|
|
Level 2 random intercept in latent re... |
|
Message/Author |
|
|
Dear Mplus experts, For the latent regression below, may I have some advice about random intercept usage? A latent trait (eta) is measured by 15 binary responses (u_1-u_15). We test a centre effect on the latent trait, controlling for individual level covariates (x_1-x_6). H0: SEM no clustering, H1: centre has a random intercept in a 2-level model. Measurement part = cluster-invariant 1-PL IRT. For item m, individual i, centre j: P(u_{ijm} = 1) = F( eta_{ij} – tau_m ) tau_m=item threshold, F=logistic. Structural part: eta_ij = beta_0j + beta_1 x_{ij;1} + ... + beta_6 x_{ij;6} + epsilon_ij H0: beta_0j = beta_0 H1: beta_0j = beta_0 + r_j Mplus input H1: (cf. ex. p. 150 of Topic 7 short course) VARIABLE: CLUSTER=centre WITHIN = x_1-x_6 CATEGORICAL = u_1-u_15; ANALYSIS: ESTIMATOR = MLR MODEL: %WITHIN% eta BY u_1-u_15@1; ! 1-PL: loadings=1 eta on x_1-x_6; %BETWEEN% etab BY u_1-u_15@1; H0: no CLUSTER, WITHIN, %WITHIN% keyword, %BETWEEN% section. 1) Is the syntax correct? 2) Mplus output gives on level 2 the intercept’s variance and the 15 thresholds, not the intercept’s mean beta_0. Why? Thanks very much for any help! Mo |
|
|
1. It looks correct. The best check is if you get the results you are looking for. 2. In a cross-sectional model, factor means are not identified when thresholds are free. |
|
|
Dear Dr Muthen, Thank you very much for answering so quickly to my inquiry. A. Regarding 2), I have followed the example of your short course, but it is unclear to me why the unique random intercept (beta_Oj) on level 2 is modeled as 15 thresholds (etab BY u1-u15). Since the measurement part is considered to be invariant across clusters, why should the thresholds (tau_m) be modeled on level 2? Curiously, under H1, the estimated 15 thresholds on level 2 are (almost) identical to the thresholds under H0, with an (almost) constant shift (0.363 to 0.367). Has the cluster-specific random intercept been somehow integrated into the cluster-invariant thresholds? B. Regarding 1), yes, the results are consistent with what I expected, and suggest a clustering effect. Am I right to use a loglikelihood difference testing with the scaling correction factor as described in your technical appendix? C. However, I get a warning (NON-POSITIVE DEFINITE FIRST-ORDER DERIVATIVE PRODUCT MATRIX... THE NONIDENTIFICATION IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE NUMBER OF CLUSTERS...). There are only 9 clusters, totaling 800 individuals, and 22 parameters. A multiple group analysis with the KNOWNCLASS option ran into other numerical problems (message ONE OR MORE MULTINOMIAL LOGIT PARAMETERS WERE FIXED TO AVOID SINGULARITY OF THE INFORMATION MATRIX). How should I address this? |
|
|
A. Saying etab BY u1-u15; does not mean that the random intercept is modeled as 15 thresholds. It means that he covariation among the u1-u15 random intercepts is explained by etab. The etab parameter is its variance. The thresholds end up on level 2 in line with multilevel modeling where all means appear on level 2, not level 1. The thresholds are indeed cluster invariant. Means/thresholds are often not different in single- and two-level analyses. B. This is a tricky topic - see "Likelihood ratio tests in linear mixed models with one variance component" Crainiceanu-Ruppert (2004) JRSSB 66, Part 1, pp. 165-185. To avoid this, I would simply report your two-level results. C. 9 clusters is smaller than we recommend. At least 20 are typically needed for good SEs and variance estimation. You should also check that you don't have more between-level parameters than clusters. An alternative is to create 8 dummy variables and use these as covariates in a single-level analysis. So changing from random to fixed mode for the clusters. |
|
|
Dear Professor Muthen, Thank you very much for your prompt and helpful answer. It's really great to have such a support from both of you and your team! A(2). This clears up the subject to me -> C(2a). Maybe it is a very basic question: what are then the free parameters at level 2? In the modeling equations of my initial post, the only unknown parameter at level 2 is the variance (+/- mean) of the cluster-specific intercept beta_0j. However, in the Mplus input, the level 2 parameters also include the 15 thresholds (and exceed nb of clusters). Was there a mistake in my equations? B(2). Thank you for the reference. The technical details are a bit hard for me, but I got it that LRT may suffer a serious bias in this multilevel setting. C(2b). Thank you for suggesting this alternative of having centre as a nominal dummy coded covariate --- well suited to our problem, since cross-group variability is on the intercept only. It works fine (results are similar to the ones I obtained in the multilevel approach). Am I right using loglikelihood difference testing in this setting to infer a centre effect? A linked question: for this latent regression with binary outcomes, should ML or MLR estimation be preferred? If I use MLR, does this setting lend itself to the scaling correction factor technique? Many thanks again for your much invaluable time and help. |
|
|
A(2). The between-level parameters are the between-level factor variance and the thresholds. B(2). It is a tough topic. For an overview, you may also want to read Chapter 3 of the second edition of Joop Hox's Multilevel book. C(2b). Center effect is simply seen as the Z test for each center dummy. I don't think you see much difference between ML and MLR for this example, but I would use MLR and it works with the scaling correction. |
|
Back to top |
|
|