Skip to contents

Binomial time-series model.

Source: R/estimator-glm.R
proportion_glm_model.Rd

This uses a generalised linear model to fit a quasi-binomial model with a time varying rate as a natural cubic spline with approx one degree of freedom per window units of the time series.

Usage

proportion_glm_model(
  d = i_proportion_input,
  ...,
  window = 14,
  frequency = "1 day",
  .progress = interactive()
)

Arguments

d

Proportion model input - a dataframe with columns:

  • denom (positive_integer) - Total test counts associated with the specified time frame

  • count (positive_integer) - Positive case counts associated with the specified time frame

  • time (ggoutbreak::time_period + group_unique) - A (usually complete) set of singular observations per unit time as a `time_period`

Ungrouped.

...

not used and present to allow proportion model to be used in a group_modify

window

a number of data points defining the bandwidth of the estimate, smaller values result in less smoothing, large value in more. The default value of 14 is calibrated for data provided on a daily frequency, with weekly data a lower value may be preferred. - default 14

frequency

the density of the output estimates as a time period such as 7 days or 2 weeks. - default 1 day

.progress

show a CLI progress bar

Value

A dataframe containing the following columns:

  • time (ggoutbreak::time_period + group_unique) - A (usually complete) set of singular observations per unit time as a time_period

  • proportion.fit (double) - an estimate of the proportion on a logit scale

  • proportion.se.fit (positive_double) - the standard error of proportion estimate on a logit scale

  • proportion.0.025 (proportion) - lower confidence limit of proportion (true scale)

  • proportion.0.5 (proportion) - median estimate of proportion (true scale)

  • proportion.0.975 (proportion) - upper confidence limit of proportion (true scale)

Any grouping allowed.

Examples


ggoutbreak::england_covid_proportion %>%
  dplyr::filter(date < "2021-01-01") %>%
  time_aggregate() %>%
  ggoutbreak::proportion_glm_model(window=5) %>%
  dplyr::glimpse()
#> Rows: 209
#> Columns: 10
#> $ time              <time_prd> 22.57143, 22.71429, 22.85714, 23.00000, 23.1428…
#> $ proportion.fit    <dbl> -4.145353, -4.168497, -4.191602, -4.214647, -4.23761…
#> $ proportion.se.fit <dbl> 0.4084512, 0.3869179, 0.3665636, 0.3475906, 0.330209…
#> $ proportion.0.025  <dbl> 0.007062264, 0.007197185, 0.007318146, 0.007421591, …
#> $ proportion.0.05   <dbl> 0.008024531, 0.008122834, 0.008206851, 0.008273554, …
#> $ proportion.0.25   <dbl> 0.011881163, 0.011780385, 0.011671731, 0.011554098, …
#> $ proportion.0.5    <dbl> 0.015590924, 0.015239662, 0.014896774, 0.014562335, …
#> $ proportion.0.75   <dbl> 0.02043506, 0.01969450, 0.01899581, 0.01833927, 0.01…
#> $ proportion.0.95   <dbl> 0.03007542, 0.02841329, 0.02689221, 0.02550831, 0.02…
#> $ proportion.0.975  <dbl> 0.03406576, 0.03197970, 0.03008592, 0.02837718, 0.02…

# TODO: deal with error conditions
# "observations with zero weight not used for calculating dispersion