= Solution
A <Gibbs sampler> draws each variable from its <full conditional distribution>, so every proposed update is accepted. At the start of a sweep define $c_s=\widetilde c_s-E_s$, $y_s=\widetilde M_s-RE_s$, and
$$
d_s=\widetilde M_s-M_0-\beta\widetilde c_s,
\qquad
v_E=\left(\frac1{\sigma_c^2}+\frac{(R-\beta)^2}{\sigma_M^2}\right)^{-1},
$$
$$
m_{E,s}=v_E\left[
\frac{\widetilde c_s-c_0}{\sigma_c^2}
+\frac{(R-\beta)d_s}{\sigma_M^2}
-\frac1\tau
\right].
$$
First update the reddenings independently as
$$
E_s\mid\text{rest}\sim TN_{[0,\infty)}(m_{E,s},v_E).
$$
Recompute $c_s$ and $y_s$, then make the Gaussian updates
$$
c_0\mid\text{rest}\sim N\!\left(\bar c,\frac{\sigma_c^2}{N}\right),
$$
$$
M_0\mid\text{rest}\sim N\!\left(
\frac1N\sum_s(y_s-\beta c_s),\frac{\sigma_M^2}{N}
\right),
$$
$$
\beta\mid\text{rest}\sim N\!\left(
\frac{\sum_sc_s(y_s-M_0)}{\sum_sc_s^2},
\frac{\sigma_M^2}{\sum_sc_s^2}
\right),
$$
and
$$
R\mid\text{rest}\sim N\!\left(
\frac{\sum_sE_s[\widetilde M_s-M_0-\beta c_s]}{\sum_sE_s^2},
\frac{\sigma_M^2}{\sum_sE_s^2}
\right).
$$
With
$$
S_c=\sum_s(c_s-c_0)^2,
\qquad
S_M=\sum_s(y_s-M_0-\beta c_s)^2,
$$
the remaining full conditionals, in the parameterization given in the question, are the <scaled inverse chi-squared distribution>[scaled inverse chi-squared laws]
$$
\sigma_c^2\mid\text{rest}
\sim\operatorname{Inv}\text{-}\chi^2\!\left(N-2,\frac{S_c}{N-2}\right),
$$
$$
\sigma_M^2\mid\text{rest}
\sim\operatorname{Inv}\text{-}\chi^2\!\left(N-2,\frac{S_M}{N-2}\right),
$$
$$
\tau\mid\text{rest}
\sim\operatorname{Inv}\text{-}\chi^2\!\left(2N-2,\frac{\sum_sE_s}{N-1}\right).
$$
Repeating these updates in the displayed order gives a complete Gibbs sweep. The formulas assume $N>2$ and nondegenerate sampled predictors; posterior propriety must be checked because the hyperpriors are improper.
Back to article page
