Specific Confounders Controlled in Higher-Quality RLC Crash Studies

confidence: high · 6 sources

⚠
Working research — not a recommendation.

This page records what the evidence says a policy does, not what anyone should do about it. It has been synthesized from the cited sources but has not been independently reviewed.

What the evidence says

Across the higher-quality studies in this set, the confounders that get explicit statistical control cluster into a small set: traffic-volume exposure, road geometry, and intersection type in general road-safety before-and-after designs; and, in red light camera (RLC)-specific work, average annual daily traffic (AADT), lane counts, median presence, right-turn channelization, left-turn signal type, posted speed limit, land use, and driving-age population. Gupta (2020) characterizes higher-quality controlled before-and-after studies as adequately controlling for "exposure effect (in traffic volume, area type, geometric design, and type of intersection)," and singles out Liu et al. (2011) as the highest-quality controlled before-and-after study because it matched treatment and comparison sites on geographic characteristics and analyzed the same time period. In RLC-specific research, Abdulsalam (2017, journal) built a geometry database from Google Earth and city traffic-operations records to control for AADT, through/left-turn lane counts, median presence, right-turn channelization, and posted speed limits across propensity-score-matched (PSM), cross-sectional, and empirical Bayes model specifications. Li (2022) reports a fourth combination: average speed limit, land use, traffic counts, and driving-age population as control variables in a geographically weighted negative binomial regression (GWNBR).

Regression to the mean (RTM) is the confounder that recurs most consistently across sources, and the one whose control changes the direction of the finding, not just its precision. Gupta (2020) reports that "almost all" road engineering intervention studies remained at risk for RTM bias because treatment sites were selected on high baseline accident rates. Cohn (2020) quantifies the consequence for right-angle crashes: studies accounting for RTM report more moderate decreases in right-angle injury crashes than studies that do not, and studies with better overall confounder control report a nonsignificant decrease in right-angle crashes, compared with a significant decrease across all studies combined. Abdulsalam (2017, thesis) states plainly that "failure to account for RTM can exaggerate the positive effects" of RLCs. Burkey (2005, tier D, citing Kulmala 1994) puts a number on this: crashes declined approximately 20% due to regression to the mean effects alone, independent of any safety measure.

Spillover (halo) effects run in the opposite direction from RTM: where RTM causes uncontrolled studies to overstate benefits, uncontrolled spillover causes studies to understate them. Abdulsalam (2017, thesis) states that ignoring spillover at intersections without RLCs "will lead to an underestimation of RLC benefits, even more so if sites with these effects are used as a comparison group." The same source treats interaction terms as a distinct control mechanism, used to identify how RLC treatment effects vary by site characteristics, and notes that the absence of such interactions in a regression specification "can affect the accuracy and interpretations of estimations."

Land use and speed limit appear as confounders with uneven treatment across the literature. Li (2022) explicitly controlled for land use alongside speed limit, traffic counts, and driving-age population, in contrast to the same source's characterization that "many existing studies using the EB method have not directly controlled for land use or other macro elements." Speed limit also surfaces as a measured effect in its own right: Li's GWNBR model finds that total crash rate increases about 12% for every one-mile-per-hour increase in average speed limit.

Where multiple estimation methods with different confounder specifications are run on the same data, the direction of the result holds up. Abdulsalam (2017, journal) reports that PSM, cross-sectional NB regression, and empirical Bayes panel analysis "yielded fairly consistent results indicating an increase in property damage collisions and a decrease in injury and fatal collisions," though the source notes some difference in magnitude across methods without quantifying it.

Data-quality confounders receive comparatively thin coverage. Burkey (2005, tier D) states that higher crash-reporting thresholds (tow-away versus traditional reporting) "tend to seriously underreport certain types of crashes," and separately characterizes prior RLC evaluations, following McGee and Eccles (2003), as "not well-designed, lacking well-selected control groups, and used little data" — a general credibility judgment on the pre-existing literature rather than a measured confounder effect.

