Fitting models to the buffalo data
Dynamic step selection functions with temporal harmonics - dry season
Load packages
Importing buffalo data
Import the buffalo data with random steps and extracted covariates that we created in the previous script: Ecography_DynamicSSF_1_Step_generation.
Here we create the sine and cosine terms that were interact with each of the covariates to fit temporally varying parameters.
Importing data
Rows: 1165406 Columns: 22
── Column specification ────────────────────────────────────────────────────────
Delimiter: ","
dbl (18): id, burst_, x1_, x2_, y1_, y2_, sl_, ta_, dt_, hour_t2, step_id_,...
lgl (1): case_
dttm (3): t1_, t2_, t2_rounded
ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.Importing data
buffalo_data_all <- buffalo_data_all %>%
mutate(t1_ = lubridate::with_tz(buffalo_data_all$t1_, tzone = "Australia/Darwin"),
t2_ = lubridate::with_tz(buffalo_data_all$t2_, tzone = "Australia/Darwin"))
buffalo_data_all <- buffalo_data_all %>%
mutate(id_num = as.numeric(factor(id)),
step_id = step_id_,
x1 = x1_, x2 = x2_,
y1 = y1_, y2 = y2_,
t1 = t1_,
t1_rounded = round_date(buffalo_data_all$t1_, "hour"),
hour_t1 = hour(t1_rounded),
t2 = t2_,
t2_rounded = round_date(buffalo_data_all$t2_, "hour"),
hour_t2 = hour(t2_rounded),
hour = hour_t2,
yday = yday(t1_),
year = year(t1_),
month = month(t1_),
sl = sl_,
log_sl = log(sl_),
ta = ta_,
cos_ta = cos(ta_),
# scale canopy cover from 0 to 1
canopy_01 = canopy_cover/100,
# here we create the harmonic terms for the hour of the day
# for seasonal effects, change hour to yday (which is tau in the manuscript),
# and 24 to 365 (which is T)
hour_s1 = sin(2*pi*hour/24),
hour_s2 = sin(4*pi*hour/24),
hour_s3 = sin(6*pi*hour/24),
hour_c1 = cos(2*pi*hour/24),
hour_c2 = cos(4*pi*hour/24),
hour_c3 = cos(6*pi*hour/24))
# to select a single year of data
buffalo_data_all <- buffalo_data_all %>% filter(t1 < "2019-07-25 09:32:42 ACST")
buffalo_ids <- unique(buffalo_data_all$id)
# Timeline of buffalo data
buffalo_data_all %>% ggplot(aes(x = t1, y = factor(id), colour = factor(id))) +
geom_point(alpha = 0.1) +
scale_y_discrete("Buffalo ID") +
scale_x_datetime("Date") +
scale_colour_viridis_d() +
theme_bw() +
theme(legend.position = "none")
Fitting the models
Creating a data matrix
First we create a data matrix to be provided to the model, and then we scale and centre the full data matrix, with respect to each of the columns. That means that all variables are scaled and centred after the data has been split into wet and dry season data, and also after creating the quadratic and harmonic terms (when using them).
We should only include covariates in the data matrix that will be used in the model formula.
Models
- 0p = 0 pairs of harmonics
- 1p = 1 pair of harmonics
- 2p = 2 pairs of harmonics
- 3p = 3 pairs of harmonics
For the dynamic models, we start to add the harmonic terms. As we have already created the harmonic terms for the hour of the day (s1, c1, s2, etc), we just interact (multiply) these with each of the covariates, including the quadratic terms, prior to model fitting. We store the scaling and centering variables to reconstruct the natural scale coefficients.
To provide some intuition about harmonic regression we have created a walkthrough script Ecography_DynamicSSF_Walkthrough_Harmonics_and_selection_surfaces that introduces harmonics and how they can be used to model temporal variation in the data. It will help provide some understand the model outputs and how we construct the temporally varying coefficients in this script.
Code
months_wet <- c(1:4, 11:12)
buffalo_ids <- unique(buffalo_data_all$id)
# comment and uncomment the relevant lines to get either wet or dry season data
# buffalo_data <- buffalo_data_all %>% filter(month %in% months_wet) # wet season
buffalo_data <- buffalo_data_all %>% filter(!month %in% months_wet) # dry season
buffalo_data_matrix_unscaled <- buffalo_data %>% transmute(
ndvi = ndvi_temporal,
ndvi_sq = ndvi_temporal ^ 2,
canopy = canopy_01,
canopy_sq = canopy_01 ^ 2,
slope = slope,
herby = veg_herby,
step_l = sl,
log_step_l = log_sl,
cos_turn_a = cos_ta)
buffalo_data_matrix_scaled <- scale(buffalo_data_matrix_unscaled)
# save the scaling values to recover the natural scale of the coefficients
# which is required for the simulations
# (so then environmental variables don't need to be scaled)
mean_vals <- attr(buffalo_data_matrix_scaled, "scaled:center")
sd_vals <- attr(buffalo_data_matrix_scaled, "scaled:scale")
scaling_attributes_0p <- data.frame(variable = names(buffalo_data_matrix_unscaled),
mean = mean_vals, sd = sd_vals)
# add the id, step_id columns and presence/absence columns to
# the scaled data matrix for model fitting
buffalo_data_scaled_0p <- data.frame(id = buffalo_data$id,
step_id = buffalo_data$step_id,
y = buffalo_data$y,
buffalo_data_matrix_scaled)Code
# buffalo_data <- buffalo_data_all %>% filter(month %in% months_wet) # wet season
buffalo_data <- buffalo_data_all %>% filter(!month %in% months_wet) # dry season
buffalo_data_matrix_unscaled <- buffalo_data %>% transmute(
# the 'linear' term
ndvi = ndvi_temporal,
# interact with the harmonic terms
ndvi_s1 = ndvi_temporal * hour_s1,
ndvi_c1 = ndvi_temporal * hour_c1,
ndvi_sq = ndvi_temporal ^ 2,
ndvi_sq_s1 = (ndvi_temporal ^ 2) * hour_s1,
ndvi_sq_c1 = (ndvi_temporal ^ 2) * hour_c1,
canopy = canopy_01,
canopy_s1 = canopy_01 * hour_s1,
canopy_c1 = canopy_01 * hour_c1,
canopy_sq = canopy_01 ^ 2,
canopy_sq_s1 = (canopy_01 ^ 2) * hour_s1,
canopy_sq_c1 = (canopy_01 ^ 2) * hour_c1,
slope = slope,
slope_s1 = slope * hour_s1,
slope_c1 = slope * hour_c1,
herby = veg_herby,
herby_s1 = veg_herby * hour_s1,
herby_c1 = veg_herby * hour_c1,
step_l = sl,
step_l_s1 = sl * hour_s1,
step_l_c1 = sl * hour_c1,
log_step_l = log_sl,
log_step_l_s1 = log_sl * hour_s1,
log_step_l_c1 = log_sl * hour_c1,
cos_turn_a = cos_ta,
cos_turn_a_s1 = cos_ta * hour_s1,
cos_turn_a_c1 = cos_ta * hour_c1)
buffalo_data_matrix_scaled <- scale(buffalo_data_matrix_unscaled)
mean_vals <- attr(buffalo_data_matrix_scaled, "scaled:center")
sd_vals <- attr(buffalo_data_matrix_scaled, "scaled:scale")
scaling_attributes_1p <- data.frame(variable = names(buffalo_data_matrix_unscaled),
mean = mean_vals, sd = sd_vals)
buffalo_data_scaled_1p <- data.frame(id = buffalo_data$id,
step_id = buffalo_data$step_id,
y = buffalo_data$y,
buffalo_data_matrix_scaled)Code
months_wet <- c(1:4, 11:12)
buffalo_ids <- unique(buffalo_data_all$id)
# buffalo_data <- buffalo_data_all %>% filter(month %in% months_wet) # wet season
buffalo_data <- buffalo_data_all %>% filter(!month %in% months_wet) # dry season
buffalo_data_matrix_unscaled <- buffalo_data %>% transmute(
ndvi = ndvi_temporal,
ndvi_s1 = ndvi_temporal * hour_s1,
ndvi_s2 = ndvi_temporal * hour_s2,
ndvi_c1 = ndvi_temporal * hour_c1,
ndvi_c2 = ndvi_temporal * hour_c2,
ndvi_sq = ndvi_temporal ^ 2,
ndvi_sq_s1 = (ndvi_temporal ^ 2) * hour_s1,
ndvi_sq_s2 = (ndvi_temporal ^ 2) * hour_s2,
ndvi_sq_c1 = (ndvi_temporal ^ 2) * hour_c1,
ndvi_sq_c2 = (ndvi_temporal ^ 2) * hour_c2,
canopy = canopy_01,
canopy_s1 = canopy_01 * hour_s1,
canopy_s2 = canopy_01 * hour_s2,
canopy_c1 = canopy_01 * hour_c1,
canopy_c2 = canopy_01 * hour_c2,
canopy_sq = canopy_01 ^ 2,
canopy_sq_s1 = (canopy_01 ^ 2) * hour_s1,
canopy_sq_s2 = (canopy_01 ^ 2) * hour_s2,
canopy_sq_c1 = (canopy_01 ^ 2) * hour_c1,
canopy_sq_c2 = (canopy_01 ^ 2) * hour_c2,
slope = slope,
slope_s1 = slope * hour_s1,
slope_s2 = slope * hour_s2,
slope_c1 = slope * hour_c1,
slope_c2 = slope * hour_c2,
herby = veg_herby,
herby_s1 = veg_herby * hour_s1,
herby_s2 = veg_herby * hour_s2,
herby_c1 = veg_herby * hour_c1,
herby_c2 = veg_herby * hour_c2,
step_l = sl,
step_l_s1 = sl * hour_s1,
step_l_s2 = sl * hour_s2,
step_l_c1 = sl * hour_c1,
step_l_c2 = sl * hour_c2,
log_step_l = log_sl,
log_step_l_s1 = log_sl * hour_s1,
log_step_l_s2 = log_sl * hour_s2,
log_step_l_c1 = log_sl * hour_c1,
log_step_l_c2 = log_sl * hour_c2,
cos_turn_a = cos_ta,
cos_turn_a_s1 = cos_ta * hour_s1,
cos_turn_a_s2 = cos_ta * hour_s2,
cos_turn_a_c1 = cos_ta * hour_c1,
cos_turn_a_c2 = cos_ta * hour_c2)
buffalo_data_matrix_scaled <- scale(buffalo_data_matrix_unscaled)
mean_vals <- attr(buffalo_data_matrix_scaled, "scaled:center")
sd_vals <- attr(buffalo_data_matrix_scaled, "scaled:scale")
scaling_attributes_2p <- data.frame(variable = names(buffalo_data_matrix_unscaled),
