Residual Analysis of spatial point process models using Bayesian methods

# Load function definitions
source(system.file(
  file.path("misc", "2d_lgcp_residuals_functions_sf.R"),
  package = "inlabru"
))

Introduction

Point processes are useful for spatial data analysis and have wide applications in ecology, epidemiology, seismology, computational neuroscience, etc. Residual analysis is an effective assessment method for spatial point processes, which commonly takes a frequentist approach.

In this vignette, we will calculate residuals of some of the models from https://inlabru-org.github.io/inlabru/articles/2d_lgcp_covars.html using a Bayesian approach to the residual methods in “Residual Analysis for spatial point processes” (Baddeley et al).

Theory of Residuals for Spatial point process models

Consider a spatial point pattern $\mathbf{x} = \{x_1, \dots, x_n \}$ of $n$ points in a bounded region $W$ of $\mathbb{R}^2$ . For a model of a spatial point process $\mathbf{X}$ with probability density $f_{\theta}$ and parameter $\theta$ , the innovation process of the model is defined by $\begin{equation}\label{eq:innovations} I_{\theta}(B) = n(\mathbf{X} \cap B) - \int_B \lambda_{\theta} (u) \, du \end{equation}$ Here, $\int_B \lambda_{\theta}(u) \, du$ is the expected number of points of the fitted model in the bounded subset of $B$ . Innovations satisfy $\mathbb{E}_{\theta} \left[ I_{\theta}(B)\right] = 0$ .

Given data $\mathbf{x}$ of the model and that the parameter estimate is $\hat{\theta}$ , the raw residuals are given by $\begin{equation}\label{eq:raw residuals} R_{\hat{\theta}}(B) = n(\mathbf{x} \cap B) - \int_B \hat{\lambda}(u) \,du \end{equation}$

Increments of the innovation process $I$ and raw residuals of the Poisson processes $R$ are analogous to errors and raw residuals of linear models respectively.

Types of Residuals

We can scale the raw residual by scaling the increments of $R_{\hat{\theta}}$ . \ The $h$ -weighted innovations are given by $\begin{equation}\label{eq:h innovations} I(B, h, \lambda) = \sum_{x_i \in \mathbf{X} \cap B} h(x_i) - \int_B h(u) \lambda(u) \,du \end{equation}$ This leads to the $h$ -weighted residuals $\begin{equation} \label{eq:h residuals} R(B, \hat{h}, \hat{\theta}) = I(B, \hat{h}, \hat{\lambda}) = \sum_{x_i \in \mathbf{x} \cap B} h(x_i) - \int_B \hat{h}(u) \hat{\lambda}(u)\,du \end{equation}$ Since the innovation has mean 0, a true model would yield $\mathbb{E}\left[R(B, \hat{h}, \hat{\theta})\right] \approx 0.$ Changing the choice of the weight function $h$ , yields different types of residuals.

Scaling Residuals

By taking

h(u) = \mathbf{1} \{ u \in B\}

, we get the raw residuals:

\begin{aligned} R(B, \hat{h}, \hat{\theta}) & = I(B, \hat{h}, \hat{\lambda}) = \sum_{x_i \in \mathbf{x} \cap B}h(x_i) - \int_B \hat{h}(u)\hat{\lambda}(u)\,du\\ & = n(\mathbf{x} \cap B) - \int_B \hat{\lambda}(u) \,du \end{aligned}

Inverse Residuals

By taking

h(u) = \frac{1}{\lambda (u)}

, we may encounter the case where

\lambda(u) = 0

. The GNZ-formula

\begin{equation} \mathbb{E}\left[ \sum_{x_i \in \mathbf{X}} h(x_i)\right] = \mathbb{E}\left[ \int_W h(u) \lambda(u) \,du \right], \end{equation}

would still hold for the first terms of the

h

-weighted residuals to exist. If we define

h(u) \lambda(u) = 0

for all

u

such that

\lambda(u) = 0

, then the second term of

h

-weighted residuals will also exist and the inverse residual is given by

\begin{aligned} R\left(B, \frac{1}{\hat{\lambda}}, \hat{\theta}\right) & = \sum_{x_i \in \mathbf{x} \cap B} \hat{h}(x_i) - \int_B \hat{h}(u) \hat{\lambda}(u) \,du \\ & = \sum_{x_i \in \mathbf{x} \cap B} \frac{1}{\hat{\lambda}(x_i)} - \int_B \mathbf{1} \{ x_i \in \mathbf{x} \} \,du \end{aligned}

