Involving local communities for effective citizen science: Determining game species' reproductive status to assess hunting effects in tropical forests

1Department of Natural Sciences, Manchester Metropolitan University, Manchester, UK; 2Mamirauá Sustainable Development Institute, Tefé, Amazonas, Brazil; 3Rede de Pesquisa para Estudos sobre Diversidade, Conservação e Uso da Fauna na Amazônia (RedeFauna), Manaus, Amazonas, Brazil; 4ComFauna, Comunidad de Manejo de Fauna Silvestre en la Amazonía y en Latinoamérica, Iquitos, Peru; 5Center for International Forestry Research (CIFOR), Situ Gede, Kota Bogor, Jawa Barat, Indonesia; 6Institute of Biological Sciences and Health, Federal University of Alagoas, Maceió, Brazil; 7Faculty of Environmental Sciences and Natural Resource Management, Norwegian University of Life Sciences, Ås, Norway; 8Museo de Culturas Indígenas Amazónicas, Iquitos, Loreto, Peru; 9Departament de Sanitat i Anatomia Animals, Universitat Autònoma de Barcelona, Bellaterra, Barcelona, Spain and 10Programa de Pós-Graduação em Saúde e Produção Animal na Amazônia, Universidade Federal Rural da Amazônia (UFRA), Belém, Pará, Brazil


| INTRODUC TI ON
Ensuring the sustainability of wildlife hunting in tropical forests is crucial to guarantee the conservation of game species and safeguard the food security and livelihoods of millions of forest dwellers (Coad et al., 2019). Since the early 1990s, a wide range of mathematical models have been developed to assess hunting sustainability (Weinbaum, Brashares, Golden, & Getz, 2013).
These models require accurate information on population abundance, reproductive and mortality patterns of the hunted species (Weinbaum et al., 2013).
Reproductive performance of wild vertebrates is usually studied by directly examining animals after capture and restrain, or from direct field observations of births (e.g. Zhang et al., 2007). Because both methods are challenging, especially for tropical forest animals (Fragoso et al., 2016), researchers are limited by the lack of basic information on age of first reproduction, breeding cycles and pregnancy rates of hunted species (Milner-Gulland & Akçakaya, 2001). Moreover, such information is also essential for the development of complex demographic models, such as how reproductive rates change in response to the removal of individuals by hunting. These models are rarely applied due to the lack of data; thus, researchers employ simpler ones that perform with much higher uncertainty levels (Weinbaum et al., 2013).
A cost-effective option for gathering large amounts of biological and ecological evidence in the field is via citizen scientists (Dickinson, Zuckerberg, & Bonter, 2010). By involving non-professionals it is possible to obtain vital information on a variety of subjects (Bonney et al., 2014;Steger, Butt, & Hooten, 2017). In the tropics, Indigenous and rural people have been involved in citizen-science projects, providing information on animal populations and trends, just as accurately as trained scientists (e.g. Danielsen et al., 2014). Since local communities in tropical forests have extensive knowledge of the environment and are the main direct users of natural resources, their participation in scientific monitoring is central (Pocock et al., 2015). Mayor, El Bizri, Bodmer, and Bowler (2017) have demonstrated the effectiveness of citizen science through a community-based collection of organs of Amazonian forest mammals to determine reproductive parameters. In this study, local hunters collected and voluntarily donated complete visceras of hunted specimens over an uninterrupted 15-year period. Using this material, Mayor et al. (2017) were able to estimate annual birth rates of female offspring. These estimates differed significantly from those obtained in sustainability assessments that often use data from captive populations.
As shown by Mayor et al. (2017) it is possible to collect accurate reproductive parameters of some hunted species over the long-term through the examination of biological materials brought back to researchers. This is possible for small-bodied animals but not for large species since their viscera are often not brought back from the forest due to their heavy weight (see Mayor et al., 2017). In the current system, local people are only responsible for sample collection, since the determination of the reproductive status of animals is done by technicians in laboratories, thus increasing survey costs, and does not involve locals in data processing and analysis, and arguably in using the collected data in decision-making. Conversely, if local people could accurately categorize the reproductive status of hunted specimens where they are hunted, data collection becomes cheaper, easier to implement and more frequent, therefore substantially increasing sample sizes. Beyond providing more precise estimates of reproductive rates, larger sample sizes would also permit the better understanding of hunting impacts, for example by determining how variation in reproductive rates over time relates to densitydependent responses of populations to hunting.
In this study, we used the lowland paca Cuniculus paca as our model species, because it is among the top three most hunted species in the Neotropics (e.g. El Bizri, Morcatty, Lima, & Valsecchi, 2015;El Bizri et al., 2020;Peres, 2000), allowing us to acquire a large number of samples. Using photographs of genitalia of paca females donated by hunter families, we tested whether an effective method could be implemented for citizen scientists to accurately diagnose the reproductive status of hunted specimens.
We first established the local peoples' capacity to determine the reproductive status of female pacas in three Amazonian sites, correlated with the percentage of immatures in the hunted population, in accordance with an expected density-dependent response to variations in hunting levels.

