Skip to content

Commit

Permalink
updates website
Browse files Browse the repository at this point in the history
  • Loading branch information
Fox Sebastian (BNSSG CCG) authored and Fox Sebastian (BNSSG CCG) committed Mar 6, 2024
1 parent 8953e0e commit 6851a22
Show file tree
Hide file tree
Showing 85 changed files with 14,607 additions and 4,556 deletions.
4 changes: 3 additions & 1 deletion .gitignore
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,4 +7,6 @@
/data/
data-raw/
/.quarto/
*.pdf
*.pdf

/ad_hoc/
34 changes: 15 additions & 19 deletions R/02_data.R
View file
Original file line number Diff line number Diff line change
Expand Up @@ -798,30 +798,26 @@ annual_gp_workforce <- tibble(
)
) |>
map(
~ summarise(
.x,
numerator = mean(numerator),
.by = any_of(
c(
"YEAR",
"ICB_CODE",
"STP_CODE",
"STAFF_GROUP"#,
#"STAFF_ROLE"
)
)
)
) |>
map_df(
~ rename(
.x,
.x,
org = any_of(
c("STP_CODE", "ICB_CODE")
)
),
year = "YEAR"
)
) |>
rename(
year = "YEAR"
bind_rows() |>
summarise(
numerator = mean(numerator),
.by = any_of(
c(
"year",
"org",
"STAFF_GROUP"#,
#"STAFF_ROLE"

)
)
)

health_populations <- quarterly_ics_populations() |>
Expand Down
173 changes: 86 additions & 87 deletions R/04_modelling_utils.R
View file
Original file line number Diff line number Diff line change
Expand Up @@ -212,6 +212,35 @@ identify_missing_vars <- function(data_train, data_validation, data_test, model_
return(vars)
}

create_lag_variables <- function(data, lagged_years, lag_variables) {
map_lag <- set_names(
seq_len(lagged_years),
nm = paste("lag", seq_len(lagged_years), sep = "_")
) |>
map(
~ purrr::partial(lag, n = .x)
)

data <- data |>
arrange(
year, org
) |>
group_by(org) |>
group_split() |>
purrr::map_df(
~ mutate(
.x,
across(
all_of(lag_variables),
.fn = map_lag,
.names = "{.fn}_{.col}"
),
)
)

return(data)
}


# plotting functions ------------------------------------------------------

Expand Down Expand Up @@ -713,8 +742,11 @@ load_data <- function(target_variable, value_type = "value", incl_numerator_rema
#' time_series_split is TRUE, set this to FALSE to remove data from the
#' predictor variables where they occur in the same year as the target
#' variable that is being predicted
#' @param remove_lag_target logical; if including lagged data, whether to
#' include the lagged target variable or not
#' @param lag_target integer; number of new fields for lagged version of the
#' target variable to include (each new field with be an extra lagged year).
#' E.g., 2 will create 2 new fields (prefixed lag_1 and lag_2) which will
#' contain data with a 1 year lag and data with a 2 year lage for the target
#' variable
#' @param time_series_split logical; order the data by year prior to splitting,
#' so training data is before validation data, which is before test data
#' @param training_years integer; number of years to use for training the model.
Expand All @@ -741,7 +773,7 @@ load_data <- function(target_variable, value_type = "value", incl_numerator_rema
#' https://www.tidyverse.org/blog/2022/05/case-weights/
#'
modelling_performance <- function(data, target_variable, lagged_years = 0,
keep_current = TRUE, remove_lag_target = TRUE,
keep_current = TRUE, lag_target = 0,
time_series_split = TRUE, training_years = NULL,
remove_years = NULL,
shuffle_training_records = FALSE,
Expand Down Expand Up @@ -799,48 +831,46 @@ modelling_performance <- function(data, target_variable, lagged_years = 0,
) {
data <- data |>
mutate(
across(
c(numerator, remainder),
~ as.integer(
round(
.x, 0
)
)
),
total_cases = frequency_weights(numerator + remainder)
) |>
select(!c("numerator", "remainder"))
}

if (lag_target > 0) {
data <- create_lag_variables(
data = data,
lagged_years = lag_target,
lag_variables = target_variable
)
}

if (lagged_years > 0) {

map_lag <- set_names(
seq_len(lagged_years),
nm = paste("lag", seq_len(lagged_years), sep = "_")
) |>
map(
~ purrr::partial(lag, n = .x)
)