Pearson Residuals

By using

h(u) = \frac{1}{\sqrt{\lambda(u)}}

, the Pearson residual is given by

\begin{aligned} R\left( B, \frac{1}{\sqrt{\hat{\lambda}}}, \hat{\theta} \right) & = \sum_{x_i \in \mathbf{x} \cap B} \hat{h}(x_i) - \int_B \hat{h}(u) \hat{\lambda}(u) \,du \\ & = \sum_{x_i \in \mathbf{x} \cap B} \frac{1}{\sqrt{\hat{\lambda}(x_i)}} - \int_B \sqrt{\hat{\lambda}(u)} \,du \end{aligned}

Here, zero values of $\hat{\lambda}(u)$ do not cause any issues with calculating residuals as we set $\hat{h}(u) \hat{\lambda}(u) = \sqrt{\hat{\lambda}(u)}$ for all $u$ .

So the three types of residuals are: $\begin{eqnarray} \text{Scaled:} \qquad & R(B, \hat{h}, \hat{\theta}) & = n(\mathbf{x} \cap B) - \int_B \hat{\lambda}(u) \,du \\ \text{Inverse:} \qquad & R\left(B, \frac{1}{\hat{\lambda}}, \hat{\theta}\right) & = \sum_{x_i \in \mathbf{x} \cap B} \frac{1}{\hat{\lambda}(x_i)} - \int_B \mathbf{1} \{ x_i \in \mathbf{x} \} \,du \\ \text{Pearson:} \qquad & R\left( B, \frac{1}{\sqrt{\hat{\lambda}}},\hat{\theta} \right) & = \sum_{x_i \in \mathbf{x} \cap B} \frac{1}{\sqrt{\hat{\lambda}(x_i)}} - \int_B \sqrt{\hat{\lambda}(u)} \,du \end{eqnarray}$

Note that $\hat{\lambda}(u)$ and $\hat{h}(u)$ are estimates of $\lambda (u)$ and $h(u)$ respectively

Motivation for choice of residuals

The scaling residuals are nothing but the raw residuals and hence depict the residual data as is. However, the Pearson residuals use a weighted function $h(u) = \frac{1}{\lambda(u)}$ so this helps to obtain normalised residuals since the denominator accounts for variance of the residuals and hence makes it reliable for any sample size.

The use of the inverse residual in this project is less clear. However, we do know that it is easier to compute this residual since we have the GNZ formula to estimate the value of $h(u) = for any value $\lambda(u)$ .

For notation’s sake in this discussion, let $h_{s}(u), h_{i}(u), h_p(u)$ indicate the weight function for scaling, inverse and Pearson residuals respectively. Then, $h_s(u) = \mathbf{1}\{u \in B\}$ , $h_i(u) = \frac{1}{\lambda(u)}$ , $h_p(u) = \frac{1}{\sqrt{\lambda(u)}}$ . We can also see that $h_s(u) = \left(h_i(u)\right)^0$ and $h_p(u) = \left(h_i(u)\right)^{1/2}$ , so the values of the Pearson residuals would lie somewhere between the other two residuals. So, in our case, the inverse residuals would not be very useful, but we do find that they are easier to compute by hand in terms of estimates.

Computation of Residuals

In this vignette, the residuals of some models from https://inlabru-org.github.io/inlabru/articles/2d_lgcp_covars.html are computed using a Bayesian approach of the residuals described above. This is done using the inlabru package. The models used are:

Model 1 (Vegetation Model): The model is defined such that the gorilla nests have vegetation type as a fixed covariate.
Model 2 (Elevation Model): The model is defined such that the gorilla nests have elevation as a continuous variable.
Model 3 (Intercept Model): The model is defined such that the gorilla nests have a constant effect.
Model 4 (SPDE Smooth type model): The model is defined such that the gorilla nests depend on an SPDE type smooth function.

The functions used to calculate the residuals of these models are given in the Code appendix at the end of this vignette.

Loading gorillas data

This vignette uses the gorillas dataset from the “inlabru” package in R for modelling spatial Poisson processes. Here, the ``inlabru” package uses Latent Gaussian Cox Processes to model the data using different model definitions.

# Load data
gorillas <- inlabru::gorillas_sf
nests <- gorillas$nests
mesh <- gorillas$mesh
boundary <- gorillas$boundary
gcov <- gorillas_sf_gcov()

Initially, the set $B$ is defines as $B=W$ , where $W$ is the object boundary. The function prepare_residual_calculations() is defined to compute a data frame of observations $x_i$ and sample points $u \in W$ to compute the residual, a matrix A_sum that helps to calculate the summation term of the residual and a matrix A_integrate that helps to calculate the integral term of the residual.

