Human activities in natural areas can impose both lethal and non-lethal impacts on animals. Furthermore, anthropogenic disturbance is analogous to predation risk and can cause animals to adjust their behaviors to avoid humans. Quantifying if disturbance-induced behavioral shifts affect individual fitness or population dynamics is needed to guide science-based conservation and management decisions. We experimentally disturbed GPS-marked mallards (Anas platyrhynchos) on sanctuaries weekly to evaluate the effects of brief pulses (1 hr) of non-lethal anthropogenic disturbance on individual survival. We used Cox proportional hazard models to examine how single and cumulative disturbance affected survival and tested whether body mass or hunting season mediated the effects of disturbance. One hundred and eighty-eight mallards were disturbed ≥1 time resulting in 629 disturbance encounters. Only 3 individuals died immediately following disturbance, representing <0.5% of encounters. Collectively, we found no effect of disturbance on daily survival, and our cumulative disturbance model showed undisturbed mallards had lower survival than disturbed mallards. Standardized body mass or hunting season did not mediate the effect of disturbance on survival. Together, we concluded no effect of our brief experimental disturbance treatments on mallard survival. Instead, diurnal sanctuary use and individual characteristics, including age, sex, and standardized body mass, affected survival. Diurnal sanctuary use was positively related to survival, and for every 20% increase in diurnal sanctuary use, the risk of mortality decreased by 15%. Additionally, female mallards were 2.7 times more likely to die compared to males, and juveniles had 53% greater risk of mortality than adults. Lastly, for every 100g heavier than average mallards were, we found a 23% lower risk of mortality during our study. If a primary goal of waterfowl sanctuary is including non-consumptive recreational use, our results suggest controlled access (e.g., ~1 hr/week) may have minimal effects on survival and be consistent with multi-use objectives on public lands with waterfowl sanctuaries. If additional recreational access to support multiple public uses is a goal on public lands managed as sanctuaries, we recommend future work identify disturbance thresholds at which point survival or other fitness metrics are impacted by disturbance related to public uses of protected areas.

Study area

Our study took place in northwestern Tennessee, USA (~6,000 km²) which encompassed 3 federally owned and 7 state-owned sanctuaries (Figure 1). Agriculture was the dominant land use in the region, and the majority of agricultural fields were harvested in late summer and autumn; however, intensive wetland management designed to attract waterfowl for recreational or conservation purposes were commonplace on both federally and state-owned sanctuaries, public hunted areas, and private property throughout our study area (Gray et al. 2021, Masto et al. 2024a, Highway et al. 2025). Managed wetlands were impounded and often had water control structures that allowed managers and landowners to produce and then flood annual seed-producing moist-soil vegetation and/ or planted crops (e.g., corn [Zea mays], millet [Echinochloa spp.; Urochloa spp.], and rice [Oryza sativa]; Gray et al. 2021). Our study area contained abundant food resources for wintering waterfowl across public and private properties. In fact, during a concurrent study, Masto (2023) estimated managed moist-soil vegetation within our study area contributed 6.7 million duck energy days and unharvested flooded corn contributed 85.6 mil duck energy days.

All sanctuaries, besides one state-managed sanctuary, contained both managed moist-soil and planted crops (i.e., predominantly corn) during our study: on average, sanctuaries contained 56.1 ha (range = 0–214.2 ha; median = 18.0 ha) of managed moist-soil and 25.9 ha (range = 0–58.0 ha; median = 25.6 ha) of flooded unharvested corn. State sanctuaries tended to contain more flooded unharvested corn than managed moist-soil (x̄ = 52.9% in corn; range = 16.1–66.7 %), and the federal sanctuaries tended to contain more managed moist-soil than corn (x̄ = 65% in moist-soil; range = 32–85 %). The remaining cover on sanctuaries were predominantly woody wetlands and harvested agriculture. Hunting and other public access (i.e., vehicular and foot traffic) was prohibited on portions of state and federal lands managed as sanctuaries on or before 15 November and until 1 March each year.

