2020-09-08-weekly-report.Rmd

---
title: "COVID-19 Forecast Hub weekly report"
author: "Nicholas G Reich, Estee Y Cramer, Martha W Zorn"
date: "report generated 2020-09-08, updated `r Sys.Date()`"
output:
  html_document:
    toc: true
    toc_float:
      collapsed: false
      smooth_scroll: false

---


```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = FALSE, message = FALSE, warning = FALSE)
library(lubridate)
library(DT)
library(zoltr) ## devtools::install_github("reichlab/zoltr")
library(scico)
source("../processing-fxns/get_next_saturday.R")
library(tidyverse)
library(htmltools)
theme_set(theme_bw())
```

```{r zoltar-setup}
## connect to Zoltar
zoltar_connection <- new_connection()
zoltar_authenticate(zoltar_connection, Sys.getenv("Z_USERNAME"), Sys.getenv("Z_PASSWORD"))

## construct Zoltar query
project_url <- "https://www.zoltardata.com/api/project/44/"
```


```{r get-date-boundaries}
next_saturday <- get_next_saturday(today())
# # use fixed date
# # use fixed date
next_saturday <-  as.Date("2020-09-12")
saturday_4_wk_ahead <- next_saturday + 7*3
saturday_4_wk_ahead_txt <- format(saturday_4_wk_ahead, "%B %d")
last_5_saturdays <- next_saturday - 7*c(5:1)
this_monday <- next_saturday - 5
```


# Background
This report provides a brief summary of the weekly ensemble forecast from the [COVID-19 Forecast Hub](https://covid19forecasthub.org/) based on forecasts submitted on `r format(this_monday, "%B %d, %Y")`. In collaboration with the US CDC, our team aggregates COVID-19 forecasts from dozens of teams around the globe. Typically on Wednesday or Thursday of each week, a summary of the week's forecasts from the COVID-19 Forecast Hub appear on the [official CDC COVID-19 forecasting page](https://www.cdc.gov/coronavirus/2019-ncov/covid-data/forecasting-us.html).


```{r nmodels-this-week}
models_this_week <- zoltr::do_zoltar_query(
  zoltar_connection, project_url, TRUE,
  targets = c("1 wk ahead cum death", "1 wk ahead inc death", "1 wk ahead inc case", "1 day ahead inc hosp"),
  timezeros = as.character(seq.Date(this_monday, this_monday-6, by="-1 day")),
  types = c("point"), 
  verbose = FALSE,
  models = c(),
  units = c()) %>%
  pull(model) %>%
  unique()
nmodels_this_week <- length(models_this_week)
```


Every week, teams submit their forecasts to the COVID-19 Forecast Hub. 
This past week, `r nmodels_this_week`* models were submitted. 

Each Monday evening or Tuesday morning, we combine the most recent forecasts from each team into a single "ensemble" forecast of reported COVID-19 cases at the county, state, and national level and deaths at the state and national level. At the moment, we only generate ensemble forecasts for four weeks into the future, as we don't have reliable evidence that the models are accurate past that horizon.

An archive of weekly reports from the COVID-19 Forecast Hub can be found at [this page](https://covid19forecasthub.org/doc/reports/). 