The function residual_df is defined to compute all three types of residuals for a given model and choice of $B$ given a set of observations.

# Define the subset B
B <- boundary

# Store the required matrices and data frames for residual computation
As <- prepare_residual_calculations(
  samplers = B, domain = mesh,
  observations = nests
)

Assessing 4 models with residuals

Vegetation Model

This model is taken from https://inlabru-org.github.io/inlabru/articles/2d_lgcp_covars.html . The model is defined such that the gorilla nests have vegetation type as a fixed covariate.

The code chunk below shows how the model is defined by lgcp() and also displays the model in the form of a plot. The residuals of the model are also computed where $B = W$ .

# Define the vegetation model
comp1 <- geometry ~
  vegetation(gcov$vegetation, model = "factor_contrast") + Intercept(1)

fit1 <- lgcp(comp1, nests,
  samplers = boundary,
  domain = list(geometry = mesh)
)

# Display the model
int1 <- predict(fit1,
  newdata = fm_pixels(mesh, mask = boundary, format = "sf"),
  ~ exp(vegetation + Intercept)
)
ggplot() +
  gg(int1, geom = "tile") +
  gg(boundary, alpha = 0, lwd = 2) +
  gg(nests, color = "DarkGreen")

# Calculate the residuals for the vegetation model
veg_res <- residual_df(
  fit1, As$df, expression(exp(vegetation + Intercept)),
  As$A_sum, As$A_integrate
)
knitr::kable(
  edit_df(veg_res, c("Type", "mean.mc_std_err", "sd.mc_std_err", "median"))
)

	mean	sd	q0.025	q0.5	q0.975
Scaling_Residuals	0.2523801	25.383651	-52.287394	2.2559526	52.870064
Inverse_Residuals	0.3347500	1.515273	-2.351127	0.3703534	3.045378
Pearson_Residuals	0.5319326	4.629636	-8.766868	0.6050649	10.250761

Note: The data frame produced by residual_df() originally contains the columns Type, mean.mc_std_err, sd.mc_std_err and median which have been removed in all the tables in the vignette to highlight only essential and non-recurring data. This editing of the data frames is done by the function edit_df.

Elevation model

This model is also taken from https://inlabru-org.github.io/inlabru/articles/2d_lgcp_covars.html . The model is defined such that the gorilla nests have elevation as a continuous variable.

The code chunk below shows how the model is defined by lgcp() and also displays the model in the form of a plot.

# Define the Elevation model
comp2 <- geometry ~ elev(elev, model = "linear") +
  mySmooth(geometry, model = matern) + Intercept(1)

fit2 <- lgcp(comp2, nests,
  samplers = boundary,
  domain = list(geometry = mesh)
)

# Display the model
int2 <- predict(
  fit2,
  fm_pixels(mesh, mask = boundary, format = "sf"),
  ~ exp(elev + mySmooth + Intercept)
)
ggplot() +
  gg(int2, geom = "tile") +
  gg(boundary, alpha = 0) +
  gg(nests, shape = "+")

The residuals of the model when $B = W$ are given below:

	mean	sd	q0.025	q0.5	q0.975
Scaling_Residuals	-27.207926	26.443445	-78.31955	-28.425138	15.271867
Inverse_Residuals	-8.636128	1.344703	-10.88233	-8.751341	-5.965917
Pearson_Residuals	-11.327663	3.149749	-17.86530	-10.915368	-5.701208

Intercept Model

This model is also taken from https://inlabru-org.github.io/inlabru/articles/2d_lgcp_covars.html . The model is defined such that the gorilla nests have a constant effect.

The code chunk below shows how the model is defined by bru_obs() and also displays the model in the form of a plot. The residuals of the model are also computed where $B = W$ .

# Define the Intercept model. Need to explicitly extend it to match the
# data size to use it in spatial predictions.
comp3 <- geometry ~ Intercept(rep(1, nrow(.data.)))
lik3 <- bru_obs(
  geometry ~ .,
  data = nests,
  samplers = boundary,
  domain = list(geometry = mesh),
  family = "cp"
)
fit3 <- bru(comp3, lik3)

# Display the model
int3 <- predict(
  fit3,
  fm_pixels(mesh, mask = boundary, format = "sf"),
  ~ exp(Intercept)
)
ggplot() +
  gg(int3, geom = "tile") +
  gg(boundary, alpha = 0) +
  gg(nests, shape = "+")

The residuals of the model when $B = W$ are given below:

	mean	sd	q0.025	q0.5	q0.975
Scaling_Residuals	-0.7315812	25.3197632	-51.044159	-0.7788689	43.350877
Inverse_Residuals	0.0074269	0.7729678	-1.452972	-0.0238888	1.428329
Pearson_Residuals	-0.0427818	4.4201763	-8.611936	-0.1364047	7.868833

Smooth Model

This model is also taken from https://inlabru-org.github.io/inlabru/articles/2d_lgcp_covars.html . The model is defined such that the gorilla nests depend on an SPDE type smooth function.

