4

Expanding Beyond the Seven Standards

The panel determined that major issues remain to be considered that are not tied to any one of the seven standards, that is, that apply to multiple standards or to larger issues that the standards do not cover. Consideration of such issues also will help to keep the standards current, reflecting improvements that might be adopted in the future through changes in technology or statistics. These include

• Adoption of new and emerging statistical methods in model-based inference for survey data;
• Adoption of new technological advances in data collection methods;
• Adoption of advances in data science, such as artificial intelligence (AI) and machine learning, to improve the efficiency and effectiveness with which estimates are produced; and
• Ensuring that Marine Recreational Information Program (MRIP) continues to meet emerging management needs.

These bespoke issues will require strategic planning and investments on the part of MRIP. In particular, the adoption and implementation of new statistical methods and technological advances in data science may require a sustained collaboration between National Oceanic and Atmospheric Administration (NOAA) and academic researchers engaged in such fields. These are generally not approaches that can be adopted immediately (though perhaps some applications of AI can be adopted quickly); rather, they will generally require a long-term planning and development effort.

These new approaches comprise a supplement to, not a replacement of, existing practices. They could expedite data collection and processing, improve data quality, and help fill in gaps where current data are weak. They also will require development work to be fully effective. Thus, they represent a long-term strategy toward staying current as statistical methodology improves and obtaining improved data quality over the future.

ENHANCED ADOPTION OF MODEL-BASED STATISTICAL INFERENCE

New data-science technologies and statistical workflows can be assimilated into MRIP’s standards for survey concepts. The survey concepts could be based on a review of new model-based frameworks being used in survey data analysis. They could also be based on the investigations, described in peer-reviewed journals, into how such frameworks might enhance statistical inference and improve the quality of survey estimates. Specifically, Standard 1 could underscore the potential benefits of new methods and data streams in expanding survey goals and objectives and include considerations of extraneous unaccounted variations and associations in analyzing survey data.

Recreational fishing surveys predominantly rely on design-based methods to generate their estimates. For example, the Access Point Angler Intercept Survey (APAIS) is conducted along the Atlantic and Gulf coasts and collects individual catch data from anglers returning to public fishing access sites such as boat ramps, piers, beaches, and jetties (National Marine Fisheries Service, 2025b). The APAIS applies a time-space sampling method in conjunction with a stratified, multi-stage cluster design that maximizes sampling efficiency and the spatiotemporal extent of the survey. The Fishing Effort Survey (FES) also follows a design-based stratified sampling approach applied to address-based sampling. The FES estimates fishing effort by residents of sampled states in numbers of angler-trips by computing the Horvitz-Thompson weighted total estimator (Horvitz & Thompson, 1952) with sample weights that reflect sample inclusion probabilities, a non-response adjustment, and a post-stratification adjustment to known population totals (National Marine Fisheries Service, 2025b). A final adjustment, derived from APAIS, accounts for non-resident (i.e., out-of-frame) fishing activity and is applied to estimate total effort by fishing mode. To summarize, these methods scale the rate of catch per trip by fishing effort to estimate the total catch. The specifics of how “fishing effort” is calculated differ between the FES and the For-Hire Survey.

Design-based vs. Model-based Approaches

Design-based inference treats the target population as a fixed entity, where the properties of the population are constant but unknown. Randomness is introduced solely through the sampling design, which dictates how units are selected from the population to form a sample. Inferences are then drawn about the fixed population parameters (e.g., population mean or total) using estimators that rely on the probabilities of inclusion for each unit in the sample. Essentially, the quality of the inference in a design-based approach is tied to the sampling procedure itself.

Model-based approaches, on the other hand, view the population as a random realization of an underlying, assumed statistical model or a superpopulation. In this case, the values of the population units are considered random variables, and the goal is often to predict the values of these random variables based on the sample data and the assumed model. While random sampling can be used in model-based approaches, it is not strictly necessary, and the focus shifts to the validity of the distributional assumptions underlying the model in generating accurate predictions.

