Webpage Views as a Proxy for Angler Pressure and Effort: Insights from Bayesian Networks

Webpage Views as a Proxy for Angler Pressure and Effort: Insights from Bayesian Networks Azar Taheri Tayebi∗1, Julia S. Schmid2, Sean Simmons3, Mark S. Poesch2, Mark A. Lewis2,4, and Pouria Ramazi1 1Department of Mathematics and Statistics, Brock University, St. Catharines, Ontario, Canada 2Department of Mathematical and Statistical Sciences, University of Alberta, Edmonton, Alberta, Canada 3Angler’s Atlas, Goldstream Publishing, Prince George, British Columbia, Canada 4Department of Biological Sciences, University of Alberta, Edmonton, Alberta, Canada 1 Abstract Reliable angler activity data inform fisheries management. Traditionally, such data are gathered through 1 surveys, but an innovative cost-effective approach involves utilizing online platforms and smartphone 2 applications. These citizen-sourced data were reported to correlate with conventional survey information. 3 However, the nature of this correlation–whether direct or mediated by intermediate variables–remains 4 unclear. We applied Bayesian networks to data from conventional surveys, the Angler’s Atlas website, 5 the MyCatch smartphone application, and environmental data across Alberta and Ontario, Canada, to 6 detect probabilistic dependencies. Using Bayesian model averaging, we quantified the strength of connec7 tions between variables. Waterbody webpage views were directly related to daily and weekly-aggregated 8 boat counts in Ontario (51% and 100% probability) and to weekly-aggregated creel survey-reported 9 fishing duration in Alberta (100%). This highlights the value of citizen-sourced data in providing unique 10 insights beyond meteorological factors, with online interest serving as a potentially reliable proxy for 11 angler pressure and effort.12 13 Keywords: Angler activity, Bayesian network, citizen-sourced data, citizen-reported data, conventional 14 surveys, machine learning15 2 Introduction16 Recreational fishing is a popular activity across the globe, with more than 10% of people participating 17 and catching billions of fish every year, contributing economic and social benefits (Arlinghaus and 18 Cooke, 2009; Kelleher et al., 2012; Hyder et al., 2018). Many recreational fisheries are facing overfishing, 19 mortality, and fish diseases (Post et al., 2002; Cooke and Cowx, 2004; Lewin et al., 2006). Comprehensive 20 assessment data on resources (e.g., fish population and species) and on fishery (e.g., effort and selectivity) 21 may enable recreational fisheries management to prevent crises through measures like size and bag 22 regulations (Post et al., 2008; Venturelli et al., 2017).23 Data can be collected through conventional surveys such as creel surveys and aerial surveys (Eero 24 et al., 2015). Creel surveys are conducted by direct interactions with anglers, either on-site while they 25 are fishing or at the conclusion of their fishing trips (Murphy et al., 1996). The collected data is limited 26 in time and space as the collection can be expensive and time-consuming (van Poorten et al., 2015; 27 Vølstad et al., 2006). Nevertheless, these conventional surveys are often considered the most reliable 28 sources of data on angler activity, providing information that is generally assumed to represent reality. 29 ∗Corresponding author (email: [email protected]) 1 Consequently, cost-effective citizen-sourced data became popular (Fischer et al., 2023; Gutowsky 30 et al., 2013). Mobile applications (apps) are modern sources of citizen-reported data. A survey on 31 experts in the recreational fishery field showed that in most countries, angler app data usage is likely to 32 increase significantly over the next 5–10 years (Skov et al., 2021). Websites also serve as an additional 33 tool for gathering data on angler activity (Martin et al., 2012). Anglers utilize these platforms to seek 34 real-time updates on fishing conditions and to share information about their fishing expeditions (Sims 35 and Danylchuk, 2017).36 Nevertheless, citizen-sourced data do not readily represent that of conventional surveys. Angler apps 37 used for collecting fishing data may be subject to biases due to non-random participation. Who uses 38 these apps can be influenced by factors such as smartphone ownership and demographics (Venturelli 39 et al., 2017). This could result in data skewed towards certain groups (Jiorle et al., 2017). There may 40 even be intentional data manipulation, because some anglers wish to influence regulations (Sullivan, 41 2003). The design of the app, location of users, and internet connectivity can also bias the data 42 (Papenfuss et al., 2015; Jiorle et al., 2017).43 Despite these limitations, several studies demonstrated the potential of citizen-sourced data to serve 44 as a supplementary or alternative source for conventional survey data, particularly with respect to catch 45 rates and angler pressure (i.e., the intensity of fishing activity, often measured by total fishing effort, 46 number of anglers, or hours fished; Table 1) (Jivthesh et al., 2022; Johnston et al., 2022; Papenfuss et al., 47 2015). For example, the iNaturalist website and app provided comprehensive data with a high level of 48 completeness compared to other reporting methods such as telephone, email, and mixed-mode approaches 49 (Taylor et al., 2022). Data from iAngler app and creel surveys conducted by the Marine Recreational 50 Information Program (MRIP) showed minimal differences, with discrepancies exhibiting an almost zero 51 central tendency (Jiorle et al., 2016). Likewise, mean catch rates reported via Fangstjournalen app 52 platform were similar to those obtained through conventional surveys on a Danish island (Gundelund 53 et al., 2021).54 However, the nature of the dependence between citizen-sourced and conventional surveys remains 55 unclear. Specifically, if the relationship is indirect and mediated by some intermediate variables, then 56 the use of citizen-sourced data to predict conventional survey outcomes would be limited, if not zero. 57 For example, if citizen-reported catch rates are indirectly related to creel-based catch rates through the 58 intermediate variable temperature (e.g., cold weather discourages angling and