The code chunk below shows how the model is defined by lgcp() and also displays the model in the form of a plot. The residuals of the model are also computed where $B = W$ .

# Define the Smooth model
comp4 <- geometry ~ mySmooth(geometry, model = matern) +
  Intercept(1)
fit4 <- lgcp(comp4, nests,
  samplers = boundary,
  domain = list(geometry = mesh)
)

# Display the model
int4 <- predict(
  fit4,
  fm_pixels(mesh, mask = boundary, format = "sf"),
  ~ exp(mySmooth + Intercept)
)
ggplot() +
  gg(int4, geom = "tile") +
  gg(boundary, alpha = 0) +
  gg(nests, shape = "+")

The residuals of the model when $B = W$ are given below:

	mean	sd	q0.025	q0.5	q0.975
Scaling_Residuals	-28.82139	25.246737	-82.92909	-24.49749	11.733245
Inverse_Residuals	-8.84339	1.466921	-10.60504	-9.14138	-5.583917
Pearson_Residuals	-12.16622	3.073856	-18.85940	-12.01163	-6.648542

Comparing models

Firstly, from the plots of the four models, we see that there are visual similarities in what the Elevation and smooth models appear to be. So, we would expect that these two models show similar trends in residuals.

Redefining the set B

Consider defining a new type of set $B$ which consists of two subpolygons within the boundary. This is done by the partition() function which divides a polygon into grids based on number of rows and columns or by a desired resolution. The code chunk below demonstrates a use of the function for dividing the polygon and calculating the residuals for this choice of $B$ .

# Create a grid for B partitioned into two
B1 <- partition(samplers = boundary, nrows = 1, ncols = 2)
plot(B1, main = "Two partitions of B")

As1 <- prepare_residual_calculations(
  samplers = B1, domain = mesh,
  observations = nests
)
# Residuals for the vegetation model
veg_res2 <- residual_df(
  fit1, As1$df, expression(exp(vegetation + Intercept)),
  As1$A_sum, As1$A_integrate
)
knitr::kable(edit_df(veg_res2, c(
  "Type", "mean.mc_std_err",
  "sd.mc_std_err", "median"
)))

	mean	sd	q0.025	q0.5	q0.975
Scaling_Residuals.1	45.579142	16.3912469	16.788694	45.665470	77.580223
Scaling_Residuals.2	-41.606583	9.7130562	-63.162011	-40.981400	-25.964260
Inverse_Residuals.1	5.429733	1.1139628	3.544744	5.292115	7.694393
Inverse_Residuals.2	-4.772610	0.4499736	-5.420918	-4.815154	-3.781790
Pearson_Residuals.1	15.987900	2.8791982	10.231827	15.696815	21.744092
Pearson_Residuals.2	-14.551927	1.6807425	-17.873380	-14.573948	-11.320371

Here, it can be seen that there are two lines of residual data for each type of residual. This signifies that the residual_df function has calculated residuals for each partition. To compare these residuals effectively, a function residual_plot() is defined that plots the residuals for each corresponding polygon of B. This is discussed later in the vignette.

Another type of partitioning is considered with three sections of the region and its residuals are as displayed below:

	mean	sd	q0.025	q0.5	q0.975
Scaling_Residuals.1	70.275529	7.6937454	55.4602823	69.378417	83.452998
Scaling_Residuals.2	-22.619243	15.3576753	-51.1742733	-23.504961	5.640258
Scaling_Residuals.3	-46.150244	4.1354585	-53.3932941	-45.868264	-39.520858
Inverse_Residuals.1	2.930862	0.6915525	1.6556138	2.895886	4.136008
Inverse_Residuals.2	3.002708	0.6847152	1.7552113	2.957585	4.104862
Inverse_Residuals.3	-5.452504	0.0492917	-5.5365311	-5.456939	-5.340719
Pearson_Residuals.1	11.665942	1.7082438	8.3161517	11.821406	14.587497
Pearson_Residuals.2	4.506311	2.0781339	0.6560517	4.577232	8.167815
Pearson_Residuals.3	-15.235021	0.8061648	-16.7090651	-15.226705	-13.873906

Residuals of models at different resolutions

Consider two more resolutions of $B$ , given in the plots below. Here we use the resolution argument instead of nrow and ncol. Due to the way terra::rast() interprets the arguments, using resolution doesn’t guarantee complete coverage of the samplers polygon.

