Stakeholder diversity matters: employing the wisdom of crowds for data-poor fisheries assessments

Abstract

Embracing local knowledge is vital to conserve and manage biodiversity, yet frameworks to do so are lacking. We need to understand which, and how many knowledge holders are needed to ensure that management recommendations arising from local knowledge are not skewed towards the most vocal individuals. Here, we apply a Wisdom of Crowds framework to a data-poor recreational catch-and-release fishery, where individuals interact with natural resources in different ways. We aimed to test whether estimates of fishing quality from diverse groups (multiple ages and years of experience), were better than estimates provided by homogenous groups and whether thresholds exist for the number of individuals needed to capture estimates. We found that diversity matters; by using random subsampling combined with saturation principles, we determine that targeting 31% of the survey sample size captured 75% of unique responses. Estimates from small diverse subsets of this size outperformed most estimates from homogenous groups; sufficiently diverse small crowds are just as effective as large crowds in estimating ecological state. We advocate for more diverse knowledge holders in local knowledge research and application.

Keywords: Collective intelligence; Fisheries management; Indigenous and local knowledge; Recreational fisheries; Wisdom of crowds.

Fig. 1

Demographics of the full crowd (n = 210 individuals). Bar plots (panels a, b) show the number of individuals across age classes and years of fishing experience. Panels c, d and e show the average number of days spent fishing each year across age classes, fishing experience categories and user groups. Points represent mean ± 95% confidence intervals.

Fig. 2

Aggregated estimates (mean ± standard deviation) of bonefish fishing quality in South Florida, USA, from 1975–2015 partitioned by age class (panel a) and fishing experience (panel c) where a value of 5 = very good and a value of 1 = very poor. In a and c the green dashed line shows the mean trend from the full sample and shaded green area shows the standard deviation from the full sample. The black line shows the mean trend from the group and shaded purple area shows the standard deviation from the group (grey where it overlaps with the results from the full sample). Panel b and d shows the mean absolute percentage error (MAPE) of estimates (compared with the full sample mean) of different groups across each time period. Black dashed lines in b and d represent the mean absolute percentage error (MAPE) of the group.

Fig. 3

Aggregated estimates (mean ± standard deviation) of bonefish fishing quality in South Florida, USA, from 1975–2015 differentiated by recreational anglers and fishing guides (panel a) where a value of 5 = very good and a value of 1 = very poor. In a the black line shows the mean trend from the group and shaded purple area shows the standard deviation from the group. The green dashed line shows the mean trend from the full sample and shaded green area shows the standard deviation from the full sample. Panel b shows the mean absolute percentage error (MAPE) of estimates (compared with the full sample mean) of different groups across each time point. Black dashed lines in b represent mean absolute percentage error (MAPE) of the group.

Fig. 4

Unique responses to a question about bonefish fishing quality in South Florida between 1975 and 2015 generated by randomly selected subsets of varying sample sizes. Panel a shows the relationship between number of unique responses generated with increasing sample size per time point. The maximum number of unique responses generated varied across periods because information was gathered from fewer respondents in early time periods (e.g., in more distant years (1975, 1985), some respondents were not yet born or not fishing). Panel b shows separate saturation curves as per panel a for each of the seven time periods (1975–2015), and dotted lines represent the mean sample size required to capture, 50% (n = 20), 75% (n = 66) and 100% (n = 130) of unique responses. To capture all unique responses, fewer individuals were needed than our full sample (n = 210).

Fig. 5

Mean absolute percentage error (MAPE) of estimates (compared with the full sample mean) of a homogenous groups from the full sample and b diverse subsets from repeated subsampling. Points in b are presented with standard error (thick line) and 95% confidence intervals (thin line). Dashed line represents an MAPE of 5% where points below this are considered very good estimates.