What the evidence says
The measured evidence on red-light camera (RLC) crash effects is dominated by one finding: the individual study estimates disagree with each other far more than sampling error alone would produce. Across the pooled studies, total crash effect estimates show I2 = 90.4% (Q = 228.4, df = 22, p < .001), and injury crash effect estimates show I2 = 93.1% (p < .001). Both are classified with high confidence in direction and magnitude, since they come directly from statistical tests reported in the meta-analysis. This level of heterogeneity means a single pooled effect size for RLCs, taken on its own, obscures more than it reveals about what any given camera program produced.
Part of that heterogeneity tracks country. In a subgroup breakdown, US studies show I2 = 86.3% (p < .001) while Australian studies show I2 = 0% (p < .001) — the Australian estimates are internally consistent while the US estimates are not. This is reported with high confidence in both direction and magnitude.
Attempts to pin down what drives the heterogeneity mostly come up empty. Country and risk-of-bias classification do not consistently explain the heterogeneity across the full set of studies (medium confidence, negative finding). Warning-sign presence or absence shows no measured effect on RLC effectiveness (medium confidence direction, low confidence magnitude), reported as an absence of evidence rather than a tested null.
Two design factors do show a patterned relationship with the reported crash effect, though neither is quantified. Studies that account for regression to the mean report more moderate decreases in right-angle injury crashes than studies that do not — direction is reported with high confidence, magnitude with medium confidence, and the size of the difference is not given. Separately, when studies are split by quality of confounder control, the subset with better control finds a decrease in right-angle crashes that was not distinguished from zero, while the full set of studies together finds a decrease that was distinguished from zero — again reported with high confidence in direction, medium in magnitude, and no quantified gap between the two.
Outside of these two moderators, the sources do not identify what else produces the very high overall heterogeneity. No claim here addresses camera dwell time, enforcement threshold, intersection type, or period effects as heterogeneity sources — those factors are simply absent from this evidence base.
Where the evidence splits
Total and injury crash effect estimates disagree sharply across studies (I2 = 90.4% and 93.1%), but the sources do not resolve this into a single explained cause — country and risk of bias are explicitly ruled out as consistent explanations, and warning-sign presence is ruled out as a moderator. The split in raw effect estimates across studies is therefore unexplained by the sources on this page beyond those two ruled-out candidates.
The US and Australian subgroups disagree on how heterogeneous their own studies are (I2 = 86.3% vs. I2 = 0%). The sources report this country-level difference directly but, in the same breath, state that country does not consistently explain heterogeneity overall — so the page contains both a clear country-level split and a statement that country isn't a reliable explanation for heterogeneity in general. The sources do not reconcile that tension; it is unexplained here.
Studies with better confounder control and studies without it disagree on right-angle crash effects: the better-controlled subset finds a decrease not distinguished from zero, while the full set of studies finds a decrease that is distinguished from zero. Here the source does supply the design difference behind the split — quality of confounder control — even though it does not quantify the gap.
Studies that account for regression to the mean and studies that do not also disagree on right-angle injury crash effects, with RTM-adjusted studies reporting more moderate decreases. As with confounder control, the source names the design difference (RTM adjustment) but does not quantify how large the resulting split is.