Generalizability of the Philadelphia Vacant Lot RCT (Branas/Moyer) to Other Cities

What the evidence says

The evidence on the Branas/Moyer-style Philadelphia vacant lot intervention itself is thin and single-city. A randomized controlled trial by Garvin et al., cited in Chowdhury (2022), found vacant lot conversion associated with reduced crime incidents, specifically decreases in gun assaults, alongside reductions in vandalism in one section of the city. A separate quasi-experimental study by Branas et al., using a difference-in-differences design, corroborates the direction on gun assaults: vacant lot greening was associated with consistent reductions in gun assaults across all four sections of Philadelphia studied. The same study found reductions in vandalism in only one section, and associations with residents reporting less stress and more exercise in select sections. None of the sources on this page state a causal mechanism connecting lot conversion to these outcomes — the claims report association, not pathway.

On the specific question this page is about — whether Philadelphia's results generalize to other cities — the only direct evidence is negative, and it concerns prediction, not the RCT's causal effect. A cross-validation experiment (Chowdhury 2022) found that predictive models trained on one city's vacant lot data do not generalize to another city: cross-city prediction accuracy was no better than random. No claim on this page tests whether the Branas/Moyer gun-assault or vandalism effect itself replicates when the intervention is run in a second city.

A separate body of evidence, from Yang (2025) and drawn from clinical-trial transportability research (an HIV trial transported to the U.S., Thailand, and Ethiopia), establishes general mechanisms relevant to any RCT-to-new-setting question, though not vacant lots specifically. Yang (2025) reports that RCT participants are typically younger and healthier than real-world patients, and that treatment effects can, in that domain, be successfully transported to external populations across different geographic and clinical settings. Several named threats to that kind of transport are described: selection bias, when the target population differs systematically from the trial population; unmeasured confounding, when key variables are missing or inconsistently captured; measurement error, when outcomes recorded in the field do not align with trial-defined endpoints; and immortal time bias, when treatment initiation and follow-up are not anchored to a common baseline. Yang (2025) also reports that unobserved conditional shifts often dominate over observed covariate shifts in determining whether a transported effect holds.

Taken together, the sources supply Philadelphia-specific outcome estimates, a general finding that spatial prediction models fail to generalize across cities, and a general methodological account of what can make treatment-effect transport succeed or fail — but no source measures whether the Philadelphia vacant lot RCT's crime effect specifically replicates elsewhere. That is an absence, not a null result: the transportability of Branas/Moyer's findings to another city has not been tested by any claim on this page.

Where the evidence splits

No contradictions in the current sources.

Thresholds and dose-response

No threshold established in the current sources.

Open empirical questions

• Whether the Philadelphia gun-assault and vandalism reductions from the Branas/Moyer-style RCT and its companion quasi-experiment persist when the intervention is run in a different city.
• Whether the cross-city predictive-model failure found for vacant lot data (no better than random) also applies to the RCT's causal effect estimates, or is specific to predictive modeling.
• What the magnitude of the age and health differential is between RCT trial populations and real-world populations, beyond the reported direction.
• What "sufficiently similar" means in operational terms for two populations to be expected to show a replicated treatment effect.