Skip to contents

Changelog

Source: NEWS.md

frmtmb.spatial 0.3.0

Needs frmtmb 0.76.0, which takes frmtmb_register_autocor(jacobian =) and accepts a bound at an end of a residual correlation parameter’s interval.

New features

  • sar() gives its whitening map’s Jacobian, (I - rho W) / sigma, as a sparse matrix. So frmtmb.sample’s log_lik() returns brms 2.23.0’s leave-one-out conditional density for both types, gaussian and student. It equals brms’s log_lik() at the same parameters to 6.3e-16 relative. One draw on a 40 x 40 lattice (1600 rows) takes 1.9 ms; brms’s dense algorithm takes 4.1 s for the same draw.

Bug fixes

  • The default_prior() table of a sar() model passes back as prior =. Its lagsar or errorsar row lists the ends of the interval as lb and ub, as brms’s does. frm() refused them as outside the interval. The fix is in frmtmb, which now accepts a bound at an end of the interval as no bound. Under frm_sample(), a bounds-only prior on lagsar keeps the flat default density; see frmtmb.sample’s NEWS.

frmtmb.spatial 0.2.0

Needs frmtmb 0.75.0, which exports frmtmb_register_autocor(), the seam through which this package registers sar().

New features

  • sar(M, type = ), brms’s spatial simultaneous autoregressive term, in its lag and error forms. As in brms 2.23.0, it is for gaussian() and student() with a constant sigma and nu. It is registered through frmtmb::frmtmb_register_autocor(). brms leaves out -N/2 log(2 pi) (-N/2 log(pi) for student). Less that constant, the log-likelihood equals brms’s Stan program to 1.4e-14 (gaussian, at frmtmb’s estimates) and 4e-15 (student, at set parameter values with nu = 5). brms fits of 64 rows agree with the maximum likelihood fits to within 0.44 standard errors in every parameter. fitted() reports brms’s posterior_epred(), (I - rho W)^-1 mu for the lag form. summary() and confint_varcorr() report lagsar or errorsar, and set_prior() takes them as classes: a density on rho, carried to the internal parameter.

  • fitted(), predict() and simulate() take newdata with one row per row of M, read as the locations in their order, as brms reads it. When the fit dropped rows for missing values, newdata covers them too, and I - rho M is then over every row of M. A counterfactual such as x + 1 everywhere then gives the lag model’s spillover, (I - rho W)^-1 mu_new. That equals brms’s posterior_epred() on the same newdata at the same parameter values. Other newdata is refused, as brms refuses it without newdata2.

  • rho is estimated on (1 / min eigenvalue, 1 / max eigenvalue) of M, where I - rho M has a positive determinant. brms bounds it by (min eigenvalue, max eigenvalue). For a row-standardized M that cuts off negative values below the smallest eigenvalue. So frmtmb can estimate rho beyond brms’s bound, and the two are then different models. An example is a row-standardized 8 x 8 queen lattice (eigenvalues -0.5006 to 1) with data drawn at rho = -0.8. frmtmb’s estimate is -1.103, with a logLik 3.74 above the best inside brms’s interval. brms’s posterior median is -0.443, with a fifth of its draws within 0.02 of its bound. For a matrix that is not row-standardized, brms’s interval reaches past 1 / max eigenvalue, where its log-determinant is not a number and Stan rejects the proposal. So brms’s support there is frmtmb’s interval, at the cost of rejected proposals.

  • A matrix whose smallest or largest eigenvalue repeats for all rows but one, such as a complete graph, gives a warning. At that end of the interval I - rho M keeps one direction only. With an intercept the likelihood has no maximum there, so rho runs to the end. The model is not refused, because without an intercept it has its maximum.

  • A row-standardized M of an irregular graph, spdep’s default, has real eigenvalues that eigen() returns with a rounding imaginary part. brms 2.23.0 passes them to Stan as they come and its fit fails (“variable does not exist; … eigenMsar”); frmtmb takes the real part when the imaginary part is below 1e-8 of the largest modulus, and refuses a matrix with complex eigenvalues.

  • spdep’s nb and listw objects are read as brms reads them (nb2mat() row-standardized, listw2mat()) without needing spdep.

frmtmb.spatial 0.1.0

The first release. Load it with library(frmtmb.spatial), which attaches frmtmb as well. Needs frmtmb 0.74.0, which exports frmtmb_register_covstruct().

New features

  • New package. Two fields moved here from frmtmb. car(M, gr = g, type = ) is brms’s conditional autoregressive field in its four types ("escar", "esicar", "icar", "bym2"). spde(fem, gr = node) is a Matern field over a finite-element mesh. Both are registered through frmtmb::frmtmb_register_covstruct(), and the package reads no frmtmb internal. Fits are unchanged. On the moved tests’ models, these are identical to frmtmb 0.73.0’s: the log-likelihood, every estimate and standard error, fitted() with and without an unseen location, ranef() conditional standard deviations, confint_varcorr() and simulate(). frmtmb.sample::frm_sample() gives the same draws from the same seed. frm_compat() gives car and spde the statuses frmtmb gave them.