mean = mean_vals, sd = sd_vals)
buffalo_data_scaled_2p <- data.frame(id = buffalo_data$id,
step_id = buffalo_data$step_id,
y = buffalo_data$y,
buffalo_data_matrix_scaled)Code
months_wet <- c(1:4, 11:12)
buffalo_ids <- unique(buffalo_data_all$id)
# buffalo_data <- buffalo_data_all %>% filter(month %in% months_wet) # wet season
buffalo_data <- buffalo_data_all %>% filter(!month %in% months_wet) # dry season
buffalo_data_matrix_unscaled <- buffalo_data %>% transmute(
ndvi = ndvi_temporal,
ndvi_s1 = ndvi_temporal * hour_s1,
ndvi_s2 = ndvi_temporal * hour_s2,
ndvi_s3 = ndvi_temporal * hour_s3,
ndvi_c1 = ndvi_temporal * hour_c1,
ndvi_c2 = ndvi_temporal * hour_c2,
ndvi_c3 = ndvi_temporal * hour_c3,
ndvi_sq = ndvi_temporal ^ 2,
ndvi_sq_s1 = (ndvi_temporal ^ 2) * hour_s1,
ndvi_sq_s2 = (ndvi_temporal ^ 2) * hour_s2,
ndvi_sq_s3 = (ndvi_temporal ^ 2) * hour_s3,
ndvi_sq_c1 = (ndvi_temporal ^ 2) * hour_c1,
ndvi_sq_c2 = (ndvi_temporal ^ 2) * hour_c2,
ndvi_sq_c3 = (ndvi_temporal ^ 2) * hour_c3,
canopy = canopy_01,
canopy_s1 = canopy_01 * hour_s1,
canopy_s2 = canopy_01 * hour_s2,
canopy_s3 = canopy_01 * hour_s3,
canopy_c1 = canopy_01 * hour_c1,
canopy_c2 = canopy_01 * hour_c2,
canopy_c3 = canopy_01 * hour_c3,
canopy_sq = canopy_01 ^ 2,
canopy_sq_s1 = (canopy_01 ^ 2) * hour_s1,
canopy_sq_s2 = (canopy_01 ^ 2) * hour_s2,
canopy_sq_s3 = (canopy_01 ^ 2) * hour_s3,
canopy_sq_c1 = (canopy_01 ^ 2) * hour_c1,
canopy_sq_c2 = (canopy_01 ^ 2) * hour_c2,
canopy_sq_c3 = (canopy_01 ^ 2) * hour_c3,
slope = slope,
slope_s1 = slope * hour_s1,
slope_s2 = slope * hour_s2,
slope_s3 = slope * hour_s3,
slope_c1 = slope * hour_c1,
slope_c2 = slope * hour_c2,
slope_c3 = slope * hour_c3,
herby = veg_herby,
herby_s1 = veg_herby * hour_s1,
herby_s2 = veg_herby * hour_s2,
herby_s3 = veg_herby * hour_s3,
herby_c1 = veg_herby * hour_c1,
herby_c2 = veg_herby * hour_c2,
herby_c3 = veg_herby * hour_c3,
step_l = sl,
step_l_s1 = sl * hour_s1,
step_l_s2 = sl * hour_s2,
step_l_s3 = sl * hour_s3,
step_l_c1 = sl * hour_c1,
step_l_c2 = sl * hour_c2,
step_l_c3 = sl * hour_c3,
log_step_l = log_sl,
log_step_l_s1 = log_sl * hour_s1,
log_step_l_s2 = log_sl * hour_s2,
log_step_l_s3 = log_sl * hour_s3,
log_step_l_c1 = log_sl * hour_c1,
log_step_l_c2 = log_sl * hour_c2,
log_step_l_c3 = log_sl * hour_c3,
cos_turn_a = cos_ta,
cos_turn_a_s1 = cos_ta * hour_s1,
cos_turn_a_s2 = cos_ta * hour_s2,
cos_turn_a_s3 = cos_ta * hour_s3,
cos_turn_a_c1 = cos_ta * hour_c1,
cos_turn_a_c2 = cos_ta * hour_c2,
cos_turn_a_c3 = cos_ta * hour_c3)
buffalo_data_matrix_scaled <- scale(buffalo_data_matrix_unscaled)
mean_vals <- attr(buffalo_data_matrix_scaled, "scaled:center")
sd_vals <- attr(buffalo_data_matrix_scaled, "scaled:scale")
scaling_attributes_3p <- data.frame(variable = names(buffalo_data_matrix_unscaled),
mean = mean_vals, sd = sd_vals)
buffalo_data_scaled_3p <- data.frame(id = buffalo_data$id,
step_id = buffalo_data$step_id,
y = buffalo_data$y,
buffalo_data_matrix_scaled)Model formula
As we have already precomputed and scaled the covariates, quadratic terms and interactions with the harmonics, we just include each parameter as a linear predictor.
Code
formula_twostep <- y ~
ndvi +
ndvi_s1 +
ndvi_c1 +
ndvi_sq +
ndvi_sq_s1 +
ndvi_sq_c1 +
canopy +
canopy_s1 +
canopy_c1 +
canopy_sq +
canopy_sq_s1 +
canopy_sq_c1 +
slope +
slope_s1 +
slope_c1 +
herby +
herby_s1 +
herby_c1 +
step_l +
step_l_s1 +
step_l_c1 +
log_step_l +
log_step_l_s1 +
log_step_l_c1 +
cos_turn_a +
cos_turn_a_s1 +
cos_turn_a_c1 +
strata(step_id) +
cluster(id)Code
formula_twostep <- y ~
ndvi +
ndvi_s1 +
ndvi_s2 +
ndvi_c1 +
ndvi_c2 +
ndvi_sq +
ndvi_sq_s1 +
ndvi_sq_s2 +
ndvi_sq_c1 +
ndvi_sq_c2 +
canopy +
canopy_s1 +
canopy_s2 +
canopy_c1 +
canopy_c2 +
canopy_sq +
canopy_sq_s1 +
canopy_sq_s2 +
canopy_sq_c1 +
canopy_sq_c2 +
slope +
slope_s1 +
slope_s2 +
slope_c1 +
slope_c2 +
herby +
herby_s1 +
herby_s2 +
herby_c1 +
herby_c2 +
step_l +
step_l_s1 +
step_l_s2 +
step_l_c1 +
step_l_c2 +
log_step_l +
log_step_l_s1 +
log_step_l_s2 +
log_step_l_c1 +
log_step_l_c2 +
cos_turn_a +
cos_turn_a_s1 +
cos_turn_a_s2 +
cos_turn_a_c1 +
cos_turn_a_c2 +
strata(step_id) +
cluster(id)Code
formula_twostep <- y ~
ndvi +
ndvi_s1 +
ndvi_s2 +
ndvi_s3 +
ndvi_c1 +
ndvi_c2 +
ndvi_c3 +
ndvi_sq +
ndvi_sq_s1 +
ndvi_sq_s2 +
ndvi_sq_s3 +
ndvi_sq_c1 +
ndvi_sq_c2 +
ndvi_sq_c3 +
canopy +
canopy_s1 +
canopy_s2 +
canopy_s3 +
canopy_c1 +
canopy_c2 +
canopy_c3 +
canopy_sq +
canopy_sq_s1 +
canopy_sq_s2 +
canopy_sq_s3 +
canopy_sq_c1 +
canopy_sq_c2 +
canopy_sq_c3 +
slope +
slope_s1 +
slope_s2 +
slope_s3 +
slope_c1 +
slope_c2 +
slope_c3 +
herby +
herby_s1 +
herby_s2 +
herby_s3 +
herby_c1 +
herby_c2 +
herby_c3 +
step_l +
step_l_s1 +
step_l_s2 +
step_l_s3 +
step_l_c1 +
step_l_c2 +
step_l_c3 +
log_step_l +
log_step_l_s1 +
log_step_l_s2 +
log_step_l_s3 +
log_step_l_c1 +
log_step_l_c2 +
log_step_l_c3 +
cos_turn_a +
cos_turn_a_s1 +
cos_turn_a_s2 +
cos_turn_a_s3 +
cos_turn_a_c1 +
cos_turn_a_c2 +
cos_turn_a_c3 +
strata(step_id) +
cluster(id)Fit the model
As we have already fitted the model, we will load it here, but if the model_fit file doesn’t exist, it will run the model fitting code. Be careful here that if you change the model formula, you will need to delete or rename the model_fit file to re-run the model fitting code, otherwise it will just load the previous model.
Due to the large number of covariates when including several pairs of harmonics, particularly when also including quadratic terms, we had difficulty fitting population-level models with the glmmTMB (Brooks et al. 2017) and INLA (Rue, Martino, and Chopin 2009) model fitting approaches (Muff, Signer, and Fieberg 2020). We therefore fitted models to all individuals using the ‘TwoStep’ approach from the TwoStepCLogit package Craiu et al. (2011), which can be considered as a computationally efficient alternative to fitting models with individual-level random effects, providing certain conditions are met, namely that all individuals visit every category level when using categorical covariates (Muff, Signer, and Fieberg 2020). The first step of the TwoStep approach is to fit individual-level conditional logistic regression models, which are then combined in the second step using the expectation-maximisation (EM) algorithm in conjunction with conditional restricted maximum likelihood to estimate the population-level parameters (Craiu et al. 2011).
Code
if(file.exists("outputs/model_twostep_0p_harms_dry.rds")) {
model_twostep_0p_harms <- readRDS("outputs/model_twostep_0p_harms_dry.rds")
} else {
tic()
model_twostep_0p_harms <- Ts.estim(formula = formula_twostep,
data = buffalo_data_scaled_0p,
all.m.1 = TRUE,
D = "UN(1)",
itermax = 10000)
toc()
# save model object
saveRDS(model_twostep_0p_harms, file = "outputs/model_twostep_0p_harms_dry.rds")
beep(sound = 2)
}Code
if(file.exists("outputs/model_twostep_1p_harms_dry.rds")) {
model_twostep_1p_harms <- readRDS("outputs/model_twostep_1p_harms_dry.rds")
} else {
tic()
model_twostep_1p_harms <- Ts.estim(formula = formula_twostep,
data = buffalo_data_scaled_1p,
all.m.1 = TRUE,
D = "UN(1)",
itermax = 10000)
toc()
# save model object
saveRDS(model_twostep_1p_harms, file = "outputs/model_twostep_1p_harms_dry.rds")
beep(sound = 2)
}Code
if(file.exists("outputs/model_twostep_2p_harms_dry.rds")) {
model_twostep_2p_harms <- readRDS("outputs/model_twostep_2p_harms_dry.rds")
} else {
tic()
model_twostep_2p_harms <- Ts.estim(formula = formula_twostep,
data = buffalo_data_scaled_2p,
all.m.1 = TRUE,
D = "UN(1)",
itermax = 10000)
toc()
# save model object
saveRDS(model_twostep_2p_harms, file = "outputs/model_twostep_2p_harms_dry.rds")
beep(sound = 2)
}Code
if(file.exists("outputs/model_twostep_3p_harms_dry.rds")) {
model_twostep_3p_harms <- readRDS("outputs/model_twostep_3p_harms_dry.rds")
} else {
tic()
model_twostep_3p_harms <- Ts.estim(formula = formula_twostep,
data = buffalo_data_scaled_3p,
all.m.1 = TRUE,
D = "UN(1)",
itermax = 10000)
toc()
# save model object
saveRDS(model_twostep_3p_harms, file = "outputs/model_twostep_3p_harms_dry.rds")
beep(sound = 2)
}Check the fitted model outputs
Create a dataframe of the coefficients with the scaling attributes that we saved when creating the data matrix. We can then return the coefficients to their natural scale by dividing by the scaling factor (standard deviation).