*(updated 9/17/2020 due to error found)
```{r count-models}
## how many models in inc_death ensemble?
inc_death_models <- read_csv(paste0("../../ensemble-metadata/", this_monday, "-inc_death-model-weights.csv")) %>%
 select(-locations) %>%
  apply(MARGIN = 2, FUN=function(x) sum(x))
n_inc_death_models <- sum(inc_death_models>0)

## how many models in cum_death ensemble?
cum_death_models <- read_csv(paste0("../../ensemble-metadata/", this_monday, "-cum_death-model-weights.csv")) %>%
  select(-locations) %>%
  apply(MARGIN = 2, FUN=function(x) sum(x))
n_cum_death_models <- sum(cum_death_models>0)

## how many models in inc_case ensemble?
inc_case_models <- read_csv(paste0("../../ensemble-metadata/", this_monday, "-inc_case-model-weights.csv"))%>%
  select(-locations) %>%
  apply(MARGIN = 2, FUN=function(x) sum(x))
n_inc_case_models <- sum(inc_case_models>0)

n_unique_models <- length(unique(c(names(inc_death_models)[inc_death_models>0],
  names(cum_death_models)[cum_death_models>0],
  names(inc_case_models)[inc_case_models>0])))
```

```{r aux-data}
## get our FIPS codes
fips_codes <- read_csv("../../data-locations/locations.csv")

## this has populations
locs <- read_csv("https://raw.githubusercontent.com/CSSEGISandData/COVID-19/master/csse_covid_19_data/UID_ISO_FIPS_LookUp_Table.csv") %>%
  mutate(FIPS = ifelse(UID==840, "US", FIPS)) %>%
  mutate(state_fips = substr(FIPS, 0, 2)) %>%
  right_join(select(fips_codes, -location_name), by=c("state_fips" = "location")) %>%
  mutate(loc_name = reorder(factor(Province_State), X=-Population)) %>%
  select(FIPS, abbreviation, Population) 
```


```{r download-ensemble-data}
inc_death_targets <- paste(1:4, "wk ahead inc death")
cum_death_targets <- paste(1:4, "wk ahead cum death")

# submit query
dat <- zoltr:: do_zoltar_query(
  zoltar_connection, project_url, TRUE,
  models = c("COVIDhub-ensemble"),
  targets = c(inc_death_targets, cum_death_targets),
  timezeros = as.character(this_monday),
  types = c("point", "quantile"), 
  units = c(),
  verbose = FALSE) %>%
  ## choose only columns we need and with data
  select(model, timezero, unit, target, class, quantile, value) %>%
  rename(fips=unit) %>%
  ## create rate variable and week-ahead
  mutate(week_ahead = as.numeric(substr(target, 0,1)),
    ## recreates the target_end_date from GitHub
    target_end_date = get_next_saturday(timezero + 7*(week_ahead-1))) 
```


# COVID-19 Mortality Forecasts 

## National level

This week, our ensemble combined forecasts from `r n_unique_models` different models. 

```{r us-summary}
us_cum_deaths <- dat %>% 
  filter(fips=="US", target=="4 wk ahead cum death", class=="point") %>% 
  pull(value) %>% 
  round(-2) %>% 
  format(big.mark = ",")
us_inc_death_range <- dat %>% 
  filter(fips=="US", target %in% inc_death_targets, class=="point") %>% 
  pull(value) %>% range() %>%
  round(-2) %>% format(big.mark = ",")

us_inc_death_wk_pi_round <- dat %>% 
  filter(fips=="US", target == "4 wk ahead inc death", quantile %in% c(0.025, 0.975)) %>% 
  pull(value) %>% sort() %>%
  round(-2) %>% format(big.mark = ",")

us_inc_death_wk_pi <- dat %>% 
  filter(fips=="US", target == "4 wk ahead inc death", quantile %in% c(0.025, 0.975)) %>% 
  pull(value) %>% sort() %>% 
  format(big.mark = ",")

```

At the national level, the ensemble model's best guess is that we will see between `r us_inc_death_range[1]` and `r us_inc_death_range[2]` deaths each week for the next four weeks with around `r us_cum_deaths` deaths by `r saturday_4_wk_ahead_txt` (Figure 1). For the week ending `r saturday_4_wk_ahead_txt`, the ensemble forecasts that reported COVID-19 deaths in the US will be between `r us_inc_death_wk_pi_round[1]` and `r us_inc_death_wk_pi_round[2]` (95% prediction interval: `r us_inc_death_wk_pi[1]` - `r us_inc_death_wk_pi[2]`).

