[Spring 2018] Meeting 2 - Jamie's Latent Class Analysis & MLM Logistic Regression

Hi all,

Thanks for all of the helpful feedback on my project today! Attached are my slides. I made some slight edits.

One thing I wanted to follow up on. We talked about group-mean centering variables at Level 1 as an alternative to including the calculated PRE group means at Level 2. One issue with this is that, if I group-mean center the POST variables, the variables will no longer be 0's and 1's if the student's mean is between 0 and 1... their Level 1 scores will have decimals. This means I will need to specify the variables as continuous. Based on our past conversations, the consequence of specifying a dichotomous variable (or one that has a non-normal distribution, as would be the case with the group-mean centered POST variables) as continuous is that you have less power to detect effects. 

I am thinking I can do a blend of the two:
1. Keep the "POST" variables as dichotomous and model them at both levels using multilevel SEM.
2. Keep the "PRE" variables as only within-level variables and group-mean center them. You can't specify them as continuous, anyway. This will also get rid of the weird autoregressive effect at Level 2.


My code now looks like this -- and the model output is happy!

  CLUSTER ARE id;

  USEVAR =    PostDefAct PostMin !POST variables
              PreMin PreDefAct; !PRE variables

  WITHIN =    PreMin PreDefAct; !PRE variables

  BETWEEN =   ;

  MISSING=ALL(-999);

  CATEGORICAL =
  PostDefAct PostMin; !POST variables

  DEFINE:
  CENTER
  PreMin PreDefAct(GROUPMEAN); !PRE variables

  ANALYSIS:
  TYPE =
  TWOLEVEL;

  MODEL:

  %WITHIN%
  PostDefAct ON PreMin PreDefAct;
  PostMin ON PreMin PreDefAct;

  %BETWEEN%
  PostDefAct WITH PostMin;

  OUTPUT: SAMPSTAT STDYX CINTERVAL;


Let me know if you have thoughts or other suggestions,
Jamie