Road effects on animal populations: a systematic review and meta-analysis

1 Introduction

Road exposure can affect wildlife through several biological pathways, including collision mortality, behavioural avoidance, movement disruption, and changes in roadside habitat (Forman and Alexander, 1998; Trombulak and Frissell, 2000). The resulting responses vary with life-history traits such as body size, reproductive rate, and movement capacity (Fahrig and Rytwinski, 2009; Rytwinski and Fahrig, 2012). Mortality, occurrence, abundance, behaviour, movement, habitat use, and connectivity therefore describe different biological processes and temporal scales rather than interchangeable indicators of one road effect. A synthesis that combines these endpoints without retaining their distinct meanings can obscure how road exposure is associated with wildlife responses.

Quantitative syntheses have addressed defined combinations of taxa, outcomes, and exposure contrasts. Existing examples include infrastructure-proximity effects on bird and mammal abundance, vertebrate abundance responses in relation to life history, bird species richness near roads, and mammalian road- crossing barriers (Benítez-López et al., 2010; Rytwinski and Fahrig, 2012; Kroeger et al., 2022; Chen and Koprowski, 2019). These syntheses harmonised evidence within separate metric families—mean species abundance, Pearson’s r, Hedges’ g, or logit-transformed crossing proportions—which preserved comparability within each analysis. Across analyses, however, the metrics represent different biological estimands. The remaining synthesis problem is to examine multiple response domains within one evidence framework while preserving the interpretation of each measurement scale.

Road and infrastructure attributes provide a second boundary for comparative inference. Prior syntheses have examined infrastructure distance and traffic intensity, road width and surface, road type, and crossing-structure design within specific evidence domains (Benítez-López et al., 2010; Chen and Koprowski, 2019; Kroeger et al., 2022; Denneboom et al., 2021). These attributes remain unevenly reported across primary studies: an avian systematic review documented inconsistent collection of traffic, habitat, and temporal variables (Johnson et al., 2022), and a crossing-structure meta-analysis lacked sufficient traffic-volume data for meta-regression (Denneboom et al., 2021). Quantifying which road characteristics are available in the eligible evidence base is therefore necessary for interpreting and extending road-effect estimates.

We therefore synthesised road-associated wildlife responses in parallel, scale-specific models. Hedges’ g, lnRR, and Fisher’s Z represent standardised group differences, proportional changes, and correlation-based associations, respectively, and were interpreted within their respective estimands. We estimated road-proximity associations, examined variation across response categories and vertebrate taxa, and quantified the availability of traffic volume and other road characteristics for comparative analysis. The review included observational comparisons of free-ranging terrestrial vertebrates between road-exposed and reference conditions, excluding aquatic-only or invertebrate studies, experimental manipulations of road features, and evaluations of mitigation structures.

2 Methods

2.1 Literature search and screening

We searched Web of Science (Core Collection) and Scopus using a structured Boolean query that combined road-infrastructure terms (road, highway, traffic*) with wildlife-response terms (abundance, density, activity, richness, genetic) and proximity or gradient descriptors. No start-date restriction was imposed; the search captured all indexed records through the date of extraction. English-language publications and records with English-language abstracts were eligible. Web of Science returned 3,240 records and Scopus returned 2,377 records, yielding 5,617 records before deduplication. The complete search strings and field codes are reproduced in Supplementary Methods S1.

Study selection was reported in accordance with the PRISMA 2020 checklist (Page et al., 2021). Deduplication by DOI matching where available, supplemented by normalised title, first-author, and publication-year matching, reduced the 5,617 records to 4,155 unique entries. A single reviewer screened titles and abstracts against predefined eligibility criteria and retained 995 records for full-text review. Of these, 32 full texts could not be retrieved. Assessment of the remaining 963 articles excluded 165 and retained 798 publications. The full-text exclusion taxonomy is defined in Supplementary Methods S2. Supplementary Table S2 reports the itemised exclusion counts, and Supplementary Table S1 summarises the screening pathway.

Data extraction excluded a further 14 publications that lacked quantitative information sufficient for effect-size computation, yielding 784 included publications. The search covered peer-reviewed literature indexed in Web of Science and Scopus. Grey literature, dissertations, conference proceedings, and records without English-language abstracts were outside the search frame.