Synthesis and applications. We show that the voluntary diagnosis of game species'
reproductive status by local people is a feasible method to obtain accurate lifehistory parameters for hunted tropical species, and to assess hunting effects on game populations. Given that almost half of the protected areas in the world are co-managed by local people, our results confirm the potential of integrating local communities in citizen-science initiatives to ensure faster, low-cost and more accurate data collection for wildlife management.

K E Y W O R D S
Amazon, citizen science, hunting, protected areas, reproduction, sustainability, tropical forests, wild meat before and after a period of training. We then used these results to correct the pregnancy rates estimated for the species from data collected by local people on hunted specimens' reproductive status over a 17-year period (2002-2018) in one study site. For this study site, we compared the corrected pregnancy rates with those determined from genitalia samples by trained researchers, and with hunting indices to describe the effects of hunting on the studied game populations.

| Study sites
We conducted our study in three Amazonian sites (Figure 1). The

| Local people interviews
Photographs were taken of all paca female genitalia collected (n = 300) from animals donated for research by local people in the YMR and ASDR (see El Bizri et al., 2018;Mayor et al., 2017;Mayor, Guimarães, & López, 2013; see Figure S1). For pregnant genitalia, we removed the conceptuses and measured their crown-rump length using a metal vernier calliper (maximum 300 mm).
From the pool of photographs, we selected a total of 42 showing complete genitalia, being seven (16.7%) photographs of non-pregnant females and 35 (83.3%) from pregnant females with conceptuses ranging from 0.5 to 26 cm in length (M = 7.51 cm; SD = 7.28 cm), depicting the increasing range in conceptus size along gestation. Using these photographs, we asked interviewees to answer whether they thought the specimen was pregnant or not. If they considered the specimen to be pregnant, the interviewee was requested to point out where he/she believed the conceptus was implanted within the uterus.
Considering the time available for the study, we used simple random sampling (Albuquerque, Cunha, De Lucena, & Alves, 2014) to select the maximum number of young and adults to interview within each sampled community. We interviewed a total of 104 people, 81 men (77.9%) and 23 (22.1%) women. The average age of the men was 38.7 ± 13.9 SD (17-72 years old), and of women 41.3 ± 14.1 SD We held two rounds of interviews for each person during a single interview day. In the first round, we showed the pictures randomly and asked people to determine the reproductive status of each specimen without giving them any clues or guidance on the reproductive biology of the species. In the second round, we used three of the pictures to train interviewees on how pregnancy occurs, explaining where the conceptuses would normally be found in the uterus, and how to determine their presence. We then shuffled the photographs and randomly presented these again for their second diagnosis.

| Hunting registers
In the ASDR, a citizen-science hunting monitoring system has been active since 2002 and is ongoing. In this site, we trained five local hunters within five separate communities to assemble a wide range of data on the daily hunting activity of community