As we can see, we have a coefficient for each covariate by itself, and then one for each of the harmonic interactions. These are the ‘weights’ that we played around with in the Ecography_DynamicSSF_Walkthrough_Harmonics_and_selection_surfaces walkthrough script, and we reconstruct them in exactly the same way. We also have the coefficients for the quadratic terms and the interactions with the harmonics, which we have denoted as ndvi_sq for instance. We will come back to these when we look at the selection surfaces.
Call:
Ts.estim(formula = formula_twostep, data = buffalo_data_scaled_0p,
all.m.1 = TRUE, D = "UN(1)", itermax = 10000)
beta coefficients:
estimate se
ndvi 0.641896 0.128190
ndvi_sq -0.687249 0.153893
canopy 0.193363 0.067529
canopy_sq -0.254871 0.067107
slope -0.193505 0.067881
herby -0.029148 0.036980
step_l -0.039132 0.051940
log_step_l 0.024948 0.029956
cos_turn_a -0.023911 0.025042
D = estimate of the between-cluster variance-covariance matrix D,
for the random coefficients only:
ndvi ndvi_sq canopy canopy_sq slope
ndvi 0.1862571 0.0000000 0.00000000 0.00000000 0.00000000
ndvi_sq 0.0000000 0.2760612 0.00000000 0.00000000 0.00000000
canopy 0.0000000 0.0000000 0.04720412 0.00000000 0.00000000
canopy_sq 0.0000000 0.0000000 0.00000000 0.04583738 0.00000000
slope 0.0000000 0.0000000 0.00000000 0.00000000 0.05672407
herby 0.0000000 0.0000000 0.00000000 0.00000000 0.00000000
step_l 0.0000000 0.0000000 0.00000000 0.00000000 0.00000000
log_step_l 0.0000000 0.0000000 0.00000000 0.00000000 0.00000000
cos_turn_a 0.0000000 0.0000000 0.00000000 0.00000000 0.00000000
herby step_l log_step_l cos_turn_a
ndvi 0.00000000 0.00000000 0.00000000 0.000000000
ndvi_sq 0.00000000 0.00000000 0.00000000 0.000000000
canopy 0.00000000 0.00000000 0.00000000 0.000000000
canopy_sq 0.00000000 0.00000000 0.00000000 0.000000000
slope 0.00000000 0.00000000 0.00000000 0.000000000
herby 0.01679267 0.00000000 0.00000000 0.000000000
step_l 0.00000000 0.03426404 0.00000000 0.000000000
log_step_l 0.00000000 0.00000000 0.01097489 0.000000000
cos_turn_a 0.00000000 0.00000000 0.00000000 0.007735816Code
# these create massive outputs for the dynamic models so we've commented them out
# model_twostep_0p_harms$beta
# model_twostep_0p_harms$se
# model_twostep_0p_harms$vcov
# diag(model_twostep_0p_harms$D) # between cluster variance
# model_twostep_0p_harms$r.effect # individual estimates
# create a dataframe of the coefficients and their scaling attributes
coefs_clr_0p <- data.frame(coefs = names(model_twostep_0p_harms$beta), value = model_twostep_0p_harms$beta)
coefs_clr_0p$scale_sd <- scaling_attributes_0p$sd
coefs_clr_0p <- coefs_clr_0p %>% mutate(value_nat = value / scale_sd)
head(coefs_clr_0p)Code
# creates a huge output due to the correlation matrix
# model_twostep_1p_harms
# model_twostep_1p_harms
# model_twostep_1p_harms$beta
# model_twostep_1p_harms$se
# model_twostep_1p_harms$vcov
# diag(model_twostep_1p_harms$D) # between cluster variance
# model_twostep_1p_harms$r.effect # individual estimates
coefs_clr_1p <- data.frame(coefs = names(model_twostep_1p_harms$beta),
value = model_twostep_1p_harms$beta)
coefs_clr_1p$scale_sd <- scaling_attributes_1p$sd
coefs_clr_1p <- coefs_clr_1p %>% mutate(value_nat = value / scale_sd)
head(coefs_clr_1p)Code
# creates a huge output due to the correlation matrix
# model_twostep_2p_harms
# model_twostep_2p_harms
# model_twostep_2p_harms$beta
# model_twostep_2p_harms$se
# model_twostep_2p_harms$vcov
# diag(model_twostep_2p_harms$D) # between cluster variance
# model_twostep_2p_harms$r.effect # individual estimates
# creating data frame of model coefficients
coefs_clr_2p <- data.frame(coefs = names(model_twostep_2p_harms$beta),
value = model_twostep_2p_harms$beta)
coefs_clr_2p$scale_sd <- scaling_attributes_2p$sd
coefs_clr_2p <- coefs_clr_2p %>% mutate(value_nat = value / scale_sd)
head(coefs_clr_2p)Code
# creates a huge output due to the correlation matrix
# model_twostep_3p_harms
# model_twostep_3p_harms$beta
# model_twostep_3p_harms$se
# model_twostep_3p_harms$vcov
# diag(model_twostep_3p_harms$D) # between cluster variance
# model_twostep_3p_harms$r.effect # individual estimates
# creating dataframe of coefficients
coefs_clr_3p <- data.frame(coefs = names(model_twostep_3p_harms$beta),
value = model_twostep_3p_harms$beta)
coefs_clr_3p$scale_sd <- scaling_attributes_3p$sd
coefs_clr_3p <- coefs_clr_3p %>% mutate(value_nat = value / scale_sd)
head(coefs_clr_3p)Reconstruct the temporally dynamic coefficients
First we reconstruct the hourly coefficients for the model with no harmonics. This step isn’t necessary as we already have the coefficients, and we have already rescaled them in the dataframe we created above. But as we are also fitting harmonic models and recover their coefficients across time, we have used the same approach here so then we can plot them together and illustrate the static/dynamic outputs of the models. It also means that we can use the same simulation code (which indexes across the hour of the day), and just change the data frame of coefficients (as it will index across the coefficients of the static model but they won’t change).
We need a sequence of values that covers a full period (or the period that we want to construct the function over, which can be more or less than 1 period). The sequence can be arbitrarily finely spaced. The smaller the increment the smoother the function will be for plotting. When simulating data from the temporally dynamic coefficients, we will subset to the increment that relates to the data collection and model fitting (i.e. one hour in this case).
Essentially, the coefficients can be considered as weights of the harmonics, which combine into a single function.
Now we can reconstruct the harmonic function using the formula that we put into our model by interacting the harmonic terms with each of the covariates, for two pairs of harmonics (2p) a single covariate, let’s say herbaceous vegetation (herby), this would be written down as:
\[ f = \beta_{herby} + \beta_{herby\_s1} \sin\left(\frac{2\pi t}{24}\right) + \beta_{herby\_c1} \cos\left(\frac{2\pi t}{24}\right) + \beta_{herby\_s2} \sin\left(\frac{4\pi t}{24}\right) + \beta_{herby\_c2} \cos\left(\frac{4\pi t}{24}\right), \]
where we have 5 \(\beta_{herby}\) coefficients, one for the linear term, and one for each of the harmonic terms.
Here we use matrix multiplication to reconstruct the temporally dynamic coefficients. We provide some background in the Ecography_DynamicSSF_Walkthrough_Harmonics_and_selection_surfaces script.
First we create a matrix of the values of the harmonics, which is just the sin and cos terms for each harmonic, and then we can multiply this by the coefficients to get the function. When we use two pairs of harmonics we will have 5 coefficients for each covariate (linear + 2 sine and 2 cosine), so there will be 5 columns in the matrix.
For matrix multiplication, the number of columns in the first matrix must be equal to the number of rows in the second matrix. The result will then have the same number of rows as the first matrix and the same number of columns as the second matrix.
Or in other words, if we have a 24 x 5 matrix of harmonics and a 5 x 1 matrix of coefficients, we will get a 24 x 1 matrix of the function, which corresponds to our 24 hours of the day.
Code
# increments are arbitrary - finer results in smoother curves
# for the simulations we will subset to the step interval
hour <- seq(0,23.9,0.1)
# create the dataframe of values of the harmonic terms over the period (here just the linear term)
hour_harmonics_df_0p <- data.frame("linear_term" = rep(1, length(hour)))
harmonics_scaled_df_0p <- data.frame(
"hour" = hour,
"ndvi" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("ndvi", coefs) & !grepl("sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"ndvi_2" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("ndvi_sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"canopy" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("canopy", coefs) & !grepl("sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"canopy_2" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("canopy_sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"slope" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("slope", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"herby" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("herby", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"sl" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("step_l", coefs) & !grepl("log", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"log_sl" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("log_step_l", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"cos_ta" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("cos", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))))
harmonics_scaled_long_0p <- pivot_longer(harmonics_scaled_df_0p,
cols = !1,
names_to = "coef")Code
# create the dataframe of values of the harmonic terms over the period
hour_harmonics_df_1p <- data.frame("linear_term" = rep(1, length(hour)),
"hour_s1" = sin(2*pi*hour/24),
"hour_c1" = cos(2*pi*hour/24))
harmonics_scaled_df_1p <- data.frame(
"hour" = hour,
"ndvi" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("ndvi", coefs) & !grepl("sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"ndvi_2" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("ndvi_sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"canopy" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("canopy", coefs) & !grepl("sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"canopy_2" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("canopy_sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"slope" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("slope", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"herby" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("herby", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"sl" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("step_l", coefs) & !grepl("log", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"log_sl" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("log_step_l", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"cos_ta" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("cos", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))))
harmonics_scaled_long_1p <- pivot_longer(harmonics_scaled_df_1p,
cols = !1,
names_to = "coef")Code
# create the dataframe of values of the harmonic terms over the period
hour_harmonics_df_2p <- data.frame("linear_term" = rep(1, length(hour)),
"hour_s1" = sin(2*pi*hour/24),
"hour_s2" = sin(4*pi*hour/24),
"hour_c1" = cos(2*pi*hour/24),
"hour_c2" = cos(4*pi*hour/24))
harmonics_scaled_df_2p <- data.frame(
"hour" = hour,
"ndvi" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("ndvi", coefs) & !grepl("sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"ndvi_2" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("ndvi_sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"canopy" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("canopy", coefs) & !grepl("sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"canopy_2" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("canopy_sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"slope" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("slope", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"herby" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("herby", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"sl" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("step_l", coefs) & !grepl("log", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"log_sl" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("log_step_l", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"cos_ta" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("cos", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))))
harmonics_scaled_long_2p <- pivot_longer(harmonics_scaled_df_2p, cols = !1,
names_to = "coef")Code
# create the dataframe of values of the harmonic terms over the period
hour_harmonics_df_3p <- data.frame("linear_term" = rep(1, length(hour)),
"hour_s1" = sin(2*pi*hour/24),
"hour_s2" = sin(4*pi*hour/24),
"hour_s3" = sin(6*pi*hour/24),
"hour_c1" = cos(2*pi*hour/24),
"hour_c2" = cos(4*pi*hour/24),
"hour_c3" = cos(6*pi*hour/24))
harmonics_scaled_df_3p <- data.frame(
"hour" = hour,
"ndvi" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("ndvi", coefs) & !grepl("sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"ndvi_2" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("ndvi_sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"canopy" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("canopy", coefs) & !grepl("sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"canopy_2" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("canopy_sq", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"slope" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("slope", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"herby" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("herby", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"sl" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("step_l", coefs) & !grepl("log", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"log_sl" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("log_step_l", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"cos_ta" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("cos", coefs)) %>%
pull(value) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))))
harmonics_scaled_long_3p <- pivot_longer(harmonics_scaled_df_3p, cols = !1,
names_to = "coef")Plot the results - scaled temporally dynamic coefficients
Here we show the temporally-varying coefficients across time (which are currently still scaled).