<!-- Throughought most of July, models have in general shown broad agreement about the trajectory of the outbreak over the coming weeks. However, the recent surge in cases has left models with quite different interpretations about what the next few weeks hold in terms of how many reported deaths from COVID-19 we will see.  -->
You can explore the full set of models, including their forecasts for past weeks online at our [interactive forecast visualization](https://viz.covid19forecasthub.org/).


```{r make-US-inc-death-plot, fig.cap="__Figure 1__: Weekly observed and forecasted COVID-19 deaths. Observed data from [JHU CSSE](https://github.com/CSSEGISandData/COVID-19/tree/master/csse_covid_19_data/) and forecasts from the [COVID-19 Forecast Hub](https://covid19forecasthub.org/).", fig.topcaption=TRUE}

quantiles_to_plot <- c(0.025, 0.1, 0.25, 0.75, 0.9, 0.975)

blues <- RColorBrewer::brewer.pal(n=length(quantiles_to_plot)/2+1, "Blues")

inc_death_forecast <- dat %>%
  filter(target %in% inc_death_targets)

## get full inc death truth for plotting
inc_death_truth <- read_csv("../../data-truth/truth-Incident Deaths.csv") %>%
  rename(target_end_date = date, fips = location) %>%
  mutate(model = "observed data (JHU)") %>%
  group_by(fips, model) %>% arrange(target_end_date) %>%
  mutate(value = RcppRoll::roll_sum(value, 7, align = "right", fill = NA)) %>%
  ungroup() %>%
  left_join(locs, by=c("fips" = "FIPS")) %>%
  filter(target_end_date %in% seq.Date(as.Date("2020-01-25"), to = this_monday, by="1 week"),
    fips %in% unique(inc_death_forecast$fips)) 

inc_death_all_points <- inc_death_truth %>%
  bind_rows(filter(inc_death_forecast, class=="point")) %>%
  bind_rows(filter(inc_death_truth, target_end_date==last_5_saturdays[5]) %>% mutate(model="COVIDhub-ensemble")) %>%
  mutate(model = relevel(factor(model), ref="observed data (JHU)"))
  
    
## inc death data for code of uncertainty
dummy_inc_death <- tibble(
  quantile = quantiles_to_plot, 
  target_end_date=last_5_saturdays[5]) %>%
  right_join(inc_death_all_points %>%
      select(-quantile) %>%
      filter(target_end_date == last_5_saturdays[5]))

inc_death_quantiles <- inc_death_forecast %>%
  dplyr::filter(class=="quantile") %>%
  bind_rows(dummy_inc_death) %>%
  dplyr::filter(quantile %in% quantiles_to_plot) %>%
  dplyr::mutate(endpoint_type = ifelse(quantile < 0.5, 'lower', 'upper'),
    alpha = ifelse(endpoint_type == 'lower',
      format(2*quantile, digits=3, nsmall=3),
      format(2*(1-quantile), digits=3, nsmall=3)),
    `Prediction Interval` = fct_rev(paste0((1-as.numeric(alpha))*100, "%"))
  ) %>%
  dplyr::filter(alpha != "1.000") %>%
  dplyr::select(-quantile) %>%
  tidyr::pivot_wider(names_from='endpoint_type', values_from='value')

ggplot() +
  geom_ribbon(data = inc_death_quantiles %>% dplyr::filter(fips=="US"),
    mapping = aes(x = target_end_date,
      ymin=lower, ymax=upper,
      fill=`Prediction Interval`)) +
  geom_line(data=inc_death_all_points %>%
      dplyr::filter(fips == "US"),
    mapping = aes(x = target_end_date, y = value, color = model)) +
  geom_point(data=inc_death_all_points %>%
      dplyr::filter(fips == "US", !(model=="COVIDhub-ensemble" & target_end_date <= this_monday)),
    mapping = aes(x = target_end_date, y = value, color = model)) +
  scale_fill_manual(values = blues[1:(length(blues)-1)]) +
  scale_color_manual(values = c("black", tail(blues,1))) +