reporting), knowing the 59 temperature alone would suffice to predict creel-based catch rates–citizen-reported catch rates provide 60 no further information.61 This defined our research goal: Is there a direct or indirect relationship between citizen-sourced 62 and conventional survey-based angler activity measured by catch rates, fishing duration, number of 63 trips, and number of boats? To answer this question, we employed Bayesian networks (BNs) (Koller 64 and Friedman, 2009), a type of probabilistic graphical model that allows the detection of dependencies 65 between variables. Our case study was the Lower Bow River and Upper Oldman River within three 66 fisheries management zones in Alberta, Canada, in 2018, as well as 98 randomly selected lakes across 67 15 management zones in Ontario, Canada, in 2018 and 2019. The target species included brook trout, 68 walleye, and lake trout. Webpage views from the Angler’s Atlas website and fishing trips and catches 69 reported by anglers through the Angler’s Atlas website and MyCatch app in Alberta and Ontario were 70 used as citizen-sourced data. Conventional survey-reported datasets included creel surveys in Alberta 71 and aerial surveys in Ontario. The environmental variables for both provinces included air temperature, 72 total precipitation, wind speed, relative humidity, solar radiation, and degree days for the waterbodies. 73 3 Materials and Methods74 First, we discretized the data collected from the mentioned variables in Alberta and Ontario. Second, 75 we learned a BN for each dataset using the Bayesian Information Criterion (BIC) score (Figure 1) 76 and analyzed the direct and indirect relationships between the variables within the BNs. Finally, we 77 employed Bayesian model averaging (BMA) to quantify the uncertainty associated with the estimated 78 2 Table 1: Examples of previous research on comparing citizen-sourced data to conventional surveys on angler activity. App or Online Platform Input Conventional Survey Studied Variables Time Span Methods Location Conclusion Fangstjournalen (Gundelund et al., 2021) date, location, effort, target species, number caught and retained or released roving creel survey, aerial survey, recall survey mean catch rate, number of anglers 2019 bootstrapping, comparative analysis, regression Danish island of Funen Citizen science data had the potential to supplement traditional surveys or act as an alternative source of catch and effort data. iNaturalist (Taylor et al., 2022) tag number, taxonomic ID, disposition of fish telephone, email, mixed survey report completeness, capture location precision 2019 Kruskal-Wallis test Oklahoma There was a significant difference in three reporting modes. MyCatch (Johnston et al., 2022) catch rate, fishing effort, and angler registration information mail, creel, gillnet survey spatial distribution, fishing landscape, catch estimates 20182019 zero-inflated negative binomial regression, linear regression Alberta General trends between the app data and conventional data were similar. iAngler (Jiorle et al., 2016) angler registration information, fishery-specific data, waterbody information NOAA’s MRIP survey mean catch rate 20122013 negative binomial distribution Florida There was a high degree of similarity between the angler app and creel survey datasets’ catch/trip estimates. iAngler (Papenfuss et al., 2015) angler registration information, fishery-specific data, waterbody information creel survey lake popularity 20102014 linear relationship Alberta There was a linear relationship between the frequency of visits reported by the iAngler app and the number of angler visits estimated by creel surveys. 3 Table 2: Summary of raw (original), daily-aggregated, and weekly-aggregated survey-based/citizensourced/combined datasets, highlighting the mean of non-zero values and the count of zeros for conventional survey-based and citizen-sourced variables. Data Number of samples Variable Mean of non-zero values Number of zeros Alberta Creel survey-based (raw data) 3714 Creel fishing duration (hour) 4.06 171 Creel number of trips (trips per day) 40.04 0 Creel number of fish caught 2.33 1787 Alberta Citizen-sourced (raw data) 1986 App fishing duration (hour) 0.54 1753 App number of trips (trips per day) 1.15 1753 App number of fish caught 0.42 60 Webpage views 2.47 0 Alberta (daily-aggregated and combined–final) 185 Creel fishing duration (hour) 3.25 4 Creel catch rate (catch per hour) 0.77 10 App fishing duration 5.95 110 App catch rate 0.63 132 App number of trips 1.47 110 Webpage views 16.80 26 Alberta (weekly-aggregated and combined–final) 66 Creel fishing duration (hour) 64.41 1 Creel catch rate (catch per hour) 0.79 3 App fishing duration 52.79 28 App catch rate 0.98 29 App number of trips 14.26 28 Webpage views 58.77 23 Ontario Aerial survey-based (raw data) 3118 Number of boats 4 1532 Ontario Citizen-sourced (raw data) 8621 App fishing duration (hour) 5.58 0 App number of trips (trips per day) 1.11 0 App number of fish caught 7.33 2532 Webpage views 1.81 0 Ontario (daily-aggregated and combined–final) 1128 Number of boats 7.99 654 App fishing duration (hour) 6.78 1110 App catch rate 1.74 1105 App number of trips 1.06 1110 Webpage views 1.68 850 Ontario (weekly-aggregated and combined–final) 320 Number of boats 55 38 App fishing duration (hour) 18 0 App catch rate 1.64 58 App number of trips 2.94 0 Webpage views 5.28 59 4 Meteorological data Conventional surveys Citizen-reported data Data preprocessing Constructing Bayesian network Figure 1: A conceptual illustration of the methodology used in this paper, that is the process of collecting data through conventional surveys, smartphone applications, and online platforms, followed by the construction of a Bayesian network to aid fishery management in decision-making. An angler utilizes online platforms to find information about waterbodies and report fishing experiences and uses a smartphone app to record fishing trip data. Creel surveys or aerial surveys gather information about the angler trips. The data is then used to construct the Bayesian network structure to reveal the conditional dependencies between the conventional survey and citizen-sourced variables in the presence of meteorological variables at as potential confounders. Fishery managers may leverage the insights derived from this network structure to implement relevant actions, such as stocking and regulations. The management decision-making did not take place in this project and was not part of the methodology. 