In the field of finite population survey sampling, the debate over whether model-based or design-based estimation is superior is long-standing and nuanced. The key distinction lies in the source of randomness: design-based inference derives randomness from the sampling design, whereas model-based inference incorporates it through assumptions about the data-generating process itself, assumptions that may be arbitrary and untestable. Little (2004) offers an excellent account of these approaches with commentary on their benefits and potential pitfalls. Little finds that rather than one approach being universally better, each approach has distinct theoretical foundations and practical strengths. Model-based methods often demonstrate superior efficiency under specific conditions, while design-based methods offer robustness against model misspecification.

A seminal article by Royall (1970) is a cornerstone of the modern model-based approach to finite population inference. Royall directly challenged the orthodoxy of design-based methods, arguing that inference should be based on the model that generated the data, not on the randomization scheme, although the paper’s more radical claims were later refined by Royall himself (and other scholars), who acknowledged the importance of robustness of estimates to misspecified models. Other relevant articles that the panel feels would benefit MRIP’s statisticians as they consider enhanced model-based approaches include Dumelle et al. (2022), who demonstrate that model-based inference tends to outperform design-based inference for spatial data analysis, especially when constructing design-based estimators for simple random sampling. That superiority is evident even when the modeling assumptions are violated, thereby revealing a degree of robustness of estimation to model misspecification.

Bayesian Hierarchical Modeling

Substantial inferential benefits are possible with the adoption of fully model-based approaches including, but not necessarily limited to, Bayesian hierarchical models for finite populations. Such a model is constructed in a manner analogous to linear or generalized linear models for the response variable. For example, the rate of catch per angler-trip is modeled as the dependent variable with design-specific and other explanatory variables appearing as covariates in a linear regression model. The model can be enriched using random effects to capture extraneous variation. In Bayesian hierarchical models, all unknown quantities (e.g., regression coefficients, random effects, variance components) are further assigned probability laws, and statistical inference proceeds by simulating the posterior distribution of unknowns given the observed data. The fitted model is then used to estimate the total catch, either using additional assumptions on the finite population units (or super-population assumptions) or using a hybrid approach where the model-fitted catch rates are scaled by the “fishing effort” metrics, as is currently used by MRIP.

Model-based approaches adopt a predictive paradigm where the unmeasured units in the population of fish are treated as “missing” and are estimated with uncertainty quantification. In particular, rather than rely solely on the design to incorporate spatial-temporal information (e.g., by stratification), the number of catches or the catch rate per angler-trip can itself be modeled using a linear mixed-effects model that accounts for spatial and/or temporal effects. Significant enrichment is possible using hierarchical models that assign structured dependence to the random effects and parameters in the model using appropriate probability distributions.

Andrew Gelman of Columbia University and his collaborators have demonstrated the advantages of Bayesian modeling over design-based estimation for finite populations in several key papers. Their work highlights that Bayesian hierarchical models can produce more stable and efficient estimates, especially in complex survey settings with small sample sizes or non-representative samples. Key papers demonstrating the effectiveness and efficiencies of Bayesian hierarchical models for finite population survey sampling include Si et al. (2015), whose authors show that their Bayesian nonparametric finite population estimator is more robust than the classical design-based estimator by regularizing (smoothing) the highly variable weights, leading to greater efficiency. In another article, Si et al. (2020) incorporate design variables into a multilevel model and demonstrate that their model-based approach outperforms classical weighting by producing more stable estimates, reducing bias, and improving coverage rates. Research has also elucidated how Bayesian methods can be used to generate accurate survey estimates from non-random samples, a common issue

where design-based methods are not applicable (Liu et al., 2023). Key advantages of Bayesian modeling shown in these papers include the following:

• Handling complex designs: Bayesian hierarchical models can coherently incorporate and account for complex design features like clustering and stratification, providing a unified framework for analysis.
• Robustness with small samples: Bayesian approaches borrow strength across different subgroups or “small areas,” which stabilizes estimates for subdomains with small sample sizes where design-based estimates would be noisy or impossible to compute.
• Efficiency: By modeling the data and borrowing information, Bayesian methods can be more efficient than design-based methods, which can suffer from noisy estimators due to highly variable inverse probability weights.
• Predictive power: Bayesian models make predictions for unsampled units, which allows for robust inference even when dealing with non-random or non-representative samples. This is a crucial advantage over design-based methods that rely on the sampling design itself. Note that the accuracy of the estimates partly depends on the proximity of the data; interpolating for missing data within the space where data has been gathered is more accurate than extrapolating beyond the bounds of the data.
• Model-based weights: Instead of relying solely on design-based weights, Bayesian models can generate more stable “model-based weights” that account for auxiliary information, smoothing estimates, and improving robustness.

