Anonymous posted on Tuesday, October 12, 2004 - 2:59 pm
I'm working on a latent profile analysis with a sample of 180 sexual aggressors. Unfortunately, I have about 10 predictors so am having problems with the models terminating normally. I don't expect that all of the predictors will be important for the final LPA but am unsure how to reduce the number of variables. Any suggestions? Thanks in advance for your response.
First, check that you are using Version 3.11 where attempts have been made to make it easier to have many x variables. An approximate exploratory approach would be to do the LPA without covariates, classify individuals into their most likely class, and then relate the covariates to the most likely class.
Anonymous posted on Friday, November 05, 2004 - 10:05 am
Here is my problem. I have a sample of 70 service programs in high-risk communities, with 4 to 6 continuous indicators or potentially 4 to 6 dichotomous indicators. I simply want to explore the structure of the data to see if there are potential latent classes. In the article by Cleland, Rothschild & Haslam 2000, Psychological Reports (87) they conclude that mixture models are less likely to produce false positives (finding a class that does not exist) at the expense of missing true latent classes. In this paper they only look at samples of 100, 300, or 600. If I am willing to accept that potential, given my study is exploratory, do you see any other drawbacks in running and LPA or LCA with a sample of 70 with potentially 4 to 6 indicators?
Is there anything else that may effect the results, given the sample size (other than the failure to detect latent classes)?
Moreover, if I find latent classes (even say only 2), how much am I taxing the model by adding one or two background variables. Many thanks in advance!!
bmuthen posted on Friday, November 05, 2004 - 11:43 am
With 70 observations it may be hard to identify many classes due to too small class sizes. If only 2 substantively meaningful classes are expected, this may be ok - even with 4-6 indicators. Finding an appropriate number of classes is not straightforward even with larger samples and depends on which statistic is used. Our simulations seem to favor BIC and particularly the sample-size adjusted BIC, but we haven't gone below n=200. The small sample size also causes the parameters and SEs to be less well estimated. Adding a background variable predicting class actually helps the analysis quite a bit in terms of adding information to give more stability to the estimation. Note also that Mplus can be used to easily do your own Monte Carlo simulation study to study these matters.