16
In other words , the modification of the algorithm consists on sampling from p(θ ∣ z, /3)
instead of from p(θ ∣ z,v,β'). A Metropolis-Hasting algorithm (see, Cowles, 1996;
Johnson and Albert 1999, p. 135) is used to simulate from the former distribution.
Once a posterior sample (∕31,θ1),..., (∕3m,0m) from p(β,θ ∣ z) is obtained the
Monte Carlo estimate of cell probability τ⅛ ≡ Pr[zf = k] for a future observation
with covariate Xf is
1 M
π* ≈ 77 = Σ K Wm - ‰) - Φ^1(∙τJ∕3m - f^)} for к = O,..., K - 1.
IVl 1
m=l
Finally, we note that Albert and Chib’s model for more than two categories re-
quires the “parallel regression assumption,” the probit version of the proportional
odds assumption of the logistic model. That is,
Pr[z ≤ к I τ] =Pr[xtβ + e ≤ 0fc+1]
=Pr[e ≤ θk+1 - xtβ] = Φ(0fc+1 - xtβ)
and, therefore,
fk(x) := Φ~1(-Pφ ≤ к I x]) = θk+1 - xtβ.
In other words, the assumption is that the functions fk for к — 0,...,7^-1 are
parallel lines.
2.3 A Phase III Clinical Trial
Studies have suggested that retinoid chemoprevention may help control second pri-
mary tumors, disease recurrence, and mortality for stage I non-small-cell lung cancer
(NSCLC) patients. A National Cancer Institute (NCI) intergroup phase III trial of
1166 patients with pathologic stage I NSCLC was conducted to validate the efficacy
of isotretinoin, a retinoid hypothesized to have chemopreventive properties. Patients
were randomly assigned to receive either placebo or isotretinoin (30 mg∕day) for 3
years in a double-blinded study. Patients were stratified at randomization by tumor