Researchers should still be open to the possibility that findings may sometimes depend on the priors that are assumed in Bayesian modeling than on the actual data.

As one example of a Bayesian approach in fisheries management, Shelton et al. (2012) devise hierarchical Bayesian models to analyze stratified sampling data from multispecies fisheries. These models allow for the integration of information across different strata and time periods, even with missing data, and provide a framework for quantifying uncertainty in species composition estimates. The hierarchical Bayesian model allows for this estimation of species composition even in strata where no sampling data are available. By providing a robust framework for analyzing data from stratified sampling and handling missing data, the hierarchical Bayesian models can improve the accuracy of stock assessments and contribute to more sustainable fisheries management.

Spatial-temporal Models

Of particular relevance is the possible adoption of spatial-temporal models for fisheries management. In contexts such as environmental monitoring, spatial location is a powerful form of auxiliary information. A 2022 study by Dumelle and others found that, while design-based and model-based approaches seem to perform similarly in estimating the finite population mean when the data do not exhibit spatial associations, for finite population spatial data the model-based approaches produce significantly lower root mean square error and better interval coverage for the finite population mean.

Spatial-temporal models are statistical and computational tools that explicitly account for the interconnectedness of data points across locations and over time. In the context of recreational fishing surveys, they can be beneficial in (a) understanding population dynamics by improving predictive estimation of fish abundance, distribution shifts, and the influence of environmental factors; (b) improving survey design and data collection by identifying areas and time periods of peak fishing activity or areas with low sampling rates; (c) developing effective management strategies by effectively evaluating the impacts of regulations, habitat changes, and climate on fish and anglers; and (d) bridging data gaps by integrating data from various sources (e.g., traditional surveys, citizen science platforms) to enhance spatial-temporal coverage and resolution. Such models are seeing increasing use to standardize data in stock assessments.

Recent publications on spatial-temporal modeling with applications to recreational fisheries include the following:

• Gundelund et al. (2023), which compares catch estimates from a citizen science platform with those from a traditional recall survey, highlighting the potential for using citizen science data in fisheries management;
• Hyman et al. (2024), which explores how fishing effort is impacted by factors like season length, fishing intensity, and weather, emphasizing the need to understand angler behavior for effective management;
• Ochwada-Doyle et al. (2024), which examines the influence of extreme weather on coastal recreational fishing patterns, suggesting that socioeconomic factors may play a larger role than initially thought;
• Brownscombe et al. (2023) and Larocque et al. (2025), which develop telemetry-based spatial-temporal fish habitat models that offer insights into fish habitat use and can help inform management strategies for recreational fishing; and

• “Recent Developments and Applications of Spatio-temporal Modelling for Fisheries Management Advice,” a symposium at the 2024 annual meeting of the American Fisheries Society, which highlighted the growing use of spatio-temporal models in fisheries and discussed advancements and applications for management.¹

In addition to the above, there is recent literature on what is known as spatial sampling in areas without reliable Census data or address information. Examples include Howell et al. (2020) and Eckman and Himelein (2020).

Conclusion 4-1: Spatial sampling is a potential tool for measuring fishing stocks as an alternative to intercept surveys.

The above list of surveys illustrates the variety of design-based, probabilistic survey approaches currently being used in recreational fishing surveys. These surveys provide the ability to statistically infer about large inaccessible populations from modest samples with measurable uncertainty. An important specific action for MRIP to consider is to adopt emerging Bayesian modeling approaches to assimilate and pool information from its different surveys. Individual surveys offer markedly disparate estimates as a result of their respective methodological biases and design attributes. Bayesian statistical frameworks have been shown to be effective in assimilating multiple surveys to arrive at pooled estimates. As one example, Tucker et al. (2024) apply a Bayesian framework to data collected from a relatively large winter fishery to integrate three traditional creel methodologies in deriving a common estimate of angler effort. These investigations report that integrating different survey types would help to lower uncertainty in the estimates.

