First release.
The package went through three internal revisions before this one. None was released anywhere, so there is no upgrade path to describe – but the decisions taken along the way are what set the defaults, and a default that was chosen against a measured alternative is worth knowing about when interpreting a fit. That is what the rest of this file records.
The continuous adaptive lasso’s omega
hyperprior now defaults to Gamma(0.1, 0.1), not
Gamma(2, 2). Gamma(a, a) has mean 1
for every a, so this holds the prior mean fixed and only
widens it – but the shape also fixes the density at the origin, and that
is the operative change: shape > 1 sends the density to
zero as omega -> 0, shape < 1 sends
it to infinity. Since small omega means strong shrinkage,
the old default forbade exactly the region the data want. Measured in
Experiment C1’s 6 x 2 factorial (3,200 paired scenarios per cell, n =
1000): the posterior places ~23% of omega’s mass below
1e-3 and Gamma(2, 2) permitted
none of it. Loosening moves RMSE
0.0895 -> 0.0843 (-5.8%), coverage
0.8008 -> 0.8234 and Winkler
0.7074 -> 0.6664; it wins at every tau and
never trades one quantile against another.
Propriety is unaffected. Proposition 1C requires
only that the hyperparameters be strictly positive – there is no lower
bound, and its normalising bound does not involve
alpha_omega or beta_omega at all.
Gamma(0.1, 0.1) is exactly as proper as
Gamma(2, 2).
Scope. This changes cbqr_alasso() only.
The continuous lasso layer’s tau hyperprior and
the binary hierarchy in prior() are unchanged: the
factorial covered the continuous adaptive lasso alone, and the binary
sensitivity study reached a different conclusion about what the binding
costs. Results from earlier versions will differ; pass
prior = list(alpha_omega = 2, beta_omega = 2) to reproduce
them.
New cbqr(), with cbqr_alasso(),
cbqr_lasso() and cbqr_none(), fitting the
three penalty layers to an observed continuous
response, against three new Fortran kernels. The binary path is
unchanged.
There is no anchor and no
sigma_fixed. A continuous response identifies the
scale, so sigma is sampled every sweep and the anchors are
misspecifications rather than competing conventions. The unpenalised
kernel therefore gains a sigma draw, which its
binary counterpart does not have at all.
New q argument: the exponent on sigma
in the local-scale prior rate, sigma^q / (2 lambda_j^2).
The layer defaults differ, and this is inherited rather than
chosen – the published lasso penalty is
lambda|beta_j| (q = 0) and the published
adaptive penalty is (sqrt(sigma)/lambda_j)|beta_j|
(q = 1). In the binary model sigma is
anchored, so the two give the same sampler and the difference is
invisible; it is not invisible here. q = 2 is the
response-scale equivariant choice, and draws sigma by slice
sampling rather than conjugately.
standardize defaults to TRUE and now
scales y as well as X. No
proper hierarchy is exactly response-scale equivariant, so fixing the
scale is part of the specification rather than a preprocessing
convenience. Draws are mapped back to the original units
exactly.
predict() on a cbqr fit returns the
fitted conditional quantile. Note that the binary predict()
returns P(y = 1 | x) from the ALD distribution function,
which on a continuous fit saturates at 1 – cbqr objects get
their own method for exactly this reason.
plot() accepts sigma as a plottable
parameter, which the binary model has nothing to show for.
Geweke’s joint-distribution test (Geweke 2004, JASA
99, 799-804) is added for the unpenalised binary kernel
and for all three continuous kernels. Every other test in the suite
checks a consequence of the sampler being right – constraints
hold, coefficients are recovered, draws are finite. A kernel with a
wrong full conditional passes all of them and still converges to the
wrong distribution. This one draws the joint p(theta, y)
twice, once ancestrally from the prior and once by a two-block Gibbs
chain that uses the package’s own sweep as its theta
transition, and the two agree only if that sweep leaves the posterior
invariant.
Largest z over eleven to fifteen moments at
tau = 0.3, 0.5, 0.75: bbqr_none 1.36 / 1.25 /
1.18; cbqr_none 2.05 / 2.78 / 2.37; cbqr_lasso
1.66 / 1.67 / 1.61; cbqr_alasso 1.42 / 1.59 / 2.62. The
last is the one the V5 correction lives in.