  scale_x_date(name = NULL, date_breaks="1 month", date_labels = "%b %d") +
  ylab("incident deaths") +
  labs(title="Weekly reported COVID-19 deaths in the US: observed and forecasted",
    caption="source: JHU CSSE (observed data), COVID-19 Forecast Hub (forecasts)") +
  theme(legend.position = c(.05,.95), legend.justification = c(0,1))
```



```{r prep-datatable}

## get last saturday observed cumulative deaths
cum_death_start <- read_csv("../../data-truth/truth-Cumulative Deaths.csv") %>%
  filter(date == last_5_saturdays[5]) %>%
  select(date, location, location_name, value) %>%
  rename(cum_deaths_at_forecast_start = value) 

## get recent observed inc deaths
recent_inc_death_totals <- read_csv("../../data-truth/truth-Incident Deaths.csv") %>%
  # filter(date > last_5_saturdays[3] & date <= last_5_saturdays[5]) %>%
  mutate(last_2wk = date > last_5_saturdays[3] & date <= last_5_saturdays[5],
    last_4wk = date > last_5_saturdays[1] & date <= last_5_saturdays[5]) %>%
  select(date, location, location_name, value, last_2wk, last_4wk) %>%
  group_by(location, location_name) %>%
  summarize(last_2wk_deaths = sum(value*last_2wk),
    last_4wk_deaths = sum(value*last_4wk)) %>%
  ungroup() %>%
  left_join(locs, by=c("location" = "FIPS")) %>%
  left_join(cum_death_start) %>% 
  rename(fips = location, target_end_date = date)
```


```{r process-ensemble-data}
ensemble_pointdat  <- dat  %>%
  filter(grepl('cum death', target)) %>%
  filter(class=="point") %>%
  select(fips, target, value, timezero)

wide_point_dat <- spread(ensemble_pointdat, target, value) %>%
  left_join(recent_inc_death_totals) %>%
  mutate(next_2wk_deaths = `2 wk ahead cum death` - cum_deaths_at_forecast_start,
    diff_2wk_deaths = next_2wk_deaths - last_2wk_deaths,
    next_4wk_deaths = `4 wk ahead cum death` - cum_deaths_at_forecast_start,
    diff_4wk_deaths = next_4wk_deaths - last_4wk_deaths,
    pop_x_1k = round(Population/1000),
    last_2wk_deaths_rate = round(last_2wk_deaths/Population*100000/14,3),
    last_4wk_deaths_rate = round(last_4wk_deaths/Population*100000/28,3),
    next_2wk_deaths_rate = round(next_2wk_deaths/Population*100000/14,3),
    next_4wk_deaths_rate = round(next_4wk_deaths/Population*100000/28, 3),
    diff_2wk_deaths_rate = round(next_2wk_deaths_rate - last_2wk_deaths_rate, 3),
    diff_4wk_deaths_rate = round(next_4wk_deaths_rate - last_4wk_deaths_rate, 3),
   next_2wk_cum_deaths = `2 wk ahead cum death` - cum_deaths_at_forecast_start) %>%
  select(location_name, Population, pop_x_1k, cum_deaths_at_forecast_start,
    last_2wk_deaths, next_2wk_deaths, diff_2wk_deaths, 
    last_4wk_deaths, next_4wk_deaths, diff_4wk_deaths,
    last_2wk_deaths_rate, next_2wk_deaths_rate,  
    last_4wk_deaths_rate, next_4wk_deaths_rate, 
    diff_2wk_deaths_rate, diff_4wk_deaths_rate, next_2wk_cum_deaths)


#filter quantile data for predicting future weeks
ensemble_quantdat <- dat %>%
  filter(target == "2 wk ahead cum death") %>%
  filter(class == "quantile")

wide_quant_dat <- spread(ensemble_quantdat, target, value) %>%
  left_join(recent_inc_death_totals %>% select(fips, location_name, last_2wk_deaths, cum_deaths_at_forecast_start)) %>% 
  mutate(next_2wk_deaths = `2 wk ahead cum death` - cum_deaths_at_forecast_start)

quant.5_cutoff <- wide_quant_dat %>% 
  filter(quantile == .5, next_2wk_deaths >= last_2wk_deaths) 

quant.25_cutoff <- wide_quant_dat %>% 
  filter(quantile == .25, next_2wk_deaths >= last_2wk_deaths) 