Using these two new choices of $B$ , the residuals are calculated for all four models at both resolutions using residual_df().

The functions set_csc() and residual_plot() are defined to plot the three types of residuals for the model for each partition in $B$ . The code chunk below demonstrates how this plotting is done for the vegetation model at resolution $B = B_3$ . Here, it is useful for each type of residual to have the same colour scale at both resolutions.

# Residuals of Vegetation model
Residual_fit1_B3 <- residual_df(
  model = fit1, df = As3$df,
  expr = expression(exp(vegetation + Intercept)), A_sum = As3$A_sum,
  A_integrate = As3$A_integrate
)

Residual_fit1_B4 <- residual_df(
  model = fit1, df = As4$df,
  expr = expression(exp(vegetation + Intercept)), A_sum = As4$A_sum,
  A_integrate = As4$A_integrate
)

# Colour scales for each model to cover both resolutions
fit1_csc <- set_csc(rbind(Residual_fit1_B3, Residual_fit1_B4), rep("RdBu", 3))

# Store plots
plotB3_fit1 <- residual_plot(B3, Residual_fit1_B3, fit1_csc, "Vegetation Model")
plotB4_fit1 <- residual_plot(B4, Residual_fit1_B4, fit1_csc, "Vegetation Model")

# comparing the vegetation model
((plotB3_fit1$Scaling | plotB3_fit1$Inverse | plotB3_fit1$Pearson) /
  (plotB4_fit1$Scaling | plotB4_fit1$Inverse | plotB4_fit1$Pearson)) +
  plot_annotation(title = "Vegetation Model") +
  plot_layout(guides = "collect") &
  theme(legend.position = "bottom")

Comparing different types of residuals for each model

Now, the plots below help to compare the residuals for each model at two resolutions of $B$ for each type of residual.

Discussion

In this section, we discuss some of the findings from the output in the code. A relevant point that would be used repeatedly in this discussion is that since the colour scales for the residuals are chosen so that negative and positive values are in hues of red and blue respectively, negative and positive residuals imply an overestimation and underestimation of gorilla nests by the model respectively. Thus, red regions in a plot would imply overestimation of gorilla nests by the model in that region and blue regions in a plot would imply underestimation of gorilla nests by the model in that region.

A common observation for all models is that the scaling residuals seem to have the highest range among the residuals since they are direct interpretations of residuals for all models. Also, the Pearson residuals seem to take values between the scaling and inverse residuals for both resolutions of all four models. However, for almost all types of residuals for each model, the positive and negative residual values are not very extreme for the $B_3$ resolution but these corresponding residual values are more extreme at the $B_4$ resolution for each type of residual. This could be because of the bias in the ``experimental” mode of which is explored at other points in this project.

For the vegetation and intercept model, the regions which have negative and positive residuals remain the same for all residuals at different levels for each type of residual. Also, it seems that the north and north-west region of both models are underestimated by the model while the remaining regions are overestimated by the model at different levels.

For the elevation model, each type of residual has slight variations in negative and positive residuals for corresponding regions in a fixed resolution. This means that different residuals suggest that a particular polygon in overestimated and underestimated by the model. However, this happens only at a few points and is not containing extremes of the residuals in the same regions. The Pearson residuals are supposedly the most reliable in this case as they are normalised residuals. All three types of residuals contain larger proportions of negative residuals which suggest that the model overestimates the gorilla population the region (at different amounts in different regions).

Comparing different models for each type of residual

Another way to compare residuals, is to consider each type of residual separately and plot the residuals for all four models at each resolution. This helps to compare the models for a given type of residual and choice of $B$ . Here, it is useful to have the same colour scale for all the 4 models.

The code chunk below demonstrates the code for plotting the “Pearson” residuals when $B = B_4$ .

