Skip to contents

Get distances and parameterise Poisson regression model for injuries

Source: R/distances_for_injury_function.R
distances_for_injury_function.Rd

Computes exposures (distances) by mode to parameterise the injury regression model, which is computed as a Poisson model and which is used in the ITHIM-Global model to predict injury fatalities at a later stage

Usage

distances_for_injury_function(journeys, dist)

Arguments

journeys

data frame with total distance (by total population) for each age and sex category and for each scenario

dist

table of (total population) distances per mode

Value

true_distances (mode distances by age and sex with all walking modes and all car modes combined and bus drivers added where relevant),

injuries_list (list of all strike, casualty, age, sex and mode distance combinations for baseline and all scenarios),

reg_model (parameterised Poisson regression model),

injuries_for_model (baseline data containing injury counts for all casualty and strike mode combinations with associated distance data)

Details

The function uses distance data and the individual injury fatality data to perform the following steps to parameterise the Poisson injury regression model:

  • taxi, shared_taxi and auto_rickshaw distances are multiplied by 2 to add drivers (note if there are multiple passengers in a vehicle apart then this will overestimate these vehicle distances but taxis also spend a significant amount of time driving around without any passenger which is not captured in the travel surveys.)

  • stage modes are aggregated such that all walk related stages (walk and walk to pt) are of the same mode (walk), similarly for all car related journeys

  • bus drivers are added to bus journeys (where relevant) to accurately represent all people on a bus

  • Takes Baseline injury tables, split into who-hit whom (whw) and no-other-vehicle (nov) parts, and adds total population distances for each strike and casualty mode (add_distance_columns.R). Distances are added by age and gender category if there exists such information for the injury counts (injuries_for_model dataframe). If there exists a fatality for some casualty and strike mode and age and sex category but no mode distance for this age and sex category, then fatalities and distances are aggregated by strike and casualty mode. If, after the aggregation there still exist fatalities for which either casualty or strike mode distance are missing, then these fatalities are removed as we cannot predict injury counts on zero distances. However, this should not happen as we should have total distances for all modes (possibly inferred from other modes) that appear in the injury data. - This data is used to parameterise the Poisson injury model.

  • A new list (injuries_list) is created containing all strike and casualty mode and age and sex combinations together with strike and casualty mode distances (add_distance_columns()) for the baseline and all scenarios. For the whw model, any strike mode and casualty pairs where strike mode equals casualty mode are removed as fatalities for these combinations have already been added to the nov matrix. Combinations which do not have a non-zero strike or casualty mode distance are also removed. This list will later be used in the injuries_function_2() function to predict fatality counts using the Poisson injury regression model.

  • The casualty and strike mode exponents used to account for the safety in number effect are added to both the injuries_for_model and injuries_list.

  • The best possible regression model is being built using Baseline injury counts and distances (injuries_for_model) such that the standard errors are small wherever possible. Strike and casualty mode pairs where cas mode = strike mode are removed if they still exist which they should not as they should have been removed by the ithim_load_data.R function. Two different forms for whw and nov matrices are defined, taking into account age and sex information where it exists. The standard errors of the newly built regression models are checked and if they are too large and if the data has not been aggregated by age and sex yet, then the data is aggregated and a new Poisson regression model is build. If the standard errors are still large after this aggregation, then a message is printed to the screen warning that the standard errors are large.