Roger Brown posted on Monday, February 16, 2015 - 9:44 am
I have been playing around with the new univariate entropy measures with an LCA using a variety of binary items. Sample size is fairly large >1500 cases. Once I seem to get larger than 17 binary items univariate entropy fails to estimate (999.000) and receive the following:
THE MODEL ESTIMATION TERMINATED NORMALLY
THE CHI-SQUARE TEST IS NOT COMPUTED BECAUSE THE FREQUENCY TABLE FOR THE LATENT CLASS INDICATOR MODEL PART IS TOO LARGE.
I am assuming a limitation on the size of the matrix? Is there a way to override this? Thanks.
Hi, I am working with a model which is similar to example 7.20 in the user guide (p. 175), including both a continuous (f1), a categorical (c) latent variable and a set of indicators (y1-y3). Is it possible to compute the univariate entropy for the continuous latent variable (f1 in example 7.20)? I have no problems obtaining the univariate entropy values for my indicators y1-y3. Thanks in advance, David
Another question on univariate entropy: My model contains a single categorical latent variable c and a number of indicators y1-y6. Some of these indicators (y1-y4) correspond to single items on a questionnaire, together they make up a factor (f), whilst other indicators (y5-y6) correspond to the domains of a scale and are based on the mean of a number of items. Is it possible to directly compare the univariate entropies of indicator y1 with indicator y6?
In example 7.20, I would be trying to compare the univariate entropy of indicator y3 with a hypothetical indicator y7 which is not part of any of the factors.
Another question based on example 7.20 in the user guide. I understand that I cannot directly compute the univariate entropy for the continuous latent variable f1. Is there another method to compute that value, e.g. summing univariate entropies for the indicators y1-y3?
You have to setup the "univariate" models, with 1 dependent variable and all parameters fixed to the parameters from the multivariate model. At that point the univariate model entropy is the univariate entropy and you can compute it manually using the cprobs from the univariate model.
I'm wondering whether univariate entropy values can be used to evaluate the quality of indicators for longitudinal mixture models? I'm trying to troubleshoot for a 4 class LCGA model for a 5-category ordinal outcome using 7 waves of data that has low entropy (0.53). I’d thought that the low entropy was largely because of substantial sample attrition. However, I requested the univariate entropy values and they’re all uniformly low (range: 0.270-0.295!). Are univariate entropy values valid for these types of models? Does this suggest that we should reconsider our latent class indicators?
Currently the univariate entropy doesn't take into account the missing data and it sort of implies what the entropy would be if all the data is observed. We will change this in the next Mplus version to account properly for the missing data. Thanks for pointing this out.
I am trying to compute the univariate entropy for 11 continuous indicators of my LPA.
I am having trouble reading the output. The value of the first variable is 999.000 and the output is not neat as the one in Asparouhov and Muthen (2018) paper but it seems shifted. In fact, the final row ends up with a value that is not in column with the correspondent variable (see below).
(NB the squared brakets are just for formatting purposes)
I am wondering therefore if I should disregard the 999.000 and shift the reading of my univariate entropy values.
Also, only two of my indicators make a substantial contribution to my overall entropy. The remaining values range between .191/.205, would you get rid of all of them? would you remove only those below .200? any suggestion or resource on this matter would be greatly appreciated.
UPDATE: I have just realized that the reason for my univariate entropy output to be shifter is that my censored variable is not fully reported. Meaning that I have 10 variable names (instead of 11) but 11 univariate entropy values. I reckon that the 999.000 value, although it is matched with another indicator, actually refers to my censored variable but I am not entirely sure if this is the proper way to read the output.