During our study, the waterfowl hunting season lasted 60 days over two segments with two days in mid- to late-November and the remaining 58 days in December through the end of January, with opening weekends in mid-November for Reelfoot Lake and late-November for the remainder of the study area (Figure 1). Following opening weekends, the waterfowl hunting season then closed until early December when it reopened and remained open until the end of January each year. Weekly aerial surveys were used to estimate hunting activity across the study area. Hunting activity was consistent throughout the 60-day waterfowl season and did not differ by week of year or between weekends and weekdays (Masto et al. 2024c). Additionally, hunting intensity was highest within river channels, near waterfowl sanctuaries (Blake-Bradshaw et al. 2023), and on public and privately hunted areas that typically had no specified rest-days (Masto et al. 2024a, c).

Animal capture and telemetry

We captured mallards using rocket-nets and swim-in traps on waterfowl sanctuaries in western Tennessee, USA during winter (1 November–28 February 2019 –2022). We aged mallards as juveniles or adults using cloacal inversion and wing plumage characteristics (Carney 1992). All mallards were weighed to index body condition (± 1 g) and banded with USGS standard aluminum tarsal bands. We equipped mallards with 20-g solar-rechargeable and remotely programmable OrniTrack Global Positioning System-Global System for Mobile transmitters (GPS-GSM; Ornitela, UAB Švitrigailos, Vilnius, Lithuania). We attached transmitters via dorsally-mounted body harnesses made of automotive moisture-wicking elastic ribbon (Masto et al. 2022, Highway et al. 2024). We programed transmitters to record GPS locations at 1-hr and 2-hr intervals. We reduced GPS-frequency to once every 36 hr when transmitter battery was <25%. We monitored individuals until transmitter failure, an individual was reported legally harvested, or we identified natural mortality using tri-axial accelerometer graphs.

Explanatory variables

To evaluate the influence of sanctuary disturbance on wintering mallard survival, we conducted three distinct disturbance treatments of increasing intensity on sanctuaries: accessing sanctuaries by covered vehicle (i.e., truck), pedestrian, or uncovered vehicle (i.e., all-terrain vehicle [ATV] or motor boat; Pease et al. 2005, Guay et al. 2019; Figure S1). For the truck disturbance, observers drove predetermined routes along roads or levees estimating waterfowl abundance from pre-specified vantage points. Route length differed among sanctuaries based on sanctuary size and vegetative cover (e.g., forested vs. open water). For the intermediate intensity pedestrian disturbance, two observers walked separate routes along levees (i.e., 4.8 km/hr) during a given visit and altered routes when normal routes were inaccessible (e.g., flooding events; Bregnballe et al. 2009, Guay et al. 2019; Figure S1). For the highest intensity disturbance treatment, we drove an ATV or a surface drive motorboat into wetlands (Havera et al. 1992, Knapton et al. 2000; Figure S1). We maintained speed of motorized vehicles to approximately 16 km/hr for 10 minutes, stopped for 5 minutes, and repeated until 1 hour elapsed.

We conducted a disturbance treatment approximately once per week, separating treatments on each sanctuary by ≥5 days to isolate disturbance events and allow GPS-marked ducks to resume normal activities (Dooley et al. 2010b). We disturbed sanctuaries for one hour between 0700 and 1100 hours. We determined an individual received a disturbance treatment when they had ≥1 GPS location on the disturbed sanctuary between sunrise and the end of the disturbance. This determination ensured that a GPS-marked mallard was exposed to a disturbance event. For instance, if we compared hourly GPS-marked mallard locations/timestamps to just the duration of the disturbance treatment (i.e., 1 hour), we could incorrectly determine whether a mallard was disturbed, especially if it left the sanctuary immediately in response to disturbance. Rather, our data and previous studies with high-frequency GPS-GSM transmitters on waterfowl show that ducks move to a single wetland or sanctuary area approximately at dawn and remain in the same location throughout the day (McDuie et al. 2019, Blake-Bradshaw et al. 2024). For instance, if a mallard was located on a sanctuary within an hour of nautical dawn (i.e., morning nautical twilight starts; ~1 hour prior to sunrise), it was highly likely the individual would remain on that same sanctuary throughout that day (Figure S2). Thus, we are confident this approach accurately represented whether a GPS-marked individual was disturbed.