Code
ggplot() +
geom_path(data = harmonics_scaled_long_0p,
aes(x = hour, y = value, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(Time-varying~parameter~values~beta)) +
scale_x_continuous("Hour") +
scale_color_discrete("Estimate") +
theme_classic() +
theme(legend.position = "bottom")
Code
ggplot() +
geom_path(data = harmonics_scaled_long_1p,
aes(x = hour, y = value, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(Time-varying~parameter~values~beta)) +
scale_x_continuous("Hour") +
scale_color_discrete("Estimate") +
theme_classic() +
theme(legend.position = "bottom")
Code
ggplot() +
geom_path(data = harmonics_scaled_long_2p,
aes(x = hour, y = value, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(Time-varying~parameter~values~beta)) +
scale_x_continuous("Hour") +
scale_color_discrete("Estimate") +
theme_classic() +
theme(legend.position = "bottom")
Code
ggplot() +
geom_path(data = harmonics_scaled_long_3p,
aes(x = hour, y = value, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(Time-varying~parameter~values~beta)) +
scale_x_continuous("Hour") +
scale_color_discrete("Estimate") +
theme_classic() +
theme(legend.position = "bottom")
Reconstructing the natural-scale temporally dynamic coefficients
As we scaled the covariate values prior to fitting the models, we want to rescale the coefficients to their natural scale. This is important for the simulations, as the environmental variables will not be scaled when we simulate steps.
Code
harmonics_nat_df_0p <- data.frame(
"hour" = hour,
"ndvi" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("ndvi", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"ndvi_2" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("ndvi_sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"canopy" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("canopy", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"canopy_2" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("canopy_sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"slope" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("slope", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"herby" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("herby", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"sl" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("step_l", coefs) & !grepl("log", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"log_sl" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("log_step_l", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))),
"cos_ta" = as.numeric(
coefs_clr_0p %>% dplyr::filter(grepl("cos", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_0p))))Code
harmonics_nat_df_1p <- data.frame(
"hour" = hour,
"ndvi" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("ndvi", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"ndvi_2" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("ndvi_sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"canopy" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("canopy", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"canopy_2" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("canopy_sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"slope" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("slope", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"herby" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("herby", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"sl" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("step_l", coefs) & !grepl("log", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"log_sl" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("log_step_l", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))),
"cos_ta" = as.numeric(
coefs_clr_1p %>% dplyr::filter(grepl("cos", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_1p))))Code
harmonics_nat_df_2p <- data.frame(
"hour" = hour,
"ndvi" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("ndvi", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"ndvi_2" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("ndvi_sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"canopy" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("canopy", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"canopy_2" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("canopy_sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"slope" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("slope", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"herby" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("herby", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"sl" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("step_l", coefs) & !grepl("log", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"log_sl" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("log_step_l", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))),
"cos_ta" = as.numeric(
coefs_clr_2p %>% dplyr::filter(grepl("cos", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_2p))))Code
harmonics_nat_df_3p <- data.frame(
"hour" = hour,
"ndvi" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("ndvi", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"ndvi_2" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("ndvi_sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"canopy" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("canopy", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"canopy_2" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("canopy_sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"slope" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("slope", coefs) & !grepl("sq", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"herby" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("herby", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"sl" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("step_l", coefs) & !grepl("log", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"log_sl" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("log_step_l", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))),
"cos_ta" = as.numeric(
coefs_clr_3p %>% dplyr::filter(grepl("cos", coefs)) %>%
pull(value_nat) %>% t() %*% t(as.matrix(hour_harmonics_df_3p))))Update the Gamma and von Mises distributions
To update the Gamma and von Mises distribution from the tentative distributions (e.g. Fieberg et al. 2021, Appendix C), we just do the calculation at each time point (for the natural-scale coefficients).
Code
# from the step generation script
tentative_shape <- 0.438167
tentative_scale <- 534.3507
tentative_kappa <- 0.1848126
hour_coefs_nat_df_0p <- harmonics_nat_df_0p %>%
mutate(shape = tentative_shape + log_sl,
scale = 1/((1/tentative_scale) - sl),
kappa = tentative_kappa + cos_ta)
# save the coefficients to use in the simulations
write_csv(hour_coefs_nat_df_0p,
paste0("outputs/TwoStep_0pDaily_coefs_dry_", Sys.Date(), ".csv"))
# turning into a long data frame
hour_coefs_nat_long_0p <- pivot_longer(hour_coefs_nat_df_0p,
cols = !1,
names_to = "coef")Code
hour_coefs_nat_df_1p <- harmonics_nat_df_1p %>%
mutate(shape = tentative_shape + log_sl,
scale = 1/((1/tentative_scale) - sl),
kappa = tentative_kappa + cos_ta)
# save the coefficients to use in the simulations
write_csv(hour_coefs_nat_df_1p,
paste0("outputs/TwoStep_1pDaily_coefs_dry_",Sys.Date(), ".csv"))
# turning into a long data frame
hour_coefs_nat_long_1p <- pivot_longer(hour_coefs_nat_df_1p,
cols = !1, names_to = "coef")Code
hour_coefs_nat_df_2p <- harmonics_nat_df_2p %>%
mutate(shape = tentative_shape + log_sl,
scale = 1/((1/tentative_scale) - sl),
kappa = tentative_kappa + cos_ta)
# save the coefficients to use in the simulations
write_csv(hour_coefs_nat_df_2p,
paste0("outputs/TwoStep_2pDaily_coefs_dry_",Sys.Date(), ".csv"))
# turning into a long data frame
hour_coefs_nat_long_2p <- pivot_longer(hour_coefs_nat_df_2p, cols = !1,
names_to = "coef")Code
hour_coefs_nat_df_3p <- harmonics_nat_df_3p %>%
mutate(shape = tentative_shape + log_sl,
scale = 1/((1/tentative_scale) - sl),
kappa = tentative_kappa + cos_ta)
# save the coefficients to use in the simulations
write_csv(hour_coefs_nat_df_3p,
paste0("outputs/TwoStep_3pDaily_coefs_dry_", Sys.Date(), ".csv"))
# turning into a long data frame
hour_coefs_nat_long_3p <- pivot_longer(hour_coefs_nat_df_3p, cols = !1,
names_to = "coef")Plot the natural-scale temporally dynamic coefficients
Now that the coefficients are in their natural scales, they will be larger or smaller depending on the scale of the covariate.
Plot just the habitat selection coefficients.
Code
ggplot() +
geom_path(data = hour_coefs_nat_long_0p %>%
filter(!coef %in% c("shape", "scale", "kappa")),
aes(x = hour, y = value, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(Time-varying~parameter~values~beta)) +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_color_discrete("Estimate") +
theme_classic() +
theme(legend.position = "bottom")
Code
ggplot() +
geom_path(data = hour_coefs_nat_long_1p %>%
filter(!coef %in% c("shape", "scale", "kappa")),
aes(x = hour, y = value, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(Time-varying~parameter~values~beta)) +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_color_discrete("Estimate") +
theme_classic() +
theme(legend.position = "bottom")
Code
ggplot() +
geom_path(data = hour_coefs_nat_long_2p %>%
filter(!coef %in% c("shape", "scale", "kappa")),
aes(x = hour, y = value, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(Time-varying~parameter~values~beta)) +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_color_discrete("Estimate") +
theme_classic() +
theme(legend.position = "bottom")
Code
ggplot() +
geom_path(data = hour_coefs_nat_long_3p %>%
filter(!coef %in% c("shape", "scale", "kappa")),
aes(x = hour, y = value, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(Time-varying~parameter~values~beta)) +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_color_discrete("Estimate") +
theme_classic() +
theme(legend.position = "bottom")
Plot only the temporally dynamic movement parameters
Code
ggplot() +
geom_path(data = hour_coefs_nat_long_0p %>%
filter(coef %in% c("shape", "kappa")),
aes(x = hour, y = value, colour = coef)) +
geom_path(data = hour_coefs_nat_long_0p %>%
filter(coef == "scale"),
aes(x = hour, y = value/1000, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(beta)) +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
ggtitle("Note that the scale parameter is divided by 1000 for plotting") +
scale_color_discrete("Estimate",
labels = c("kappa" = "von Mises kappa",
"scale" = "Gamma scale / 1000",
"shape" = "Gamma shape")) +
theme_classic() +
theme(legend.position = "right")
Code
ggplot() +
geom_path(data = hour_coefs_nat_long_1p %>%
filter(coef %in% c("shape", "kappa")),
aes(x = hour, y = value, colour = coef)) +
geom_path(data = hour_coefs_nat_long_1p %>%
filter(coef == "scale"),
aes(x = hour, y = value/1000, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(beta)) +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
ggtitle("Note that the scale parameter is divided by 1000 for plotting") +
scale_color_discrete("Estimate",
labels = c("kappa" = "von Mises kappa",
"scale" = "Gamma scale / 1000",
"shape" = "Gamma shape")) +
theme_classic() +
theme(legend.position = "right")
Code
ggplot() +
geom_path(data = hour_coefs_nat_long_2p %>%
filter(coef %in% c("shape", "kappa")),
aes(x = hour, y = value, colour = coef)) +
geom_path(data = hour_coefs_nat_long_2p %>%
filter(coef == "scale"),
aes(x = hour, y = value/1000, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous("Value of parameter") +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
ggtitle("*Note that the scale parameter is divided by 1000 for plotting") +
scale_color_discrete("Estimate",
labels = c("kappa" = "von Mises kappa",
"scale" = "Gamma scale / 1000",
"shape" = "Gamma shape")) +
theme_classic() +
theme(legend.position = "right")
Code
ggplot() +
geom_path(data = hour_coefs_nat_long_3p %>%
filter(coef %in% c("shape", "kappa")),
aes(x = hour, y = value, colour = coef)) +
geom_path(data = hour_coefs_nat_long_3p %>%
filter(coef == "scale"),
aes(x = hour, y = value/1000, colour = coef)) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_y_continuous(expression(beta)) +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
ggtitle("Note that the scale parameter is divided by 1000 for plotting") +
scale_color_discrete("Estimate",
labels = c("kappa" = "von Mises kappa",
"scale" = "Gamma scale / 1000",
"shape" = "Gamma shape")) +
theme_classic() +
theme(legend.position = "right")
Sample from temporally dynamic movement parameters
Here we sample from the movement kernel to generate a distribution of step lengths for each hour of the day, to assess how well it matches the observed step lengths. This is the ‘selection-free’ movement kernel, so the step lengths and turning angles from the simulations will be different, as the steps will be conditioned on the habitat, but this is a useful diagnostic to assess whether the harmonics are capturing the observed movement dynamics.
Code
`summarise()` has grouped output by 'id'. You can override using the `.groups`
argument.Code
# number of samples at each hour (more = smoother plotting, but slower)
n <- 1e5
gamma_dist_list <- vector(mode = "list", length = nrow(hour_coefs_nat_df_0p))
gamma_mean <- c()
gamma_median <- c()
gamma_ratio <- c()
for(hour_no in 1:nrow(hour_coefs_nat_df_0p)) {
gamma_dist_list[[hour_no]] <- rgamma(n, shape = hour_coefs_nat_df_0p$shape[hour_no],
scale = hour_coefs_nat_df_0p$scale[hour_no])
gamma_mean[hour_no] <- mean(gamma_dist_list[[hour_no]])
gamma_median[hour_no] <- median(gamma_dist_list[[hour_no]])
gamma_ratio[hour_no] <- gamma_mean[hour_no] / gamma_median[hour_no]
}
gamma_df_0p <- data.frame(model = "0p",
hour = hour_coefs_nat_df_0p$hour,
mean = gamma_mean,
median = gamma_median,
ratio = gamma_ratio)
mean_sl_0p <- ggplot() +
geom_path(data = movement_summary_buffalo,
aes(x = hour, y = mean_sl, colour = factor(id))) +
geom_path(data = gamma_df_0p,
aes(x = hour, y = mean), colour = "red", linetype = "dashed") +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_y_continuous("Mean step length") +
scale_colour_viridis_d("Buffalo") +
ggtitle("Observed and modelled mean step length",
subtitle = "No harmonics") +
theme_classic() +
theme(legend.position = "right")
mean_sl_0p
Code
median_sl_0p <- ggplot() +
geom_path(data = movement_summary_buffalo,
aes(x = hour, y = median_sl, colour = factor(id))) +
geom_path(data = gamma_df_0p, aes(x = hour, y = median),
colour = "red", linetype = "dashed") +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_y_continuous("Median step length") +
scale_colour_viridis_d("Buffalo") +
ggtitle("Observed and modelled median step length",
subtitle = "No harmonics") +
theme_classic() +
theme(legend.position = "right")
median_sl_0p
Code
Code
Code
Code
gamma_dist_list <- vector(mode = "list", length = nrow(hour_coefs_nat_df_1p))
gamma_mean <- c()
gamma_median <- c()
gamma_ratio <- c()
for(hour_no in 1:nrow(hour_coefs_nat_df_1p)) {
gamma_dist_list[[hour_no]] <- rgamma(n,
shape = hour_coefs_nat_df_1p$shape[hour_no],
scale = hour_coefs_nat_df_1p$scale[hour_no])
gamma_mean[hour_no] <- mean(gamma_dist_list[[hour_no]])
gamma_median[hour_no] <- median(gamma_dist_list[[hour_no]])
gamma_ratio[hour_no] <- gamma_mean[hour_no] / gamma_median[hour_no]
}
gamma_df_1p <- data.frame(model = "1p",
hour = hour_coefs_nat_df_1p$hour,
mean = gamma_mean,
median = gamma_median,
ratio = gamma_ratio)
mean_sl_1p <- ggplot() +
geom_path(data = movement_summary_buffalo,
aes(x = hour, y = mean_sl, colour = factor(id))) +
geom_path(data = gamma_df_1p,
aes(x = hour, y = mean),
colour = "red", linetype = "dashed") +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_y_continuous("Mean step length") +
scale_colour_viridis_d("Buffalo") +
ggtitle("Observed and modelled mean step length",
subtitle = "One pair of harmonics") +
theme_classic() +
theme(legend.position = "none")
mean_sl_1p
Code
median_sl_1p <- ggplot() +
geom_path(data = movement_summary_buffalo,
aes(x = hour, y = median_sl, colour = factor(id))) +
geom_path(data = gamma_df_1p,
aes(x = hour, y = median),
colour = "red", linetype = "dashed") +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_y_continuous("Median step length") +
scale_colour_viridis_d("Buffalo") +
ggtitle("Observed and modelled median step length",
subtitle = "One pair of harmonics") +
theme_classic() +
theme(legend.position = "none")
median_sl_1p
Code
Code
Code
Code
gamma_dist_list <- vector(mode = "list", length = nrow(hour_coefs_nat_df_2p))
gamma_mean <- c()
gamma_median <- c()
gamma_ratio <- c()
for(hour_no in 1:nrow(hour_coefs_nat_df_2p)) {
gamma_dist_list[[hour_no]] <- rgamma(n,
shape = hour_coefs_nat_df_2p$shape[hour_no],
scale = hour_coefs_nat_df_2p$scale[hour_no])
gamma_mean[hour_no] <- mean(gamma_dist_list[[hour_no]])
gamma_median[hour_no] <- median(gamma_dist_list[[hour_no]])
gamma_ratio[hour_no] <- gamma_mean[hour_no] / gamma_median[hour_no]
}
gamma_df_2p <- data.frame(model = "2p",
hour = hour_coefs_nat_df_2p$hour,
mean = gamma_mean,
median = gamma_median,
ratio = gamma_ratio)
mean_sl_2p <- ggplot() +
geom_path(data = movement_summary_buffalo,
aes(x = hour, y = mean_sl, colour = factor(id))) +
geom_path(data = gamma_df_2p,
aes(x = hour, y = mean),
colour = "red", linetype = "dashed") +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_y_continuous("Mean step length") +
scale_colour_viridis_d("Buffalo") +
ggtitle("Observed and modelled mean step length",
subtitle = "Two pairs of harmonics") +
theme_classic() +
theme(legend.position = "none")
mean_sl_2p
Code
median_sl_2p <- ggplot() +
geom_path(data = movement_summary_buffalo,
aes(x = hour, y = median_sl, colour = factor(id))) +
geom_path(data = gamma_df_2p,
aes(x = hour, y = median),
colour = "red", linetype = "dashed") +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_y_continuous("Median step length") +
scale_colour_viridis_d("Buffalo") +
ggtitle("Observed and modelled median step length",
subtitle = "Two pairs of harmonics") +
theme_classic() +
theme(legend.position = "none")
median_sl_2p
Code
Code
Code
Code
gamma_dist_list <- vector(mode = "list", length = nrow(hour_coefs_nat_df_3p))
gamma_mean <- c()
gamma_median <- c()
gamma_ratio <- c()
for(hour_no in 1:nrow(hour_coefs_nat_df_3p)) {
gamma_dist_list[[hour_no]] <- rgamma(n,
shape = hour_coefs_nat_df_3p$shape[hour_no],
scale = hour_coefs_nat_df_3p$scale[hour_no])
gamma_mean[hour_no] <- mean(gamma_dist_list[[hour_no]])
gamma_median[hour_no] <- median(gamma_dist_list[[hour_no]])
gamma_ratio[hour_no] <- gamma_mean[hour_no] / gamma_median[hour_no]
}
gamma_df_3p <- data.frame(model = "3p",
hour = hour_coefs_nat_df_3p$hour,
mean = gamma_mean,
median = gamma_median,
ratio = gamma_ratio)
mean_sl_3p <- ggplot() +
geom_path(data = movement_summary_buffalo,
aes(x = hour, y = mean_sl, colour = factor(id))) +
geom_path(data = gamma_df_3p,
aes(x = hour, y = mean),
colour = "red", linetype = "dashed") +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_y_continuous("Mean step length") +
scale_colour_viridis_d("Buffalo") +
ggtitle("Observed and modelled mean step length",
subtitle = "Three pairs of harmonics") +
theme_classic() +
theme(legend.position = "none")
mean_sl_3p
Code
median_sl_3p <- ggplot() +
geom_path(data = movement_summary_buffalo,
aes(x = hour, y = median_sl, colour = factor(id))) +
geom_path(data = gamma_df_3p,
aes(x = hour, y = median),
colour = "red", linetype = "dashed") +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_y_continuous("Median step length") +
scale_colour_viridis_d("Buffalo") +
ggtitle("Observed and modelled median step length",
subtitle = "Three pairs of harmonics") +
theme_classic() +
theme(legend.position = "none")
median_sl_3p
Code
Code
Code
Creating selection surfaces
As we have both quadratic and harmonic terms in the model, we can reconstruct a ‘selection surface’ to visualise how the animal’s respond to environmental features changes through time.
To illustrate, if we don’t have temporal dynamics (as is the case for this model), then we have a coefficient for the linear term and a coefficient for the quadratic term. Using these, we can plot the selection curve at the scale of the environmental variable (in this case NDVI).
Using the natural scale coefficients from the model:
Code
# first get a sequence of NDVI values,
# starting from the minimum observed in the data to the maximum
ndvi_min <- min(buffalo_data$ndvi_temporal, na.rm = TRUE)
ndvi_max <- max(buffalo_data$ndvi_temporal, na.rm = TRUE)
ndvi_seq <- seq(ndvi_min, ndvi_max, by = 0.01)
# take the coefficients from the model and calculation the selection value
# for every NDVI value in this sequence
# we can separate to the linear term
ndvi_linear_selection <- hour_coefs_nat_df_0p$ndvi[1] * ndvi_seq
plot(x = ndvi_seq, y = ndvi_linear_selection,
main = "Selection for NDVI - linear term",
xlab = "NDVI", ylab = "Estimated selection")
lines(ndvi_seq, rep(0,length(ndvi_seq)), lty = "dashed")
Code
Code
# and the sum of both
ndvi_sum_selection <- ndvi_linear_selection + ndvi_quadratic_selection
plot(x = ndvi_seq, y = ndvi_sum_selection,
main = "Selection for NDVI - sum of linear and quadratic terms",
xlab = "NDVI", ylab = "Estimated selection")
lines(ndvi_seq, rep(0,length(ndvi_seq)), lty = "dashed")
When there are no temporal dynamics, then this quadratic curve will be the same throughout the day, but when we have temporally dynamic coefficients for both the linear term and the quadratic term, then we will have a curves that vary continuously throughout the day, which we can visualise as a selection surface.
Here we illustrate for the model with 2 pairs of harmonic terms.
For brevity we won’t plot the linear and quadratic terms separately, but we can do so if needed.
First for Hour 3
Code
hour_no <- 3
# we can separate to the linear term
ndvi_linear_selection <-
hour_coefs_nat_df_1p$ndvi[which(hour_coefs_nat_df_1p$hour == hour_no)] * ndvi_seq
# plot(x = ndvi_seq, y = ndvi_linear_selection,
# main = "Selection for NDVI - linear term",
# xlab = "NDVI", ylab = "Estimated selection")
# and the quadratic term
ndvi_quadratic_selection <-
(hour_coefs_nat_df_1p$ndvi_2[which(hour_coefs_nat_df_1p$hour == hour_no)] * (ndvi_seq ^ 2))
# plot(x = ndvi_seq, y = ndvi_quadratic_selection,
# main = "Selection for NDVI - quadratic term",
# xlab = "NDVI", ylab = "Estimated selection")
# and the sum of both
ndvi_sum_selection <- ndvi_linear_selection + ndvi_quadratic_selection
plot(x = ndvi_seq, y = ndvi_sum_selection,
main = "Selection for NDVI - sum of linear and quadratic terms",
xlab = "NDVI", ylab = "Estimated selection")
lines(ndvi_seq, rep(0,length(ndvi_seq)), lty = "dashed")
We can see that the coefficient at hour 3 shows highest selection for NDVI values slightly above 0.2, and the coefficient is mostly negative.