The test is shown to have power rather than merely to pass: a prior shape perturbed from 4 to 5 reads 19.6 to 27.4, a quantile off by 0.05 reads 24.7, against controls of 1.8 to 2.9.
The binary penalised kernels are still not
covered. Their sweeps carry lambda^2, s,
omega and delta between sweeps with no
argument to initialise them, so a run cannot continue a chain.
cbqr takes sigma_init, delta_init
and omega_init/tau_init as prior components,
which is why the continuous penalised kernels can be tested and their
binary twins cannot.
Both test files are skip_on_cran().
benchmarks/geweke_joint.R and
benchmarks/geweke_joint_continuous.R are the full-power
versions.
The release decision is taken: continuous and binary stay in
one package rather than splitting into siblings,
because they share prior(), the hierarchies, the derivation
and the kernels’ structure. DESCRIPTION’s
Title drops “Binary” and its Description now
covers both response types; ?bbqr-package gains a Two
response families section; the README opens with the two entry
points and gains a continuous section; and the vignette gains When
the response is observed. Four of the package’s eight exported
fitters used to appear nowhere outside NEWS.md.
An audit of every help page against that change found sixteen
statements that were wrong rather than merely thin, and they are
corrected here. Among them: plot.cbqr() had no
alue; the continuous kernels were described as having “no
latent draw” when they still draw z and s
every sweep (it is the latent response that goes);
prior() was said to serve both families when
cbqr() rejects it; and the continuous adaptive lasso at its
default q = 1 reproduces Alhamzawi, Yu and Benoit (2012),
not the 2023 thesis. The four cbqr pages gained examples,
references and cross-links, so R CMD check now exercises
the continuous code it previously skipped.
prior() defaults to model = "v5" for
every penalty. The published adaptive-lasso and lasso hierarchies
(model = "v3") have no posterior: the log-uniform
hyperprior on the global shrinkage scale (omega, or
tau_h for the lasso) is not integrable along the ray where
the scale, the lambda_j^2 and the latent s_j
vanish together, for every prior on delta and at either
anchor. "v5" replaces every improper component with a
proper one – Gamma(2, 2) on the global scale and, for the
lasso, on delta; a tau-calibrated normal prior
on the intercept – so the posterior is proper with no condition on the
design, on k, or on the anchor. "v3" remains
available for reproducing earlier results and is labelled
IMPROPER wherever it is printed.
The default anchor is "sigma1" for every penalty
(previously "free" for the adaptive lasso and the lasso).
Holding sigma at 1 closes the scale ray the likelihood
never identifies, gives the coefficients a stated scale, keeps the
credible intervals from inheriting the width of an unidentified
direction, and is the convention under which the "v5"
intercept prior is exact. "free" – sigma
sampled from its full conditional, as the derivations also allow – is
one argument away and is a derived posterior under every
model. bbqr_alasso() and
bbqr_lasso() changed accordingly.
A "v5" lasso prior now requires
beta_delta > 0, as the adaptive lasso already did: a
flat prior on delta is admissible only under
"v3", where it is the published specification.
print() and summary() report which
hierarchy a fit sampled and flag a chain that does not target a derived
posterior (a projection anchor combined with "v4" or
"v5"). Objects saved by earlier builds carry no
model and print no hierarchy line.prior() documents model for all three
penalties and the lasso’s alpha_tau/beta_tau;
bbqr_alasso() documents kappa_lam;
bbqr() gains a Derivation version section and
documents the model, derived and
b0_prior components of the fit. The clip-site count is
eighteen, not fifteen.
The README and the vignette describe the two hierarchies and the new defaults.
The help pages now render their code spans, emphasis and cross-links (roxygen markdown was never switched on, so these appeared as literal backticks, asterisks and brackets), and the Kozumi and Kobayashi (2011) DOI is corrected.
Experiment_v4.R through Experiment_v9.R, the
real-data and churn benchmarks, and diagnose_hyperparams.R)
now pass model = "v3" explicitly, so re-running them
reproduces what they reported rather than picking up the new
default.