bma() named its first two elements identically, both
"Table with the binomial model prior results". The second element holds the
binomial-beta results and is now named accordingly. Only the name changes;
the contents were always correct.model_space() gains mc3, draws and burn. With mc3 = TRUE the model
space is explored by MC^3 sampling (Madigan and York, 1995) instead of
exhaustive enumeration, which makes a large number of regressors feasible:
enumeration means fitting 2^K models, already 131,072 at K = 17. Reduced
model spaces are supported, with the proposal correction the size constraint
requires. The chain reports its acceptance rate, the number of distinct
models visited and a convergence diagnostic, and warns when the diagnostic
falls below 0.99 or when the chain has barely moved.
Extreme Bounds Analysis is withheld for a sampled model space. Extreme bounds
are the smallest and largest coefficient estimates across all models, and a
sampler does not visit the extremes of a distribution it is exploring by
posterior mass, so bounds from a visited subset are systematically too
narrow. Element 3 of the bma object is NULL and a message explains why.
print, summary and coef methods for the model_space and bma
objects. summary() on a bma object gives the coefficient table ordered by
posterior inclusion probability, for either model prior; coef() returns the
posterior means as a named vector, optionally conditional on inclusion.
model_sizes() and model_pmp() gain type, selecting between the existing
line display and a histogram.
The estimators no longer depend on the BLAS for their treatment of a
perfectly fitting model. With a constant dependent variable the residual sum
of squares is zero in exact arithmetic, but whether it came out as exactly
zero or as a rounding-level residual decided whether the log marginal
likelihood was Inf or a large finite number. Rounding-level sums of squares
are now snapped to zero, so the degenerate case behaves identically on any
BLAS. Reported by CRAN's BLIS check machine.
The same guard removes a cancellation in the g-prior estimators. The centered
sum of squares was computed as a difference of two nearly equal numbers and
could come out slightly negative, which would have produced a NaN log
marginal likelihood and propagated silently into the posterior model
probabilities.
coef_hist(), model_pmp(), model_sizes() and posterior_dens() now have
\usage sections in their help files. A utils::globalVariables() call
between the roxygen block and the function had attached the block to that
call, so no usage could be derived. Reported as a NOTE by CRAN's r-devel
check machines.
model_pmp() misspelled "ranking" on the x axis of two of its three
graphs, and reported the wrong quantity when top
exceeded the size of the model space.
An enumerated model_space object now always carries a sixth element,
info, recording how the space was built. Elements 1 to 5 are unchanged.
The bma object gains a sixteenth element holding the chain diagnostics,
NULL for an enumerated model space.
Results obtained from an enumerated model space are unchanged from 0.1.2.