Released on CRAN.
@sev) of MGWRSAR()
(GWR, mixed GWR, SAR variants) and of TDS_MGWR() without
get_AIC now use the variance of the local smoother with a
local residual variance. They ignored the effective local sample size of
the normalised kernel weights, and the fixed part of the residual in
mixed models, so that they were too small (2.8 to 4.5 times on a
2,667-observation tds_mgtwr fit). t-values, p-values and
significance maps change. Affected: all previous versions. Coefficients,
fitted values, hat matrices, AICc and the standard errors of
TDS_MGWR(get_AIC = TRUE) are unchanged.predict(method_pred = "TP") now uses the control of the
training call (kernel type, adaptive bandwidth, W), and for
a mixed model (Model = "MGWR") keeps the fixed coefficients
and re-estimates the varying ones on the partial residual. Affected:
1.3.1 to 1.4.3, where the defaults were used instead, and every earlier
version for mixed models.Not released on CRAN.
TDS_MGWR() returns the terminal sweep of the
backfitting, as MGWR backfitting does, instead of the sweep with the
smallest in-sample RMSE (1.3.1 to 1.4.2).
control_tds$retained_state = "best" restores the former
behaviour.TDS_MGWR(): after the search, the coefficients are
refitted at fixed bandwidths until convergence
(control_tds$polish, default 3 sweeps; 0 disables it).TDS_MGWR(): the safeguard against saturated candidate
fits is a unitless ratio of residual variances
(control_tds$saturation_ratio, default
exp(-2)), with counts in the new slot
saturation. Up to 1.4.2 it depended on the unit of the
response and on get_AIC.TDS_MGWR(): the importance-driven update order uses the
standard deviation of beta_k x_k and no longer depends on
the units or the level of the covariates (up to 1.4.2 it did). The path
of the descent may change.predict() for tds_mgwr /
tds_mgtwr: a target with very small prediction weights no
longer gets a non-finite prediction, and a target without any positive
weight takes the coefficient of its nearest training point, with a
warning.control_tds:
ordering = "random", besides "cyclic" and the
default "Importance"; importance_order
(default "decreasing"); update_last (default
NULL, no covariate forced last).Not released on CRAN.
TDS_MGWR() with pinned bandwidths, an adaptive kernel
and NN < n: a coefficient pinned at the global value is
fitted as a global component, as predict() assumes.
Affected: 1.4.1.TDS_MGWR(control = list(SE = TRUE)): @sev
and @edf_k are those of the returned sweep, not of the last
one. Affected: 1.4.1 (relative errors up to about 18%).search_bandwidths() no longer discards the true minimum
of a grid round, which could return a worse local minimum. Affected:
1.4.1. The AICc of a saturated binomial GWR is now
Inf.Released on CRAN.
gtwr_HWB2010(): the GTWR of Huang, Wu and Barry
(2010) with its (h_ST, tau) parameterisation, and
converters bw_hwb2gdt(), bw_gdt2hwb() and
as_gtwr().summary() gains fdr_method = "spatial_BY",
a Benjamini-Yekutieli correction calibrated on the bandwidth.control_tds levers: per-coefficient and per-axis
pinned bandwidths (H, Ht, now working for
Type = 'GDT'), a grid in distances (V_dist), a
floor of the spatial grid for panel data (panel_floor) and
TRUEBETA for simulations.TDS_MGWR() is about 4 to 5 times faster than in 1.3.2 and
search_bandwidths() about 1.5 to 2 times, with unchanged
results.options(mgwrsar.matprod = )).Inf for a saturated fit, which previous
versions rewarded with a penalty of the wrong sign.TDS_MGWR(): after a sweep that raises the RMSE the
descent continues from that sweep; 1.3.2 restarted from an inconsistent
state that froze the descent.MGWRSAR(), search_bandwidths(),
TDS_MGWR() and the other functions that draw internally.
Affected: previous versions.search_bandwidths() no longer closes every connection
of the session, and its golden-section refinement can no longer loop
forever.W or fixed_vars).fitted() is exported for mgwrsar objects;
vignettes and citation() render correctly; new adversarial
test campaign.Released on CRAN.
tds_mgtwr model for Multiscale GTWR with
additive or multiplicative (cyclic or acyclic) spatio-temporal
kernels.TDS_MGWR().NN is now capped to n in both
MGWRSAR() and TDS_MGWR() to avoid invalid k-NN
requests.TDS_MGWR(), control_tds$nns is now
capped to min(round(n/8), round(maxit/2)); if a
user-provided value is too large, it is truncated with a warning.Released on CRAN.
