Message/Author 


Hi, does anyone have any input on dealing with ceiling effects by, instead of running censored analyses, using a MAR strategy for the ceiling values? I am modeling GPA data and the ceiling is 4.0. Does anyone have experience with this situation? The largest issue I can think of is the fact that, because of this ceiling, my intercept and slope factors are very highly and negatively correlated, which may cause improper value imputation. This is bad because the purpose of imputation would be to remove the restriction of range problem. Any input is appreciated! 

bmuthen posted on Friday, June 10, 2005  12:38 am



Do you mean replacing the ceiling value with missing data and let MAR kick in? I haven't seen this done. Does anyone have any input on this? 


Yes, that's exactly what I mean. One could simply specify the ceiling as a missing value. It came to me the other day and I thought it might be an interesting way to attempt to unrestrict range. Please, if anyone has any thoughts, I'd really like to hear them. If not, I'd love to collaborate on a project that explores this idea. 

bmuthen posted on Friday, June 10, 2005  12:46 pm



The question is if it is better to know that the observed value is at least ymax  as we do when treating it as censored from above  or ignore this information and infer the value from correlated variables  as we do in MAR. I would guess that the former is better since the y* value that censored works with is a form of an "imputation" of the values above ymax using correlated information as well as the ymax information. 


This is an interesting topic. I wasn't aware that the "uncensoring" process was informed by correlates. If this is true, then is it a good idea to avoid "unconditional models" when working with censored data? 


Hi Bengt, I cannot seem to find anything in the technical appendix on the censoring technique employed in Mplus 3.12. Do you know where I could find something on this? Thanks. 

bmuthen posted on Saturday, June 11, 2005  12:00 pm



A good source is the Maddala book that we refer to in our reference list on the web. The censoredinflated version, however, is not in there, but it something we came up with. 

Callie Burt posted on Saturday, March 15, 2008  8:00 pm



I am about to write up a description of the censoredinflated model (for a multilevel model that is right censored and inflated). I have two questions: (1) I am using the censoredinflated model in a large part because approx 25% of the sample is censored at 75 and the model seems to fit better when I add the inflation component than without. Does that seem ok? I don't know that there is a true always>75 group as the descriptions of inflation models (e.g., Long) seems to imply (2) As I understand it, the inflation portion relaxes the usual censored requirement that the covariates' influence on the prob of censoring is proportional to the covariates influence on the amount observed when not censored. Is that right? Thanks, apologies if I am being dense. 


(1) I don't know that one has to believe that there is a group of always > censoring point. Seems enough that in this case, such a group has manifested itself. (2) That is true  although it is not as clearcut as with 2part modeling. Unlike 2part, the inflation part does not predict the total probability of being at the censoring point because the regular censorednormal part also contributes to this probability. Good to have applications of this modeling out in the literature. Let me know how it works out. 


Thanks much for your response. In the above post from 2005, you noted that there were not any references for the censoredinflated technique which you created. I didn't see any on the site's reference list, but I may have missed something. Have you all written anything on the censoredinflated technique? Or are you aware of anything that has come out since this post? 


No we haven't written on this. We should, but have been to busy creating new things. We came up with the idea and programmed it, but left it at that so far. I don't think anybody has published on this. I am not aware of anything having come out on it since we put it into Mplus. 


One last question. In the abovecensoredinflated model, for the inflated portion of the model how would one interpret a positive coefficient? 


See Example 3.3. 

Back to top 