In reconhub/epichange: Detect Changes in Temporal Trends

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.align = "center",
  fig.path = "man/figures/README-",
  echo = TRUE,
  fig.width = 8,
  fig.height = 6
)

trendbreaker

The trendbreaker package implements tools for detecting changes in temporal trends of a single response variable. It implements the Automatic Selection of Models and Outlier Detection for Epidemmics (ASMODEE), an algorithm originally designed for detecting changes in COVID-19 case incidence.

ASMODEE proceeds by:

1. defining a training set excluding the last k data points
2. identifying the temporal trend in the training set by fitting a range of (user-specified) models to the data and retaining the best predicting / fitting model
3. calculating the prediction interval (PI) of the temporal trend
4. classifying any data point outside the PI as outlier

The algorithm can be applied with fixed, user-specified value of k, so as to monitor potential changes in this recent time period. Alternatively, the optimal value of k can be determined automatically.

Disclaimer: this is work in progress. Please reach out to the authors before using this package. Also note this package may soon be renamed to avoid clashes with other projects and reflect a more general scope.

Getting started

Once it is released on CRAN, you will be able to install the stable version of the package with:

install.packages("trendbreaker")

The development version can be installed from GitHub with:

if (!require(remotes)) {
  install.packages("remotes")
}
remotes::install_github("reconhub/trendbreaker")

The best place to start for using this package is to read the documentation of the function asmodee and run its example:

library(trendbreaker)
?asmodee
example(asmodee)

Main features

ASMODEE

We illustrate ASMODEE using publicly available NHS pathways data recording self-reporting of potential COVID-19 cases in England (see ?nhs_pathways_covid19 for more information).

library(trendbreaker) # for ASMODEE
library(dplyr)        # for data manipulation
library(future)
plan("multisession")

# load data
data(nhs_pathways_covid19)

# select last 6 weeks of data
first_date <- max(nhs_pathways_covid19$date, na.rm = TRUE) - 6*7
pathways_recent <- nhs_pathways_covid19 %>%
  filter(date >= first_date)

# define candidate models
models <- list(
  regression = lm_model(count ~ day),
  poisson_constant = glm_model(count ~ 1, family = "poisson"),
  negbin_time = glm_nb_model(count ~ day),
  negbin_time_weekday = glm_nb_model(count ~ day + weekday)
)

# analyses on all data
counts_overall <- pathways_recent %>%
  group_by(date, day, weekday) %>%
  summarise(count = sum(count))

# results with fixed 'k' = 7
res <- asmodee(
  counts_overall,
  models,
  k = 7,
  date_index = "date",
  method = evaluate_aic,
  simulate_pi = TRUE
)
res
plot(res, "date")

ASMODEE would typically be more useful to investigate shifts in temporal trends from a large number of time series (e.g. at a fine geographic scale). To make this sort of analysis easier trendbreaker also works with incidence2 objects. To illustrate this we can consider trends over NHS regions.

library(incidence2)   # for `incidence()` objects

# select last 6 weeks of data
first_date <- max(nhs_pathways_covid19$date, na.rm = TRUE) - 6*7
pathways_recent <- filter(nhs_pathways_covid19, date >= first_date)

# create incidence object with extra variables
lookup <- select(pathways_recent, date, day, weekday) %>%  distinct()

dat <-
  pathways_recent %>%
  incidence(date_index = date, groups = nhs_region, count = count) %>%
  left_join(lookup, by = c("date_index" = "date"))

# define candidate models
models <- list(
  regression = lm_model(count ~ day),
  poisson_constant = glm_model(count ~ 1, family = "poisson"),
  negbin_time = glm_nb_model(count ~ day),
  negbin_time_weekday = glm_nb_model(count ~ day + weekday)
)

# analyses on all data
res <- asmodee(dat, models, method = evaluate_aic, k = 7)

plot(res)

reconhub/epichange documentation built on April 28, 2021, 2 p.m.