Where the evidence splits

Cohn (2020) reports two figures for the same outcome — right-angle crash counts — that point in different directions depending on study quality: studies with better overall confounder control report a nonsignificant decrease in right-angle crashes, while all studies combined report a significant decrease. The source ties this to confounder control generally and to RTM control specifically, but it does not state whether the shift from significant to nonsignificant reflects removal of confounding bias or simply a smaller sample size in the higher-quality subset — the claim's own annotation flags this as unresolved. The split is unexplained beyond that ambiguity in the sources on this page.

Thresholds and dose-response

Li (2022) establishes a dose-response rate for one confounder-turned-covariate: total crash rate increases about 12% for every one-mile-per-hour increase in average speed limit, based on a GWNBR model. Burkey (2005, tier D) reports a related magnitude for regression to the mean — approximately 20% of observed crash reductions attributable to RTM alone when uncontrolled — though this is an average bias share rather than a rate tied to a specific level or cutoff, and the source itself flags that the 20% figure traces to Kulmala (1994) and was not independently verified against that primary source.

Open empirical questions

Findings

Studies that fail to control for regression to the mean report larger apparent safety benefits than studies that control for it

confidence — direction: highconfidence — magnitude: mediumconsensus: contested on right-angle crash significance, settled on direction of bias (uncontrolled RTM inflates apparent benefit) — weight leans toward RTM causing overstatement of benefit
mechanism
treatment sites for road-safety and RLC interventions are selected because of high baseline crash rates, so crash counts fall toward the mean on their own regardless of the intervention; controlling for RTM removes this component from the estimated effect
effect
decreases in right-angle injury crashes are "more moderate" when RTM is controlled; the significant decrease in right-angle crashes reported across all studies becomes nonsignificant in studies with better confounder control; uncontrolled RTM alone was estimated to account for about 20% of observed crash reductions (tier D)
conditions / moderators
almost all road engineering intervention studies reviewed by Gupta (2020) remained at risk for RTM bias because treatment sites were selected on high accident rates

Higher-quality before-and-after studies control for traffic-volume exposure, geometric design, and intersection type

confidence — direction: highconfidence — magnitude: high (Gupta's characterization) / medium (time-series subset)consensus: settled that these confounders are the baseline expectation for higher-quality design
mechanism
matching or including exposure and road-design variables so that pre/post differences are not attributed to changes in traffic volume, area type, geometry, or intersection type rather than the intervention itself
effect
not a measured outcome magnitude — this is a methodological characterization of which confounders were controlled, not an effect size
conditions / moderators
the highest-quality example (Liu et al., 2011, as characterized by Gupta 2020) additionally matched treatment and comparison sites on geographic characteristics and the same time period; some time-series studies controlled for traffic volume changes only by limiting the design to one period before and one after

RLC-specific regression studies control for site geometry and traffic-exposure variables through model specification

confidence — direction: highconfidence — magnitude: highconsensus: settled that these variables are treated as necessary controls in RLC crash regression
mechanism
intersection-level covariates (AADT, lane counts, median presence, right-turn channelization, left-turn signal type, posted speed limit, land use, traffic counts, driving-age population) are entered directly into PSM, cross-sectional NB, empirical Bayes, or GWNBR models to isolate the RLC effect from site differences
effect
not a measured outcome magnitude — describes which variables were controlled, not a resulting effect size
conditions / moderators
the two studies used different confounder sets — Abdulsalam (2017, journal) controlled road-geometry and signal-timing variables; Li (2022) controlled speed limit, land use, traffic counts, and driving-age population