# create lag variables

not_lag_variables <- c("year", "org", "nhs_region", "pandemic_onwards")
if (remove_lag_target) {
not_lag_variables <- c(not_lag_variables, target_variable)
}

not_lag_variables <- c("year", "org", "nhs_region", "pandemic_onwards", target_variable)

if (model_type == "logistic_regression") {
not_lag_variables <- c(not_lag_variables, "total_cases")
}

data <- data |>
arrange(
year, org
) |>
group_by(org) |>
group_split() |>
purrr::map_df(
~ mutate(
.x,
across(
!all_of(not_lag_variables),
.fn = map_lag,
.names = "{.fn}_{.col}"
),
)
)
lag_variables <- setdiff(
names(data),
not_lag_variables
)

data <- create_lag_variables(
data = data,
lagged_years = lagged_years,
lag_variables = lag_variables
)

if (time_series_split) {
data <- data |>
Expand Down Expand Up @@ -1044,9 +1074,6 @@ modelling_performance <- function(data, target_variable, lagged_years = 0,

if (model_type %in% c("logistic_regression")) {
model_recipe <- model_recipe |>
step_zv(
all_of(predictor_variables)
) |>
step_range(
all_numeric_predictors(),
clipping = FALSE
Expand Down Expand Up @@ -1090,6 +1117,7 @@ modelling_performance <- function(data, target_variable, lagged_years = 0,
by = 10
) * 10
)
tuning_grid <- 30
}

extract_input <- NULL
Expand Down Expand Up @@ -1221,13 +1249,6 @@ modelling_performance <- function(data, target_variable, lagged_years = 0,
} else if (model_type == "random_forest") {
case_wts <- FALSE
}
# browser()
# train_metrics <- calculate_train_metric(
# final_fit = model_fit,
# training_data = data_train,
# target_variable = target_variable,
# case_weights = case_wts
# )

train_metrics <- model_fit |>
extract_workflow() |>
Expand Down Expand Up @@ -1322,7 +1343,7 @@ modelling_performance <- function(data, target_variable, lagged_years = 0,
`Training years` = training_years,
`Lagged years` = lagged_years,
`Current year included` = ifelse(lagged_years > 0, keep_current, NA),
`Lagged target variable` = ifelse(lagged_years > 0, !remove_lag_target, NA)
`Lagged target variable` = paste(lag_target, "lagged years")
)

output <- list(
Expand Down Expand Up @@ -1397,10 +1418,9 @@ pick_model <- function(modelling_outputs, evaluation_metric, id = "best") {
#' 2 years of data using a linear trend)
#' @param area_code string; code for the area(s) of interest. Can take a vector
#' containing multiple area code
#' @param extrapolation_method string; "linear", "average" or "spline". Linear
#' will extrapolate the last n points linearly. Average will take a rolling
#' average of the last n points. Spline will extend the last n points with a
#' natural spline.
#' @param extrapolation_method string; "linear" or "spline". Linear will
#' extrapolate the last n points linearly. Spline will extend the last n
#' points with a natural spline.
#' @param model_fit an object returned from the last_fit function
#' @param target_variable string; name of target variable
#' @param scenario tibble with two columns, metric and multiplier. The
Expand All @@ -1414,7 +1434,7 @@ apply_model_to_scenario <- function(input_data, number_year_to_extrapolate,

extrapolation_method <- match.arg(
extrapolation_method,
c("linear", "spline", "average")
c("linear", "spline")
)

# expand dataset to include future years by org and metric
Expand Down Expand Up @@ -1464,19 +1484,7 @@ apply_model_to_scenario <- function(input_data, number_year_to_extrapolate,
)$y,
.by = c(org, metric)
)
} else if (extrapolation_method == "average") {
browser()
predictions <- predictions |>
mutate(
prediction = slider::slide_index_mean(
x = value,
i = year,
before = number_year_to_extrapolate,

)
)

}
}

if (!is.null(scenario)) {
predictions <- predictions |>
Expand Down Expand Up @@ -1507,42 +1515,33 @@ apply_model_to_scenario <- function(input_data, number_year_to_extrapolate,
.default = value
)
) |>
select(!c("prediction"))

# visualiste the linear interpolations

# dc_data_future |>
# filter(metric %in% names(input_data)[4:24]) |>
# ggplot(
# aes(
# x = year
# )
# ) +
# geom_point(
# aes(y = value,
# colour = type)
# ) +
# facet_wrap(
# facets = vars(metric),
# scales = "free_y"
# )

# create predictor matrix
predictions <- predictions |>
select(!c("prediction", "type")) |>
select(!c("type")) |>
pivot_wider(
names_from = metric,
values_from = value
)

# are there lagged years in the modelling
lagged_years <- calculate_lag_years_from_last_fit(model_fit)
if (lagged_years > 0) {
# calculate lagged values for input variables
predictions <- create_lag_variables(
data = predictions,
lagged_years = lagged_years,
lag_variables = c("ENTER FIELDS HERE")
)
}

filter(
year > max(input_data[["year"]])
) |>
add_prediction_to_data(
final_fit = model_fit
) |>
select(
"year", "org", "nhs_region", ".pred"
) |>
filter(
year > max(input_data[["year"]])
) |>
rename_with(
~ target_variable,
".pred"
Expand Down
Loading

0 comments on commit 6851a22

Please sign in to comment.