5 connections between the variables.79 3.1 Data80 The Alberta case study comprised the lower section of the Bow River and the upper section of the 81 Oldman River and its major tributaries in southwestern Alberta, from June to October 2018. The 82 Ontario case study included 98 lakes in Ontario. The data sources included the Angler’s Atlas website 83 and correspondent mobile app MyCatch, two creel surveys, an aerial survey, and a weather simulation 84 model, described below.85 86 Citizen-sourced data87 Citizen-sourced data were collected by the Angler’s Atlas website and MyCatch mobile app. The 88 Angler’s Atlas platform, established in 1999 by Goldstream Publishing Incorporated based in Prince 89 George, British Columbia, Canada, provides detailed information on 330,000 waterbodies across Canada 90 through its website ( www.anglersatlas.com ). Each waterbody has its own webpage featuring details 91 such as location, popular species, and markers for hotspots and boat launches. The webpages can be 92 viewed by anyone and do not require a membership on the platform. Anglers can use these webpages to 93 gather trip-planning information or to report details about their fishing experiences. Variable “webpage 94 views” was defined as the number of unique visits on a day. Several webpage views by the same individual 95 were counted as one view. “Webpage views” can be considered as an indicator of angler pressure, because 96 visiting a specific waterbody’s webpage reflects people’s interest in obtaining information about the 97 waterbody and planning trips (Xiang et al., 2015). Webpage visits in the last few days (seven in our 98 case) were considered since trips are typically planned a few days in advance (Nyblom, 2014).99 On April 1, 2018, the MyCatch mobile app was launched by the Angler’s Atlas team and app data 100 collection began on May 11, 2018, through email campaigns targeting the subscribers of Angler’s Atlas 101 in Alberta (Johnston et al., 2022). Through both the Angler’s Atlas online platform and the MyCatch 102 app, anglers could report details about their fishing trips, such as date, location, duration, and number 103 of fish caught by species. Anglers could also report trips with no catches. The primary caught species104 reported in Alberta and Ontario were brook trout, walleye, and lake trout. The data were anonymized 105 by daily aggregation and included only completed trips, i.e., those with fully reported information on the 106 trip. When a trip was reported for a specific waterbody, it indicated that the trip took place anywhere 107 on that waterbody. We refer to the collected data as “citizen-sourced data,” with no distinction between 108 the app and online platform.109 We used citizen-sourced data from the entire Bow River and Oldman River systems, which span all 110 three fisheries management zones in Alberta. There were 233 reports of fishing trips from these two 111 river systems between May 11 and October 31, 2018. For Ontario, a total of 8,621 fishing trips were 112 reported over 1,670 unique waterbodies and 310 unique days between May 19, 2018 and August 29, 113 2019 (Figure 2).114 For each waterbody and day i, the data included the total number of trips, denoted Ni , and for each 115 trip j the number of fish caught Cij and the time spent fishing Dij . We computed the total fishing 116 duration Dtotal i [hour], total number of fish caught Ctotal i [number of fish], and mean catch rate CP UEi 117 (“catch per unit effort”) [fish/hour], for each day i, as follows:118 (1) Dtotal i= Ni X j=1 Dij, 119 (2) Ctotal i= Ni X j=1 Cij, 120 (3) CP UEi=Ctotal i Dtotal i . 6 For simplicity, we refer to app-reported Dtotal i as “app fishing duration” and CP UEi as “app catch 121 rate.” These formed the two app variables used in this study. The other citizen-sourced variable was 122 “webpage views” presenting the number of times a waterbody’s webpages was visited in the last seven days. 123 124 Creel surveys125 Data from two creel surveys in Alberta in 2018 were used (Johnston et al., 2022). The first survey 126 was conducted by Alberta Environment and Parks (AEP) along the approximately 100 kilometers of 127 Lower Bow River, from June through the end of November 2018 (Christensen et al., 2020). The second 128 creel survey was carried out by the Alberta Conservation Association (ACA) along a stretch of 199 129 kilometers encompassing the Upper Oldman River and its principal tributaries, the Livingstone River,130 Dutch Creek, and Racehorse Creek, from June to October 2018 (Hurkett and Fitzsimmons, 2019). We 131 used creel data on both rivers which were collected from angler interviews during roving surveys at the 132 public access points (Hurkett and Fitzsimmons, 2019; Ripley and Council, 2006; Malvestuto, 1996). 133 The surveys were conducted across spatial and temporal strata. Spatially, this contained four reaches 134 on the Bow River and three on the Oldman River system. Temporally, this happened during shifts in 135 the morning (8:00–15:00) or evening (15:00–22:00)–although the exact hour and date of each interview 136 was recorded. During interviews, 2751 anglers from the Bow River (across 122 days) and 963 anglers 137 from the Oldman River (across 88 days) were requested to provide information about their fishing trips, 138 including the date, time duration of the trip, and quantity of fish caught. Angler interviews included 139 both complete and incomplete trip data. We excluded two samples due to daily fishing durations 140 exceeding 24 hours.141 We obtained the following creel-survey variables in the same way as with the corresponding app 142 variables, using Equations 1 to 3: “total fishing duration of anglers sampled on day i ” Dtotal, sampled i , 143 “number of fish caught by anglers sampled on day i ” Ctotal, sampled i , and “catch per unit effort of anglers 144 sampled on day i ” CP UEi . As creel surveys were not designed to measure total fishing duration, we 145 instead used the mean value of the variable, denoted Dmean