I am trying to run a two-level growth model for a continuous outcome (3 level analysis). In addition to the clustering in the data, we also have sampling weights to account for unequal probability of selection. The outcome is measured at 7 time points, and the sampling weights are different for each time point. Is there any way to use all 7 weights in m-plus?
bmuthen posted on Tuesday, November 01, 2005 - 9:17 am
Can you tell me why the sampling weights are different across time? Is it due to poststratification?
The study was not originally set up to be longitudinal. It is looking at the effectiveness of a program for teachers. There were 7 years of data collection, but the samples and response rates were not necessarily equivalent at each point. For example, one year we may have sampled 500 out of 2000 teachers in the program and got responses from 450 of them, while another year we may have sampled 300 and got responses from 225 of them. There are several teachers with multiple time points, but the data are very sparse over time (this is another problem-- no one has all 7 data points, but many have at least 2 or 3 time points). Is it possible to incorporate different probabilities of selection for each time point?
Mplus doesn't allow multiple weight variables in the analysis; however I think that this would not be appropriate for the situation you describe anyway. Missing data generally is not a reason to introduce weights. Simply including type=missing in the analysis section should be sufficient. Maximum likelihood estimation for multivariate models will be able to extrapolate the missing observations to some extent from the observed data. Weights included in the analysis simply because of non-response may actually lead to incorrect estimates.
The weights were not primarily intended for non-response. Not all teachers were sampled at each time point (because it was not originally set up as a longitudinal study). Because different number of teachers were sampled at each timepoint, there were unequal probabilities of selection across time.
bmuthen posted on Thursday, November 03, 2005 - 9:09 am
This sounds like an example of missing by design. This can be handled via a multiple-group analysis where group corresponds to pattern of missing data by design.
Djangou C posted on Monday, November 14, 2016 - 5:02 pm
We are using a longitudinal nationally representative samples of young people derived from PISA. The sample scheme is a two-stage stratified sample. Schools are sampled using PPS in the first stage while individuals were sample in the second stage. For schools that have more than the fixed number of student required, equal probability selection is used. But all students were selected for schools having less than the fixed number required. As for the stratification, schools were ordered by size within strata and individual schools were selected using PPS within strata. Moreover, the data base contains weighting variables at each time point to ensure that the sample matches the original population, given both the use of complex sampling scheme and attrition. Actually, at each time point we have 2 weighting variables, one at the beginning and one at the end (this one take also the attrition into account). Another aspect of the data set is that data were collected through a combination of school achievement tests and background information. Five plausible values are available for each achievement test whereas a single variable is available for the background information. We are interested in estimating a multi-group path model at the individual level testing for the invariance of the regression coefficients between the groups.
Djangou C posted on Monday, November 14, 2016 - 5:04 pm
Below are my questions:
1) In the analysis we are intendent to include the stratification and the weighting information. How can using the USEOBS option will impact the estimates given that the weighting is done for the entire sample and not for each variable separately? 2) Is it possible to include 10 weighting variables in Mplus? If not how can I deal with the weighting variables then? 3) How to conduct the analysis with the five plausible values for the achievement tests and the derived variable for the background information in a path model? Our intention here is to perform different analyses for each plausible value in Mplus and to combine the information for all regression parameters involving a plausible value outside Mplus, using Rubin’s formula. Is there a way to do this within mplus (apart from creating fictive plausible values for derived variables)? 4) I read in Mplus discussions the thread related to weighting. From my understanding the weight shouldn’t incorporate the missing information. If this is correct, what is the appropriate way to conduct these analyses? Any advice would be appreciated. Thank you.