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.