that the neighborhood clusters are to be represented as strata or PSUs. If so, the sampling design can be represented in a svyset command. If you did not collect the data (which I assume is the case) refer to the survey statisticians from organization who did or the data documentation for information on how to set up the svyset command. sem in Stata is svy-prefix compatible and will incorporate the strata or PSUs appropriately into the sampling variance.
If, however, the neighborhoods are not to represent strata or PSUs, you can try to represent the neigborhoods in sem by estimating the clustering effect as a latent variable (see Bauer, 2003; Curran, 2003) or, using gllamm from SSC (see http://www.gllamm.org/). There are some complications for two-level models with sampling/design/probability weights which are discussed in papers available from the GLLAMM website. GLLAMM would also allow you to represent the Likert-type scales as ordinal logistic, instead of as normally distributed, if you care to go that route.
- joe
Bauer, D. J. (2003). Estimating multilevel linear models as structural equation models. Journal of Educational and Behavioral Statistics, 28(2), 135-167.
Curran, P. J. (2003). Have multilevel models been structural equation models all along?. Multivariate Behavioral Research, 38(4), 529-569.
http://www.gllamm.org/
- Badges
- Science method
- Similar topics
- Mathematical Sciences
- Statistics
- Sampling Design