LCA p-values, CI's and bootstrapping ... PreviousNext
Mplus Discussion > Structural Equation Modeling >
Message/Author
 Jean Flores posted on Wednesday, October 31, 2018 - 10:50 pm
Hello, I am attempting to conduct an LCA analysis for complex survey data with weights, strata and clusters using the following syntax:
Analysis:
Type=COMPLEX MIXTURE;
Tech10;
Output: tech10 cinterval;

For the OR comparisons of the different classes, I noticed that some of my p-values are >0.05, but the 95% confidence interval is above 1.0. My understanding is that the distribution of my data may be asymmetrical, and I should use the bootstrap procedure to get adjusted CI's. I added the two additional options at the bottom of this post to the Analysis procedure, but each time I run the program, it always terminates prematurely, and no errors are indicated in the log. This happens regardless of the # of iterations I request for bootstrap = xx. Any advice on how proceed?
REPSE=BOOTSTRAP;
BOOTSTRAP=2000;

Thanks
 Tihomir Asparouhov posted on Saturday, November 03, 2018 - 11:39 am
Bootstrap is not available for type=complex mixture.

The confidence interval significance of the OR matches the significance of the beta coefficient.
 Jean Flores posted on Thursday, August 29, 2019 - 4:36 pm
Hello, I have a follow-up question regarding this analysis. I found that the 4-class model is the best fit, and for each variable, I would like to do a test of difference across the four classes for each variable and report the resulting p-value (see sample table at bottom of message).

Is there an output Tech option that does this test? If yes, what option should I use, and what would the heading of the output be?
Second, in light of the issue described above where I found that the CIs didn't always align with the p-values for this complex analysis (I believe due to asymmetry issues), is it fair to assume that a test of difference across the 4 classes and resulting p-value would be reliable?

Probability of responding "Y" to survey question

--------CL1 CL2 CL3 CL4 p-val
VAR1 0.82 0.77 0.35 0.28 p=
VAR2 0.52 0.63 0.11 0.05 p=
VAR3 0.63 0.50 0.32 0.54 p=
VAR4 0.37 0.50 0.68 0.46 p=
etc..

Thank you for your help!
 Bengt O. Muthen posted on Saturday, August 31, 2019 - 5:30 pm
With mixtures, you can test hypotheses using the Model Test command. For instance, if you want to test that the mean is the same in 3 classes, you say:

Model Test:

0 = b-a;
0 = c-a;

where a, b, c are labels given in the Model command to the means in the 3 classes. You do this for each variable at a time in separate runs.

Yes, probabilities have non-symmetric sampling distributions. I don't know how strongly that affects Model Test which does assume approx. normally distributed estimates a - c. With binary variables, you could consider instead testing the corresponding estimated logits which are closer to normally distributed.
 Jean Flores posted on Wednesday, September 11, 2019 - 11:11 am
Thank you Dr. Muthen for the response. My analysis is actually an exploratory LCA, and I didn't use the model command. In you opinion, is it even advisable to do tests of differences across the variables for an exploratory analysis (i.e. fitting and testing a model with the same data)? If this is a valid approach, can you suggest some syntax for the model statement? My current syntax is
TITLE: EXPLORATORY ANALYSIS

DATA: FILE IS C:\...\FILEINP.CSV;
FORMAT IS FREE;
VARIABLE: NAMES ARE WEIGHT STRATUM
PSU VAR1 VAR2 VAR3 ETC. ;
missing are .;

USEVARIABLES ARE VAR1 VAR2 VAR3 ETC;

CATEGORICAL ARE VAR1 VAR2 VAR3 ETC;

STRATIFICATION IS STRATUM;
CLUSTER IS PSU;
WEIGHT IS WEIGHT;
CLASSES = c(4);

Analysis:
Type=COMPLEX MIXTURE;

Output: tech10 cinterval;
 Bengt O. Muthen posted on Wednesday, September 11, 2019 - 1:31 pm
I think it is ok to test in exploratory settings - although more typically you just look at the different mean/probability profiles using the Plot command.

To test, you would have to add a Model command where you list your variables' thresholds and give them labels that are then used in Model Test. Along these lines for binary variables and 2 classes:

%c#1%
[var1$1-var5$1] (thresh11-thresh15);
%c#2%
[var1$1-var5$1] (thresh21-thresh25);

Model Test can then test differences in these thresholds or the corresponding probabilities P = 1/[1+exp(thresh)].
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: