trendeval aims to provide a coherent interface for evaluating models fit with the trending package. Whilst it is useful in an interactive context, it’s main focus is to provide an intuitive interface on which other packages can be developed (e.g. trendbreaker).

Installing the package

You can install the stable version of this package from CRAN with:

install.packages("trendeval")

The development version can be installed from GitHub with:

if (!require(remotes)) {
  install.packages("remotes")
}
remotes::install_github("reconhub/trendeval", build_vignettes = TRUE)

Model selection example

library(dplyr)      # for data manipulation
library(outbreaks)  # for data
library(trending)   # for trend fitting
library(trendeval)  # for model selection

# load data
data(covid19_england_nhscalls_2020)

# define a selection of model in a named list
models  <- list(
  simple = lm_model(count ~ day),
  glm_poisson = glm_model(count ~ day, family = "poisson"),
  glm_poisson_weekday = glm_model(count ~ day + weekday, family = "quasipoisson"),
  glm_quasipoisson = glm_model(count ~ day, family = "poisson"),
  glm_quasipoisson_weekday = glm_model(count ~ day + weekday, family = "quasipoisson"),
  glm_negbin = glm_nb_model(count ~ day),
  glm_negbin_weekday = glm_nb_model(count ~ day + weekday),
  will_error = glm_nb_model(count ~ day + nonexistant)
)

# select 8 weeks of data (from a period when the prevalence was decreasing)
last_date <- as.Date("2020-05-28")
first_date <- last_date - 8*7
pathways_recent <-
  covid19_england_nhscalls_2020 %>%
  filter(date >= first_date, date <= last_date) %>%
  group_by(date, day, weekday) %>%
  summarise(count = sum(count), .groups = "drop")

# split data for fitting and prediction
dat <-
  pathways_recent %>%
  group_by(date <= first_date + 6*7) %>%
  group_split()
fitting_data <- dat[[2]]
pred_data <- select(dat[[1]], date, day, weekday)

# assess the models using the evaluate_resampling and a variety of metrics
results <- evaluate_models(
  models,
  fitting_data, 
  method = evaluate_resampling,
  metrics = list(yardstick::rmse, yardstick::huber_loss, yardstick::mae)
)

results
#> # A tibble: 8 x 8
#>   model_name       model    data        warning   error   huber_loss   mae  rmse
#>   <chr>            <named > <list>      <named l> <named>      <dbl> <dbl> <dbl>
#> 1 simple           <trndng… <tibble [4… <NULL>    <NULL>       6902. 6903. 6903.
#> 2 glm_poisson      <trndng… <tibble [4… <NULL>    <NULL>       5193. 5193. 5193.
#> 3 glm_poisson_wee… <trndng… <tibble [4… <NULL>    <NULL>       5166. 5166. 5166.
#> 4 glm_quasipoisson <trndng… <tibble [4… <NULL>    <NULL>       5193. 5193. 5193.
#> 5 glm_quasipoisso… <trndng… <tibble [4… <NULL>    <NULL>       5166. 5166. 5166.
#> 6 glm_negbin       <trndng… <tibble [4… <NULL>    <NULL>       5238. 5238. 5238.
#> 7 glm_negbin_week… <trndng… <tibble [4… <NULL>    <NULL>       5223. 5224. 5224.
#> 8 will_error       <trndng… <tibble [4… <NULL>    <chr […        NA    NA    NA

Example of how this output could then be used

library(tidyr)      # for data manipulation
library(purrr)      # for data manipulation
library(ggplot2)    # for plotting

# Pull out the model with the lowest RMSE
best_by_rmse <- 
    results %>% 
    filter(map_lgl(warning, is.null)) %>%  # remove models that gave warnings
    filter(map_lgl(error, is.null))  %>%   # remove models that errored
    slice_min(rmse) %>% 
    select(model) %>% 
    pluck(1,1)

best_by_rmse
#> $model_class
#> [1] "glm"
#> 
#> $fit
#> function (data) 
#> {
#>     model <- glm(formula = count ~ day + weekday, family = "quasipoisson", 
#>         data = data, ...)
#>     model_fit(model, data)
#> }
#> <environment: 0x55e7220706f0>
#> 
#> attr(,"class")
#> [1] "trending_glm"   "trending_model"

# Now let's look at the following 14 days as well
new_dat <-
  covid19_england_nhscalls_2020 %>% 
  filter(date > "2020-05-28", date <= "2020-06-11") %>% 
  group_by(date, day, weekday) %>%
  summarise(count = sum(count), .groups = "drop")

all_dat <- bind_rows(pathways_recent, new_dat)

out <- 
  best_by_rmse %>%  
  fit(pathways_recent) %>% 
  predict(all_dat) %>%  
  as_tibble()

out
#> # A tibble: 71 x 9
#>    date         day weekday   count estimate lower_ci upper_ci lower_pi upper_pi
#>    <date>     <int> <fct>     <int>    <dbl>    <dbl>    <dbl>    <dbl>    <dbl>
#>  1 2020-04-02    15 rest_of_… 71917   58301.   53120.   63988.    40268    79935
#>  2 2020-04-03    16 rest_of_… 63365   56280.   51398.   61625.    38771    77291
#>  3 2020-04-04    17 weekend   52412   47107.   41966.   52878.    30644    67450
#>  4 2020-04-05    18 weekend   54014   45474.   40576.   50962.    29459    65284
#>  5 2020-04-06    19 monday    78996   62206.   54362.   71181.    41350    87955
#>  6 2020-04-07    20 rest_of_… 62026   48871.   45017.   53054.    33258    67650
#>  7 2020-04-08    21 rest_of_… 51692   47176.   43540.   51116.    31991    65458
#>  8 2020-04-09    22 rest_of_… 40797   45540.   42108.   49253.    30765    63346
#>  9 2020-04-10    23 rest_of_… 33946   43961.   40718.   47463.    29580    61313
#> 10 2020-04-11    24 weekend   32269   36796.   33092.   40916.    23141    53834
#> # … with 61 more rows

# plot output
ggplot(out, aes(x = date, y = count)) +
  geom_line() +
  geom_ribbon(mapping = aes(x = date, ymin = lower_ci, ymax = upper_ci),
              data = out, alpha = 0.5, fill = "#BBB67E") +
  geom_ribbon(mapping = aes(x = date, ymin = lower_pi, ymax = upper_pi),
              data = out, alpha = 0.5, fill = "#BBB67E") +
  geom_vline(xintercept = as.Date("2020-05-28") + 0.5) +
  theme_bw()

Resources

Getting help online

Bug reports and feature requests should be posted on github using the issue system. All other questions should be posted on the RECON slack channel see https://www.repidemicsconsortium.org/forum/ for details on how to join.