| Interview responses
For each interview, responses were scored as 1 if correct and 0 if incorrect. We considered an incorrect answer when the interviewee identified a non-pregnant paca as pregnant (false positive) and a pregnant paca as non-pregnant (false negative), also when the interviewee incorrectly indicated the location where the conceptus was implanted in the uterus.
We analysed all responses using GLMM with a binomial distribution to obtain logistic regressions of the probability of interviewees giving correct answers according to a set of predictor variables. We generated separate models for non-pregnant and pregnant females.
For non-pregnant specimens, we considered the number of correct and incorrect responses (scores: 0-incorrect and 1-correct) as the dependent variables, and the interviewees' sex (man/woman) and age (continuous) as predictor variables. For pregnant specimens, we considered the number of correct and incorrect responses (scores: 0-incorrect and 1-correct) as dependent variables, and the interviewees' sex and age and the length of the conceptuses (continuous) as predictor variables. We built a null model (no effect of predictor variables) and models with different combinations of predictor variables, from simple ones (only one predictor variable) to a more complex one (all variables in the model). We independently tested each interview round. Interviews from the same community may not be independent since there were a different number of interviewees for each community, and because we consider that the learning process and sharing of information on the biological traits of game species among local people may be nested at the community level. As a result of these two issues, the communities sampled were included as a random categorical factor.
Final models from GLMMs were compared based on Akaike information criterion (AIC) values, and all models with the ΔAIC < 2 in relation to the model with lowest AIC were considered as with strong support (see Burnham & Anderson, 2004). To avoid selecting a model that was overfitting due to a large sample size, we also used a likelihood ratio test to compare the significance of these models; if two models were similar (p > 0.05), we considered the best-fitted model the one with lowest number of parameters, that is fewest degrees of freedom. However, we present all models with ΔAIC < 2 in Table S1. GLMMs were conducted using the r package lme4.
Finally, we assessed the effect of training on the probability of giving more correct answers in the second interview. For both non-pregnant and pregnant specimens, we calculated the difference between the percentages of correct and incorrect responses in the two interview rounds using a chi-squared test. In addition, for pregnant specimens, we calculated the difference between the average logistic regression formulas produced in the GLMMs for each round of interview.

| Corrections of reproductive status and density dependence effects
Interview results were then used to correct the data on the reproductive status of hunted females collected over the 17-year (2002-2018) monitoring period in the ASDR. Since no systematic training on pregnancy diagnosis was conducted with local people during the monitoring period, we considered they could be categorizing pregnant females as non-pregnant at a similar rate as obtained in the first interview round. On this basis, we therefore applied a correction factor based on two steps. First, we used the formula of the relationship between conceptus length and gestational age in El Bizri  Table S2). The gestation length of 149 days for the paca was based on the study by Guimarães et al. (2008).
We used these corrected reproductive data to validate the effectiveness of a long-term citizen-science collection of reproductive performance data to: (1) provide reliable information on the reproductive rates for the species; and (2) understand hunting effects. For the first aim (1), we selected females with a body mass ≥5.5 kg, which we considered as mature, following El Bizri, Fa, Valsecchi, Bodmer, and Mayor (2019). We then calculated the monthly percentage of pregnant specimens among the mature females sampled each month (hereafter monthly pregnancy rate), independently for each year and assessed the relationship between monthly pregnancy rate and the expected monthly pregnancy rate of pacas. For this purpose, we considered the published metrics of pregnancy rates in El Bizri et al. (2018), calculated from pregnancy diagnoses from genitalia collected in the ASDR by researchers and examined in the laboratory, as the expected monthly pregnancy rates (see Table S2). In this analysis, the month was treated as a fixed factor, while the year of collection was treated as a random factor.
For our second aim (2), we hypothesized that any hunting impact on any animal population would lead to an increase in pregnancy rates over time as a density-dependent response to lower population abundance (e.g. Lima & Jaksic, 1998). This increase in pregnancy rate would then be related to an increase in immature individuals in the population due to higher birth rates. We tested for all these relationships by calculating the yearly pregnancy rates of hunted specimens, correlating these with the mean annual catch-per-uniteffort (CPUE, in ind hunter −1 hr −1 ) of hunting events as a proxy of paca abundance (see Marrocoli et al., 2019;Rist, Milner-Gulland, Cowlishaw, & Rowcliffe, 2010;Valsecchi, Bizri, & Figueira, 2014).
Then, we correlated the yearly pregnancy rates with the annual percentage of immature individuals (number of individuals < 5.5 kg/total number of individuals hunted; see El Bizri, Fa, Valsecchi, et al., 2019;Mendes-Oliveira et al., 2012). Although 848 paca hunting events were recorded, we calculated CPUE for only nocturnal hunts using the 'spotlighting' method (n = 519), since paca is strictly nocturnal and spotlighting is a technique that is specifically used to hunt this and no other species, therefore guaranteeing that this index indeed reflected the species' abundance in the study area (see Valsecchi et al., 2014).
To assess the relationship between the monthly pregnancy rate and the expected monthly pregnancy rate of pacas, as well as densitydependent effects, we used a generalized additive model (GAM), which consists of univariate regression analysis that allows each regression variable to have a linear or nonlinear relationship with the dependent variable. The type of nonlinearity is tested through the use of several smoothers/additive terms in the modelling process. These terms normally adds penalizations in the model to prevent overfitting.
For the conduction of GAMs, we first tested the best family of distribution fitted to the response variable, comparing these based on QQ-plots of the residuals, and on the difference between AIC values among them. After selecting the best-fitted distribution, we built a null model (no relationship with the predictor variable), and models with the predictor variables in linear and nonlinear forms, testing penalized splines and cubic splines as additive terms. The final model, both in terms of retention of predictor variables and the type of relationship (whether linear or nonlinear), was selected based on AIC values (Burnham & Anderson, 2004), in which the model with the lowest AIC was selected. For these analyses, the r package gamlss was used (Stasinopoulos & Rigby, 2007). An advantage of using the gamlss package is its flexibility in terms of families of distribution (over 100 continuous, discrete and mixed distributions for modelling the response variable are available), and the various additive terms available compared to other packages. We used r 3.3.3 software for all statistical procedures, and a p-value < 0.05 was considered significant.

