Define the treatment

Turns a panel into a treatment definition that the estimators can use: a continuous stop intensity, a binary intervention in named areas and months, staggered adoption dates by area, or a dated event applied to every area. The result records how the treatment was built, is stored in the contract of every estimate made with it, and is printed by the estimators alongside their identifying assumption.

Usage

lamp_treatment(
  panel,
  type = c("continuous", "binary", "staggered", "event"),
  measure = c("stop_rate", "stops", "stops_per_100_crimes"),
  transform = c("identity", "log", "ihs"),
  areas = NULL,
  window = NULL,
  adoption = NULL,
  event = NULL,
  date = NULL,
  scope = NULL
)

Arguments

panel

A lamp_panel from lamp_panel().

type

"continuous", "binary", "staggered" or "event".

measure, transform

For continuous: which intensity and which transformation.

areas

For binary: the treated areas.

window

For binary: the first and last treated month. For event: the event window in months, default c(-12, 12).

adoption

For staggered: a table with columns area and adoption_month.

event

For event: a name in lamp_shocks(), or NULL when date is given.

date

For event: the event month, overriding the shocks table.

scope

For event: areas the event applies to; NULL for all.

Value

A lamp_treatment object: a list with type, column (the treatment variable's name), data (area, month, the treatment column and, where applicable, rel_time and cohort), details, n_treated_areas, n_control_areas and created.

Details

Types:

• continuous: measure is "stops", "stop_rate" (per 1,000 residents, the default) or "stops_per_100_crimes" (searches per 100 crimes recorded in the area's previous twelve months, which removes the mechanical dependence on population). transform is "identity", "log" (log of one plus the measure) or "ihs" (inverse hyperbolic sine, which handles zeros).

• binary: areas are treated during window, a pair of months, and untreated outside it; every other area is a control.

• staggered: adoption is a table of area and adoption_month. Areas absent from it are never treated. rel_time (months since adoption) and cohort are added.

• event: a dated shock applied to every area in scope, from lamp_shocks() by event name or from date, with an event window in months around it. rel_time is months since the event.

A continuous treatment is a dose, not an experiment: the allocation of searches to areas responds to crime, so estimates from it describe an association unless the variation is argued to be exogenous. See lamp_allocation().

See also

Other treatment: lamp_detect_surges()

Examples

sim <- lamp_simulate(n_areas = 16, n_months = 18, design = "event", seed = 1)

# a dated event, twelve months either side
lamp_treatment(sim, "event", date = "2018-10", window = c(-6, 6))
#> 
#> ── streetlamp treatment 
#> Type: "event"; column treat
#> Areas: 16 treated, 0 control
#> Event: user-supplied event on 2018-10-01, window -6 to 6 months

# stop intensity as a continuous dose
lamp_treatment(sim, "continuous", measure = "stop_rate", transform = "ihs")
#> 
#> ── streetlamp treatment 
#> Type: "continuous"; column treat
#> Areas: 16 treated, 0 control
#> Measure: stop_rate, transform ihs

# a surge in named areas
lamp_treatment(sim, "binary", areas = unique(sim$area)[1:4], window = c("2018-06", "2018-12"))
#> 
#> ── streetlamp treatment 
#> Type: "binary"; column treat
#> Areas: 4 treated, 12 control
#> Never-treated areas: 12