In summary, peer-reviewed research demonstrates the significant benefits of using Bayesian hierarchical models for meta-analysis, especially when pooling estimates from diverse sources like MRIP surveys. These models offer a robust framework for handling heterogeneity, sharing information, incorporating prior knowledge, and providing comprehensive uncertainty assessments, which can lead to more robust and informative conclusions in various fields, including fisheries management and environmental science.

Recommendation 4-1: Marine Recreational Information Program should consider broadening the allowed data collection and analysis approaches to incorporate other types of data to supplement survey statistics.

___________________

¹ https://afsannualmeeting2024.fisheries.org/2024/03/22/recent-developments-and-applications-of-spatio-temporal-modelling-for-fisheries-management-advice/

Recommendation 4-2: Marine Recreational Information Program should explore and, where appropriate, promote the use of model-based inference as a way of improving data quality and timeliness.

Although not the mandated goal of MRIP, its catch and effort data prove valuable as input to economic and other models. There is a substantial body of literature on fisheries management using utility-based decision-theoretic modeling. These models are based on the premise that fishing effort and catch per trip jointly inform about the estimated total catch and subsequently feed into a population dynamics model. Implicitly, there is an underlying production function for which catch is proportional to effort. However, such assumptions may be controversial (Zhang & Smith, 2011), as are assumptions that effort is exogenous, which also often fail to hold empirically (Smith et al., 2008; Zhang, 2011).

Statistical models for survey data that rely on distributional assumptions on a superpopulation may fail to identify fundamental causal drivers of behavior (and the system) such that the parameters can be used to simulate new policies. This is a version of the Lucas critique in economics. Structural bioeconomic models (Homans & Wilen, 1997) can overcome this problem, but at the cost of making strong assumptions (in the model) and perhaps at the cost of statistical model complexity. Researchers and scholars of recreational fishing are increasingly highlighting the need to look at institutions, economic incentives, and population dynamics jointly (Abbott et al., 2018, 2022; Arlinghaus et al., 2019; Fujiwara et al., 2018). How bioeconomic models will, if at all, effectuate improved estimates from MRIP’s surveys in practice merits further investigation and should be explored further.

The bioeconomics literature has important implications for setting statistical standards. First, as discussed under Standard 1, defining statistical standards only in terms of survey research may be too narrow (though standards for survey research are definitely needed in order to produce quality national data). Second, the quality of survey statistics may sometimes be enhanced through the use of such model-based approaches. Finally, it is difficult to set standards for model-based statistics, but one definite need is for all underlying assumptions within such approaches to be clearly specified so that researchers can better evaluate how those assumptions affect the data and resulting conclusions.

ADOPTION OF NEW DATA-SCIENCE TECHNOLOGIES

Current MRIP survey concepts are predominantly devoted to probability sampling. The rapid emergence of data-science technologies, however, affords new opportunities for incorporating extraneous sources of

information and non-probability data streams into surveys. These include, but are not limited to, angler apps, cell-phone mobility data, and citizen science.

Based on rapid developments in statistical methodologies, computational frameworks, and data-science technologies, there are opportunities to integrate data analytic workflows into the standard. As one example, Monnahan (2024) offers a statistical workflow for Bayesian modeling and data analysis that could be particularly beneficial for assessing fisheries stocks. Such a workflow should be considered carefully for possible integration into the standards. Monnahan’s work, which adapts the modern Bayesian workflow proposed by Gelman et al. (2020), should form the basis for proposed good practices in fisheries sciences.

The Bayesian workflow is clearly explained in the form of a flowchart, as depicted in Figure 1 of Monnahan (2024). We offer a brief outline here and refer the reader to Section 2, Table 1, of Monnahan (2024). The essential idea of a Bayesian workflow relies on simulating data sets from generative models (essentially any valid joint probability model for the data and the parameters) and carrying out Bayesian data analysis on these simulated data sets in a streamlined manner. This data analysis involves model building, inference by sampling from the posterior distribution of parameters (that could also include random effects, missing data), evaluating these models in terms of validation with respect to the data being analyzed, incorporating possible improvements, and comparing different models to better understand how they relate to the data.

While there is some flexibility with regard to the specific components of a flowchart of a Bayesian workflow, one would typically begin by simulating data from a generative probability model, including a prior distribution for the parameters and a probability model for the data (likelihood). “Prior predictive checking” of this model is carried out using simulation experiments to check that the priors and model data generating process produce expected quantities and simulated data that match physical or biological limitations or other expert opinion (if available).