| The diagnosis of the reproductive status by local people
For non-pregnant females, there was a proportion of 92.6% and 90.4% of correct responses in the first and second round of interviews, respectively, with no significant difference between rounds (χ 2 = 0.17; p = 0.68); no significant effect was observed for any of the predictor variables on the probability of giving correct responses (Table 1).
For pregnant females, the proportion of correct responses was 72.5% for the first interview round. The probability of correctly diagnosing pregnancy varied according to the conceptus size, with lower probabilities for smaller conceptuses (Table 1; Figure 2a). This probability also varied according to the age of the interviewee, with youngers answering correctly more often than older individuals; this difference being greater for small-sized conceptuses TA B L E 1 Details of the best-fit GLMMs for the diagnoses of mature paca Cuniculus paca females by local people through genitalia pictures, and generalized additive models (GAMs) for the relationship between the corrected monthly pregnancy rate from hunting registers with the expected pregnancy rate from genitalia, the trend in annual pregnancy rates along the hunting monitoring period (2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018), and between annual pregnancy rates with the catch-per-unit-effort (CPUE, ind hunter −1 hr −1 ) and proportion of immature individuals in the Amanã Sustainable Development Reserve (ASDR). The interviewees' communities were included as a random categorical factor in the GLMMs Best-fit model p < 0.0001). This improvement occurred for all conceptus sizes, but was higher for pregnant females with smaller conceptuses, between 0.5 and 10 cm (32-85 gestation days, 21.5%-57.0% of total gestation length respectively), with peak at around 5 cm (57 gestation days, 38.3% of total gestation length), for which diagnoses improved by around 25% (Figure 3).

| Estimates of reproductive rates of game species and hunting effects
Out of the total number of hunted individuals recorded in our monitoring program (n = 1,236 pacas), 634 were females out of which 554 were mature (87.4%, ≥5.5 kg). Among these, local dwellers classified 445 as non-pregnant and 109 as pregnant. Without any correction, the average monthly pregnancy rate from hunting registers was 22.7% (Table 2). Using the formula obtained from the  females that were in fact pregnant but incorrectly diagnosed as non-pregnant (Table 2).
Corrected monthly pregnancy rates were positively and significantly correlated with the expected monthly pregnancy rates calculated from genitalia collections (El Bizri et al., 2018; Table 1; Figure 4). The annual pregnancy rates calculated from hunting registers were on average 39.1%, presenting an inverted U-shaped pattern of changes across years, with a peak in 2012 (Table 3; Figure 5).
Pregnancy rates were negatively correlated with the annual CPUE over time (Table 1; Figure 6), and years with higher pregnancy rates had a higher proportion of immature individuals (Table 1; Figure 6).

