The evidence base is weaker than the slogans
The Cochrane systematic review — the gold standard of evidence synthesis — looked at speed cameras across 35 studies and concluded the direction was positive but the size could not be trusted:
"…an overall magnitude of this effect is currently not deducible due to heterogeneity and lack of methodological rigour." The included studies were "of overall moderate quality at best." — Wilson, Willis et al., Cochrane Database of Systematic Reviews (2010)
Crucially, Cochrane found no randomised controlled trials — every study is an observational before-and-after, the design most vulnerable to statistical distortion. The first systematic review (Pilkington & Kinra, BMJ 2005) reviewed 14 studies, rated none high-quality, and concluded the "level of evidence is relatively poor."
Sources: Cochrane review (2010); Pilkington & Kinra, BMJ 2005. Note: "positive direction, untrustworthy magnitude" is the fair reading — not "cameras do nothing."
Regression to the mean: the illusion in the numbers
Cameras get placed where crashes recently spiked. Crashes then fall back toward normal on their own — and the fall gets credited to the camera. This is regression to the mean (RTM), and governments' own evaluations show it inflates the results.
Roads with a high number of crashes in one period "are likely to have fewer during the following period, even if no measures are taken… the effects of the intervention may be overestimated." — European Commission road-safety directorate, on RTM in speed enforcement
The meta-analyst Høye flagged that a headline −51% fatal-crash figure "could partly be explained by regression to the mean." And Erke's red-light-camera meta-analysis found results are "more favourable when there is a lack of control for regression to the mean" — i.e. the sloppier the study, the better cameras look.
Sources: UK 4-year camera evaluation (2005), Appendix H (Mountain & Maher); Høye, EU SafetyCube synopsis; Erke (2009), Accident Analysis & Prevention.
Red-light cameras trade one crash for another
Red-light cameras reduce dangerous side-on ("T-bone") crashes — but reliably increase rear-end collisions as drivers slam on the brakes to avoid a fine.
The critical meta-analysis (Erke 2009)
Overall crashes up ~15%; rear-end collisions up ~40%; right-angle crashes down ~10% (effects non-significant). Conclusion: "on the whole RLCs do not seem to be a successful safety measure."
The pro-camera rebuttal (FHWA 2005)
Right-angle crashes down ~25%, rear-end up ~15%, but a net +$18.5M economic benefit because the crashes prevented are more severe than the ones caused. Both sides agree on the trade-off — they disagree on the net.
A separate 2018 meta-analysis: overall crashes −12%, right-angle −24%, but rear-end +32%. The honest summary: red-light cameras shift the crash mix; whether that's a net win depends on the intersection — and the cleaner the statistics, the smaller the claimed benefit.
Sources: Erke 2009; FHWA red-light camera evaluation (2005); Accident Analysis & Prevention (2018).
The evaluators who mark their own homework
Much of Australia's pro-camera evidence comes from the Monash University Accident Research Centre (MUARC), whose evaluation of Victoria's fixed cameras found casualty-crash reductions of up to 47% on the camera-monitored approach. MUARC argues its estimates were "not inflated by regression to the mean," and that's an empirical argument worth taking seriously.
But two things are fair to note: its design is comparison-site regression, not the full empirical-Bayes method used to strip out RTM in the UK study above; and MUARC is funded by the road-safety agencies whose programs it evaluates. When the body that certifies the policy is paid by the body that runs it, independence is a reasonable question — not an accusation.
The bottom line Cameras probably do reduce crashes where they're properly sited — average-speed ("section") control looks the strongest of all. The defensible criticism is not "enforcement is useless." It's that (1) the benefit's size is routinely overstated by weak, RTM-contaminated studies, (2) red-light cameras create new crashes they rarely account for, and (3) the money and the siting — see The Goldmines — follow revenue, not risk.
Source: MUARC Report 307.