V0.7.0 includes various adjustments to speed things up, and the DIC is now provided for model comparison (in addition to WAIC).
New features:
Changes in the background that should improve the user experience:
centerx
argument).spdep
’s creation of neighbors objects
(https://github.com/ConnorDonegan/geostan/issues/19). This will speed up
the shape2mat
function in some cases.Other changes:
prep_car_data
and prep_sar_data
have
changed somewhat, but the user workflow is the same.geostan was removed from CRAN for a moment due to an issue with the StanHeaders R package. This should be resolved now. This release puts geostan back on CRAN with only minimal internal changes to geostan.
Two updates:
geostan::predict
.There are three updates related to spatial connectivity matrices:
browseVignettes('geostan')
), written for new users.geostan::edges
can now
return a simple features object; this can be used to visualize (map) the
graph structure of the spatial connectivity matrix. There is an example
in the new vignette.geostan::shape2mat
: an option for k-nearest
neighbors has been added, the queen
argument is being
replaced by method
, and the function now prints a summary
of the matrix to the console (using the new geostasn::n_nbs
function)There was one change to the geostan::predict
method:
Updates:
prep_icar_data
has been fixedThe model fitting functions (stan_glm
,
stan_car
, etc.) now allow for missing data in the outcome
variable. This is explained in the geostan::stan_glm
documentation, next to the discussion of handling censored observations.
When missing observations are present, there will (only) be a warning
issued. This functionality is available for any GLM
(stan_glm
), any ESF model (stan_esf
), and any
model for count data (Poisson and binomial models including CAR and SAR
models). The only models for which this functionality is not currently
available are CAR and SAR models that are being been fit to continuous
outcome variables.
The prep_icar_data
function, which is used inside
stan_icar
, did not have the expected behavior in all cases
- this has been fixed thanks to this pull
request.
The package home page now has instructions for installing from github
using devtools::install_github
https://connordonegan.github.io/geostan/
Minor updates to the vignettees and documentation, also re-compiled geostan models using the latest StanHeaders (fixing an error on CRAN).
The gamma
function (which is available to help set prior
distributions) has been renamed to geostan::gamma2
to avoid
conflict with base::gamma
.
Some code for geostan::stan_car
was cleaned up to avoid
sending duplicate variables to the Stan model when a spatial ME
(measurement error) model was used:
https://github.com/ConnorDonegan/geostan/issues/17. This should not
change any functionality and there is no reason to suspect that results
were ever impacted by the duplicate variables.
This release was built using rstan 2.26.23, which incorporates Stan’s new syntax for declaring arrays. Some models seems to run a little bit faster, but otherwise there are no changes that users should notice.
The warnings issued about the sp package can be ignored; these are due to geostan’s dependence on spdep, which imports sp but does not use any of the deprecated functions.
A new vignette shows how to implement some of geostan’s spatial models directly in Stan, using the custom Stan functions that make the CAR and SAR models sample quickly, and using some geostan functions that make the data cleaning part easy.
This release fixes some issues that were introduced with the
slim
and drop
arguments (in v0.5.0).
The package now provides some support for spatial regression with raster data, including for layers with hundreds of thousands of observations (possibly more, depending on one’s computational resources). Two new additions make this possible.
slim = TRUE
The model fitting functions
(stan_glm
, stan_car
, stan_sar
,
stan_esf
, stan_icar
) now provide the option to
trim down the parameters for which MCMC samples are collected. For large
N and/or many N-length vectors of parameters, this option can speed up
sampling considerably and reduce memory usage. The new drop
argument provides users control over which parameter vectors will be
ignored. This functionality may be helpful for any number of purposes,
including modeling large data sets, measurement error models, and Monte
Carlo studies.prep_sar_data2
and prep_car_data2
These
two functions can quickly prepare required data for SAR and CAR models
when using raster layers (observations on a regularly spaced grid). The
standard and more generally applicable functions
prep_car_data
and prep_sar_data
are limited in
terms of the size of spatial weights matrices they can handle.These new functions are discussed in a new vignette titled “Raster regression.”
