2-1-1 path analysis with binary media...
Message/Author
 Ahyoung Song posted on Tuesday, April 15, 2014 - 11:54 pm
Hi,

I am analyzing 2-1-1 path model with binary mediator (intent2) and outcome (condom) by using MLR. In the results, odds ratios were only given for intent2->condom and no odds ratios were given for independent variables -> intent2. May I ask how I obtain the odds ratios of independent variables? and in this case, is it ok to use MLR instead of other estimators?
I copied my syntax below:
VARIABLE:
NAMES ARE caseid weight cdattit
hivsus norm intent primary
condom intent2;
USEVARIABLES ARE caseid weight cdattit
hivsus norm intent primary
condom intent2 inter;
CATEGORICAL IS condom intent2;
BETWEEN IS cdattit hivsus norm;
WITHIN IS primary inter;
CLUSTER IS caseid;
WEIGHT IS weight;
DEFINE:inter = primary*intent2;
ANALYSIS: TYPE IS TWOLEVEL;
MODEL:
%WITHIN%
condom ON intent2(b)
primary
inter;
%BETWEEN%
cdattit hivsus norm intent2 condom;
intent2 ON cdattit(a);
intent2 ON norm(c);
intent2 ON hivsus(d);
condom ON cdattit;
condom ON norm;
condom ON hivsus;
MODEL CONSTRAINT:
NEW(indb indb1 indb2);
indb=a*b;
indb1=c*b;
indb2=d*b;

 Bengt O. Muthen posted on Wednesday, April 16, 2014 - 1:38 pm
You can create odds ratios by Model constraint, e.g.

or = exp(b);

MLR should be fine here.
 Ahyoung Song posted on Wednesday, April 16, 2014 - 9:02 pm
Thank you so much!
 Nicki Keating posted on Tuesday, August 04, 2015 - 2:34 am
Hello
I am interested in a 2 (cdep) -2 (ccoll)-1 (fiver) path, with moderation between the level 2 variables by another variable (cbot). Covariates age, male and ethnic are all categorical at L1, weight (continuous) at level 2. My drinking outcome is binary, but my level 2 mediator and predictor are both continuous. My code is:
categorical = fiver;
usevariables are ngh fiver male ses ethnic ccoll cbot cdep wgtbot INT2;
missing are all (-9999);
cluster is ngh;
between is ccoll cbot cdep wgtbot INT2; !centered except wgtbot
within = male ses ethnic ;
define:CENTER wgtbot(GRANDMEAN) ;
define: int2 =cbot*cdep;
ANALYSIS:
TYPE = twolevel;
estimator=wlsmv;
MODEL: %WITHIN%
fiver on male ses ethnic ;
%BETWEEN%
fiver ON ccoll(b);
fiver on cdep (cp1);

ccoll on cdep (a1);
ccoll on cbot ;
ccoll on wgtbot;
ccoll ON INT2 (bb);

MODEL CONSTRAINT:
new (dp_col_int wmodval dep_coll depdir);
dp_col_int = (a1+bb*wmodval)*b;
wmodval = 0;
dep_coll = a1*b;
depDIR = cp1;

Is this the correct estimator? I have assumed I use the raw data, and perhaps it is this assumption I have incorrect.
Thank you so much
 Bengt O. Muthen posted on Tuesday, August 04, 2015 - 6:43 pm
Looks right. Raw data are needed.
 Nicki Keating posted on Thursday, September 03, 2015 - 2:38 pm
Hello again,

Thank you so much for your prompt response.

I am getting very different results between MLR and WLSMV. I have read through many of the discussion pieces on estimators and read the choice of estimators pdf.

My binary outcome has a 90/10 split (only 10% have the outcome) and my mediators and independent variables (both at level 2) are very non-normal. Is it likely that it is these characteristics which are producing different results? Some discussion pages say WSLMV is more robust to non-normality, other pages say less. My sample size is 4267.

I hope you can help!
Many thanks
Nicki
 Linda K. Muthen posted on Thursday, September 03, 2015 - 3:41 pm
MLR is logistic regression as the default. WLSMV is probit regression. The comparison should be of the patterns of significance.
 Nicki Keating posted on Thursday, September 03, 2015 - 4:01 pm
Thank you so much for your incredibly prompt response. It is wonderful, especially being in a small country with not many MPlus users.

Sorry, I meant to say that my comparisons were relating to the patterns of significance. My results with my non-normal independent variables are all non-significant in WLSMV and all significant in ML and MLR using integration. I am unsure which to trust. I do have one very normally distributed independent variable which produces identical results using all 3 estimators, but my non-normal indicators produce very contrasting findings.

Any further advice would be greatly appreciated. I am so appreciative.

Nicki
 Linda K. Muthen posted on Thursday, September 03, 2015 - 5:03 pm
 Nicki Keating posted on Wednesday, December 16, 2015 - 12:00 pm
Festive greetings,

I have been reading everyone's posts who are interested in 2-1-1 models and I see many differences across them in their approach.

If x is the contextual predictor at L2 and m is the mediator at L1 and y is the outcome, I see the following codes:

%WITHIN%
y ON m (b) ;
%BETWEEN%
m on x (a);
y ON x(cp);
MODEL CONSTRAINT:
NEW(indb);
indb=a*b;

Or I see
%WITHIN%
y ON m;
%BETWEEN%
y ON m (b);
m on x (a);
y ON x (cp);
MODEL CONSTRAINT:
NEW(indb);
indb=a*b;

Or the Preacher Appendix for MLM is essentially the same as the first:
%WITHIN%
y ON m (b) ;
%BETWEEN%
m ON x (a)
y on m (b) constrained to be equal to b;
y ON x(cp);
MODEL CONSTRAINT:
NEW(indb);
indb=a*b;

Help! Thank you and wishing you a very merry Xmas.

Nicki
 Bengt O. Muthen posted on Wednesday, December 16, 2015 - 7:03 pm
I don't care for the first one because why would you not have y ON m on Between? If you leave it out you get the the y ON x slope distorted.

I like the second one because here it is clear that the effect of the between-level x is expressed in a good mediation model on Between. But I don't know if that's what the originators of the 2-1-1 language had in mind.

Number 3 is ok if one can make that equality assumption.