|
|
3-level model (binary outcome) |
|
Message/Author |
|
|
I have a three-level dataset: reports (1-7 "pa" reports per event) events (1-29 "condom" events per indiv) individual (49 individuals) report is continuous, event is binary. I need to use the univariate format and cluster statement since there are varying numbers of reports per event and varying number of events per individual. I can 1.) run using indiv as the cluster, BUT this incorrectly assumes that each report is indep w/in indiv (1-7 reports per event all refer to the same event). %WITHIN% condom on pa; %BETWEEN% condom on age gender; 2.)I can assume a cluster of indiv*event, this appropriately assumes relationship among reports from the same event within individual. BUT, I cannot have the relationship b/t event and report be non-linear (logit), the relationship between the event and the individual intercept be non-linear (logit) and the relationship between the continuous individual intercept and the individual-level covariates be linear which is what I want. I get errors that say the threshold should be at the between level as well. Any suggestions? here is what I tried that gives errors. %WITHIN% fpa by pa@1; condom on fpa; [condom$1@0]; %BETWEEN% condom on age gender; [condom@0]; |
|
|
I think it is ok to have varying number of reports per event - why not use a wide, 7-variable format where many have missing data on some reports? That would make it a two-level model. |
|
Emily Blood posted on Friday, March 18, 2011 - 11:37 am
|
|
|
I thought that if I had 7 report ("pa") variables and for most observations some were missing that in the wide, multivariate format the whole observation is dropped. So, if an event only had 6 reports, pa7 would be missing and therefor the observation would be dropped. If this is not the case, I will do that. That would be great. Please let me know. Thank you for your response! Emily |
|
|
You have a factor influencing the pa's, so the pa's are DVs. People are not dropped due to missing on DVs. |
|
Back to top |
|
|