Description
1. Consider the two different hyperprior formulations for the binomial hierarchical model of
Lesson 3.2: Hierarchical Modeling Fundamentals. This exercise shows how different those
priors are.
Note: Consult help(distributions) in R for the random number generators you will need.
(You do not need JAGS.)
(a) The first prior formulation was
θj | α, β ∼ Beta(α, β)
α, β ∼ iid Expon(0.001)
(i) [2 pts] Independently simulate 1000 pairs (α, β) from their hyperprior, and
produce a scatterplot of log(β) versus log(α).
(ii) [2 pts] Using the simulated pairs (α, β), forward-simulate θj , and produce a
histogram of the result (an approximation of its marginal prior).
(b) The second prior formulation was
θj | α, β ∼ Beta(α, β)
α = φ1/φ2
2 β = (1 − φ1)/φ2
2
φ1 ∼ U(0, 1) φ2 ∼ U(0, 1000)
(i) [2 pts] Independently simulate 1000 pairs (α, β) from their hyperprior, and
produce a scatterplot of log(β) versus log(α).
(ii) [2 pts] Using the simulated pairs (α, β), forward-simulate θj , and produce a
histogram of the result (an approximation of its marginal prior).
2. Twelve separate case-control studies were run to investigate the potential link between
presence of a certain genetic trait (the PlA2 polymorphism of the glycoprotein IIIa subunit
of the fibrinogen receptor) and risk of heart attack.1 For the j
th study, an estimated
log-odds ratio, ψˆ
j , and its (estimated) standard error, σj , were computed:
j ψˆ
j σj j ψˆ
j σj j ψˆ
j σj
1 1.055 0.373 5 1.068 0.471 9 0.507 0.186
2 -0.097 0.116 6 -0.025 0.120 10 0.000 0.328
3 0.626 0.229 7 -0.117 0.220 11 0.385 0.206
4 0.017 0.117 8 -0.381 0.239 12 0.405 0.254
1From Burr, et al. (2003), Statistics in Medicine, 22: 1741–1760.
1
Consider this Bayesian hierarchical model:
ψˆ
j | ψj ∼ indep. N
ψj , σ2
j
j = 1, …, 12
ψj | ψ0, σ0 ∼ iid N
ψ0, σ2
0
j = 1, …, 12
ψ0 ∼ N
0, 10002
σ0 ∼ U(0, 1000)
with ψ0 and σ0 independent, and the values σj , j = 1, . . . , 12, regarded as fixed and known.
(a) [2 pts] Specify improper densities that the proper hyperpriors given above are
apparently intended to approximate. (Which parameters are the hyperparameters?)
(b) [5 pts] Draw a directed acyclic graph (DAG) appropriate for this model. (Use the
notation introduced in lecture, including “plates.”) You may draw it neatly by hand or
use software.
(c) [5 pts] Using the template asgn2template.bug provided on the course website, form a
JAGS model statement (consistent with your graph). Also, set up any R (rjags)
statements appropriate for creating a JAGS model. Be careful to name your data
variables correctly. [Also remember: JAGS “dnorm” uses precisions, not variances!]
(d) [5 pts] Run at least 10,000 iterations of burn-in, then 100,000 iterations to use for
inference. For both ψ0 and σ
2
0
(not σ0), produce a posterior numerical summary and
also graphical estimates of the posterior densities. Explicitly give the approximations of
the posterior expected values, posterior standard deviations, and 95% central posterior
intervals. (Just showing R output is not enough!)
(e) Suppose a new case-control study is to be performed, and assume that its log-odds
standard error (new σ) will be 0.125. Assume the ψ for the new study is exchangeable
with those for the previous studies.
(i) [2 pts] Re-draw your DAG, adding new nodes to represent the new ψˆ and new ψ.
(ii) [2 pts] Correspondingly modify your JAGS model to answer the following parts.
Show the modified JAGS and R code and output that you used.
(iii) [3 pts] Estimate the posterior mean and posterior standard deviation, and form a
95% central posterior predictive interval for the estimated log-odds ratio that the
new study will obtain. (Remember, this new estimated log-odds ratio will be the
new ψˆ, not the new ψ.)
(iv) [1 pt] Estimate the posterior predictive probability that the new estimated
log-odds ratio will be at least twice its standard error, i.e., at least two standard
errors (2σ) greater than zero. (This is roughly the posterior probability that the
new study will find a statistically significant result, and in the positive direction.)
Suggestion: Add an indicator variable to your JAGS model – one that equals 1
when the event occurs, and 0 otherwise. (What is its mean?)
Use at least 10,000 iterations of burn-in, and 100,000 for inference as before.
Total: 33 pts