```


## State-by-state Breakdown

The ensemble model estimates that `r nrow(quant.5_cutoff)` states and territories of the US have a greater than 50% chance of having more deaths in the next two weeks compared to the past two weeks (Table 1).The model forecasts that `r nrow(quant.25_cutoff)` states and territories have a greater than 75% chance of an increase over the next two weeks (`r paste(quant.25_cutoff$location_name, collapse=", ")`). 


The sortable and searchable table below shows the total number of reported COVID-19 deaths at the US level and by state over the last two weeks (ending Saturday, `r format(last_5_saturdays[5], "%B %d, %Y")`) and the forecasted counts for the subsequent two weeks (ending `r format(last_5_saturdays[5]+14, "%B %d, %Y")`).

```{r make-datatable-inc-death-counts}
death_max_2wk <- max(abs(wide_point_dat$diff_2wk_deaths))

brks <- seq(-death_max_2wk, death_max_2wk, length.out = 100) #quantile(df, probs = seq(.05, .95, .05), na.rm = TRUE)
#clrs <- scico(n=length(brks)+1, palette="roma")
clrs <- colorRampPalette(colors = rev(RColorBrewer::brewer.pal(n=3, "RdBu")))(length(brks)+1)

table1_cap <- paste0("Table 1: US national and state-level observed deaths to date and for the previous two weeks (ending ", format(last_5_saturdays[5], "%B %d, %Y") ,") and the next two weeks (ending ", format(last_5_saturdays[5]+14, "%B %d, %Y"), ").")

datatable(wide_point_dat %>% 
    select(location_name, Population,
      cum_deaths_at_forecast_start, 
      last_2wk_deaths, next_2wk_deaths, diff_2wk_deaths) %>%
    arrange(desc(diff_2wk_deaths)),
  caption = table1_cap,
  options = list(
    autoWidth = TRUE,
    columnDefs = list(list(width = '100px', targets = c(0, 1, 2, 3, 4, 5)))
  ),  #width=paste0(c(10, 100, 100, 100), 'px'),
  rownames=FALSE,
  colnames=c('state'='location_name', 
    #'Population, \'000'='pop_x_1k',
    'Population' = 'Population',
    'Total COVID-19 deaths'='cum_deaths_at_forecast_start',
    'COVID-19 deaths, last 2 weeks'='last_2wk_deaths',
    'COVID-19 deaths, next 2 weeks'='next_2wk_deaths',
    'Difference' = 'diff_2wk_deaths')) %>%
##  formatStyle("Daily deaths, last 2 weeks", backgroundColor = styleInterval(brks, clrs)) %>%
##  formatStyle("Daily deaths, next 2 weeks", backgroundColor = styleInterval(brks, clrs))  %>%
  formatStyle("Difference", backgroundColor = styleInterval(brks, clrs)) %>%
  formatCurrency('Total COVID-19 deaths',currency = "", interval = 3, mark = ",", digits=0) %>%
  formatCurrency('Population',currency = "", interval = 3, mark = ",", digits=0) %>%
  formatCurrency('COVID-19 deaths, last 2 weeks',currency = "", interval = 3, mark = ",", digits=0) %>%
  formatCurrency('COVID-19 deaths, next 2 weeks',currency = "", interval = 3, mark = ",", digits=0)