i[hour/trip] on day ias follows:146 (4) Dmean i=Dtotal, sampled i Nsampled i . where Nsampled i is the total number of trips reported by anglers sampled on day i . For simplicity, we 147 refer to the creel survey-reported Dmean i as “creel fishing duration” and CP UEi as “creel catch rate”, 148 which formed the two survey variables in the Alberta dataset.149 150 Aerial surveys151 We utilized data from Ontario’s inland lake ecosystem collected through the Broad-scale Monitoring 152 (BsM) program. The dataset included 98 randomly selected lakes over 50 ha, stratified by spatial (15 153 fisheries management zones), temporal (weekend/holiday and weekday), and lake-size categories. Data 154 were gathered during 101 aerial surveys conducted between 9:00 and 17:00 over 78 randomly selected 155 days from June to August in 2018 and 2019 (BsM cycle 3). The variable “number of boats,” was defined 156 as the number of actively fishing boats observed during the surveys, serving as an indicator of angler 157 pressure and the only survey variable in the Ontario dataset.158 159 Meteorological data160 Daily averages of meteorological data, considered as potential intermediate variables, were obtained 161 using the BioSIM tool (Régnière et al., 1996). BioSIM is a software developed by Natural Resources 162 Canada that runs weather-driven simulation models using geographic weather data. Based on the four 163 nearest weather stations within a radius of 300 km from the centroid of the waterbody, the software 164 adjusted the data for elevation, latitude, and longitude differences, and restored variability to long-term 165 averages. Historical daily weather observations were utilized, and a bi-linear interpolation method was 166 applied (Régnière et al., 2017). We used the software to obtain data for the waterbodies, focusing on 167 variables known to be related to or influencing angler’s behavior, such as the daily air temperature 168 7 (Gundelund et al., 2022; Kendall et al., 2021), precipitation (Shaw et al., 2021), wind speed (Gundelund 169 et al., 2022; Agmour et al., 2020), relative humidity (Shaw et al., 2021), solar radiation (Shaw et al., 170 2021; Cooke et al., 2017; Speers and Gillis, 2012), and degree days (Speers and Gillis, 2012).171 A temporal binary day-type variable “is weekend” was used to indicate whether the fishing trip 172 occurred on a weekday or a weekend (Kendall et al., 2021).173 174 The Alberta and Ontario datasets175 We constructed two final (daily-aggregated) datasets: the Alberta dataset and the Ontario dataset 176 (Table 2). The Alberta dataset was created by combining the creel survey data and app data for the two 177 river systems on days when creel data were available. Each sample of the dataset included six angler 178 activity variables: “app catch rate,” “app fishing duration,” “app number of trips,” “webpage views,” 179 “creel catch rate,” and “creel fishing duration,” along with the seven auxiliary variables described earlier. 180 The dataset had a total of 185 daily samples: 101 from Bow River and 84 from Oldman River, and an 181 overall of 75 reported fishing trips from citizen-sourced data. There were initially 233 reported trips in 182 Alberta. However, during the merging of the app data with the creel data, zero values were assigned to 183 app variables that did not have a value at the specified date and waterbody in the creel dataset. These 184 zero assignments to citizen-sourced variables were not data imputation. The reason why there was no 185 data for a certain waterbody or day was that no one reported on that day and for that waterbody. Thus, 186 by definition, the corresponding app variables would be zero. In the same way, the Ontario dataset was 187 formed by combining aerial survey data with app data for the waterbodies and dates with available 188 aerial data, resulting in 1,128 daily samples over 98 waterbodies and 18 citizen-sourced reported trips. 189 Each sample included five angler activity variables: “app catch rate,” “app fishing duration,” “app 190 number of trips,” “webpage views,” and “number of boats,” and the same seven auxiliary variables used 191 in the Alberta data (Table 3; Supplementary Material Figures S1, S2, and S3).192 193 Discretization194 Continuous variables were discretized because the applied package for the BN learning algorithm 195 required discrete data. Non-binary variables were discretized into three bins using the equal quantile 196 method, ensuring that each bin contained approximately an equal number of samples (Sun and Shenoy, 197 2007) (Supplementary Material, Figures S4, S5, and S6). Consequently, each variable had the statuses 198 “low,” “medium,” and “high” (Figures S4, S5, and S6). Discretization into three bins was shown to be 199 optimal in previous applications of BNs in ecological systems (Milns et al., 2010; Yu et al., 2004). The 200 discretization was performed separately for the Alberta and Ontario datasets.201 If the majority of the values of a variable were zero, the first bin included only the zero values, and 202 the remaining values were discretized into two bins with an equal number of samples. In the Alberta 203 dataset, this was the case with the variables “app number of trips” (59%), “app fishing duration” (59% 204 zeros), and “app catch rate” (71% zeros). In the Ontario dataset, this applied to the variables “number 205 of boats” (58% zeros), “app catch rate” (99% zeros), “app number of trips” (98% zeros), “app fishing 206 duration” (98% zeros), “webpage views” (75% zeros), and “total precipitation” (57% zeros).207 Due to the high proportion of zero values in the datasets, we conducted an additional experiment 208 on two weekly-aggregated datasets for waterbodies in Alberta and Ontario. Each sample was assigned 209 the calendar week number based on its calendar date. Then citizen-sourced and survey variables were 210 aggregated on a weekly basis, e.g., “number of boats” would indicate the total number of boats surveyed 211 during the calendar week. Meteorological variables were averaged over the days of the week for which any 212 of the survey-reported variables were available, and “is weekend” was removed. In Ontario, although the 213 aggregation reduced the dataset to 320 samples, it decreased the