The PDF documentation has been improved—previously, multi-line equations were not rendered properly. Now they render correctly, and a mistake in the description of Binomial CAR models has been corrected.
sp_diag
) will now take
a spatial connectivity matrix from the fitted model object provided by
the user. This way the matrix will be the same one that was used to fit
the model. (All of the model fitting functions have been updated to
support this functionality.)residuals
, fitted
, spatial
,
etc.) were previously packed into one page. Now, the documentation is
spread over a few pages and the methods are grouped together in a more
reasonable fashion.The simultaneously-specified spatial autoregressive (SAR)
model—referred to as the spatial error model (SEM) in the spatial
econometrics literature—has been implemented. The SAR model can be
applied directly to continuous data (as the likelihood function) or it
can be used as prior model for spatially autocorrelated parameters.
Details are provided on the documentation page for the
stan_sar
function.
Previously, when getting fitted values from an auto-normal model
(i.e., the CAR model with family = auto_gaussian()
) the
fitted values did not include the implicit spatial trend. Now, the
fitted.geostan_fit
method will return the fitted values
with the implicit spatial trend; this is consistent with the behavior of
residuals.geostan_fit
, which has an option to
detrend
the residuals. This applies to the SAR and CAR
auto-normal specifications. For details, see the documentation pages for
stan_car
and stan_sar
.
The documentation for the models (stan_glm
,
stan_car
, stan_esf
, stan_icar
,
stan_sar
) now uses Latex to typeset the model
equations.
bridge_sampler(geostan_fit$stanfit)
). By default, geostan
only collects MCMC samples for parameters that are expected to be of
some interest for users. To become compatible with bridgesampling, the
keep_all
argument was added to all of the model fitting
functions. For important background and details see the bridgesampling
package documentation and vignettes on CRAN.lisa
function would automatically
center and scale the variate before computing local Moran’s I. Now, the
variate will be centered and scaled by default but the user has the
option to turn the scaling off (so the variate will be centered, but not
divided by its standard deviation). This function also row-standardized
the spatial weights matrix automatically, but there was no reason why.
That’s not done anymore.The distance-based CAR models that are prepared by the
prep_car_data
function have changed slightly. The
conditional variances were previously a function of the sum of
neighboring inverse distances (in keeping with the specification of the
connectivity matrix); this can lead to very skewed frequency
distributions of the conditional variances. Now, the conditional
variances are equal to the inverse of the number of neighboring sites.
This is in keeping with the more common CAR model specifications.
geostan now supports Poisson models with censored count data, a
common problem in public health research where small area disease and
mortality counts are censored below a threshold value. Model for
censored outcome data can now be implemented using the
censor_point
argument found in all of the model fitting
functions (stan_glm, stan_car, stan_esf, stan_icar).
The measurement error models have been updated in three important respects:
?prep_me_data
.?prep_me_data
for usage.stan_car
, ME models
automatically employed the CAR model as a prior for the modeled
covariates. That has changed, so that the default behavior for the ME
models is the same across all stan_*
models (CAR, GLM, ESF,
ICAR).The second change listed above is particularly useful for variables
that are highly skewed, such as the poverty rate. To determine whether a
transformation should be considered, it can be helpful to evaluate
results of the ME model (with the untransformed covariate) using the
me_diag
function. The logit transform is done on the
‘latent’ (modeled) variable, not the raw covariate. This transformation
cannot be applied to the raw data by the user because that would require
the standard errors of covariate estimates (e.g., ACS standard errors)
to be adjusted for the transformation.
A predict
method has been introduced for fitted geostan
models; this is designed for calculating marginal effects. Fitted values
of the model are still returned using fitted
and the
posterior predictive distribution is still accessible via
posterior_predict
.
The centerx
argument has been updated to handle
measurement error models for covariates. The centering now happens
inside the Stan model so that the means of the modeled covariates
(latent variables) are used instead of the raw data mean.
geostan’s first release.