Secondly for Hour 12
Code
hour_no <- 12
# we can separate to the linear term
ndvi_linear_selection <-
hour_coefs_nat_df_1p$ndvi[which(hour_coefs_nat_df_1p$hour == hour_no)] * ndvi_seq
# plot(x = ndvi_seq, y = ndvi_linear_selection,
# main = "Selection for NDVI - linear term",
# xlab = "NDVI", ylab = "Estimated selection")
# and the quadratic term
ndvi_quadratic_selection <-
(hour_coefs_nat_df_1p$ndvi_2[which(hour_coefs_nat_df_1p$hour == hour_no)] * (ndvi_seq ^ 2))
# plot(x = ndvi_seq, y = ndvi_quadratic_selection,
# main = "Selection for NDVI - quadratic term",
# xlab = "NDVI", ylab = "Estimated selection")
# and the sum of both
ndvi_sum_selection <- ndvi_linear_selection + ndvi_quadratic_selection
plot(x = ndvi_seq, y = ndvi_sum_selection,
main = "Selection for NDVI - sum of linear and quadratic terms",
xlab = "NDVI", ylab = "Estimated selection")
lines(ndvi_seq, rep(0,length(ndvi_seq)), lty = "dashed")
Whereas for hour 12, the coefficient shows highest selection for NDVI values slightly above 0.4, and the coefficient is positive for NDVI values above 0.
We can imagine viewing these plots for every hour of the day, where each hour has a different quadratic curve, but this would be a lot of plots. We can also see it as a 3D surface, where the x-axis is the hour of the day, the y-axis is the NDVI value, and the z-axis (colour) is the coefficient value.
We simply index over the linear and quadratic terms and calculate the coefficient values at every time point.
NDVI selection surface
Code
ndvi_min <- min(buffalo_data$ndvi_temporal, na.rm = TRUE)
ndvi_max <- max(buffalo_data$ndvi_temporal, na.rm = TRUE)
ndvi_seq <- seq(ndvi_min, ndvi_max, by = 0.01)
# Create empty data frame
ndvi_fresponse_df <- data.frame(matrix(ncol = nrow(hour_coefs_nat_df_0p),
nrow = length(ndvi_seq)))
# loop over each time increment, calculating the selection values for each NDVI value
# and storing each time increment as a column in a dataframe that we can use for plotting
for(i in 1:nrow(hour_coefs_nat_df_0p)) {
# Assign the vector as a column to the dataframe
ndvi_fresponse_df[,i] <- (hour_coefs_nat_df_0p$ndvi[i] * ndvi_seq) +
(hour_coefs_nat_df_0p$ndvi_2[i] * (ndvi_seq ^ 2))
}
ndvi_fresponse_df <- data.frame(ndvi_seq, ndvi_fresponse_df)
colnames(ndvi_fresponse_df) <- c("ndvi", hour)
ndvi_fresponse_long <- pivot_longer(ndvi_fresponse_df,
cols = !1, names_to = "hour")
ndvi_contour_max <- max(ndvi_fresponse_long$value) # 0.7890195
ndvi_contour_min <- min(ndvi_fresponse_long$value) # -0.7945691
ndvi_contour_increment <- (ndvi_contour_max-ndvi_contour_min)/10
ndvi_quad_0p <- ggplot(data = ndvi_fresponse_long,
aes(x = as.numeric(hour), y = ndvi)) +
geom_point(aes(colour = value)) +
geom_contour(aes(z = value),
breaks = seq(ndvi_contour_increment,
ndvi_contour_max,
ndvi_contour_increment),
colour = "black", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value),
breaks = seq(-ndvi_contour_increment,
ndvi_contour_min,
-ndvi_contour_increment),
colour = "red", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value), breaks = 0, colour = "black", linewidth = 0.5) +
scale_x_continuous("Hour", breaks = seq(0,24,6)) +
scale_y_continuous("NDVI value", breaks = seq(-1, 1, 0.25)) +
scale_colour_viridis_c("Selection") +
ggtitle("Normalised Difference Vegetation Index (NDVI)") +
theme_classic() +
theme(legend.position = "none")
ndvi_quad_0p
Code
ndvi_min <- min(buffalo_data$ndvi_temporal, na.rm = TRUE)
ndvi_max <- max(buffalo_data$ndvi_temporal, na.rm = TRUE)
ndvi_seq <- seq(ndvi_min, ndvi_max, by = 0.01)
# Create empty data frame
ndvi_fresponse_df <- data.frame(matrix(ncol = nrow(hour_coefs_nat_df_1p),
nrow = length(ndvi_seq)))
for(i in 1:nrow(hour_coefs_nat_df_1p)) {
# Assign the vector as a column to the dataframe
ndvi_fresponse_df[,i] <- (hour_coefs_nat_df_1p$ndvi[i] * ndvi_seq) +
(hour_coefs_nat_df_1p$ndvi_2[i] * (ndvi_seq ^ 2))
}
ndvi_fresponse_df <- data.frame(ndvi_seq, ndvi_fresponse_df)
colnames(ndvi_fresponse_df) <- c("ndvi", hour)
ndvi_fresponse_long <- pivot_longer(ndvi_fresponse_df, cols = !1, names_to = "hour")
ndvi_contour_max <- max(ndvi_fresponse_long$value) # 0.7890195
ndvi_contour_min <- min(ndvi_fresponse_long$value) # -0.7945691
ndvi_contour_increment <- (ndvi_contour_max-ndvi_contour_min)/10
ndvi_quad_1p <- ggplot(data = ndvi_fresponse_long,
aes(x = as.numeric(hour), y = ndvi)) +
geom_point(aes(colour = value)) + # colour = "white"
geom_contour(aes(z = value),
breaks = seq(ndvi_contour_increment,
ndvi_contour_max,
ndvi_contour_increment),
colour = "black", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value),
breaks = seq(-ndvi_contour_increment,
ndvi_contour_min,
-ndvi_contour_increment),
colour = "red", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value), breaks = 0, colour = "black", linewidth = 0.5) +
scale_x_continuous("Hour", breaks = seq(0,24,6)) +
scale_y_continuous("NDVI value", breaks = seq(-1, 1, 0.25)) +
scale_colour_viridis_c("Selection") +
# ggtitle("Normalised Difference Vegetation Index (NDVI)") +
theme_classic() +
theme(legend.position = "none")
ndvi_quad_1p
Code
ndvi_min <- min(buffalo_data$ndvi_temporal, na.rm = TRUE)
ndvi_max <- max(buffalo_data$ndvi_temporal, na.rm = TRUE)
ndvi_seq <- seq(ndvi_min, ndvi_max, by = 0.01)
# Create empty data frame
ndvi_fresponse_df <- data.frame(matrix(ncol = nrow(hour_coefs_nat_df_2p),
nrow = length(ndvi_seq)))
for(i in 1:nrow(hour_coefs_nat_df_2p)) {
# Assign the vector as a column to the dataframe
ndvi_fresponse_df[,i] <- (hour_coefs_nat_df_2p$ndvi[i] * ndvi_seq) +
(hour_coefs_nat_df_2p$ndvi_2[i] * (ndvi_seq ^ 2))
}
ndvi_fresponse_df <- data.frame(ndvi_seq, ndvi_fresponse_df)
colnames(ndvi_fresponse_df) <- c("ndvi", hour)
ndvi_fresponse_long <- pivot_longer(ndvi_fresponse_df, cols = !1,
names_to = "hour")
ndvi_contour_max <- max(ndvi_fresponse_long$value) # 0.7890195
ndvi_contour_min <- min(ndvi_fresponse_long$value) # -0.7945691
ndvi_contour_increment <- (ndvi_contour_max-ndvi_contour_min)/10
ndvi_quad_2p <- ggplot(data = ndvi_fresponse_long,
aes(x = as.numeric(hour), y = ndvi)) +
geom_point(aes(colour = value)) + # colour = "white"
geom_contour(aes(z = value),
breaks = seq(ndvi_contour_increment,
ndvi_contour_max,
ndvi_contour_increment),
colour = "black", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value),
breaks = seq(-ndvi_contour_increment,
ndvi_contour_min,
-ndvi_contour_increment),
colour = "red", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value), breaks = 0, colour = "black", linewidth = 0.5) +
scale_x_continuous("Hour", breaks = seq(0,24,6)) +
scale_y_continuous("NDVI value", breaks = seq(-1, 1, 0.25)) +
scale_colour_viridis_c("Selection") +
# ggtitle("Normalised Difference Vegetation Index (NDVI)") +
theme_classic() +
theme(legend.position = "right")
ndvi_quad_2p
Code
ndvi_min <- min(buffalo_data$ndvi_temporal, na.rm = TRUE)
ndvi_max <- max(buffalo_data$ndvi_temporal, na.rm = TRUE)
ndvi_seq <- seq(ndvi_min, ndvi_max, by = 0.01)
# Create empty data frame
ndvi_fresponse_df <- data.frame(matrix(ncol = nrow(hour_coefs_nat_df_3p),
nrow = length(ndvi_seq)))
for(i in 1:nrow(hour_coefs_nat_df_3p)) {
# Assign the vector as a column to the dataframe
ndvi_fresponse_df[,i] <- (hour_coefs_nat_df_3p$ndvi[i] * ndvi_seq) +
(hour_coefs_nat_df_3p$ndvi_2[i] * (ndvi_seq ^ 2))
}
ndvi_fresponse_df <- data.frame(ndvi_seq, ndvi_fresponse_df)
colnames(ndvi_fresponse_df) <- c("ndvi", hour)
ndvi_fresponse_long <- pivot_longer(ndvi_fresponse_df, cols = !1,
names_to = "hour")
ndvi_contour_max <- max(ndvi_fresponse_long$value) # 0.7890195
ndvi_contour_min <- min(ndvi_fresponse_long$value) # -0.7945691
ndvi_contour_increment <- (ndvi_contour_max-ndvi_contour_min)/10
ndvi_quad_3p <- ggplot(data = ndvi_fresponse_long,
aes(x = as.numeric(hour), y = ndvi)) +
geom_point(aes(colour = value)) + # colour = "white"
geom_contour(aes(z = value),
breaks = seq(ndvi_contour_increment,
ndvi_contour_max,
ndvi_contour_increment),
colour = "black", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value),
breaks = seq(-ndvi_contour_increment,
ndvi_contour_min,
-ndvi_contour_increment),
colour = "red", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value), breaks = 0, colour = "black", linewidth = 0.5) +
scale_x_continuous("Hour", breaks = seq(0,24,6)) +
scale_y_continuous("NDVI value", breaks = seq(-1, 1, 0.25)) +
scale_colour_viridis_c("Selection") +
# ggtitle("Normalised Difference Vegetation Index (NDVI)") +
theme_classic() +
theme(legend.position = "none")
ndvi_quad_3p
Canopy cover selection surface
Code
canopy_min <- min(buffalo_data$canopy_01, na.rm = TRUE)
canopy_max <- max(buffalo_data$canopy_01, na.rm = TRUE)
canopy_seq <- seq(canopy_min, canopy_max, by = 0.01)
# Create empty data frame
canopy_fresponse_df <- data.frame(matrix(ncol = nrow(hour_coefs_nat_df_0p),
nrow = length(canopy_seq)))
for(i in 1:nrow(hour_coefs_nat_df_0p)) {
# Assign the vector as a column to the dataframe
canopy_fresponse_df[,i] <- (hour_coefs_nat_df_0p$canopy[i] * canopy_seq) +
