signs of covariates change using multiple membership model

[morning03]

I am estimating a residential mobility (logistic) model whose covariate effects vary according to marital status: whether single or whether couples. One of the covariates is the number of years living in current address. Intuitively, we would expect negative duration effects on the probability of a residential move. In other words, the longer a person lives in current residence, the less likely that he/she will move. When estimating standard multilevel models (level 2: person, level 1: current wave), I observed this expected pattern of relationship between X and Y. However, when I estimated a multiple-membership model where a couple's covariate effect is a weighted function of each partner's unobserved effect, the duration effects for couples become positive (although the duration effects for singles stay negative). Initially, I have thought that moving from standard multilevel models to multiple-membership model will not have a dramatic impact on the fixed part parameters and only the random part parameters will be affected significantly. Apparently, this doesn't seem to be case. Any thoughts?


Stata code:
sort pid wave
capture noisily runmlwin Ymove_b cons ${x_singles} ${x_couples} if units_included_foranalysis == 1, level2(pid: cons) level1(wave:) discrete(distribution(binomial) link(logit) ///
denominator(cons) pql2) nopause mlwinpath(C:\Program Files\MLwiN v2.25\mlwin.exe)

sort pid wave
capture noisily runmlwin Ymove_b cons ${x_singles} ${x_couples} if units_included_foranalysis == 1, level2(pid: cons, mmids(pid pid_mem2) mmweights(weight_mem1b weight_mem2b)) ///
level1(wave:) mcmc(burnin(1000) chain(20000)) initsprevious discrete(distribution(binomial) link(logit) denominator(cons)) nopause mlwinpath(C:\Program Files\MLwiN v2.25\mlwin.exe)

[GeorgeLeckie]

Hi,

This question is not really a question specific to runmlwin (or MLwiN) in that you would expect to see the same result had you fitted your model in other software (e.g. WinBUGS).

You are right that changes to the random part of the model often (but certainly not always) do not lead to practically large changes to the point estimates of the fixed part parameter estimates (usually it is the standard errors which change more dramatically).

I suspect the explanation will be more a substantive one and will most likely relate to the nature of how your problematic covariate is distributed at each level.

I am sorry not to be of more help.

Best wishes

George