Select GWR predictors by forward selection on AICc
Source:R/model-selection-gwr.R
gwr_model_selection.RdGrows a GWR model one predictor at a time and scores every intermediate
model with a corrected Akaike information criterion, the sweep
GWmodel::gwr.model.selection() defines, and returns a ranked table
with each model's distance from the best. The model without predictors is
scored too. Every model is fitted by the package's own local solve (see
"The local solve" in fit_gwr_model), so a candidate
that is constant inside some windows, a regional indicator say, is scored
on what those windows do estimate and neither stops the sweep nor scores
as rounding error.
Usage
gwr_model_selection(
data_sf,
response_var,
candidate_vars,
bandwidth = NULL,
adaptive = TRUE,
kernel = c("bisquare", "gaussian", "tricube", "boxcar", "exponential"),
bw_approach = c("AICc", "CV"),
max_models = 200L,
dmat_max_n = 2000L,
quiet = TRUE,
.engine = .gwr_ms_engine
)
# S3 method for class 'gwr_model_selection'
print(x, n = 10L, ...)Arguments
- data_sf
An
sfobject with response, predictors and geometry.- response_var
Response column name.
- candidate_vars
Character vector naming at least two numeric predictors to choose among. Factor, character and logical candidates are refused: a factor is several model-matrix columns while this sweep counts it as one variable, so the criteria would not be comparable, and
fit_gwr_model(), the documented next step, takes only numerics. Encode them as numeric indicators first.- bandwidth
Bandwidth held fixed across all candidate models. If
NULL(default) it is searched for on the model containing every candidate, by the grid-then-refine search offit_gwr_model()(see its Bandwidth selection); what the search scored is returned inbandwidth_search. Integer neighbour count whenadaptive = TRUE; otherwise a distance in the units of the projected CRS the sweep runs in, whichprep_model_data()may have chosen for you. Geographic input is projected before the bandwidth is used, so a value in degrees would be read as metres; a fixed bandwidth below a ten-thousandth of the data's extent raises a warning saying so, as infit_gwr_model(). An adaptive count too small for the full model is raised, with a warning, to the number of candidates plus 3 for the bisquare and tricube kernels (which give the farthest neighbour in a window weight 0), plus 2 for the others. That floor is enough unless several neighbours tie at the kernel's edge (a regular grid), which leaves a window fewer weighted points; then use a larger bandwidth. An adaptive count below 1 or above R's largest integer is refused, as infit_gwr_model(). One above the number of observations is capped at it, with a warning.- adaptive
Logical; adaptive (nearest-neighbour) bandwidth. Default
TRUE.- kernel
Weighting kernel: one of
"bisquare"(default),"gaussian","tricube","boxcar"or"exponential", as infit_gwr_model().- bw_approach
Criterion for the bandwidth search on the full model:
"AICc"(default, the criterionfit_gwr_model()uses) or"CV", the leave-one-out sum of squared prediction errors.fit_gwr_model()has no CV option, so with"CV"pass the selected bandwidth to it explicitly. When observations share a location the search leaves out whole locations whichever is asked for; see Observations that share a location.- max_models
Refuse the call if the sweep would exceed this many model fits. Default 200, which admits up to 19 candidates.
- dmat_max_n
Not used. The sweep takes its distances window by window and builds no
nxndistance matrix, so a fixed bandwidth works at anyn. The argument is accepted so that calls written with it run.- quiet
Print no progress. Default
TRUE, unlikeselect_features_forward(). WithFALSEone line is written (withcat()) as each model is fitted, to watch a long sweep progress.- .engine
Internal; leave at the default.
- x
A
gwr_model_selectionobject.- n
Number of top-ranked models
print()shows. Default 10.- ...
Not used.
Value
An object of class gwr_model_selection, a list with:
bestCharacter vector of the selected predictors, the model with the lowest criterion;
character(0)when that is the model without predictors.tableRanked data.frame of every model evaluated, with columns
rank,n_vars,variables,criterionanddelta, the criterion minus the best model's. The model without predictors is the row<none>.criterionisNA, and the model is ranked last, where it could not be evaluated or where AICc is undefined because the model's effective number of parameters \(tr(S)\) is not below \(n - 2\) (with a warning).within_2,smallest_within_2The rows of
tablewhosedeltais at most 2, and the predictors of the smallest model among them.criterion"AICc".criterion_by_name,criterion_column,criterion_verifiedWhere the criterion was read from the diagnostic table:
TRUE, 3 andTRUEwith the built-in engine.response_var,candidate_varsThe response and the full candidate set the sweep ran over, both echoed by
print().bandwidth,bandwidth_source,adaptive,kernelThe smoothing held fixed across the sweep, and where it came from:
"supplied", or the criterion's name followed by"search on the full model", the criterion being"leave-one-location-out CV"when observations share a location.bandwidth_searchWhat the search scored, as
fit_gwr_modelkeeps it;NULLwhen the bandwidth was supplied.n_obs,n_modelsThe number of observations, and the number of rows of
table.used_dmatAlways
FALSE.rawThe forward sweep in the layout
GWmodel::gwr.model.selection()returns: the two-element list of its model list and its diagnostic table, without the model that has no predictors.
Scope of the criterion
AICc penalises the effective number of parameters but is computed on the
data the model was fitted to, and under spatial autocorrelation an
in-sample criterion is optimistic in a way that a spatially blocked
estimate is not. Use this function as a
fast screen and confirm with select_features_forward(),
which runs the same forward search on a spatially blocked cross-validated
score at a higher cost.