| D ISCUSS I ON
We showed that local people can participate in the voluntary diag-  F I G U R E 3 Improvement rate, in percentage, of the probability of local people in Amazonia at correctly diagnosing pregnancy in pictures of genitalia of lowland pacas Cuniculus paca after a training session in the butchering and cooking of animals (Bonwitt et al., 2017;MacDonald, 2007). This TEK on the reproductive cycles of species play a particularly important role in influencing the younger generation's capacity to detect and hunt animals, and can be used in local management strategies to avoid overhunting (Berkes, Colding, & Folke, 2000). Vieira, von Muhlen, and Shepard (2015), in Canada corresponded to the reproductive status of the animals after dissection (see Lyver & Gunn, 2004). Thus, if pregnant animals can be identified during hunting events, it may be possible to develop more effective management strategies that take into account the reproductive biology of the game species.
Several interviewees were unable to identify small conceptuses in pictures of pregnant genitalia. Though we acknowledge that a number of interviewees may have guessed their answers, we argue that they may have classified them as non-pregnant actually believing they were correct, given the difficulty in detecting early pregnancies. This is supported by the rate of correct responses for non-pregnant specimens, which was much higher than expected in a random guessing situation (90% vs. 50% respectively).
In our study, before the training, younger people were better at diagnosing pregnancy than older people. This could be explained by two factors: hunting frequency decreases with age, and therefore younger people may be more frequently in contact with hunted animals; and older people claimed that their weaker vision impaired them for properly seeing small conceptuses in the pictures, so younger people may have benefited from the method used in the study. Conversely, as exemplified in other citizen scientist studies (e.g. Ratnieks et al., 2016), we showed that after a short training session the difference between these age groups was overcome, and we significantly improved our interviewees' ability at diagnosing pregnancy, especially for early pregnancy stages.
Thus, we claim that local peoples' ability at diagnosing pregnancy after training is similar overall, and will be even higher when they are handling the specimens themselves, since they would be able to palpate the uterus in search for any sign of pregnancy, as they do while processing game. A previous training also allows for a direct calculation of pregnancy rates without the need for extensively correcting the data as we did here, considering that only a small proportion of pregnant individuals (around 10% from very early gestations) would Year Annual pregnancy rate (%) be incorrectly diagnosed, which is similar to the proportion of underdetected pregnancies by transabdominal ultrasonography (Mayor, López-Gatius, & López-Béjar, 2005). Our claim is supported by the strong match between local and laboratory diagnoses of pregnancy, which shows that citizen scientists in rural communities are able to collect accurate information of the natural cycle of the species in the wild.
Our citizen-science monitoring of annual reproductive rates has also proved to be reliable for assessing density-dependent responses of populations to hunting, which is difficult to obtain in the field. To our knowledge, this technique has not been applied for tropical game species before our study. In addition, the highest pregnancy rate value obtained in our study (64.5%) is similar to that obtained by Mayor et al. (2017)  per biological sample obtained through community-based collections. Therefore, in the present study, the citizen-science diagnosis of pregnancy generated savings of US $1,743.50 for the paca alone.
Accordingly, we advocate that the method presented here can be useful for faster, easier, low-cost and accurate assessments of sustainability. Using this method to properly assess hunting impacts would aid more effective strategies to protect wildlife from overexploitation, and decision-makers and local leaders can be provided with accurate tools to implement tangible policies aimed at minimizing food insecurity.
Globally, around 45% of the protected areas are fully or comanaged by local people (Garnett et al., 2018;UNEP-WCMC & IUCN, 2016), and thus, subjected to direct human use of fauna. We believe that the present method shows great potential to be applied in several contexts around the world, and even improved when integrated into new technologies currently used in participatory monitoring systems, such as smartphones and specific monitoring software (see van Vliet et al., 2017).
Very few studies on life-history parameters of tropical species have relied on data obtained in the wild, and even fewer have obtained such data with the collaboration of local people, even though these are the main actors responsible to manage game populations. Here, we provide a tool to overcome the challenge of obtaining reproductive data on game species in the wild, offering a practicable method that improves sample size for wildlife research. We confirmed that the diagnosis of pregnancy in game specimens undertaken by local people integrated in a citizen-science program is useful to assess hunting effects and provides accurate data for evaluating hunting sustainability for tropical mammals. Finally, this study offers the opportunity for locally produced data to be integrated into educational programs and policies, providing significant information that can be used against food insecurity and wildlife overexploitation world-wide.

ACK N OWLED G EM ENTS
We sincerely thank the people from the Yavarí-Mirín River, Amanã Sustainable Development Reserve and Juruá River, without whom this work would not be possible. We also thank T.Q. Morcatty, L. Luiselli and an anonymous reviewer for useful comments on the manuscript, and C.L.B. Franco for building the map. This work was supported by the Gordon and Betty Moore Foundation, the National Council for Scientific and Technological Development (CNPq), the IVITA,