percentage of zeros in the app-reported 214 and survey variables to a maximum of 18% and 12%, respectively. For data discretization, the quantile 215 method was applied to all variables except for “total precipitation,” which had a high proportion of 216 zeros (43%). For this variable, the first bin was designated to represent zero values, while the nonzero 217 values were divided into two additional bins (bins 2 and 3) using the quantile method. In Alberta, the 218 aggregated data included 66 weekly samples, with a maximum of 44% zeros in app variables and 35% in 219 8 Table 3: List of variables included the Alberta and Ontario datasets. Variable Description Alberta dataset Ontario dataset App catch rate (fish/hour) Mean citizen-reported number of fish caught per hour ✓ ✓ App number of trips Total citizen-reported number of trips per day ✓ ✓ App fishing duration (hour) Total citizen reported fishing hours per day ✓ ✓ Webpage views Number of views of the waterbody webpage over the last 7 days ✓ ✓ Creel catch rate (fish/hour) Mean number of fish caught per hour recorded in the creel data ✓× Creel fishing duration (hour/trip) Mean fishing hours per trip per day recorded in the creel data ✓× Number of boats Number of boats per day recorded in the aerial survey ×✓ Air temperature (◦C) Mean daily air temperature of the waterbody ✓ ✓ Total precipitation (mm) Total daily both liquid and solid precipitations ✓ ✓ Wind speed (km/hour) Mean daily wind speed ✓ ✓ Relative humidity (%) Mean daily relative humidity ✓ ✓ Solar radiation (Watt/m2) Mean daily amount of solar energy received per square meter ✓ ✓ Degree days (◦C) Summing days with ≥ 5 ◦ C from April 1st each year ✓ ✓ Is weekend Whether the date is a weekend ✓ ✓ “webpage views.” For data discretization, the quantile method was performed for all variables, except for 220 “app catch rate” (44% zeros values), “app fishing duration” (42% zeros), “app number of trips” (42% 221 zeros), and “webpage views” (35% zeros). Similarly, for these variables, the first bin was allocated to 222 represent zero values, and nonzero values were divided into two bins using the quantile method.223 3.2 Bayesian network structure and analysis224 While the concept of direct and indirect relatedness can be challenging to investigate, a directed acyclic 225 graph (DAG) can be used to make definitions precise and undertake analysis (Ramazi et al., 2022). In a 226 causal model that is represented by a DAG, the parents of each node are considered as the direct causes 227 of that variable and the ancestors as the indirect causes. More generally and out of the framework 228 of causality, one can define direct and indirect relationships without DAGs by using the notion of 229 probabilistic conditional independence. Given a specified set of random variables, two variables X and 230 Y are said to be directly related if they neither are marginally independent nor become independent 231 conditioned on any subset of the other variables. In contrast, variables X and Y are indirectly related if 232 they are dependent but become independent once conditioned on a subset of the other variables, which 233 we refer to as intermediate variables. Variables X and Y are not (probabilistically) related if they are 234 marginally/statistically independent (Nadkarni and Shenoy, 2001; Eden et al., 1992).235 A BN is a tool to identify conditional independencies among specified variables. A BN consists of a 236 structure that is a DAG over the variables considered as nodes and parameters that are the conditional 237 probability distributions of each variable conditioned on its parents in the DAG (Koller and Friedman, 238 2009; Ramazi and Kalantari, 2023). In the structure, an edge between two variables implies a direct 239 relationship: a dependence that does not turn into independence regardless of what other variables are 240 observed (conditioned on). Conditional independence in a BN is captured by the notion of d-separation 241 (directed separation) in the network structure. A collider on a path is the subgraph A→B←C on the 242 path, that is, a triple of nodes A , B , and C where A and C are linked to B . The node B is called the 243 9 certain variables (Kitson and Constantinou, 2024). For example, “wind speed” and “degree days” were 364 indirectly connected through intermediate variable “creel catch rate” in the Alberta dataset, but the 365 two variables were directly connected in the Ontario dataset where “creel catch rate” was absent. The 366 differences could also stem from unobserved intermediate variables. Environmental factors such as water 367 temperature and air pressure, socioeconomic factors such as angler income, and social events occurring 368 in the waterbodies can shape angler activity (Meyer et al., 2023; Dabrowksa et al., 2017; Cantrell et al., 369 2004).370 Secondly, differences in survey methods could introduce bias. Creel surveys recorded fishing durations 371 of incomplete fishing trips, likely leading to underestimation (Eckelbecker, 2019). Conversely, app372 reported fishing durations corresponded to completed trips, but it is unclear if anglers reported the time 373 actively spent fishing only or the entire trip duration, possibly causing overestimation. Additionally, 374 weather variables were averaged for entire river systems or sections, but weather conditions can vary 375 significantly at individual waterbodies, which may influence angler activities (Dippold et al., 2020).376 Thirdly, spatial variability within and between datasets can influence angler activities and, conse377 quently, the BN structures. For example, the Bow River is dominated by brown trout, while the Oldman 378 River is home to cutthroat trout (Johnston et al., 2022). Anglers may be drawn to specific fish species 379 and genus, affecting catch rates and pressure (Hunt et al., 2019, 2002). Other spatial factors, such 380 as water quality, aesthetics, privacy, and accessibility impact angler choices (Hunt, 2005; Moeller and 381 Engelken, 1972; Hunt and Moore, 2006). Local income levels, fishing skills, and motivation of anglers 382 further influence behavior, but these factors are often difficult to measure (Gundelund and Skov, 2021; 383 Hunt et al., 2019).384 This study has several limitations. First, it