# Calculate the residuals for all models at B4
Residual_fit1_B4 <- residual_df(
  model = fit1, df = As4$df,
  expr = expression(exp(vegetation + Intercept)), A_sum = As4$A_sum,
  A_integrate = As4$A_integrate
)

Residual_fit2_B4 <- residual_df(
  model = fit2, df = As4$df,
  expr = expression(exp(elev + mySmooth + Intercept)),
  A_sum = As4$A_sum, A_integrate = As4$A_integrate
)

Residual_fit3_B4 <- residual_df(
  model = fit3, df = As4$df,
  expr = expression(exp(Intercept)), A_sum = As4$A_sum,
  A_integrate = As4$A_integrate
)

Residual_fit4_B4 <- residual_df(
  model = fit4, df = As4$df,
  expr = expression(exp(mySmooth + Intercept)), A_sum = As4$A_sum,
  A_integrate = As4$A_integrate
)

# Set the colour scale
B4res <- rbind(
  Residual_fit1_B4, Residual_fit2_B4,
  Residual_fit3_B4, Residual_fit4_B4
)
B4csc <- set_csc(B4res, rep("RdBu", 3))

# Plots for residuals of all 4 models with resolution B4
plotB4_veg <- residual_plot(B4, Residual_fit1_B4, B4csc, "Vegetation Model")
plotB4_elev <- residual_plot(B4, Residual_fit2_B4, B4csc, "Elevation Model")
plotB4_int <- residual_plot(B4, Residual_fit3_B4, B4csc, "Intercept Model")
plotB4_smooth <- residual_plot(B4, Residual_fit4_B4, B4csc, "Smooth Model")

# Comparing all models for B4 Pearson residuals
((plotB4_veg$Pearson | plotB4_elev$Pearson) /
  (plotB4_int$Pearson | plotB4_smooth$Pearson)) +
  plot_annotation("B4 Pearson Residuals") +
  plot_layout(guides = "collect") &
  theme(legend.position = "bottom")

Now, the plots below help to compare all 4 models at a particular choice of $B$ and for a chosen residual type.

Discussion

Firstly, a common feature in each of the three figures is that the residual plots for the vegetation and intercept models are relatively similar and those of the elevation and smooth models are relatively similar as well for both resolutions for any type of residual. Also, for a fixed type of residual and a chosen model, the $B_3$ residuals are less extreme compared to the $B_4$ residuals, plausibly owing to the bias in the ``experimental” mode in . Another feature in all three types of residuals is that the elevation and intercept model have less extreme residual values compared to the vegetation and intercept models for a given type of residual. This could suggest that the former two are better models for gorilla nesting in the given region. The residual values have a wider range when $B=B_4$ than when $B = B_3$ .

For the scaling residuals and the Pearson residuals, the elevation and smooth models have less extreme values of residuals relative the vegetation and intercept models. This suggests that the elevation and smooth models are more suited to estimating the gorilla nesting locations in $W$ . Also, there is a significant underestimation in the north-west areas by the vegetation and intercept models while the same areas are not underestimated by the other two models. However, as explored in , these residuals are calculated in the ``experimental” mode of which produces some overestimates in these two models, so this must be kept in mind while reaching to conclusions about a suitable model for nesting locations.

For the inverse residuals, the same observations hold, except the residuals for the elevation and smooth models have residuals that are not significantly less extreme relative to the vegetation and intercept models as seen in the case of the Pearson and scaling residuals for both resolutions $B_3$ and $B_4$ .

Code appendix

Function definitions

# Niharika Reddy Peddinenikalva
# Vacation Scholarship project
# Updated for sf by Finn Lindgren 2024

suppressPackageStartupMessages(library("RColorBrewer"))
suppressPackageStartupMessages(library("ggplot2"))
suppressPackageStartupMessages(library("dplyr"))
suppressPackageStartupMessages(library("lwgeom"))
suppressPackageStartupMessages(library("patchwork"))
suppressPackageStartupMessages(library("terra"))
suppressPackageStartupMessages(library("sf"))
suppressPackageStartupMessages(library("INLA"))
suppressPackageStartupMessages(library("inlabru"))
theme_set(theme_bw())



#' ----------------------------------
#' prepare_residual_calculations
#' ----------------------------------
#'
#' Computes the A_sum, A_integrate matrices and the data frame used for
#' calculating the residuals for the given set of polygons B

#' Input:
#' @param samplers `sf` polygons containing partitions for which
#' residuals are to be calculated
#' @param domain A mesh object
#' @param observations `sf` points containing observed data
#'
#' Output:
#' @return A_sum - matrix used to compute the summation term of the residuals
#' @return A_integrate - matrix used to compute the integral term
#' @return df - `sf` containing all the locations 'u' for
#' calculating residuals
#'
prepare_residual_calculations <- function(samplers, domain, observations) {
  # Calculate the integration weights for A_integrate
  ips <- fm_int(domain = domain, samplers = samplers)