Failure to control spillover (halo) effects underestimates RLC safety benefits

confidence — direction: highconfidence — magnitude: high (per source's own confidence annotation), though no numeric magnitude is givenconsensus: thin — one source only, no independent replication in this set
mechanism
RLCs can reduce crashes at nearby, non-treated intersections; if those spillover sites are used as an untreated comparison group, the comparison group's crash counts also fall, shrinking the measured treatment effect
effect
direction only — underestimation of true treatment benefit; magnitude not quantified in the excerpted text
conditions / moderators
effect is worse specifically when spillover sites are used as the comparison group; spillover distance and timing were reportedly determined but not shown in the excerpt

Propensity score matching corrects site-selection bias between treatment and comparison intersections

confidence — direction: highconfidence — magnitude: highconsensus: thin — one source
mechanism
matches RLC and non-RLC intersections on observed characteristics before comparing outcomes, addressing the tendency for RLC sites to be chosen non-randomly
effect
not a measured outcome magnitude — described as bias correction, not quantified in the excerpt (covariate balance table referenced but not shown)
conditions / moderators
not established — whether matching achieved adequate covariate balance is not stated in the excerpted text

Increasing posted speed limit increases total crash rate

confidence — direction: highconfidence — magnitude: mediumconsensus: thin — single-source, single-location estimate
mechanism
not stated beyond speed limit being modeled as a control/risk variable in the GWNBR specification
effect
total crash rate increases about 12% for every one-mile-per-hour increase in average speed limit
conditions / moderators
based on one study location; not established whether this rate generalizes elsewhere

Fixed and random-effects panel regression controls for spatial and temporal correlation in collision data

confidence — direction: highconfidence — magnitude: high (per source annotation), though unquantifiedconsensus: thin — one source
mechanism
panel structure accounts for correlation between observations at the same site over time and between nearby sites, rather than treating each observation as independent
effect
not a measured outcome magnitude — the degree of correlation removed is not quantified in the excerpt
conditions / moderators
not established

Land use is inconsistently controlled across empirical Bayes RLC studies

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin — based on a literature characterization, not an empirical test
mechanism
land use as a macro-level confounder is often omitted from empirical Bayes model specifications, which focus on intersection geometric characteristics instead
effect
no measured effect of the omission is given — this is a characterization of methodological practice, not a quantified bias
conditions / moderators
at least one study (Li 2022) has since controlled directly for land use alongside speed limit, traffic counts, and driving-age population

Different confounder specifications across estimation methods yield consistent directional results for RLC collision effects

confidence — direction: mediumconfidence — magnitude: lowconsensus: thin — one source, qualitative consistency claim only
mechanism
not stated beyond the three methods (PSM, cross-sectional NB regression, empirical Bayes) each including their own confounder sets
effect
all three methods indicate an increase in property damage collisions and a decrease in injury and fatal collisions; the source notes "some differences in magnitude" across methods without quantifying them
conditions / moderators
not established how much of the magnitude difference is attributable to differing confounder control versus other model differences

Interaction terms identify how RLC treatment effects vary by site characteristics, and their absence reduces estimate accuracy

confidence — direction: highconfidence — magnitude: mediumconsensus: thin — one source
mechanism
interacting site characteristics with the treatment indicator lets the model capture treatment-effect heterogeneity rather than forcing a single average effect across all sites
effect
no measured effect magnitude given — described as affecting "accuracy and interpretations of estimations" when absent
conditions / moderators
not established whether identified interactions reflect true causal moderation or confounding by unmeasured variables

Higher crash-reporting thresholds and inadequate control groups in prior RLC studies compromise data and design quality

confidence — direction: highconfidence — magnitude: lowconsensus: thin — tier D source, citing secondary reviews (Zegeer et al. 1998; McGee and Eccles 2003) not independently verified here
mechanism
raising the reporting threshold (for example, tow-away versus traditional reporting) excludes lower-severity crashes from the record; poorly selected comparison groups leave treatment effects confounded with site differences
effect
"seriously underreport[ing]" of certain crash types under higher thresholds; no magnitude given. Prior RLC studies characterized as "not well-designed, lacking well-selected control groups, and used little data"
conditions / moderators
not established — general literature characterization, not a quantified test