andkov / emosa

JAGS estimation/modeling EMOSA data

EMOSA

Programming support to estimation of EMOSA (Epidemic Models of Onset of Social Activities) models as applied to religious attendance of emergent adults.

See the full list of manuscript figures, including online appendix, in ./figures/

Abstract

Epidemic Models of the Onset of Social Activities (EMOSA Models) have been used in the past to describe behavioral processes that spread within social networks through social influence. Two theoretically distinct methods of social influence are represented within EMOSA models, social contagion (a direct one-to-one interpersonal social influence process) and general diffusion (the kind of social influence that passes through media and other broad cultural channels). Past EMOSA models explain adolescent problem behaviors such as smoking, drinking, precocious sexuality, and delinquency. In the current paper, we adapt the EMOSA framework to model transitions in religious participation during late adolescence and early adulthood. We fit the model to data from the National Longitudinal Survey of Youth 1997 survey, based on a U.S. household probability sample spanning 10 years of longitudinal observation, and examine the relative influence of social contagion versus general diffusion. We present several methodological innovations, including EMOSA applications in a new theoretical arena using a three-stage bi-directional model, a Bayesian Markov Chain Monte Carlo (MCMC) estimation approach, graphical innovations to represent different parts of the EMOSA process, and an empirical validation exercise. Results indicate that diffusion processes are dominant during periods of late-adolescent rapid reduction in church attendance, whereas both diffusion and contagion models explain the period of relative stability in church attendance following in early adulthood. Both methodological and substantive implications are discussed.

Model Specification

Model Results

• Diffusion Only
• Contagion Only
• Hybrid

emosa's Issues

Manipulate raw data to get observed proportions & transitions

Write contagion, diffusion & hybrid equations in Word

One single document with three series of equations.

Also save the Word document as a PDF, and have them in the /Documentation directory in the repository. Be really specific with subscripts, so to minimize translation errors for the JAGS model.

Write hybrid contagion/diffusion model in JAGS

Replace Gaussian likelihoods with beta likelihoods

So the proportions can't exceed [0, 1]

Example code for chain analysis

Most likely focusing histograms in ggplot2. Kernel of the code that Andrey will complete.

Validation of parameters against observed transition counts

(...or possibly proportions).

Andrey's self-prompt: "...Proportions of what?"

Save chains from MCMC results to a CSV.

It's a little more complicated than I thought.

Consider ggmcmc

It uses ggplot2 to graph the MCMC results.

http://martynplummer.wordpress.com/2012/10/21/ggmcmc-diagnostic-plots-for-mcmc-with-ggplot2/#more-189

Reparameterize C so it's binomial

change the distribution from ~dcat to ~dbin.

Write Contagion model in JAGS

Check beta parameterization with David or somebody

Make sure it's not too wacky. (It helps that it's consistent with the versions that used Gaussian likelihood).

knitr ERROR - cannot open the connection

"Diffusion.R" executes fine, reaching the data at the end of the links.
however, when run from "Diffusion.rmd" the connection fails:

Warning: cannot open file
'C:/Users/inspirion/Documents/GitHub/EMOSA/DiffusionOnly/Data/SummaryBirthYearByTime.csv':
No such file or directory
Error: cannot open the connection

Incorporate DIC

Try to use existing run to estimate the DIC.

If that doesn't work, the worst case is to uncomment the DIC line of code (which will take twice as long to run)

Write R code so that reads from dataset

Instead of the proportions hard-coded in the *.R file.