excessmort (version 0.8.2)

excess_model: Fit excess count model

Description

This function estimates the increase in the rate for a count time series relative to the rate for a typical year. Two options are available: 1 - model the rate increase as a smooth function and estimate this function or 2 - estimate the total excess in intervals. For 1 an `event` date can be provided and a discontinuity included in the model. You can do either 1 or 2 or both.

Usage

excess_model(
  counts,
  start = NULL,
  end = NULL,
  knots.per.year = 12,
  event = NULL,
  intervals = NULL,
  discontinuity = TRUE,
  model = c("quasipoisson", "poisson", "correlated"),
  exclude = NULL,
  include.trend = TRUE,
  trend.knots.per.year = 1/7,
  harmonics = 2,
  frequency = NULL,
  weekday.effect = FALSE,
  control.dates = NULL,
  max.control = 5000,
  order.max = 14,
  aic = TRUE,
  maxit = 25,
  epsilon = 1e-08,
  alpha = 0.05,
  min.rate = 1e-04,
  keep.counts = FALSE,
  keep.components = TRUE,
  verbose = TRUE
)

Value

If only `intervals` are provided a data frame with excess estimates described below for `excess`. if `start` and `end` are provided the a list with the following components are included:

date

The dates for which the estimate was computed

observed

The observed counts

expected

The expected counts

fitted

The fitted curve for excess counts

se

The point-wise standard error for the fitted curve

population

The population size

sd

The standard deviation for observed counts on a typical year

cov

The estimated covariance matrix for the observed counts

x

The design matrix used for the fit

betacov

The covariance matrix for the estimated coefficients

dispersion

The estimated overdispersion parameter

detected_intervals

Time intervals for which the 1 - `alpha` confidence interval does not include 0

ar

The estimated coefficients for the autoregressive process

excess

A data frame with information for the time intervals provided in `itervals`. This includes start, end, observed death rate (per 1,000 per year), expected death rate, standard deviation for the death rate, observed counts, expected counts, excess counts, standard deviation

Arguments

counts

A data frame with date, count and population columns.

start

First day of interval to which model will be fit

end

Last day of interval to which model will be fit

knots.per.year

Number of knots per year used for the fitted smooth function

event

If modeling a discontinuity is desired, this is the day in which it happens

intervals

Instead of `start` and `end` a list of time intervals can be provided and excess is computed in each one

discontinuity

Logical that determines if discontinuity is allowed at `event`

model

Which version of the model to fit

exclude

Dates to exclude when computing expected counts

include.trend

Logical that determines if a slow trend is included in the model.

trend.knots.per.year

Number of knots per year used by `compute_expected` to estimate the trend for the expected counts

harmonics

Number of harmonics used by `compute_expected` to estimate seasonal trend

frequency

Number of observations per year. If not provided an attempt is made to calculate it

weekday.effect

Logical that determins if a day of the week effects is included in the model. Should be `FALSE` for weekly or monthly data.

control.dates

When `model` is set to `correlated`, these dates are used to estimate the covariance matrix. The larger this is the slower the function runs.

max.control

If the length of `control.dates` is larger than `max.control` the function stops.

order.max

Larges order for the Autoregressive process used to model the covariance structure

aic

A logical that determines if the AIC criterion is used to selected the order of the AR process

maxit

Maxium number of iterations for the IRLS algorithm used when `model` is `correlated`

epsilon

Difference in deviance requried to declare covergenace of IRLS

alpha

Percentile used to define what is outside the normal range

min.rate

The estimated expected rate is not permited to go below this value

keep.counts

A logical that if `TRUE` forces the function to include the data used to fit the expected count model.

keep.components

A logical that if `TRUE` forces the function to return the estimated trend, seasonal model, and weekday effect, if included in the model. Ignored if expected counts already provided or `keep.counts` is `FALSE`.

verbose

Logical that determines if messages are displayed

Details

Three versions of the model are available: 1 - Assume counts are Poisson distributed, 2 - assume counts are overdispersed Poisson, or 3 - assume a mixed model with correlated errors. The second is the default and recommended for weekly count data. For daily counts we often find evidence of correlation and recommend the third along with setting `weekday.effect = TRUE`.

If the `counts` object includes a `expected` column produced by `compute_expected` these are used as the expected counts. If not, then these are computed.

Examples

data(cdc_state_counts)
counts <- cdc_state_counts[cdc_state_counts$state == "Massachusetts", ]
exclude_dates <- c(seq(as.Date("2017-12-16"), as.Date("2018-01-16"), by = "day"),
seq(as.Date("2020-01-01"), max(cdc_state_counts$date), by = "day"))
f <- excess_model(counts, 
exclude = exclude_dates,
start = min(counts$date),
end = max(counts$date),
knots.per.year = 12)
data(new_jersey_counts)
exclude_dates <- as.Date("2012-10-29") + 0:180
control_dates <- seq(min(new_jersey_counts$date), min(exclude_dates) - 1, by = "day")
f <- excess_model(new_jersey_counts,
start = as.Date("2012-09-01"), 
end = as.Date("2013-09-01"), 
exclude = exclude_dates,
model = "correlated",
weekday.effect = TRUE,
control.dates = control_dates)