Message/Author 

patelrj posted on Wednesday, August 10, 2011  4:03 am



Dear readers, I am slightly confused how to define a multilevel model, especially what to put in the between model. So far, I've copied the within model and added a between level predictor. I'm analysing a model in which nigh all variables are measured at the within level, and only one variable is measured at the between. All within variables have high ICC's (> .10). I'm wondering if my syntax is correct when I want to model: (1) random intercepts for all within variables (2) no random slopes (3) equal models for all groups VARIABLE: NAMES ARE Clus Group w X1 X2 X3 Y M1 M2 M3; GROUPING IS Group (0=g1 1=g2 2=g3); BETWEEN = w; CLUSTER = Clus ANALYSIS: TYPE=TWOLEVEL; ITERATIONS = 5000; ESTIMATOR = MUML; MODEL: %WITHIN% Y ON M1 M2 M3; M1 ON X2; M2 ON X2 X3; M3 ON X3; X2 ON X1; X3 ON X1; X2 with X3; M1 WITH M2 M3; M2 WITH M3; %BETWEEN% Y ON M1 M2 M3; M1 ON X2 w; M2 ON X2 X3 w; M3 ON X3; X2 ON X1 w; X3 ON X1 w; X2 with X3; M1 WITH M2 M3; M2 WITH M3; I get no error messages, and nice fit measures, but I'm not sure if I defined my model properly 


The models specified on within and between should be guided by your research questions and hypotheses. (1) Yes. (2) Yes. Example 9.2 shows a random slope model. (3) There are no equalities in your model. Structural parameters are not held equal across groups as the default. I would not use MUML. I would use the default estimator. 

patelrj posted on Thursday, August 11, 2011  3:14 am



Dear dr. Muthen, thank you very much for your clear reply. I've added equality constraints to my model, and found all hypothetical links homogeneous between groups, except for one; but the scale used there is very unreliable, so I'm inclined to point to power problems before I theorize a lot about it. Due to very unequal cluster sizes (unequal classroom composition) I used MUML to handle that type of data. Is it ok to use it in this instance? 


MUML has no advantage with unequal cluster sizes. In fact, it has a disadvantage. When cluster sizes are equal, MUML is ML. When they are not, it is not ML. 


Dear Dr Muthen, I am confused how to define a 3level model (L1 withinperson, L2 betweenperson, L3 between classroom) with a grouping command. I am interested in gender differences in effects of X on Y on Level 2. CLUSTER = id class; GROUPING=gender (0=boy, 1=girl); between (id) = X_centered gender; between = (class) = X_samplecentered; within=time X_classcentered; ANALYSIS: TYPE = THREELEVEL RANDOM; MODEL: %within% Y; Y ON X_centered; s1 Y ON time; %between id% Y; s1; s1 WITH Y; %between boy% Y ON X_samplecentered; s1 ON X_samplecentered; %between girl% Y ON X_samplecentered; s1 ON X_samplecentered; %between class% Y; s1; s1 WITH Y; Y ON X_classcentered; 


The easiest way to do this is this  instead of %between boy% Y ON X_samplecentered; s1 ON X_samplecentered; %between girl% Y ON X_samplecentered; s1 ON X_samplecentered; use Y S1 on gender X_samplecentered interaction; where the interaction variable would be defined by Define: interaction=gender*X_samplecentered; 


Dear Dr. Asparuhov, Thank you for your response. 1. Is this way correct even though my moderator is dichotomous (boys/girls)? The interpretation would be then, the more someone is a girl, the more predictor affects the outcome...? 2. For the moderation to take place, the interaction term has to be significant. But what about the model fit? Can I compare the fit of a model with freely estimated path from the interaction term to the outcome with a model where this path is constrained to zero? Thank you again for your guidance! 


1. If gender=1 for girl and 0 for boy then the intercept Y is interpreted as the intercept for group boy, while the intercept for group girl would be [Y] + Y on gender. Similarly, Y on X is the coefficient for boy and the coefficient for girl is Y on X + Y on interaction. 2. Yes. Such a test would be equivalent to testing equality of the regression coefficient Y on X between the two groups. 

Back to top 