focuses on only two provinces. Evaluating the methodology 385 in other regions or with broader datasets would likely improve its generalizability. Second, more detailed 386 spatial, temporal, and demographic data (e.g., waterbody size, water temperature, population, and 387 income) could further clarify the relationships between conventional surveys and citizen-sourced data. 388 Third, the high proportion of zeros in the citizen-sourced dataset—particularly in Ontario—suggests 389 low app participation, probably due to limited promotion following its launch. Because the app 390 was introduced earlier and more extensively in Alberta, this discrepancy is understandable. A higher 391 percentage of reporting anglers may uncover relationships to conventionally surveyed variables. Increasing 392 engagement and accurate data collection will require user-friendly interfaces, interactive features, well393 designed prompts, and training for anglers Venturelli et al. (2017).394 Dataset availability395 The data underlying this article were provided by Ontario Ministry of Natural Resources under licence / 396 by permission. The data are available in Taheri Tayebi (2024).397 Acknowledgements398 This project, including the use of the dataset, was reviewed and approved by the Research Ethics 399 Board of the Alberta Research Information Services (ARISE, University of Alberta), study ID MS5 400 Pro00102610. We acknowledge the support of the Government of Canada’s New Frontiers in Research 401 Fund (NFRF), NFRFR-2021-00265. We acknowledge Joel Knudsen, Jamie Svendsen, and Clayton 402 Green for their assistance in gathering and preparing the data. We acknowledge Fiona D. Johnston for 403 her valuable support with the creel data. PR acknowledges funding from an NSERC Discovery Grant.404 Competing Interests Statement405 The authors declare there are no competing interests.406 16 References407 Agmour, I., Bentounsi, M., Baba, N., El Foutayeni, Y., Achtaich, N., and Aouiti, C. (2020). Impact of 408 wind speed on fishing effort. https://doi.org/10.1007/s40808-020-00736-7.409 Arlinghaus, R. and Cooke, S. J. (2009). Recreational fisheries: socioeconomic importance, conservation 410 issues and management challenges. Recreational hunting, conservation and rural livelihoods: science 411 and practice, pages 39–58.412 Baptista, V., Campos, C. J., and Leitão, F. (2016). The influence of environmental factors and fishing 413 pressure on catch rates of diplodus vulgaris. https://doi.org/10.1007/s12237-015-9990-y.414 Brooks, S. (1998). Markov chain monte carlo method and its application. https://doi.org/10.1111/415 1467-9884.00117.416 Broom, B. M., Do, K.-A., and Subramanian, D. (2012). Model averaging strategies for structure learning 417 in bayesian networks with limited data. BMC bioinformatics, 13:1–18.418 Cantrell, R. N., Garcia, M., Leung, P., and Ziemann, D. (2004). Recreational anglers’ willingness to pay 419 for increased catch rates of pacific threadfin (polydactylus sexfilis) in hawaii. https://doi.org/10.420 1016/j.fishres.2004.01.003.421 Christensen, P., SULLIVAN, M., and PATTERSON, B. (2020). Survey to estimate angler effort on the 422 bow river. open.alberta.ca.423 Cooke, S. J. and Cowx, I. G. (2004). The role of recreational fishing in global fish crises. https:424 //doi.org/10.1641/0006-3568(2004)054[0857:TRORFI]2.0.CO;2.425 Cooke, S. J., Lennox, R. J., Bower, S. D., Horodysky, A. Z., Treml, M. K., Stoddard, E., Donaldson, 426 L. A., and Danylchuk, A. J. (2017). Fishing in the dark: the science and management of recreational 427 fisheries at night. https://doi.org/10.5343/bms.2015.1103.428 Dabrowksa, K., Hunt, L. M., and Haider, W. (2017). Understanding how angler characteristics and 429 context influence angler preferences for fishing sites. https://doi.org/10.1080/02755947.2017.430 1383325.431 Dippold, D. A., Adams, G. D., and Ludsin, S. A. (2020). Spatial patterning of walleye recreational 432 harvest in lake erie: role of demographic and environmental factors. https://doi.org/10.1016/j.433 fishres.2020.105676Getrightsandcontent.434 Eckelbecker, R. W. (2019). Comparing angler effort and catch rate estimates across creel survey methods 435 at three alabama reservoirs. Master’s thesis, Auburn University.436 Eden, C., Ackermann, F., and Cropper, S. (1992). The analysis of cause maps. https://doi.org/10.437 1111/j.1467-6486.1992.tb00667.x.438 Eero, M., Strehlow, H. V., Adams, C. M., and Vinther, M. (2015). Does recreational catch impact the 439 tac for commercial fisheries? https://doi.org/10.1093/icesjms/fsu121.440 Ellison, A. M. (2004). Bayesian inference in ecology. https://doi.org/10.1111/j.1461-0248.2004.441 00603.x.442 Fischer, S. M., Ramazi, P., Simmons, S., Poesch, M. S., and Lewis, M. A. (2023). Boosting propag443 ule transport models with individual-specific data from mobile apps. https://doi.org/10.1111/444 1365-2664.14356.445 Franzin, A., Sambo, F., and Di Camillo, B. (2017). bnstruct: an r package for bayesian network structure 446 learning in the presence of missing data. https://doi.org/10.1093/bioinformatics/btw807.447 17 Friedman, N. and Koller, D. (2003). Being bayesian about network structure. a bayesian approach to 448 structure discovery in bayesian networks. Machine learning, 50:95–125.449 Gundelund, C., Arlinghaus, R., Birdsong, M., Flávio, H., and Skov, C. (2022). Investigating angler 450 satisfaction: The relevance of catch, motives and contextual conditions. https://doi.org/10.1016/451 j.fishres.2022.106294.452 Gundelund, C. and Skov, C. (2021). Changes in angler demography and angling patterns during the 453 covid-19 lockdown in spring 2020 measured through a citizen science platform. https://doi.org/10.454 1016/j.marpol.2021.104602.455 Gundelund, C., Venturelli, P., Hartill, B. W., Hyder, K., Olesen, H. J., and Skov, C. (2021). Evaluation 456 of a citizen science platform for collecting fisheries data from coastal sea trout anglers. https:457 //doi.org/10.1139/cjfas-2020-0364.458 Gutowsky, L. F., Gobin, J., Burnett, N. J., Chapman, J. M., Stoot, L. J., and Bliss, S. (2013). 