I have three predictor variables that counted how many times an offender used: 1)change talk, 2)was resistant, or 3)responded neutrally during a supervision session with a probation officer. I wanted to determine whether the number of evidence-based practices an officer uses during a supervision session changes as a function of the proportion of time the offender is using change talk vs. resistance vs. responds neutrally. Therefore I have three variables (change talk, resistance, neutral response) and I divided each one by the total number of offender reactions to get a proportion. For example if the offender used change talk 5 times, resistance 2 times, and responded neutrally 10 times, I created three proportions that add up to 1 (e.g., 5/17 = 0.294 would be the proportion I would use to represent the proportion of time used change talk). What would be the best way to use three variables that are proportions and add up to 1 when running a multi-level model analysis. I tried treating the 3 predictors like multinomial data, but Mplus gave me an error message notifying me that they are not even integers. Does anyone have any ideas of other ways I can model the proportion of time an offender used change talk, was resistant, or was neutral during each supervision session? I do not want to have to dichotomize these three variables, because then I lose information.
I do have one more quick question. Can you specify in Mplus that a independent variable is a proportion and ranges from 0 to 1 in a Crossed Random Effects Multilevel Model? The syntax I am currently using looks like this and I just want mplus to know that Change Resist and Neutral are proportions that add up to 1: VARIABLE: NAMES ARE ID Change Resist Neutral NumEBPs; USEVARIABLES ARE ID Time Change Resist Neutral NumEBPs x1 x2 x3; WITHIN=Change Resist Neutral x1 x2 x3; COUNT IS NumEBPs; CLUSTER=ID; ANALYSIS: TYPE=TWOLEVEL RANDOM; ESTIMATOR=MLF MODEL: %WITHIN% NumEBPs ON Change Resist Neutral;