Addressing Bias in Non‐Probability Fisheries Surveys Using Multilevel Regression and Poststratification Approaches

ABSTRACT

Quantifying harvest of fish stocks is challenging as census data are often unavailable, so surveys are required. Traditional probability‐based surveys use random sampling to obtain representative data that can be scaled to estimate total impact. However, these surveys are costly, so cheaper, non‐probability, citizen science surveys of self‐selected participants are becoming commonplace. Respondents in these surveys may not be representative and inclusion probabilities are unknown, making weighting difficult. This resembles exit polls, which successfully predict election outcomes using Multi‐level Regression and Post‐stratification (MRP), but this has not been tested in fisheries. This study applied MRP to quantify UK recreational sea angling (RSA) participation and catches using data from the UK Sea Angling Diary Project. This represents a ‘worst‐case’ scenario, as there is no sampling frame and no mandatory reporting. Five Bayesian multi‐level models were fitted to two survey datasets: one quantifying participation and days fished, and another location, numbers, and weight of fish caught. Predictions were post‐stratified using UK census data. Estimates from the traditional reweighting and MRP were compared, and a simulation approach was used to assess the accuracy and precision of the two methods. MRP reduced self‐selection bias and improved estimates in under‐represented groups, producing a 40% and 61% improvement for participation and catch, respectively. Average annual UK RSA participation was estimated to be 688,000 anglers, with 6.1 million days fished. Annual catch estimates were 32.4 million fish, weighing 14,876 t. This study demonstrates the value of MRP for reducing bias when analysing and scaling non‐probabilistically collected survey data.