459 Smartphones and digital tablets: emerging tools for fisheries professionals. https://doi.org/10.460 1080/03632415.2013.838133.461 Guy, C. S. and Willis, D. W. (1991). Seasonal variation in catch rate and body condition for four fish 462 species in a south dakota natural lake. https://doi.org/10.1080/02705060.1991.9665305.463 Heggerud, C. M., Xu, J., Wang, H., Lewis, M. A., Zurawell, R. W., Loewen, C. J., Vinebrooke, R. D., 464 and Ramazi, P. (2024). Predicting imminent cyanobacterial blooms in lakes using incomplete timely 465 data. https://doi.org/10.1029/2023WR035540.466 Hinne, M., Gronau, Q. F., van den Bergh, D., and Wagenmakers, E.-J. (2020). A conceptual introduction 467 to bayesian model averaging. https://doi.org/10.1177/2515245919898657.468 Hoeting, J. A., Madigan, D., Raftery, A. E., and Volinsky, C. T. (1999). Bayesian model averaging: 469 a tutorial (with comments by m. clyde, david draper and ei george, and a rejoinder by the authors. 470 https://doi.org/10.1214/ss/1009212519.471 Hunt, L., Haider, W., and Armstrong, K. (2002). Understanding the fish harvesting decisions by anglers. 472 https://doi.org/10.1080/10871200290089355.473 Hunt, L. and Moore, J. (2006). The potential impacts of climate change on recreational fishing in 474 northern ontario. CABI Databases.475 Hunt, L. M. (2005). Recreational fishing site choice models: insights and future opportunities. https:476 //doi.org/10.1080/10871200591003409.477 Hunt, L. M., Camp, E., van Poorten, B., and Arlinghaus, R. (2019). Catch and non-catch-related 478 determinants of where anglers fish: a review of three decades of site choice research in recreational 479 fisheries. https://doi.org/10.1080/23308249.2019.1583166.480 Hurkett, B. and Fitzsimmons, K. (2019). Angler survey in the upper oldman river watershed, 2018. 481 Data Report, produced by the Alberta Conservation Association, Sherwood Park, Alta., Canada.482 Hwang, K.-B. and Zhang, B.-T. (2005). Bayesian model averaging of bayesian network classifiers 483 over multiple node-orders: application to sparse datasets. https://doi.org/10.1109/TSMCB.2005.484 850162.485 Hyder, K., Weltersbach, M. S., Armstrong, M., Ferter, K., Townhill, B., Ahvonen, A., Arlinghaus, 486 R., Baikov, A., Bellanger, M., Birzaks, J., et al. (2018). Recreational sea fishing in europe in a 487 global context—participation rates, fishing effort, expenditure, and implications for monitoring and 488 assessment. https://doi.org/10.1111/faf.12251.489 18 Jiorle, R., Ahrens, R., and Allen, M. (2017). Determining the utility of electronic, self-reported 490 recreational data for fisheries stock assessment. In 147th Annual Meeting of the American Fisheries 491 Society. AFS.492 Jiorle, R. P., Ahrens, R. N., and Allen, M. S. (2016). Assessing the utility of a smartphone app for 493 recreational fishery catch data. https://doi.org/10.1080/03632415.2016.1249709.494 Jivthesh, M., Gaushik, M., Shibu, N., Raj, D., and Rao, S. N. (2022). A comprehensive survey of web 495 and mobile apps for fishermen. Data, Engineering and Applications, pages 199–211.496 Johnson, S. and Mengersen, K. (2012). Integrated bayesian network framework for modeling complex 497 ecological issues. https://doi.org/10.1002/ieam.274.498 Johnston, F. D., Simmons, S., Poorten, B. v., and Venturelli, P. (2022). Comparative analyses with 499 conventional surveys reveal the potential for an angler app to contribute to recreational fisheries 500 monitoring. https://doi.org/10.1139/cjfas-2021-0026.501 Kelleher, K., Westlund, L., Hoshino, E., Mills, D., Willmann, R., de Graaf, G., and Brummett, R. 502 (2012). Hidden harvest: The global contribution of capture fisheries. Worldbank; WorldFish.503 Kendall, M. S., Williams, B. L., Winship, A. J., Carson, M., Grissom, K., Rowell, T. J., Stanley, J., 504 and Roberson, K. W. (2021). Winds, waves, warm waters, weekdays, and which ways boats are 505 counted influence predicted visitor use at an offshore fishing destination. https://doi.org/10.1016/506 j.fishres.2021.105879.507 Kitson, N. K. and Constantinou, A. C. (2024). The impact of variable ordering on bayesian network 508 structure learning. https://doi.org/10.1007/s10618-024-01044-9.509 Koivisto, M. and Sood, K. (2004). Exact bayesian structure discovery in bayesian networks. The Journal 510 of Machine Learning Research, 5:549–573.511 Koller, D. and Friedman, N. (2009). Probabilistic graphical models: principles and techniques. MIT 512 press.513 Lavine, M. (2019). Frequentist, bayes, or other? https://doi.org/10.1080/00031305.2018.1459317 . 514 Lewin, W.-C., Arlinghaus, R., and Mehner, T. (2006). Documented and potential biological impacts 515 of recreational fishing: insights for management and conservation. https://doi.org/10.1080/516 10641260600886455.517 Liao, Z. (2023). Improved bayesian network structure learning in the model averaging paradigm. UW.518 Madigan, D. and Raftery, A. E. (1994). Model selection and accounting for model uncertainty in 519 graphical models using occam’s window. https://doi.org/10.1080/01621459.1994.10476894.520 Malvestuto, S. P. (1996). Sampling the recreational creel. Fisheries techniques, 2nd edition. American 521 Fisheries Society, Bethesda, Maryland, pages 591–623.522 Martin, D. R., Pracheil, B. M., DeBoer, J. A., Wilde, G. R., and Pope, K. L. (2012). Using the internet 523 to understand angler behavior in the information age. https://doi.org/10.1080/03632415.2012.524 722875.525 Meyer, K. A., McCormick, J. L., Kozfkay, J. R., and Dillon, J. C. (2023). Effects of elevated water 526 temperature on cutthroat trout angler catch rates and catch-and-release mortality in idaho streams.527 https://doi.org/10.1111/fme.12605.528 Midway, S. R. and Miller, P. W. (2023). Heat, hurricanes, and health: Effects of natural disturbances 529 on angling effort. https://doi.org/10.1371/journal.pone.0291126.530 19 Milns, I., Beale, C. M., and Smith, V. A. (2010). Revealing ecological networks using bayesian network 531 inference algorithms. https://doi.org/10.1890/09-0731.1.532 Moeller, G. H. and Engelken, J. H. (1972). What fishermen look for in a fishing experience. https:533 //doi.org/10.2307/3799256.534 Mohammadi, R. and Wit, E. C. (2015). Bdgraph: An r package for bayesian structure learning in 535 graphical models. https://doi.org/10.18637/jss.v089.i03.536 Murphy, B. R., Willis, D. W., et al. (1996). Fisheries techniques. Citeseer.537 Nadkarni, S. and Shenoy, P. P. (2001). A bayesian network approach to making inferences in causal 