We conducted a full-text critical appraisal of 246 studies. Quality sensitivity analyses included appraised studies represented in the three scale-specific models. One reviewer assessed eight prespecified domains adapted to heterogeneous observational and quasi-experimental designs: sample selection, exposure measurement, outcome measurement, confounding control, replication structure, statistical analysis, missing data, and selective reporting. Overall judgments followed a prespecified hierarchical rule: any high-risk domain yielded an overall high-risk judgment; otherwise, any domain judged as having some concerns or unclear information yielded an overall judgment of some concerns; overall low risk required low-risk judgments across all eight domains. No numeric quality score was calculated. Supplementary Methods S7 provides the complete decision rules.

2.2 Effect-size extraction and coding

We extracted effect sizes on three separate metric scales to accommodate heterogeneous reporting formats. Group means and standard deviations were converted to Hedges’ g. Treatment and control means with associated dispersion were converted to the log response ratio (lnRR), defined as the natural-log proportional change relative to a control condition. Correlation coefficients were transformed to Fisher’s Z. Additional raw formats, including odds ratios, proportions, and complete 2 × 2 tables, were converted to the corresponding scale using standard formulae. Sampling variances were computed with the analytical expressions implemented in the metafor package (Viechtbauer, 2010). Initial extraction yielded 9,825 effect-size records from 784 publications, reflecting multiple species, sites, or response metrics within individual studies.

Effect direction followed the source contrast and predictor coding. Road-exposed and reference groups were explicitly oriented; distance, proximity, traffic-volume, and road-density gradients followed their stated direction, with reversal only when an increasing-distance coefficient was re-expressed as increasing proximity. Unsigned omnibus statistics entered the dataset only when source values supported one uniquely reconstructable road-specific contrast. Signs retain endpoint-specific meaning, including greater roadkill, lower abundance or activity, and greater genetic differentiation. Full decision rules are provided in Supplementary Methods S5.

The allocation unit was one extractable road-specific contrast. Each contrast entered at most one metric family: Hedges’ g for standardized independent-group differences or an explicitly supported binary log-odds mapping, lnRR for a positive road-exposed/reference mean or rate ratio, and Fisher’s Z for a signed Pearson correlation. Complete means, standard deviations, and sample sizes took priority over approximate conversions. Point estimates and sampling variances referred to the same contrast and scale, and the three metric families were analysed separately.

These metric families encompassed different reported response variables rather than fixed biological domains: Hedges’ g was used for standardized group contrasts such as abundance, activity, behaviour, mortality, or habitat-use responses; lnRR was used for positive response means or rates such as abundance, activity, and roadkill; and Fisher’s Z was used for signed correlations between animal responses and continuous road-exposure gradients. The full metric-to-response mapping and examples are provided in Supplementary Methods S6.

Cross-scale and approximate conversions followed prespecified rules and source-reported inputs. Incompatible estimands, descriptive quantities, and records without recoverable point estimates or positive sampling variances were excluded. The resulting dataset contained 1,732 records from 309 studies, including 1,007 estimable records from 222 studies. Restriction to Hedges’ g, lnRR, or Fisher’s Z with absolute effect sizes no greater than 10 and sampling variances no greater than 25 yielded 643 analysed effects from 160 studies (Table 1). Supplementary Methods S4 defines these analytic sample stages, and Supplementary Methods S6 provides the full conversion mapping.

Table 1. Analytic effect-size flow from extraction to the three scale-specific models.

Each retained effect was coded for response type, taxonomic group, and road characteristics. Annual average daily traffic volume and road-feature descriptors were extracted where reported. Road-characteristic extraction and harmonisation are described in Supplementary Methods S3. Of the 643 analysed effects, 449 from 120 studies included a reported total sample size (Supplementary Table S6).

Each effect was assigned to one of eight mutually exclusive response categories: Roadkill and mortality, Abundance and density, Activity, Behavioural avoidance, Habitat use and selection, Genetic differentiation, Fitness, and Other endpoints. When a response matched more than one category, categories were applied in this order; unmatched endpoints entered Other endpoints. A study could contribute effects to multiple categories, but each effect entered only one. Supplementary Table S3 defines the categories and overlap rules.

2.3 Statistical models

We fitted three-level random-effects models in which individual effect sizes (level 1) were nested within studies (level 2), thereby partitioning total heterogeneity into within-study and between-study variance components (Konstantopoulos, 2011; Van den Noortgate et al., 2013; Cheung, 2014). This structure accommodates the non-independence that arises when multiple effect sizes are extracted from the same publication (Cheung, 2019; Rodgers and Pustejovsky, 2021). To construct approximate variance–covariance matrices for correlated outcomes within studies, we assumed a within-study correlation of ρ = 0.5, a conventional default when true correlations among outcomes are unknown. Because the three effect-size scales (Hedges’ g, lnRR, Fisher’s Z) differ in distributional properties and interpretive units, separate models were fitted for each scale.

