In rplzzz/vdhcovid: COVID-19 data by health district from Virginia Department of Health

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(vdhcovid)
library(dplyr)
library(ggplot2)

The daily number of tests has a strong periodic signal.

apr15 <- as.Date('2020-04-15')
dailytotal <- 
  group_by(vadailytests, date) %>% 
  summarise(ntest=sum(ntest)) %>%
  filter(date >= apr15)
trendfit <- lm(ntest~date, dailytotal)
dailytotal$ntest_detrend <- dailytotal$ntest - predict(trendfit)
N <- nrow(dailytotal)
Nhalf <- floor(N/2 + 1)
ggplot(dailytotal, aes(x=date, y=ntest)) + geom_line(size=1.2, color='MidnightBlue') +
  theme_bw()

We can do a spectral analysis to work out what frequencies are represented.

ps <- fft(dailytotal$ntest_detrend)[1:Nhalf]
f <- (seq_along(ps) - 1) / N
pltps <- data.frame(f=f, ps=Mod(ps))
ggplot(pltps, aes(x=f, y=ps)) + 
  geom_vline(color='red', xintercept = 1/7, linetype=2) +
  geom_col(width=0.25*f[2], fill='MidnightBlue') + 
  ylab('Spectral density') +
  xlab('Frequency (1/day)') +
  theme_bw()

top3 <- order(pltps$ps, decreasing=TRUE)[1:3]
f3 <- f[top3]
per3 <- signif(1/f3, 2)

The top 3 modes have periods of r paste(per3, collapse=', ') days. The first of these confirms what we expected; there is a weekly periodicity that is probably due to less testing and test processing on weekends. The third is probably due to weekends being two days long. The middle one is about half a week; it's probably related to the fact that the character of the periodicity is more like $\sin^2 t$ than $\sin t$.

rplzzz/vdhcovid documentation built on July 2, 2021, 7:43 a.m.