The next step analyzes the simulated data using samples drawn from the posterior distribution of the Bayesian model. This model fitting exercise is followed by model validation using “posterior predictive checks,” which ascertain how well the fitted model replicates the data being analyzed. Posterior predictive checks are often summarized in the form of model performance scores (often using cross-validation metrics based on different probabilistic information criteria) that can be used to compare across different competing models to select one that has the best predictive performance for the data. The estimates and standard errors for the quantities of interest from this selected model are officially reported. Monnahan (2024) applies

these workflows to fisheries stock assessments, but such workflows can also be adapted to survey data for reporting catches.

Monnahan’s article provides a cogent argument that the Bayesian and frequentist paradigms complement each other, with both helping analysts better understand different aspects of their models and data (Monnahan, 2024). Wider adoption of Bayesian methods using the good practices proposed here would therefore lead to improved scientific advice used to manage fisheries. Such statistical workflows are enabled using software development such as “fishSTAN” (Erickson et al., 2022) that can be integrated into MRIP’s standards for survey methodologies.

New technologies, particularly remote sensing and AI, offer the potential to improve recreational fisheries data collection and management. These technologies may provide more timely, accurate, and cost-efficient data, leading to better management decisions and more sustainable fisheries. While NOAA has invested in and adopted aerial and remote sensing methods in diverse enterprises, it is not clear to what extent such technologies can be adopted by MRIP and what their potential benefits would be for enhancing and improving survey estimates. While it may be somewhat premature to formally integrate such data-science technologies into current standards, the potential benefits of these technologies should be evaluated. Part of that evaluation might include talking with fisheries assessment scientists, ecosystem analysts, and oceanographers about what information would be most helpful. The information produced by such technologies could lead to enhanced predictive inference for fish populations in model-based frameworks.

Recent articles have examined remote sensing and AI. These include Kim et al. (2024), which introduced the Artificial Intelligence-based Real-time Catch Analysis System (AI-RCAS) for sustainable fisheries management. AI-RCAS employs fish recognition, tracking, and counting, demonstrating improved species recognition rates in fishing environments. It aims to offer real-time analysis to enhance the efficiency of the total allowable catch system, addressing the limitations of existing electronic monitoring systems. Another recent article (Wing & Woodward, 2024) discusses the potential of AI in fisheries for improved monitoring, catch and bycatch estimates, and identification of illegal fishing, while noting that innovation is hindered by challenges such as regulation and complex policy. Advancing in AI will require collaboration.

Other studies that have explored the prospects of using AI and emerging data-science technologies in fishing data collection and management include a study of image classification models for monitoring recreational catches (Baker et al., 2025); the use of AI and computer vision for monitoring coastal recreational activities and detecting and classifying fish and estimating fishing effort (Baker et al., 2025); and the use of remote sensing

and AI for tracking recreational boats, which could lead to automatic monitoring of recreational fishing effort (Signaroli et al., 2025). MRIP’s standards may be enhanced by considering prudent integration of technologies like acoustic cameras and AI, which are already being employed by NOAA Fisheries to collect and analyze data and improve data reliability.²

While many of these technologies are still developing and face challenges related to data and algorithm accuracy in various environments, in considering its standards more broadly MRIP should begin exploring the possibilities and potential for adopting these technologies, in conjunction with model-based statistical and probabilistic machine learning methods, to build more efficient statistical workflows.

As discussed in Chapter 2, states sometimes have limited statistical resources. MRIP can help the states by coordinating with them on what their needs are, sharing what it has learned through its own investigations, and sharing tools that it has developed. A common use of these shared tools may also promote greater consistency in data collection and processing.

Conclusion 4-2: New technologies are available that may be helpful in combining data from multiple sources and reconciling differences.

Recommendation 4-3: Given the limited resources available to many state and regional agencies, and also to avoid duplication of effort, Marine Recreational Information Program should provide guidance to the agencies on how to use technology to promote higher data quality.

___________________

² https://www.fisheries.noaa.gov/feature-story/future-red-snapper-data-using-active-acoustic-monitoring-and-artificial-intelligence; https://www.obawebsite.com/noaa-uses-ai-and-sonar-to-track-red-snapper-in-the-gulf

This page intentionally left blank.