We evaluated the sensitivity of the pooled estimates through three complementary approaches. First, robust variance estimation (RVE) with the CR2 small-sample correction and study-level clustering reduced reliance on the assumed within-study correlation structure (Tanner-Smith and Tipton, 2014). Second, a study-averaged random-effects model collapsed multiple effect sizes per study to a single arithmetic mean, eliminating within-study dependence and providing a structural sensitivity check. Third, moderator meta-regressions tested response type and taxon as categorical moderators. Additional complete-case analyses tested log-transformed annual average daily traffic (AADT), road type, and publication year among studies reporting the relevant variable.

Study quality entered the analysis through prespecified sensitivity models. We refitted each scale-specific three-level model after excluding studies with an overall high-risk judgment, then repeated the procedure for each appraisal domain after excluding studies rated high risk in that domain. The scale-specific models included all eligible effects, and the sensitivity estimates quantified how internal-validity limitations affected the pooled results.

Publication bias was assessed using a multi-method battery applied to each scale’s pooled dataset. Egger’s regression test (Egger et al., 1997) evaluated funnel-plot asymmetry through the association between effect size and standard error. Begg’s rank-correlation test (Begg and Mazumdar, 1994) provided a non-parametric complement. Trim-and-fill (Duval and Tweedie, 2000) estimated the number of potentially missing studies and produced bias-adjusted pooled estimates. The precision-effect estimate with standard error (PEESE) offered a conditional estimator less susceptible to small-study effects. All models were implemented with the metafor package (Viechtbauer, 2010), following current guidance for ecology and evolution meta-analyses (Nakagawa et al., 2021). Supplementary Table S8 reports the diagnostic estimates.

3 Results

3.1 Evidence base and overall effects

Road proximity was associated with lower responses in the standardized mean-difference and ratio analyses, while the correlation estimate was positive but uncertain. The three scale-specific models included 643 effects from 160 studies (Table 1): Hedges g (k = 461, 119 studies), lnRR (k = 104, 33 studies), and Fisher’s Z (k = 78, 19 studies; Figure 1). Figure 2 reports the search and screening pathway.

The pooled estimates were g = −0.1131 (95% CI [−0.2608, 0.0346], p = 0.133), lnRR = −0.3872 (95% CI [−0.7293, −0.0452], p = 0.027), and Z = 0.1691 (95% CI [−0.1915, 0.5297], p = 0.353; Figure 1). Only the ratio-scale confidence interval excluded zero; the standardized and correlation intervals included zero.

Heterogeneity remained substantial. Within- and between-study variance components were 0.4313 and 0.4246 for g, 0.3011 and 1.2854 for lnRR, and 0.5248 and 0.1685 for Z, respectively (Supplementary Table S5).

3.2 Response-type heterogeneity

Response domains did not show one common pattern across the three estimands. Between-response tests detected heterogeneity for Hedges’ g (Q_between = 16.55, p = 0.021) and Fisher’s Z (Q_between = 9.61, p = 0.022), but not for lnRR (Q_between = 5.96, p = 0.114). Category definitions are provided in Supplementary Table S3.

On the standardized mean-difference scale, the estimate for Other endpoints was negative (g = −0.184, 95% CI [−0.357,−0.011]; k = 213, 70 studies). All other response-category confidence intervals included zero. Subgroup estimates are presented in Figure 3.

For the same broad endpoint family, the ratio analysis indicated lower road-exposed mortality or roadkill rates (lnRR = −0.893, 95% CI [−1.577,−0.210], p = 0.010; k = 46, 15 studies), whereas the correlation estimate was positive (Z = 0.822, 95% CI [0.393,1.251]; k = 35, 6 studies). Subgroup estimates are reported in Table 2.

Table 2. Response-type subgroup estimates by effect-size scale.

3.3 Taxon-specific Responses

Between-taxon tests were Q_between = 1.14 (p = 0.889, five groups) for Hedges’ g and Q_between = 0.28 (p = 0.963, four groups) for lnRR. For mammals, the Hedges’ g confidence interval included zero (g = −0.086, 95% CI [−0.245,0.074]; k = 253, 65 studies), whereas the lnRR interval excluded zero (lnRR = −0.371, 95% CI [−0.700,−0.041]; k = 68, 18 studies). On the correlation scale, the mammal estimate was positive (Z = 0.474, 95% CI [0.060,0.888]; k = 40, 10 studies), the bird confidence interval included zero, and the two-group taxon test was not significant (Q_between = 2.38, p = 0.123).