Every model is fitted at one bandwidth, as Lu et al. (2014) define the
method: comparing criteria across models needs the smoothing held fixed.
The bandwidth chosen for the full model is not optimal for smaller ones,
so refit the selected model with bandwidth = NULL once the set is
settled.
The model without predictors, and the gaps
The forward sweep starts from one predictor, so by itself it always names
at least one, whether or not any candidate helps. The model without
predictors, a local mean at the sweep's bandwidth, is therefore scored as
well and ranked with the others as <none>. When it ranks first,
best is character(0): no candidate improves on it.
The delta column of table is each model's criterion minus
the best one's. Differences of AICc below about 2 do not separate two
models (Burnham and Anderson 2002), so within_2 lists the models
that close to the best and smallest_within_2 names the one among
them with the fewest predictors, which is the one to prefer when several
are that close. best is still the model with the lowest criterion.
Observations that share a location
When rows of data_sf have the same coordinates, the bandwidth
search leaves out whole locations, for either bw_approach (see
fit_gwr_model, Observations that share a
location); bandwidth_source then reads
"leave-one-location-out CV search on the full model". The models
are still ranked by AICc, which counts the rows at one location as
independent, so the gaps overstate what a predictor adds when a location
has an effect of its own; a warning of class
"spatialkit_gwr_repeated_locations" says so with the counts. A
model whose fit reproduces the mean of the observations at each location
(every such model scores the same, whatever its predictors) has criterion
NA and is listed last, as is one with a term that no local window
can estimate; a warning names them.
Cost
The sweep fits p * (p + 1) / 2 GWR models for p candidates
(55 at p = 10, 210 at p = 20), and the one without predictors, each over
all n locations.
max_models stops the call before it runs for hours.
Using $raw with GWmodel directly
raw has the layout of GWmodel's own list(model.list, GWR.df),
so its two elements have to be unpacked before GWmodel's helpers will take
them: GWmodel::gwr.model.view() takes
(DeVar, InDeVars, model.list), so the call is
GWmodel::gwr.model.view(sel$response_var, sel$candidate_vars, sel$raw[[1]])Pass sel$raw[[1]], not sel$raw. The diagnostic table is
sel$raw[[2]], a numeric matrix whose columns are
bandwidth, AIC, AICc, RSS in that order (and
are named so); the criterion column of $table is its third
column.
Printing
print() shows the response, the candidate predictors, and the
top-ranked models with their criterion values and, in the delta
column, how far each is from the best. Below the table it lists the
models within 2 of the best, which a difference in AICc does not separate
(Burnham and Anderson 2002), and names the smallest of them. When several
models are that close, treat the choice of predictors as unsettled; when
the model without predictors (<none>) is among them, no candidate
has been shown to help. print() returns x invisibly.
References
Burnham, K. P. and Anderson, D. R. (2002). Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach, 2nd edition. Springer.
Lu, B., Harris, P., Charlton, M. and Brunsdon, C. (2014). The GWmodel R package: further topics for exploring spatial heterogeneity using geographically weighted models. Geo-spatial Information Science 17(2), 85–101. doi:10.1080/10095020.2014.917453
See also
select_features_forward() for the blocked
cross-validated counterpart, plot.gwr_model_selection() to
draw the models, and fit_gwr_model() to fit the selected
model.
Other predictor selection:
plot.feature_selection(),
plot.gwr_model_selection(),
select_features_forward()
Examples
library(sf)
set.seed(1)
n <- 80
dat <- st_as_sf(
data.frame(x = 5e5 + runif(n, 0, 1000), y = 5e6 + runif(n, 0, 1000),
a = rnorm(n), b = rnorm(n), noise = rnorm(n)),
coords = c("x", "y"), crs = 32632
)
# The slope on a rises from 1 in the west to 3 in the east.
east <- (st_coordinates(dat)[, 1] - 5e5) / 1000
dat$z <- (1 + 2 * east) * dat$a - dat$b + rnorm(n, 0, 0.3)
sel <- gwr_model_selection(dat, "z", c("a", "b", "noise"), bandwidth = 30)
sel # every model tried, by AICc, with the gaps
#> Geographically weighted regression - forward model selection
#>
#> response : z
#> candidates : 3 (a, b, noise)
#> observations: 80
#> models : 7
#> bandwidth : 30 (adaptive; supplied)
#> kernel : bisquare
#> criterion : AICc
#>
#> rank n_vars variables criterion delta
#> 1 2 a + b 70.55026 0.00000
#> 2 3 a + b + noise 81.27942 10.72916
#> 3 1 a 262.02384 191.47358
#> 4 2 a + noise 268.26330 197.71304
#> 5 1 b 371.33475 300.78449
#> 6 0 <none> 374.24430 303.69404
#> 7 1 noise 384.67142 314.12116
#>
#> Selected: z ~ a + b
#>
#> No other model is within 2 of the best; the runner-up is 10.73 away.
#>
#> The criterion is in-sample and every model shares one bandwidth.
#> Confirm with select_features_forward() before relying on this.
sel$best # the winning predictor set
#> [1] "a" "b"
# Refit with the bandwidth chosen afresh for the selected set.
fit_gwr_model(dat, "z", sel$best)
#> <GWR> spatial model fit
#> Formula : z ~ a + b
#> n : 80
#> CRS : EPSG:32632
#> Bandwidth: 35 neighbours (adaptive, bisquare kernel)
#> AICc : 70.30
#> Eff. parameters: 22.0 of 80 observations