  # Set-up the A_integrate matrix
  # A_integrate has as many rows as polygons in the samplers,
  # as many columns as mesh points
  A_integrate <- inla.spde.make.A(
    mesh = domain, ips, weights = ips$weight,
    block = ips$.block, block.rescale = "none"
  )


  # Set-up the A_sum matrix
  # A_sum has as many rows as polygons in the samplers,
  # as many columns as observed points
  # each row has 1s for the points in the corresponding polygon
  idx <- sf::st_within(
    observations,
    samplers,
    sparse = TRUE
  )
  A_sum <- sparseMatrix(
    i = unlist(idx),
    j = rep(
      seq_len(nrow(observations)),
      vapply(idx, length, 1L)
    ),
    x = rep(1, length(unlist(idx))),
    dims = c(nrow(samplers), nrow(observations))
  )


  # Setting up the data frame for calculating residuals
  observations$obs <- TRUE
  df <- st_as_sf(
    data.frame(
      obs = rep(FALSE, domain$n),
      x = domain$loc[, 1],
      y = domain$loc[, 2]
    ),
    coords = c("x", "y"),
    crs = fm_crs(domain)
  )
  df <- dplyr::bind_rows(df, observations)

  # Return A-sum, A_integrate and the data frame for predicting the residuals
  list(A_sum = A_sum, A_integrate = A_integrate, df = df)
}



#' ------------
#' residual_df
#' ------------
#'
#' Computes the three types if residuals and returns a data frame containing
#' information about all 3 residuals for each partition of 'B'
#'
#' Inputs:
#' @param model fitted model for which residuals need to be calculated
#' @param df sf object containing all the locations 'u'
#' for calculating residuals
#' @param expr an expression object containing the formula of the model
#' @param A_sum matrix used to compute the summation term of the residuals
#' @param A_integrate matrix that computes the integral term of the residuals
#'
#' Outputs:
#' @return Data frame containing residual information for each of the
#' partitions of the subset 'B'
#'

residual_df <- function(model, df, expr, A_sum, A_integrate) {
  # Compute residuals
  res <- predict(
    object = model,
    newdata = df,
    ~ {
      lambda <- eval(expr)
      h1 <- lambda * 0 + 1
      h2 <- 1 / lambda
      h3 <- 1 / sqrt(lambda)
      data.frame(
        Scaling_Residuals =
          as.vector(A_sum %*% h1[obs]) -
            as.vector(A_integrate %*% (h1 * lambda)[!obs]),
        Inverse_Residuals =
          as.vector(A_sum %*% h2[obs]) -
            as.vector(A_integrate %*% (h2 * lambda)[!obs]),
        Pearson_Residuals =
          as.vector(A_sum %*% h3[obs]) -
            as.vector(A_integrate %*% (h3 * lambda)[!obs])
      )
    },
    used = bru_used(expr)
  )

  # Label the three types of residuals
  res$Scaling_Residuals$Type <- "Scaling Residuals"
  res$Inverse_Residuals$Type <- "Inverse Residuals"
  res$Pearson_Residuals$Type <- "Pearson Residuals"
  do.call(rbind, res)
}





#' --------------
#' set_csc
#' --------------
#'
#' Sets the colour scale for the three types of residuals
#'
#' Inputs:
#' @param residuals frame containing residual information for each of the
#' partitions of the subset 'B'
#' @param col_theme vector of themes for each type of residual
#'
#' Outputs:
#' @return a list of 3 colour scales for each type of residual
#'

set_csc <- function(residuals, col_theme) {
  # Store data for the colour scale of the plots for each type of residual
  cscrange <- data.frame(
    residuals %>%
      group_by(Type) %>%
      summarise(maxabs = max(abs(mean)))
  )

  # Set the colour scale for all three types of residuals
  scaling_csc <-
    scale_fill_gradientn(
      colours = brewer.pal(9, col_theme[1]),
      name = "Scaling Residual",
      limits =
        cscrange[cscrange$Type == "Scaling Residuals", 2] *
          c(-1, 1)
    )

  inverse_csc <-
    scale_fill_gradientn(
      colours = brewer.pal(9, col_theme[2]),
      name = "Inverse Residual",
      limits =
        cscrange[cscrange$Type == "Inverse Residuals", 2] *
          c(-1, 1)
    )

  pearson_csc <-
    scale_fill_gradientn(
      colours = brewer.pal(9, col_theme[3]),
      name = "Pearson Residual",
      limits =
        cscrange[cscrange$Type == "Pearson Residuals", 2] *
          c(-1, 1)
    )

  list("Scaling" = scaling_csc,
       "Inverse" = inverse_csc,
       "Pearson" = pearson_csc)
}