(hour_coefs_nat_df_0p$canopy_2[i] * (canopy_seq ^ 2))
}
canopy_fresponse_df <- data.frame(canopy_seq, canopy_fresponse_df)
colnames(canopy_fresponse_df) <- c("canopy", hour)
canopy_fresponse_long <- pivot_longer(canopy_fresponse_df,
cols = !1,
names_to = "hour")
canopy_contour_min <- min(canopy_fresponse_long$value) # 0
canopy_contour_max <- max(canopy_fresponse_long$value) # 2.181749
canopy_contour_increment <- (canopy_contour_max-canopy_contour_min)/10
canopy_quad_0p <- ggplot(data = canopy_fresponse_long, aes(x = as.numeric(hour),
y = canopy)) +
geom_point(aes(colour = value)) +
geom_contour(aes(z = value),
breaks = seq(canopy_contour_increment, canopy_contour_max,
canopy_contour_increment),
colour = "black", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value),
breaks = seq(-canopy_contour_increment, canopy_contour_min,
-canopy_contour_increment),
colour = "red", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value), breaks = 0, colour = "black", linewidth = 0.5) +
scale_x_continuous("Hour", breaks = seq(0,24,6)) +
scale_y_continuous("Canopy cover", breaks = seq(0, 1, 0.25)) +
scale_colour_viridis_c("Selection") +
ggtitle("Canopy Cover") +
theme_classic() +
theme(legend.position = "none")
canopy_quad_0p
Code
canopy_min <- min(buffalo_data$canopy_01, na.rm = TRUE)
canopy_max <- max(buffalo_data$canopy_01, na.rm = TRUE)
canopy_seq <- seq(canopy_min, canopy_max, by = 0.01)
# Create empty data frame
canopy_fresponse_df <- data.frame(matrix(ncol = nrow(hour_coefs_nat_df_1p),
nrow = length(canopy_seq)))
for(i in 1:nrow(hour_coefs_nat_df_1p)) {
# Assign the vector as a column to the dataframe
canopy_fresponse_df[,i] <- (hour_coefs_nat_df_1p$canopy[i] * canopy_seq) +
(hour_coefs_nat_df_1p$canopy_2[i] * (canopy_seq ^ 2))
}
canopy_fresponse_df <- data.frame(canopy_seq, canopy_fresponse_df)
colnames(canopy_fresponse_df) <- c("canopy", hour)
canopy_fresponse_long <- pivot_longer(canopy_fresponse_df, cols = !1,
names_to = "hour")
canopy_contour_min <- min(canopy_fresponse_long$value) # 0
canopy_contour_max <- max(canopy_fresponse_long$value) # 2.181749
canopy_contour_increment <- (canopy_contour_max-canopy_contour_min)/10
canopy_quad_1p <- ggplot(data = canopy_fresponse_long,
aes(x = as.numeric(hour), y = canopy)) +
geom_point(aes(colour = value)) +
geom_contour(aes(z = value),
breaks = seq(canopy_contour_increment,
canopy_contour_max,
canopy_contour_increment),
colour = "black", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value),
breaks = seq(-canopy_contour_increment,
canopy_contour_min,
-canopy_contour_increment),
colour = "red", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value), breaks = 0, colour = "black", linewidth = 0.5) +
scale_x_continuous("Hour", breaks = seq(0,24,6)) +
scale_y_continuous("Canopy cover", breaks = seq(0, 1, 0.25)) +
scale_colour_viridis_c("Selection") +
# ggtitle("Canopy Cover") +
theme_classic() +
theme(legend.position = "none")
canopy_quad_1p
Code
canopy_min <- min(buffalo_data$canopy_01, na.rm = TRUE)
canopy_max <- max(buffalo_data$canopy_01, na.rm = TRUE)
canopy_seq <- seq(canopy_min, canopy_max, by = 0.01)
# Create empty data frame
canopy_fresponse_df <- data.frame(matrix(ncol = nrow(hour_coefs_nat_df_2p),
nrow = length(canopy_seq)))
for(i in 1:nrow(hour_coefs_nat_df_2p)) {
# Assign the vector as a column to the dataframe
canopy_fresponse_df[,i] <- (hour_coefs_nat_df_2p$canopy[i] * canopy_seq) +
(hour_coefs_nat_df_2p$canopy_2[i] * (canopy_seq ^ 2))
}
canopy_fresponse_df <- data.frame(canopy_seq, canopy_fresponse_df)
colnames(canopy_fresponse_df) <- c("canopy", hour)
canopy_fresponse_long <- pivot_longer(canopy_fresponse_df, cols = !1,
names_to = "hour")
canopy_contour_min <- min(canopy_fresponse_long$value) # 0
canopy_contour_max <- max(canopy_fresponse_long$value) # 2.181749
canopy_contour_increment <- (canopy_contour_max-canopy_contour_min)/10
canopy_quad_2p <- ggplot(data = canopy_fresponse_long,
aes(x = as.numeric(hour), y = canopy)) +
geom_point(aes(colour = value)) +
geom_contour(aes(z = value),
breaks = seq(canopy_contour_increment,
canopy_contour_max,
canopy_contour_increment),
colour = "black", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value),
breaks = seq(-canopy_contour_increment,
canopy_contour_min,
-canopy_contour_increment),
colour = "red", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value), breaks = 0, colour = "black", linewidth = 0.5) +
scale_x_continuous("Hour", breaks = seq(0,24,6)) +
scale_y_continuous("Canopy cover", breaks = seq(0, 1, 0.25)) +
scale_colour_viridis_c("Selection") +
# ggtitle("Canopy Cover") +
theme_classic() +
theme(legend.position = "none")
canopy_quad_2p
Code
canopy_min <- min(buffalo_data$canopy_01, na.rm = TRUE)
canopy_max <- max(buffalo_data$canopy_01, na.rm = TRUE)
canopy_seq <- seq(canopy_min, canopy_max, by = 0.01)
# Create empty data frame
canopy_fresponse_df <- data.frame(matrix(ncol = nrow(hour_coefs_nat_df_3p),
nrow = length(canopy_seq)))
for(i in 1:nrow(hour_coefs_nat_df_3p)) {
# Assign the vector as a column to the dataframe
canopy_fresponse_df[,i] <- (hour_coefs_nat_df_3p$canopy[i] * canopy_seq) +
(hour_coefs_nat_df_3p$canopy_2[i] * (canopy_seq ^ 2))
}
canopy_fresponse_df <- data.frame(canopy_seq, canopy_fresponse_df)
colnames(canopy_fresponse_df) <- c("canopy", hour)
canopy_fresponse_long <- pivot_longer(canopy_fresponse_df, cols = !1,
names_to = "hour")
canopy_contour_min <- min(canopy_fresponse_long$value) # 0
canopy_contour_max <- max(canopy_fresponse_long$value) # 2.181749
canopy_contour_increment <- (canopy_contour_max-canopy_contour_min)/10
canopy_quad_3p <- ggplot(data = canopy_fresponse_long,
aes(x = as.numeric(hour), y = canopy)) +
geom_point(aes(colour = value)) +
geom_contour(aes(z = value),
breaks = seq(canopy_contour_increment,
canopy_contour_max,
canopy_contour_increment),
colour = "black", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value),
breaks = seq(-canopy_contour_increment,
canopy_contour_min,
-canopy_contour_increment),
colour = "red", linewidth = 0.25, linetype = "dashed") +
geom_contour(aes(z = value), breaks = 0, colour = "black", linewidth = 0.5) +
scale_x_continuous("Hour", breaks = seq(0,24,6)) +
scale_y_continuous("Canopy cover", breaks = seq(0, 1, 0.25)) +
scale_colour_viridis_c("Selection") +
# ggtitle("Canopy Cover",
# subtitle = "Three pairs of harmonics") +
theme_classic() +
theme(legend.position = "none")
canopy_quad_3p
Combining the plots
Movement parameters
Code
gamma_df <- rbind(gamma_df_0p, gamma_df_1p, gamma_df_2p, gamma_df_3p)
gamma_df <- gamma_df %>% mutate(model_f = as.numeric(factor(model)))
mean_sl <- ggplot() +
geom_path(data = movement_summary_buffalo,
aes(x = hour, y = mean_sl, group = factor(id)),
alpha = 0.25) +
geom_path(data = gamma_df, aes(x = hour, y = mean, linetype = model)) +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_y_continuous("Mean step length") +
scale_linetype_manual("Model", breaks=c("0p","1p", "2p", "3p"),
values=c(4,3,2,1)) +
ggtitle("Observed and modelled mean step length") +
theme_classic() +
theme(legend.position = "bottom")
mean_sl
Code
# ggsave(paste0("outputs/plots/manuscript_figs_R1/mean_sl_",
# Sys.Date(), ".png"),
# width=150, height=90, units="mm", dpi = 1000)
median_sl <- ggplot() +
geom_path(data = movement_summary_buffalo,
aes(x = hour, y = median_sl, group = factor(id)),
alpha = 0.25) +
geom_path(data = gamma_df, aes(x = hour, y = median, linetype = model)) +
scale_x_continuous("Hour", breaks = seq(0,24,2)) +
scale_y_continuous("Median step length") +
scale_linetype_manual("Model", breaks=c("0p","1p", "2p", "3p"),
values=c(4,3,2,1)) +
ggtitle("Observed and modelled median step length") +
theme_classic() +
theme(legend.position = "bottom")
median_sl
Habitat selection
Code
harmonics_scaled_long_0p <- harmonics_scaled_long_0p %>% mutate(model = "0p")
harmonics_scaled_long_1p <- harmonics_scaled_long_1p %>% mutate(model = "1p")
harmonics_scaled_long_2p <- harmonics_scaled_long_2p %>% mutate(model = "2p")
harmonics_scaled_long_3p <- harmonics_scaled_long_3p %>% mutate(model = "3p")
harmonics_scaled_long_Mp <- rbind(harmonics_scaled_long_0p,
harmonics_scaled_long_1p,
harmonics_scaled_long_2p,
harmonics_scaled_long_3p)
coef_titles <- unique(harmonics_scaled_long_Mp$coef)
ndvi_harms <- ggplot() +
geom_path(data = harmonics_scaled_long_Mp %>%
filter(coef == "ndvi"),
aes(x = hour, y = value, linetype = model)) +
geom_hline(yintercept = 0, linetype = "dashed", colour = "red") +
scale_y_continuous(expression(beta)) +
scale_x_continuous("Hour") +
scale_linetype_manual("Model", breaks=c("0p","1p", "2p", "3p"),
values=c(4,3,2,1)) +
ggtitle("NDVI") +
theme_classic() +
theme(legend.position = "bottom")
ndvi_harms
Code
ndvi_2_harms <- ggplot() +
geom_path(data = harmonics_scaled_long_Mp %>%
filter(coef == "ndvi_2"),
aes(x = hour, y = value, linetype = model)) +
geom_hline(yintercept = 0, linetype = "dashed", colour = "red") +
