How stop and search activity follows recorded crime

Regresses searches on the crime recorded in previous months, with area and month fixed effects. This is not an effect of crime on searching in any causal sense; it is a description of how activity is allocated, and it is here because without it the other estimators cannot be read honestly.

Usage

lamp_allocation(
  panel,
  stops = "stops",
  crime = "crime_total",
  crime_lags = 1:3,
  cluster = c("area", "force"),
  family = c("poisson", "ols_log")
)

Arguments

panel

A lamp_panel.

stops

The activity column, default "stops".

crime

The crime column, default "crime_total".

crime_lags

Lags of crime to include, default 1 to 3.

cluster

"area" (default) or "force".

family

"poisson" (default) or "ols_log".

Value

A lamp_estimate of class lamp_allocation, with one coefficient per lag and diagnostics$total for their sum. The identifying assumption field states plainly that the result is descriptive.

Why this matters for every other estimate

If police search more where crime has just risen, then searching is a response to crime as well as a possible cause of it. A regression of crime on searches then mixes the two directions: the deterrent effect pushes the coefficient down, the allocation response pushes it up, and the estimate is the net of them. A positive coefficient here is the size of the channel that has to be argued away before a stop-crime elasticity or a difference-in-differences estimate can be called causal. A coefficient near zero does not establish exogeneity or rule out other allocation channels.

See also

Other estimators: lamp_crimes_prevented(), lamp_did_staggered(), lamp_displacement_quotient(), lamp_elasticity(), lamp_event_study(), lamp_spillover(), lamp_synth(), lamp_twfe(), tidy.lamp_estimate()

Examples

sim <- lamp_simulate(n_areas = 24, n_months = 30, design = "continuous", seed = 1)
lamp_allocation(sim, crime_lags = 1:2)
#> 
#> ── streetlamp estimate: lamp_allocation 
#> Outcome: stops; family: Poisson pseudo-likelihood on counts
#> Treatment: lagged crime; clustered by area
#> Identifying assumption: None: this is descriptive. It measures how searching
#> has tracked recorded crime, not an effect of crime on searching.
#> Sample: 672 area-months in 24 areas; 0 rows dropped for coverage.
#> Dispersion: 0.925
#>   Term             Estimate  Std. error            95% CI      p
#>   ───────────────  ────────  ──────────  ────────────────  ─────
#>   .log_crime_lag1   −0.0637      0.0507  [−0.163, 0.0356]  0.208
#>   .log_crime_lag2    0.0566       0.056   [−0.053, 0.166]  0.312
#> Sum over lags: -0.0071 [-0.145, 0.131]
#> No clear allocation response was detected in this panel. This does not
#> establish exogeneity or rule out reverse allocation and confounding.