Anonymous posted on Sunday, August 07, 2005 - 7:31 am
I have data that is in a two level structure, children nested within families.
The variable of interest is latent with 3 indicators. The indicators are censored at the bottom (ie, zero inflated).
However, when I attempt to run the two level SEM model in Mplus, I get the following message:
*** ERROR in Analysis command TYPE=TWOLEVEL is not available when outcomes are censored, unordered categorical (nominal), or count variables.
So, may this be an error on my part in the model commands?
Or if not, could you offer a suggestion as to how to go about using the censored command while still taking into account the nesting structure of the data?
Thank you for your time.
bmuthen posted on Monday, August 08, 2005 - 1:24 pm
No, not an error on your part. It is simply the case that the current Mplus version does not cover 2-level censored analysis. You can, however, use type=complex to correct your SEs for the nested data structure.
Tashia Abry posted on Monday, September 29, 2014 - 11:24 am
I have noticed that in an analysis using a censored (from above) dependent variable , the adjusted mean for the outcome is slightly greater than the scale maximum. When running the exact same model without the Censored command, the adjusted mean is within scale limits leading me to believe the discrepancy is a function of censoring the DV. I would like to present the results from the Censored model, but wanted to make sure this was appropriate even though the adjusted mean is greater than the scale maximum. If so, how could we most accurately interpret the adjusted mean in the censored model?
Under the censored model the DV is y*, the latent continuous response variable underlying y. Because you have censoring from above, y* stretches beyond the max observed value and can therefore get a high y* mean. Please see the censored-normal literature, also mentioned in our Topic 2 handout and video.
Is there a way to test whether modeling an outcome as censored provides a better fit than modeling it as not censored and the comparison between censored and inflated?
I have been running some multilevel models (two levels) and although the fit indices for some outcomes appear different, I do not see how to compare them.
For example, for var1 (which does not seem to have a censored or an inflated distribution) I get the following: not censored log: -5055 BIC/ADJ: 10124 censored log: -5135 BIC/ADJ: 10285 cens/inflated log: -5042 BIC/ADJ: 10103 Although these fit indices are different, they are not that different.
In comparison, for var2 (which seems to have an inflated distribution) I get the following: not censored log: -4990 BIC/ADJ: 9996 censored log: -4162 BIC/ADJ: 8339 cens/inflated log: -4010 BIC/ADJ: 8045 The fit indices of these three models vary much more than for var1, and are all a lot lower for the cens/inflated model.
I would like to justify using the censored/inflated model for var2, whereas I would like to say that it is not justified for var1.
Also, for a continuous outcome can Mplus model it as inflated but not censored? Or is censoring a requirement for inflation?
Starting with your last question, "inflation" implies a piling up at the floor (or ceiling), so the variable is then by its nature censored. UG ex 7.25 shows how to do explicit inflation modeling using a 2-class model but that is for counts, not continuous outcomes.
Regarding your early questions, they are answered in the latest workshop videos just posted from the Johns Hopkins August 16 workshop on Regression and Mediation Analysis using Mplus,