scale_y_continuous(expression(beta)) +
scale_x_continuous("Hour") +
scale_linetype_manual("Model", breaks=c("0p","1p", "2p", "3p"),
values=c(4,3,2,1)) +
ggtitle(expression(NDVI^2)) +
theme_classic() +
theme(legend.position = "bottom")
ndvi_2_harms
Code
canopy_harms <- ggplot() +
geom_path(data = harmonics_scaled_long_Mp %>%
filter(coef == "canopy"),
aes(x = hour, y = value, linetype = model)) +
geom_hline(yintercept = 0, linetype = "dashed", colour = "red") +
scale_y_continuous(expression(beta)) +
scale_x_continuous("Hour") +
scale_linetype_manual("Model", breaks=c("0p","1p", "2p", "3p"),
values=c(4,3,2,1)) +
ggtitle("Canopy cover") +
theme_classic() +
theme(legend.position = "bottom")
canopy_harms
Code
canopy_2_harms <- ggplot() +
geom_path(data = harmonics_scaled_long_Mp %>%
filter(coef == "canopy_2"),
aes(x = hour, y = value, linetype = model)) +
geom_hline(yintercept = 0, linetype = "dashed", colour = "red") +
scale_y_continuous(expression(beta)) +
scale_x_continuous("Hour") +
scale_linetype_manual("Model", breaks=c("0p","1p", "2p", "3p"),
values=c(4,3,2,1)) +
ggtitle(expression(Canopy~cover^2)) +
theme_classic() +
theme(legend.position = "bottom")
canopy_2_harms
Code
herby_harms <- ggplot() +
geom_path(data = harmonics_scaled_long_Mp %>%
filter(coef == "herby"),
aes(x = hour, y = value, linetype = model)) +
geom_hline(yintercept = 0, linetype = "dashed", colour = "red") +
scale_y_continuous(expression(beta), limits = c(-0.4,0.15)) +
scale_x_continuous("Hour") +
scale_linetype_manual("Model", breaks=c("0p","1p", "2p", "3p"),
values=c(4,3,2,1)) +
ggtitle("Herbaceous vegetation") +
theme_classic() +
theme(legend.position = "bottom")
herby_harms
Code
slope_harms <- ggplot() +
geom_path(data = harmonics_scaled_long_Mp %>%
filter(coef == "slope"),
aes(x = hour, y = value, linetype = model)) +
geom_hline(yintercept = 0, linetype = "dashed", colour = "red") +
scale_y_continuous(expression(beta), limits = c(-0.4,0.15)) +
scale_x_continuous("Hour") +
scale_linetype_manual("Model", breaks=c("0p","1p", "2p", "3p"),
values=c(4,3,2,1)) +
ggtitle("Slope") +
theme_classic() +
theme(legend.position = "bottom")
slope_harms
Code
Combining selection surfaces
NDVI
- A = 0p model
- B = 1p model
- C = 2p model
- D = 3p model
Code
ggarrange(ndvi_quad_0p + theme(plot.title = element_blank(),
axis.title.x = element_blank(),
axis.text.x = element_blank()),
ndvi_quad_1p + theme(plot.title = element_blank(),
axis.title.x = element_blank(),
axis.text.x = element_blank(),
axis.title.y = element_blank(),
),
ndvi_quad_2p,
ndvi_quad_3p + theme(plot.title = element_blank(),
axis.title.y = element_blank(),
),
labels = c("A", "B", "C", "D"),
ncol = 2, nrow = 2,
legend = "bottom",
common.legend = TRUE)
Canopy cover
- A = 0p model
- B = 1p model
- C = 2p model
- D = 3p model
Code
ggarrange(canopy_quad_0p + theme(plot.title = element_blank(),
axis.title.x = element_blank(),
axis.text.x = element_blank()),
canopy_quad_1p + theme(plot.title = element_blank(),
axis.title.x = element_blank(),
axis.text.x = element_blank(),
axis.title.y = element_blank(),
),
canopy_quad_2p,
canopy_quad_3p + theme(plot.title = element_blank(),
axis.title.y = element_blank(),
),
labels = c("A", "B", "C", "D"),
ncol = 2, nrow = 2,
legend = "bottom",
common.legend = TRUE)
Adding all selection surfaces to the same plot
We combine these plots into the plot that is in the paper. On the top is the NDVI selection surface, and on the bottom is the canopy cover selection surface.
Code
surface_plots_0p <- ggarrange(ndvi_quad_0p +
ggtitle("0p") +
theme(axis.title.x = element_blank(),
axis.text.x = element_blank()),
canopy_quad_0p +
scale_x_continuous("Hour", breaks = c(0,12,24)) +
theme(plot.title = element_blank()),
ncol = 1, nrow = 2,
align = "v",
legend = "none",
common.legend = TRUE)Scale for x is already present.
Adding another scale for x, which will replace the existing scale.
Code
surface_plots_1p <- ggarrange(ndvi_quad_1p +
ggtitle("1p") +
theme(axis.title.x = element_blank(),
axis.text.x = element_blank(),
axis.title.y = element_blank(),
axis.text.y = element_blank()),
canopy_quad_1p +
theme(plot.title = element_blank(),
axis.title.y = element_blank(),
axis.text.y = element_blank()),
ncol = 1, nrow = 2,
align = "v",
legend = "none",
common.legend = TRUE)
surface_plots_1p
Code
surface_plots_2p <- ggarrange(ndvi_quad_2p +
ggtitle("2p") +
theme(axis.title.x = element_blank(),
axis.text.x = element_blank(),
axis.title.y = element_blank(),
axis.text.y = element_blank()),
canopy_quad_2p +
theme(plot.title = element_blank(),
axis.title.y = element_blank(),
axis.text.y = element_blank()),
ncol = 1, nrow = 2,
align = "v",
legend = "none",
common.legend = TRUE)
surface_plots_2p
Code
surface_plots_3p <- ggarrange(ndvi_quad_3p +
ggtitle("3p") +
theme(axis.title.x = element_blank(),
axis.text.x = element_blank(),
axis.title.y = element_blank(),
axis.text.y = element_blank()),
canopy_quad_3p +
theme(plot.title = element_blank(),
axis.title.y = element_blank(),
axis.text.y = element_blank()),
ncol = 1, nrow = 2,
align = "v",
legend = "none",
common.legend = TRUE)
surface_plots_3p
All selection surfaces
Code
References
Brooks, Mollie E, Kasper Kristensen, Koen J van Benthem, Arni Magnusson, Casper W Berg, Anders Nielsen, Hans J Skaug, Martin Mächler, and Benjamin M Bolker. 2017. “glmmTMB Balances Speed and Flexibility Among Packages for Zero-Inflated Generalized Linear Mixed Modeling.” The R Journal 9 (2): 378–400.
Craiu, Radu V, Thierry Duchesne, Daniel Fortin, and Sophie Baillargeon. 2011. “Conditional Logistic Regression with Longitudinal Follow-up and Individual-Level Random Coefficients: A Stable and Efficient Two-Step Estimation Method.” Journal of Computational and Graphical Statistics 20 (3): 767–84. https://doi.org/10.1198/jcgs.2011.09189.
———. 2016. “R Package ’TwoStepCLogit’. Conditional Logistic Regression: A Two-Step Estimation Method.”
Fieberg, John, Johannes Signer, Brian Smith, and Tal Avgar. 2021. “A ’How to’ Guide for Interpreting Parameters in Habitat-Selection Analyses.” The Journal of Animal Ecology 90 (5): 1027–43. https://doi.org/10.1111/1365-2656.13441.
Muff, Stefanie, Johannes Signer, and John Fieberg. 2020. “Accounting for Individual-Specific Variation in Habitat-Selection Studies: Efficient Estimation of Mixed-Effects Models Using Bayesian or Frequentist Computation.” Edited by Eric Vander Wal. The Journal of Animal Ecology 89 (1): 80–92. https://doi.org/10.1111/1365-2656.13087.
Rue, Håvard, Sara Martino, and Nicolas Chopin. 2009. “Approximate Bayesian Inference for Latent Gaussian Models by Using Integrated Nested Laplace Approximations.” Journal of the Royal Statistical Society. Series B, Statistical Methodology 71 (2): 319–92. https://doi.org/10.1111/j.1467-9868.2008.00700.x.
Session info
R version 4.4.1 (2024-06-14 ucrt)
Platform: x86_64-w64-mingw32/x64
Running under: Windows 11 x64 (build 22631)
Matrix products: default
locale:
[1] LC_COLLATE=English_Australia.utf8 LC_CTYPE=English_Australia.utf8
[3] LC_MONETARY=English_Australia.utf8 LC_NUMERIC=C
[5] LC_TIME=English_Australia.utf8
time zone: Europe/Zurich
tzcode source: internal
attached base packages:
[1] stats graphics grDevices utils datasets methods base
other attached packages:
[1] formatR_1.14 scales_1.3.0 sjPlot_2.8.16
[4] glmmTMB_1.1.9 patchwork_1.2.0 MASS_7.3-60.2
[7] ggpubr_0.6.0 clogitL1_1.5 Rcpp_1.0.13
[10] beepr_2.0 ecospat_4.1.1 TwoStepCLogit_1.2.5
[13] tictoc_1.2.1 terra_1.7-78 survival_3.6-4
[16] lubridate_1.9.3 forcats_1.0.0 stringr_1.5.1
[19] dplyr_1.1.4 purrr_1.0.2 readr_2.1.5
[22] tidyr_1.3.1 tibble_3.2.1 ggplot2_3.5.1
[25] tidyverse_2.0.0
loaded via a namespace (and not attached):
[1] gridExtra_2.3 rlang_1.1.4 magrittr_2.0.3
[4] e1071_1.7-14 compiler_4.4.1 mgcv_1.9-1
[7] vctrs_0.6.5 pkgconfig_2.0.3 crayon_1.5.3
[10] fastmap_1.2.0 backports_1.5.0 labeling_0.4.3
[13] utf8_1.2.4 rmarkdown_2.28 tzdb_0.4.0
[16] nloptr_2.1.1 bit_4.0.5 xfun_0.47
[19] jsonlite_1.8.8 sjmisc_2.8.10 ggeffects_1.7.0
[22] broom_1.0.6 parallel_4.4.1 R6_2.5.1
[25] stringi_1.8.4 car_3.1-2 boot_1.3-30
[28] numDeriv_2016.8-1.1 iterators_1.0.14 knitr_1.48
[31] audio_0.1-11 Matrix_1.7-0 splines_4.4.1
[34] timechange_0.3.0 tidyselect_1.2.1 abind_1.4-5
[37] yaml_2.3.10 TMB_1.9.14 codetools_0.2-20
[40] sjlabelled_1.2.0 lattice_0.22-6 withr_3.0.1
[43] evaluate_0.24.0 proxy_0.4-27 isoband_0.2.7
[46] pillar_1.9.0 carData_3.0-5 KernSmooth_2.23-24
[49] foreach_1.5.2 insight_0.20.3 generics_0.1.3
[52] vroom_1.6.5 hms_1.1.3 munsell_0.5.1
[55] minqa_1.2.8 class_7.3-22 glue_1.7.0
[58] tools_4.4.1 lme4_1.1-35.5 ggsignif_0.6.4
[61] cowplot_1.1.3 grid_4.4.1 datawizard_0.12.2
[64] colorspace_2.1-1 nlme_3.1-164 performance_0.12.2
[67] cli_3.6.3 fansi_1.0.6 viridisLite_0.4.2
[70] sjstats_0.19.0 gtable_0.3.5 rstatix_0.7.2
[73] digest_0.6.37 classInt_0.4-10 htmlwidgets_1.6.4
[76] farver_2.1.2 htmltools_0.5.8.1 lifecycle_1.0.4
[79] bit64_4.0.5