```


The sortable and searchable table below shows the total number of reported COVID-19 deaths at the US level and by state as of Saturday, `r format(last_5_saturdays[5], "%B %d, %Y")` ("Total COVID-19 Deaths") as well as the rate of reported COVID-19 deaths in the population (standardized per 100,000 population) over the last two weeks and over the next two weeks. Looking at the rates allows for easier comparison across states, where you can see which states have had or are predicted to have propoportionally higher rates in comparison to other states. These tables calculate an average daily number of deaths per 100,000 population across the last two weeks (ending Saturday, `r format(last_5_saturdays[5], "%B %d, %Y")`) and forecasted for the following two weeks (ending `r format(last_5_saturdays[5]+14, "%B %d, %Y")`).

```{r make-datatable-inc-death-rates}

## color for rates
death_rate_max_2wk <- max(c(wide_point_dat$last_2wk_deaths_rate, wide_point_dat$next_2wk_deaths_rate))
brks <- seq(0, death_rate_max_2wk, length.out = 100) #quantile(df, probs = seq(.05, .95, .05), na.rm = TRUE)
clrs <- round(seq(255, 40, length.out = length(brks) + 1), 0) %>%
  {paste0("rgb(255,", ., ",", ., ")")}

## colors for rate difference
death_rate_diff_2wk <- max(abs(wide_point_dat$diff_2wk_deaths_rate))
brks1 <- seq(-death_rate_diff_2wk, death_rate_diff_2wk, length.out = 100) #quantile(df, probs = seq(.05, .95, .05), na.rm = TRUE)
clrs1 <- colorRampPalette(colors = rev(RColorBrewer::brewer.pal(n=3, "RdBu")))(length(brks1)+1)


table2_cap <- paste0("Table 2: US national and state-level observed and predicted daily death rates for the previous two weeks (ending ", format(last_5_saturdays[5], "%B %d, %Y") ,") and the next two weeks (ending ", format(last_5_saturdays[5]+14, "%B %d, %Y"), ").")

datatable(wide_point_dat %>% 
    select(location_name, Population,
      cum_deaths_at_forecast_start, 
      last_2wk_deaths_rate, next_2wk_deaths_rate, diff_2wk_deaths_rate) %>%
    arrange(desc(diff_2wk_deaths_rate)),
  caption = table2_cap,
  options = list(
    autoWidth = TRUE,
    columnDefs = list(list(width = '100px', targets = c(0, 1, 2, 3, 4, 5)))
  ),  #width=paste0(c(10, 100, 100, 100), 'px'),
  rownames=FALSE,
  colnames=c('state'='location_name', 
    #'Population, \'000'='pop_x_1k',
    'Population' = 'Population',
    'Total COVID-19 deaths'='cum_deaths_at_forecast_start',
    'Daily deaths per 100k, last 2 weeks'='last_2wk_deaths_rate',
    'Daily deaths per 100k, next 2 weeks'='next_2wk_deaths_rate',
    'Death rate difference' = 'diff_2wk_deaths_rate')) %>%
  ## formatStyle("Daily deaths per 100k, last 2 weeks", backgroundColor = styleInterval(brks, clrs)) %>%
  ## formatStyle("Daily deaths per 100k, next 2 weeks", backgroundColor = styleInterval(brks, clrs))  %>%
  formatStyle('Death rate difference', backgroundColor = styleInterval(brks1, clrs1))  %>%
  formatCurrency('Total COVID-19 deaths',currency = "", interval = 3, mark = ",", digits=0) %>%
  formatCurrency('Population',currency = "", interval = 3, mark = ",", digits=0)
```


This report was reproducibly and dynamically generated using RMarkdown. The code for the report can be found [here](https://github.com/reichlab/covid19-forecast-hub/tree/master/code/reports).


```{r}
htmltools::includeScript("statcounter.js") 
```