Not released on CRAN.
Released on CRAN. * Introducing top-down scale/multiscale GWR (tds_mgwr), adaptive top-down scale/multiscale GWR (atds_mgwr) and regular multiscale GWR (multiscale_gwr) * Introducing generalized GWR for binomial (bionomial and quasibinomial families). * Introducing GAM/GWR with gradient descent boosting. * Improving kernels for spatio-temporal GWR to introduce seasonnality (GDT) * Removing unused experimental General kernel Product functions (GDX,GDC) * Correcting a bug for MGWRSAR_x_x_x models with spatial autocorrelation when SE=TRUE. * Correcting a bug for GWR model with a single explanatory variable when doMC=TRUE. * Correcting a bug for formula without explicit names of variable (like “~.”). * Correcting a bug in GWR with parallel computation (split error in gwr_beta)
Released on CRAN. * Removing dependency to qlcMatrix * Introducing experimental multiscale GWR Model * Introducing experimental GWR with glm family * Introducing experimental GWR with glmboost Model * Adding the ability to calculate the trace of S without calculating the standard deviation matrix (control$get_ts parameter) * bandwidths_mgwrsar subroutines improved * AICc criteria added for bandwidth search * Removing the remove_local_outlier method * Rename all variables “coord” to “coords” * Rename the coordinates in the data example to c(‘x’,‘y’)
Released on CRAN. * Improved kernel_matW function by reintroducing possibility of island with no neighbours for non adaptie kernel * Introduced an experimental version of a backfitting algorithm for the estimation of the Multiscale GWR with a selection of the bandiwth of each variable by cross validation (LOOCV). The predict_mgwrsar function allows to make predictions from a model of the ‘multiscale_gwr’ class.
Not released on CRAN. * Fixed a bug for plot_mgwrsar * Fixed a bug for computing SE and edf when system is computationally singular * Fixed a bug for the function plot_mgwrsar coming from the way of naming the columns of the coordinates matrix returned in the MGWRSAR function. * Added a warning for cases where isgcv=TRUE and SE=TRUE when calling the MGWRSAR function. * Improvement of plot_effects function to plot also effect of spatially non-varying parameters * Improvement of plot_mgwrsar function to allow control of leaflet tile and number of quantile in legend. * WARNING: function to predict with target points using TP options have to be checked for model with spatial autocorrelation.
Not released on CRAN.
Not released on CRAN.
All models of mgwrsar package based on local linear regression can now be estimated using a target points set. Several functions that allows to choose an optimal set of target points to obtain a faster approximation of GWR coefficients has been added.
Predictions on new data can now be done using the jacknife estimation method instead of spatial extrapolation of local coefficients from a preliminarily estimated model. Only the optimal value of the bandwidth modeled from the initial data is then used. In the function ‘predict_mgwrsar’, if the parameter method_pred = ‘TP’ (default), the prediction is done by recalculating a MGWRSAR model with the new data as target points keeping the bandwidth at the optimal value chosen with the training data, otherwise if method_pred= (‘tWtp_model’, ‘model’, ‘shepard’) then a matrix is used for the spatial extrapolation of the estimated coefficients, and prediction are done using these extrapolated coefficients (as in the previous version of mgwrsar package).
‘KNN’ function is deprecated and replaced by ‘kernel_matW’ function that allows to build spatial weight matrix and interaction matrix based on General Kernel Product. In kernel_matW function it’s possible to specify the maximum number of neighbors to consider in gaussian kernel (rough gaussian kernel) to increase speed and sparsity of weights matrix.
Fast computation of local OLS coefficients in the previous version of mgwrsar package (0.1) uses non pivotal computation that may provides undesirable results in presence of strong colinearity. The RCCP ‘fastlmLLT_C’ function has been replaced by R native lm.fit function in this realease.
A new ploting function has been added: plot_effect is a function that plots the effect of a variable X_k with spatially varying coefficient, i.e X_k * Beta_k(u_i,v_i) for comparing the magnitude of effects of between variables
Released on CRAN. * First release.