538 maps. https://doi.org/10.1016/S0377-2217(99)00368-9.539 Nyblom, Å. (2014). Making plans or “just thinking about the trip”? understanding people’s travel 540 planning in practice. https://doi.org/10.1016/j.jtrangeo.2014.01.003.541 Papenfuss, J. T., Phelps, N., Fulton, D., and Venturelli, P. A. (2015). Smartphones reveal angler 542 behavior: a case study of a popular mobile fishing application in alberta, canada. https://doi.org/543 10.1080/03632415.2015.1049693.544 Pearl, J. (1988). Probabilistic reasoning in intelligent systems: networks of plausible inference. Morgan 545 kaufmann.546 Post, J., Persson, L., Parkinson, E. v., and Kooten, T. v. (2008). Angler numerical response across 547 landscapes and the collapse of freshwater fisheries. https://doi.org/10.1890/07-0465.1.548 Post, J. R., Sullivan, M., Cox, S., Lester, N. P., Walters, C. J., Parkinson, E. A., Paul, A. J., 549 Jackson, L., and Shuter, B. J. (2002). Canada’s recreational fisheries: the invisible collapse? https:550 //doi.org/10.1577/1548-8446(2002)027%3C0006:CRF%3E2.0.CO;2.551 Ramazi, P., Fischer, S. M., Alexander, J., James, C. T., Paul, A. J., Greiner, R., and Lewis, M. A. 552 (2022). Myxobolus cerebralis establishment and spread: a graphical synthesis. https://doi.org/10.553 1139/cjfas-2020-0352.554 Ramazi, P. and Kalantari, H. (2023). Bayesian and Causal Bayesian Networks. https://555 openlibrary-repo.ecampusontario.ca/jspui/handle/123456789/1790. Accessed: 2024-12-26.556 Ramazi, P., Kunegel-Lion, M., Greiner, R., and Lewis, M. A. (2021). Exploiting the full potential of 557 bayesian networks in predictive ecology. https://doi.org/10.1111/2041-210X.13509.558 Régnière, J., Logan, J. A., Shore, T., and MacLean, D. (1996). Landscape-wide projection of temperature559 driven processes for seasonal pest management decision support: a generalized approach. Decision 560 Support Systems for Forest Pest Management, Proceedings. Entomological Society of Canada, FRDA 561 Report, 260:43–55.562 Régnière, J., Saint-Amant, R., Béchard, A., and Moutaoufik, A. (2017). Biosim 11–manuel d’utilisation. 563 Quebec, QC, Canada: Natural Resources Canada, Canadian Forest Services, Laurentian Forestry 564 Center.565 Ripley, T. and Council, T. (2006). Bow river sport fish angler survey, 2006. data report. Technical 566 report, D-2007-003, produced by the Alberta Conservation Association, Lethbridge . ...567 Shaw, S. L., Renik, K. M., and Sass, G. G. (2021). Angler and environmental influences on walleye 568 sander vitreus and muskellunge esox masquinongy angler catch in escanaba lake, wisconsin 2003–2015. 569 https://doi.org/10.1371/journal.pone.0257882.570 20 Silander, T. and Myllymaki, P. (2012). A simple approach for finding the globally optimal bayesian 571 network structure. https://doi.org/10.48550/arXiv.1206.6875.572 Sims, B. and Danylchuk, A. J. (2017). Characterizing information on best practice guidelines for 573 catch-and-release in websites of angling-based non-government organizations in the united states. 574 https://doi.org/10.1016/j.fishres.2016.09.019.575 Skov, C., Hyder, K., Gundelund, C., Ahvonen, A., Baudrier, J., Borch, T., DeCarvalho, S., Erzini, K., 576 Ferter, K., Grati, F., et al. (2021). Expert opinion on using angler smartphone apps to inform marine 577 fisheries management: status, prospects, and needs. https://doi.org/10.1093/icesjms/fsaa243.578 Speers, J. and Gillis, D. (2012). Catch and effort variation in the commercial gillnet fishery of lake 579 winnipeg, canada, in relation to environmental factors. https://doi.org/10.1016/j.jglr.2011.580 05.006.581 Sullivan, M. G. (2003). Exaggeration of walleye catches by alberta anglers. https://doi.org/10.1577/582 1548-8675(2003)023%3C0573:EOWCBA%3E2.0.CO;2.583 Sun, L. and Shenoy, P. P. (2007). Using bayesian networks for bankruptcy prediction: Some method584 ological issues. https://doi.org/10.1016/j.ejor.2006.04.019.585 Sylvia, A., Maahs, B., and Weber, M. J. (2020). Influence of largemouth bass behaviors, angler behaviors, 586 and environmental conditions on fishing tournament capture success. https://doi.org/10.1002/587 tafs.10216.588 Taheri Tayebi, A. (2024). Webpage views as a proxy for angler pressure and effort: Insights from 589 bayesian networks [data set], https://doi.org/10.5281/zenodo.10881645. Accessed: 2024-12-10.590 Taylor, A. T., Peeper, A. M., Chapagain, B., Joshi, O., and Long, J. M. (2022). Modern reporting methods 591 for angler tag-return studies: Trends in data quality, choice of method, and future considerations. 592 https://doi.org/10.1002/nafm.10738.593 Tian, J., He, R., and Ram, L. (2012). Bayesian model averaging using the k-best bayesian network 594 structures. https://doi.org/10.48550/arXiv.1203.3520.595 van Poorten, B. T., Carruthers, T. R., Ward, H. G., and Varkey, D. A. (2015). Imputing recreational 596 angling effort from time-lapse cameras using an hierarchical bayesian model. https://doi.org/10.597 1016/j.fishres.2015.07.032.598 Venturelli, P. A., Hyder, K., and Skov, C. (2017). Angler apps as a source of recreational fisheries data: 599 opportunities, challenges and proposed standards. https://doi.org/10.1111/faf.12189.600 Vølstad, J. H., Pollock, K. H., and Richkus, W. A. (2006). Comparing and combining effort and catch 601 estimates from aerial–access designs as applied to a large-scale angler survey in the delaware river. 602 http://dx.doi.org/10.1577/M04-146.1.603 Wintle, B. A., McCarthy, M. A., Volinsky, C. T., and Kavanagh, R. P. (2003). The use of bayesian 604 model averaging to better represent uncertainty in ecological models. https://doi.org/10.1111/j.605 1523-1739.2003.00614.x.606 Xiang, Z., Wang, D., O’Leary, J. T., and Fesenmaier, D. R. (2015). Adapting to the internet: trends in 607 travelers’ use of the web for trip planning. Journal of travel research, 54(4):511–527.608 Yu, J., Smith, V. A., Wang, P. P., Hartemink, A. J., and Jarvis, E. D. (2004). Advances to bayesian 609 network inference for generating causal networks from observational biological data. https://doi.610 org/10.1093/bioinformatics/bth448.611 Zischke, M. T., Griffiths, S. P., and Tibbetts, I. R. (2012). Catch and effort from a specialised recreational 612 pelagic sport fishery off eastern australia. https://doi.org/10.1016/j.fishres.2012.04.011.613 21