We collapsed disturbance treatments into a single binary variable for model convergence (i.e., yes/no; Table 1). Additionally, we calculated cumulative disturbance as the number of disturbances experienced during winter. Few mallards had >3 disturbance encounters so we binned the upper end of the cumulative disturbance covariate to 3 (Table 1). We included age and sex as independent variables (Table 1). Support for body condition indices (i.e., function of mass and structural components) is mixed and may not accurately represent of energetic stores in waterfowl (Klimas et al. 2020); thus, we standardized individual body mass by averaging mass for the four age-sex cohorts (i.e., adult and juvenile males, adult and juvenile females) and subtracted this value from each individual’s weight depending on their age-sex cohort, uncorrected for a structural component (Veon et al. 2024; Table 1). Veon et al. (2024) found standardized body mass and structurally corrected body condition were highly correlated (|r| ≥ 0.9), so we use standardized body mass as a proxy for body condition. To evaluate if mortality risk (i.e., hunting season) mediated the effects of disturbance on mallard survival, we created a binary variable representing hunting and non-hunting periods. Hunting periods occurred during December and January and non-hunt periods during November and February (Table 1). Last, we calculated a 3-day rolling average of proportion of diurnal locations on a sanctuary for each individual to represent short-term propensity of an individual mallard to use sanctuary (Table 1). This time-dependent variable was represented by a proportion (i.e., 0–1) and updated daily with individual’s mean diurnal sanctuary use over the three preceding days, increasing towards 1 over time as diurnal sanctuary use increased and towards 0 with decreased diurnal sanctuary use. We defined the diurnal period as 30 minutes prior to sunrise until sunset, mirroring legal hunting hours.

Survival analyses

We estimated daily survival during winter 1 November–28 February 2019–2022 via known-fate Cox proportional hazard models using the survival package (Therneau and Grambsch 2000, Therneau 2022). We estimated robust standard errors around parameter estimates to account for repeated sampling of the same individual (Therneau and Grambsch 2000). We generated daily encounter histories for each individual using a staggered entry design due to differences in capture dates (Pollock et al. 1989). If a transmitter did not report GPS data during a given day, we right-censored the individual for that date and re-entered them into the daily encounter histories on the next date with GPS locations. We left-censored each mallard for 4 days after they were equipped with a GPS-transmitter to allow acclimation to the transmitter (Cox and Afton 1998) and right-censored individuals that were still alive at the end of the study or stopped providing data. If a transmitter stopped reporting data, we assumed transmitter failure and right-censored the individual on the last day data were received. We assumed unbiased censoring (Benson et al. 2014). To model daily and 120-day winter survival rates, we generated daily survival estimates from a Kaplan-Meier survival curve using ‘survfit’ in the survival package (Karp and Gehr 2020, Therneau 2022).

Statistical analyses

Prior to model formulation, we checked for multicollinearity and removed one of a pair of correlated variables (|r| ≥ 0.6; Dormann et al. 2013). We tested whether covariates met the proportional hazards assumption by examining Schoenfeld residuals (Therneau 2022). We standardized continuous variables by z-scoring. To estimate daily survival, we fitted separate models during the hunt and non-hunt periods, both included the binary disturbance variable. We used the log-rank χ² test with an a priori α = 0.05 to test whether disturbance influenced daily survival during hunt and non-hunt periods. We fit a model to evaluate whether cumulative disturbance(s) affected daily survival and another model to evaluate whether survival differed more generally between mallards that received ≥1 experimental disturbance and mallards that did not receive any experimental disturbances (i.e., undisturbed mallards). We fit the remaining a priori alternative hypotheses (Table 1) and implemented model selection using Akaike’s Information Criterion corrected for small sample sizes (AIC_c; Burnham and Anderson 2002). We present predictions for variables which 85% confidence intervals did not overlap zero and associated hazard ratios (Therneau and Grambsch 2000). We evaluated competing models for parsimony to avoid uninformative parameters and performed model averaging across models ≤2 ΔAIC_c if there was model selection uncertainty (Burnham and Anderson 2002: 151–153, Arnold 2010). We evaluated goodness-of-fit using Harrell’s Concordance index (Harrell et al. 1982, Thernau 2022). All analyses were conducted in program R v.4.2.2 (R Core Team 2022).