Reptile estimates were g = −0.150 (95% CI [−0.774,0.476]; k = 49, 15 studies) and lnRR = −0.640 (95% CI [−1.555,0.275]; k = 7, 4 studies). Study-level support within response-specific reptile cells was sparse (Figure 4). Complete taxon subgroup estimates are reported in Table 3.

Table 3. Taxon subgroup estimates by effect-size scale.

3.4 Road variables and reporting gaps

Road-characteristic reporting was sparse across the effect-size dataset. Annual average daily traffic (AADT) was available for 21.9% of records (379 of 1,732; 49 studies), number of lanes for 13.8% (239 records; 38 studies), road width for 6.6% (115 records; 20 studies), and speed limit for 5.3% (91 records; 21 studies; Table 4; Figure 5; Supplementary Table S4). Among the 643 analysed effects, 91 effects from 17 studies reported AADT and 202 effects from 65 studies reported road type (Supplementary Table S6).

Table 4. Road-feature reporting rates across converted and estimable effect-sizerecords.

The AADT meta-regression for Hedges’ g included 72 effects from 15 studies and gave a slope of β = 0.0491 (SE = 0.0704, 95% CI [−0.0888,0.1871], p = 0.488; Table 5; Supplementary Table S7).

Table 5. Continuous moderator meta-regression results.

3.5 Robustness diagnostics

Dependence-structure sensitivity analyses preserved the direction of all three pooled estimates. Under robust variance estimation with the CR2 correction, the confidence interval excluded zero for lnRR and included zero for Hedges’ g and Fisher’s Z; after averaging effects within studies, all three confidence intervals included zero (Table 6).

Table 6. Dependence-structure sensitivity estimates by effect-size scale.

Small-study-effect tests differed across scales and methods: Egger’s test detected asymmetry for lnRR, whereas Begg’s test detected asymmetry for Hedges’ g. Trim-and-fill imputed no effects on any scale, and PEESE was estimated only for Fisher’s Z (Supplementary Table S8).

Excluding 113 studies with an overall high-risk judgment retained 72 Hedges’ g studies, 17 lnRR studies, and 13 Fisher’s Z studies. The three point estimates retained their directions, and all confidence intervals included zero (Supplementary Table S9).

4 Discussion

4.1 Overall road-associated responses and ecological pathways

Road exposure was associated with lower wildlife responses, with the clearest evidence on the proportional scale. The approximately 32% reduction in the lnRR analysis indicates that road-associated changes are detectable across studies, even though the standardized and correlation-based estimates were less precise.

The response-type results identify the ecological pathways behind this pattern. Lower ratio-scale responses for roadkill or mortality and Activity are consistent with roads changing both how animals use space and how often they encounter vehicles. These changes can reduce local abundance or activity, while the accumulated loss of individuals can alter population processes; earlier quantitative syntheses similarly reported lower mammal and bird abundance near infrastructure (Benítez-López et al., 2010).

Behavioural avoidance changes where animals move and forage, whereas road mortality removes individuals directly. Population consequences become strongest when deaths are concentrated in reproductive adults or exceed recruitment (Moore et al., 2023). Our meta-analysis captures this mortality-related pressure, but the available studies rarely measure recruitment or age-specific survival well enough to reconstruct the full population trajectory.

Variation among sites reflects different species–road interactions. Species with large movement ranges or low reproductive rates can be strongly affected by habitat loss and mortality, whereas species attracted to roadside resources or released from predators may show increased use near roads. Road studies have documented negative, positive, and null responses under these different combinations of movement behaviour and life history (Fahrig and Rytwinski, 2009; Rytwinski and Fahrig, 2012). The wide prediction intervals in the present synthesis therefore mark a mixture of ecological pathways: road exposure can reduce habitat use and population performance in some settings while producing weaker or positive responses in others.

Reptiles illustrate how these mechanisms can operate together. Their small body size, ectothermy, and often slow locomotion make road surfaces both potential thermal resources and high-risk movement corridors. Individuals may enter roads to bask, nest, or move between seasonal habitats, while slow crossing and freezing responses can increase collision exposure. Roads can therefore generate simultaneous signals of attraction, avoidance, roadkill, and reduced connectivity. The demographic consequences are likely to be greatest for turtles and snakes that mature late, reproduce slowly, and depend on high adult survival; even modest road mortality can then reduce population viability (Andrews et al., 2015; Langen and Colino-Rabanal, 2026).