#' ---------------
#' residual_plot
#' ---------------
#'
#' plots the three types of residuals for each polygon
#'
#' Input:
#' @param samplers A sf containing partitions for which
#' residuals are to be calculated
#' @param residuals frame containing residual information for each of the
#' partitions of the subset 'B'
#' @param csc list of three colour scales for the three types of residuals
#' @param model_name string containing the name of the model being assessed
#'
#' Output:
#' @return a list of three subplots scaling, inverse and Pearson residuals
#' for the different partitions of samplers


residual_plot <- function(samplers, residuals, csc, model_name) {
  # Initialise the scaling residuals plot
  samplers$Residual <- residuals %>%
    filter(Type == "Scaling Residuals") %>%
    pull(mean)
  scaling <- ggplot() +
    gg(samplers, aes(fill = Residual), alpha = 1, colour = NA) +
    csc["Scaling"] +
    theme(legend.position = "bottom") +
    labs(subtitle = paste(model_name, "Scaling"))

  # Initialise the inverse residuals plot
  samplers$Residual <- residuals %>%
    filter(Type == "Inverse Residuals") %>%
    pull(mean)
  inverse <- ggplot() +
    gg(samplers, aes(fill = Residual), alpha = 1, colour = NA) +
    csc["Inverse"] +
    theme(legend.position = "bottom") +
    labs(subtitle = paste(model_name, "Inverse"))

  # Initialise the Pearson residuals plot
  samplers$Residual <- residuals %>%
    filter(Type == "Pearson Residuals") %>%
    pull(mean)
  pearson <- ggplot() +
    gg(samplers, aes(fill = Residual), alpha = 1, colour = NA) +
    csc["Pearson"] +
    theme(legend.position = "bottom") +
    labs(subtitle = paste(model_name, "Pearson"))

  # Return the three plots in a list
  list(
    Scaling = scaling, Inverse = inverse,
    Pearson = pearson
  )
}




#' ------------------
#' partition
#' ------------------
#'
#' Partitions the region based on the given criteria for calculating residuals
#' in each partition. Parts of this function are taken from concepts in
#' https://rpubs.com/huanfaChen/grid_from_polygon
#'
#' Input:
#' @param samplers A sf polygon containing region for which
#' partitions need to be created
#' @param resolution resolution of the grids that are required
#' @param nrows number of rows of grids that are required
#' @param ncols number of columns of grids that are required
#'
#' Output:
#' @return a partitioned sf with polygons as required
#'
#'
partition <- function(samplers, resolution = NULL, nrows = NULL, ncols = NULL) {
  # Create a grid for the given boundary
  if (is.null(resolution)) {
    grid <- terra::rast(
      terra::ext(samplers),
      crs = fm_proj4string(samplers),
      nrows = nrows, ncols = ncols
    )
  }

  if (is.null(c(nrows, ncols))) {
    grid <- terra::rast(
      terra::ext(samplers),
      crs = fm_proj4string(samplers),
      resolution = resolution
    )
  }

  gridPolygon <- terra::as.polygons(grid)

  # Extract the boundary with subpolygons only
  sf::st_as_sf(
    terra::intersect(
      gridPolygon,
      terra::vect(samplers)
    )
  )
}


#' ------------------
#' edit_df
#' ------------------
#'
#' Edits the residual data frames to remove columns that need not be displayed
#'
#' Input:
#' @param df the data frame which needs to be edited
#' @param columns a vector of columns that need to be deleted from df
#'
#' Output:
#' @return the edited data frame with only the desired columns
#'
#'
edit_df <- function(df, columns) {
  # Remove the columns that are not required
  df[, !(colnames(df) %in% columns), drop = FALSE]
}

Niharika Reddy Peddinenikalva, with sf updates by Finn Lindgren

2024-12-18

Introduction

Theory of Residuals for Spatial point process models

Types of Residuals

Scaling Residuals

Inverse Residuals

Pearson Residuals

Motivation for choice of residuals

Computation of Residuals

Loading gorillas data

Assessing 4 models with residuals

Vegetation Model

Elevation model

Intercept Model

Smooth Model

Comparing models

Redefining the set B

Residuals of models at different resolutions

Comparing different types of residuals for each model

Discussion

Comparing different models for each type of residual

Discussion

Code appendix

Function definitions