Hi and any help greatly appreciated specifying this model!
I have a binary outcome, measured at the person level. There are 9 'risk event' predictors (also binary coded as occurred/not). Some are correlated, some are not. ppl can experience a discrete number of these events coded as 1;s, the rest are coded as zeroes. I think of them as a subset of risk events from some hypothetical universe of risks that might have been experienced, so I am thinking of them as a random effect.
A person can have experienced some discrete number of the events, so I think that risks are nested within people. However I am having a hard time specifying the model as a multilevel model because the outcome is only measured once.
Can someone help me think through this model? I would like to specify a random effects model, but at a minimum I need to estimate the ICC for proportion of variance in the outcome explained by within person risk events vs between people variability.
I am probably not understanding this but here goes. Do you have a multivariate binary response model with the 9 outcomes nested within the person?
see Chapter 14.5 of this
Rasbash, J., Steele, F., Browne, W.J. and Goldstein, H. (2017) A User's Guide to MLwiN, v3.00 (PDF, 4,395kB), Centre for Multilevel Modelling, University of Bristol.
http://www.bristol.ac.uk/cmm/media/software/mlwin/downloads/manuals/3-01/manual-web.pdf
Kelvyn, thanks so much for your response... I did read chap 14 in the attached guide. However, that is not the model I am struggling with! Yes, I view the events as nested within people (L1) , but I only have one outcome measured per person (e.g. measured at L2).
I am predicting the binary outcome as a function on the nested 'exposures'. Any thoughts on this model?
Hi. It is unusual to have the outcome measured at L2 and to attempt to predict it using L1 variables. However, there are some authors who discussed such models.
Article Predicting Group-Level Outcome Variables From Variables Meas...
Ok - sorry not to understand. It would help to give more of an example of what is going on eg why is there always the same outcome on all exposures?
That being said there is this entry on MLwin resources
http://www.bristol.ac.uk/cmm/software/mlwin/mlwin-resources.html
Responses at more than one level & of different type see REALCOM
which links to
http://www.bristol.ac.uk/cmm/software/realcom/
But I have never done it.
Many thanks to you both for your responses. I am considering a multi-level latent model and have read the paper you suggest Laurentiu.
Kelvyn - the model is for predicting a cardiac event - (I have retrospective cross sectional data). So for each i there is only one outcome measure (cardiac event Y/N). The linear model is composed of covariates and a set of 'event predictors' - certain events that may be associated with the outcome (they are coded dichotomously...the events happened or not). It is this model I am having a hard time specifying. The events are clearly clustered within individuals. I know this specification will lend itself to a multilevel latent model, and I will probably go that route.
My issue, though, is that I would like to specify 'events' as a random variable because these are a (somewhat random) sample of all events that might have been evaluated but were not. If I cant specify the model as a random effects model I would like to try to estimate the proportion of the outcome accounted for between people vs the within person events.
Hope that is clearer and thanks!
I think I now understand.
I think that this is just a one-level model as the responses are not clustered. But I still may not be understanding - if that is so, please tell me.
So I would have a binary outcome and just fit a standard logit/ probit with co-variates plus dummies of whether the exposure has happened or not. If things are more complicated than that (sequential exposure? ) then recode the separate exposures into something more meaningful.,
Actually because I would be concerned with interactions and the worry of over-fitting I would use a tree classification model which has inbuilt complexity penalization. I find them useful as an exploratory even when I subsequently revert to classical - logit-like procedures. And they can help clinicians understand what is going on as a decision-like tree.
https://en.wikipedia.org/wiki/Decision_tree_learning
" The events are clearly clustered within individuals ". If I understand correctly, you have more than one measurement moments from each participant, and in some cases the cardiac event occurred.
If this is the case, you can predict the occurrence of the cardiac event using both intra-personal and inter-personal variables in a Bernoulli multilevel model.
- Badges
- Science topic
- Similar topics
- Engineering
- Materials Engineering
- Powders