4.2 Road exposure gradients

The evidence supports road-presence effects more clearly than a general traffic-volume gradient. AADT was reported for 21.9% of the full dataset (379 records from 49 studies), but only 91 of the 643 analysed effects, from 17 studies, included AADT. The Hedges’ g AADT model did not resolve a consistent cross-study gradient (Supplementary Table S7; Table 5). The synthesis therefore captures the composite effect of living near a road more reliably than the way that effect changes with traffic volume (Denneboom et al., 2024).

Traffic volume does not operate as a single biological dose. High- and low-traffic roads can generate different mortality patterns because animals alter crossing and avoidance behaviour as vehicle encounters change (Denneboom et al., 2024; Jacobson et al., 2016). Species may approach, wait, cross, avoid, or be killed at different traffic levels, so the same increase in vehicles can alter different components of the road effect in different species.

Road width and surface add physical structure to this exposure gradient. Across mammal crossing studies, these characteristics explained part of the heterogeneity in barrier effects, while traffic volume could not be modelled consistently because it was rarely available (Chen and Koprowski, 2019). Joint reporting of traffic volume, road geometry, speed, surface, and habitat context would allow future syntheses to separate road-presence avoidance, traffic-mediated disturbance, and collision risk, following the standardised reporting framework proposed for road-ecology studies (Johnson et al., 2022).

4.3 Inference and management implications

Road-associated responses were directionally stable across dependence-structure and study-quality analyses, with the clearest average reduction on the ratio scale. This consistency indicates that roads exert a broad ecological pressure across the evidence base, while the magnitude of response varies among endpoints and sites. Precision depended on how within-study dependence was handled, especially for the ratio-scale estimate, but the central direction persisted.

The management implication is pathway-specific. Crossing structures can address movement barriers, but their effectiveness should be evaluated against pre-road or no-mitigation baselines rather than assumed from their presence alone (Soanes et al., 2024). Speed management and other road-kill measures target collision mortality, for which meta-analytic evidence shows that effectiveness varies among intervention types and study designs (Rytwinski et al., 2016). At the network scale, graph-based prioritisation can identify roads that divide high-quality habitat or make the largest contribution to lost connectivity (Ascensao et al., 2019). These measures should be paired with monitoring of movement, mortality, and population performance in species exposed to high road-associated risk (Grilo et al., 2021).

Future syntheses should treat body size and life history as mechanistic moderators rather than generic taxonomic descriptors. Species-level data on body mass, locomotor speed, maturation, adult survival, sex-biased movement, and road-crossing behaviour would allow researchers to distinguish barrier effects from mortality effects and to test why the same road produces avoidance in some species but attraction or repeated crossings in others. Such data would also improve the selection of reptile mitigation measures, including rerouting, seasonal closures, verge management, fencing, and passage structures (Andrews et al., 2015; Langen and Colino-Rabanal, 2026). The present dataset supports this mechanistic interpretation, while a quantitative body-size moderation analysis requires harmonised species-level trait data that were not available across the included studies.

Most included studies were observational, so the synthesis estimates road-associated responses rather than the effect of any single intervention. Longitudinal or before–after controlled studies should combine traffic exposure, habitat use, movement, mortality, and demographic rates to test which pathway drives local population change.

5 Conclusions

Road exposure was associated with lower wildlife responses, with the clearest average signal on the proportional scale: road-exposed responses were approximately 32% lower in the lnRR analysis. The synthesis indicates that roads affect multiple components of wildlife ecology, including movement, activity, habitat use, mortality, and population performance. The direction and magnitude of these responses vary among species and sites because road effects combine behavioural avoidance, barrier effects, mortality, and life-history differences.

The evidence currently identifies road-presence effects more clearly than a transferable traffic-volume threshold. Effective mitigation should therefore be matched to the dominant pathway at each site: maintaining permeability and prioritising fragmented road links for barrier effects, and using speed management or road-kill measures where collision mortality dominates. Future road studies should combine traffic exposure, road structure, movement, mortality, and demographic rates so that mitigation can be evaluated against the process it is intended to change.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fevo.2026.1940830/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Pedro Gonçalves Vaz, National Institute for Agricultural and Veterinary Research (INIAV), Portugal

Reviewed by: Tom Langen, Clarkson University, United States

Ravi Bandara Dissanayake, Food and Agriculture Organization of the United Nations (Italy), Italy

Diana Cupsa, University of Oradea, Romania