I'm looking for the best method to analyze data from a 36 month follow up study of substance abuse. We want to model the predictive value of different variables on relapse in substance abuse of a cohort of participants that received treatment while being in jail. Any suggestions?
Logistic Regression. Set up a categorical dependent variable. You also may want to set up lag variables if you want to look at certain time periods.
I'm not sure that logistic regression captures the time component, because it focuses only on the ultimate outcome (relapse or not) rather than the outcomes at each time point. You may want to look at survival analysis methods, which can help you model *different* points in time of relapse that are a function of your predictors.
You can have relapse as the dependent variable and use logistic regression. Relapse can be a dichotomous variable in this model, yes or no. The other possible variable is the number of relapses which can be an interval variable with real zero value. In this case you can utilize linear regression model. The other variable can be length of being relapsed, the period of time that takes the patient return to start the next detoxification. This could be the same as the later one.
Depends on whether you want to model time to relapse (survival analysis) or presence of absence of use over time. The latter is longitudinal logistic regresssion which can be handled with a non-linear mixed model or a generalized estimating equation. What software do you have access to?
Thanks, Paul. We are interested in the latter option. Presence of substance abuse or dependency have been assessed with he CIDI interview module at diferent times. We have access to SPSS, Stata and R.
I figure you can do this in any of those packages. I am a SAS user but SPSS has a generalized linear models program that will do generalized estimating equations (GEE). I'm sure the others do as well. Good luck. Sounds like an interesting analysis.
I can help you in SPSS. Both logistic and linear models are applicable in SPSS. Please justify your hypothesis, then let me know how I can help you.
I don't doubt it's possible to do a mixed effects logistic regression model in SPSS; however, I've found the analysis to be a bit more straight forward in R. If you're interested in using R to run the analysis, Juan, try looking up the glmer( ) function. You can specify which predictors are longitudinal in the regression formula part of the function, and you can specify a logistic model by setting the "family" argument to family=binomial(link = "logit"). Documentation on glmer( ) can be found here: http://hosho.ees.hokudai.ac.jp/~kubo/Rdoc/library/lme4/html/lmer.html
There is a very comprehensive websie at
http://gseacademic.harvard.edu/alda/
which accompanies Applied Longitudinal Data Analysis:Modeling Change and Event Occurrence by Judith D. Singer & John B. Willett
Discrete time survival models are very flexible - the trick is to create the data in a long form that is as person -period data with a binary response if the event occured to a person in a particular time interval ; this is well explained with clear examples in the Alda book and also in
Judith D. Singer, John B. Willett (1993) It's About Time: Using Discrete-Time Survival Analysis to Study Duration and the Timing of Events , Journal of Educational and Behavioral Statistics -vol. 18, no. 2, pp. 155-195, DOI: 10.3102/10769986018002155
http://ccsr.ac.uk/qmss/summer/Paris09/documents/SingerWillet.pdf
They use SAS while the UCLA site gives code in a number of packages - see http://www.ats.ucla.edu/stat/examples/alda.htm
For more advanced issues including:
Multilevel models for recurrent events and unobserved heterogeneity
Modelling transitions between multiple states
Modelling competing risks
Multiprocess models for correlated histories
there is material at the University of Bristol Centre for Multilevel modelling website
http://www.bristol.ac.uk/cmm/software/support/workshops/materials/eha.html
Hi Juan, Kelvyn's answer is outstanding in terms of the resources it provides. I had one problem with your question, which is I couldn't tell if you just did one follow up at 36 months, or if you followed the subjects. If you only did the one follow up, you options may be limited to the logistic regression approach. There is one middle ground. If you asked when they relapsed, you may be able to create a retrospective person-period data set. There are instructions in this article under my publications: "Estimating multi-level discrete-time hazard models using cross-sectional data: Neighborhood effects on the onset of adolescent cigarette use." There is a similarly named conference session. The full text is under the paper in Multivariate Behavioral Research. So basically, I see three options. Multiple time points, a survival approach, if you have one time point with time of relapse information, retrospective survival, or just relapses yes/no, logistic approach. Bob
The research team conduted 3 follow ups at 12, 14 and 36 months. Thanks to all of you. I'll check the references.
Hi Juan, I am going to assume that's 12, 24 (not 14), and 36 months. I am assuming that at each time period you are asking whether there as been a relapse or not, but not more about when. So you have a pretty sold (if short in terms of the number or time points) survival (aka event history) analysis. Kelvyn's links will be very helpful! Many people like the Singer and Willet book because of the great writing. They were pretty important figures in bringing survival analysis from medicine to other disciplines. You will see they have several interesting papers. Since your example seems pretty straightforward, you will find that lots of statistical software can handle this. Bob
- Badges
- Science topic
- Similar topics
- Biological Science
- Microbiology
- Virology
- Virus