Descriptive inference using large, unrepresentative nonprobability samples: An introduction for ecologists.

Biodiversity monitoring usually involves drawing inferences about some variable of interest across a defined landscape from observations made at a sample of locations within that landscape. If the variable of interest differs between sampled and nonsampled locations, and no mitigating action is taken, then the sample is unrepresentative and inferences drawn from it will be biased. It is possible to adjust unrepresentative samples so that they more closely resemble the wider landscape in terms of "auxiliary variables." A good auxiliary variable is a common cause of sample inclusion and the variable of interest, and if it explains an appreciable portion of the variance in both, then inferences drawn from the adjusted sample will be closer to the truth. We applied six types of survey sample adjustment-subsampling, quasirandomization, poststratification, superpopulation modeling, a "doubly robust" procedure, and multilevel regression and poststratification-to a simple two-part biodiversity monitoring problem. The first part was to estimate the mean occupancy of the plant Calluna vulgaris in Great Britain in two time periods (1987-1999 and 2010-2019); the second was to estimate the difference between the two (i.e., the trend). We estimated the means and trend using large, but (originally) unrepresentative, samples from a citizen science dataset. Compared with the unadjusted estimates, the means and trends estimated using most adjustment methods were more accurate, although standard uncertainty intervals generally did not cover the true values. Completely unbiased inference is not possible from an unrepresentative sample without knowing and having data on all relevant auxiliary variables. Adjustments can reduce the bias if auxiliary variables are available and selected carefully, but the potential for residual bias should be acknowledged and reported.

Treating gaps and biases in biodiversity data as a missing data problem.

Big biodiversity data sets have great potential for monitoring and research because of their large taxonomic, geographic and temporal scope. Such data sets have become especially important for assessing temporal changes in species' populations and distributions. Gaps in the available data, especially spatial and temporal gaps, often mean that the data are not representative of the target population. This hinders drawing large-scale inferences, such as about species' trends, and may lead to misplaced conservation action. Here, we conceptualise gaps in biodiversity monitoring data as a missing data problem, which provides a unifying framework for the challenges and potential solutions across different types of biodiversity data sets. We characterise the typical types of data gaps as different classes of missing data and then use missing data theory to explore the implications for questions about species' trends and factors affecting occurrences/abundances. By using this framework, we show that bias due to data gaps can arise when the factors affecting sampling and/or data availability overlap with those affecting species. But a data set per se is not biased. The outcome depends on the ecological question and statistical approach, which determine choices around which sources of variation are taken into account. We argue that typical approaches to long-term species trend modelling using monitoring data are especially susceptible to data gaps since such models do not tend to account for the factors driving missingness. To identify general solutions to this problem, we review empirical studies and use simulation studies to compare some of the most frequently employed approaches to deal with data gaps, including subsampling, weighting and imputation. All these methods have the potential to reduce bias but may come at the cost of increased uncertainty of parameter estimates. Weighting techniques are arguably the least used so far in ecology and have the potential to reduce both the bias and variance of parameter estimates. Regardless of the method, the ability to reduce bias critically depends on knowledge of, and the availability of data on, the factors creating data gaps. We use this review to outline the necessary considerations when dealing with data gaps at different stages of the data collection and analysis workflow.

We need to talk about nonprobability samples.

In most circumstances, probability sampling is the only way to ensure unbiased inference about population quantities where a complete census is not possible. As we enter the era of 'big data', however, nonprobability samples, whose sampling mechanisms are unknown, are undergoing a renaissance. We explain why the use of nonprobability samples can lead to spurious conclusions, and why seemingly large nonprobability samples can be (effectively) very small. We also review some recent controversies surrounding the use of nonprobability samples in biodiversity monitoring. These points notwithstanding, we argue that nonprobability samples can be useful, provided that their limitations are assessed, mitigated where possible and clearly communicated. Ecologists can learn much from other disciplines on each of these fronts.

An operational workflow for producing periodic estimates of species occupancy at national scales.

Policy makers require high-level summaries of biodiversity change. However, deriving such summaries from raw biodiversity data is a complex process involving several intermediary stages. In this paper, we describe an operational workflow for generating annual estimates of species occupancy at national scales from raw species occurrence data, which can be used to construct a range of policy-relevant biodiversity indicators. We describe the workflow in detail: from data acquisition, data assessment and data manipulation, through modelling, model evaluation, application and dissemination. At each stage, we draw on our experience developing and applying the workflow for almost a decade to outline the challenges that analysts might face. These challenges span many areas of ecology, taxonomy, data science, computing and statistics. In our case, the principal output of the workflow is annual estimates of occupancy, with measures of uncertainty, for over 5000 species in each of several defined 'regions' (e.g. countries, protected areas, etc.) of the UK from 1970 to 2019. This data product corresponds closely to the notion of a species distribution Essential Biodiversity Variable (EBV). Throughout the paper, we highlight methodologies that might not be applicable outside of the UK and suggest alternatives. We also highlight areas where the workflow can be improved; in particular, methods are needed to mitigate and communicate the risk of bias arising from the lack of representativeness that is typical of biodiversity data. Finally, we revisit the 'ideal' and 'minimal' criteria for species distribution EBVs laid out in previous contributions and pose some outstanding questions that should be addressed as a matter of priority. Going forward, we hope that this paper acts as a template for research groups around the world seeking to develop similar data products.

ROBITT: A tool for assessing the risk-of-bias in studies of temporal trends in ecology.

Aggregated species occurrence and abundance data from disparate sources are increasingly accessible to ecologists for the analysis of temporal trends in biodiversity. However, sampling biases relevant to any given research question are often poorly explored and infrequently reported; this can undermine statistical inference. In other disciplines, it is common for researchers to complete 'risk-of-bias' assessments to expose and document the potential for biases to undermine conclusions. The huge growth in available data, and recent controversies surrounding their use to infer temporal trends, indicate that similar assessments are urgently needed in ecology.We introduce ROBITT, a structured tool for assessing the 'Risk-Of-Bias In studies of Temporal Trends in ecology'. ROBITT has a similar format to its counterparts in other disciplines: it comprises signalling questions designed to elicit information on the potential for bias in key study domains. In answering these, users will define study inferential goal(s) and relevant statistical target populations. This information is used to assess potential sampling biases across domains relevant to the research question (e.g. geography, taxonomy, environment), and how these vary through time. If assessments indicate biases, then users must clearly describe them and/or explain what mitigating action will be taken.Everything that users need to complete a ROBITT assessment is provided: the tool, a guidance document and a worked example. Following other disciplines, the tool and guidance document were developed through a consensus-forming process across experts working in relevant areas of ecology and evidence synthesis.We propose that researchers should be strongly encouraged to include a ROBITT assessment when publishing studies of biodiversity trends, especially when using aggregated data. This will help researchers to structure their thinking, clearly acknowledge potential sampling issues, highlight where expert consultation is required and provide an opportunity to describe data checks that might go unreported. ROBITT will also enable reviewers, editors and readers to establish how well research conclusions are supported given a dataset combined with some analytical approach. In turn, it should strengthen evidence-based policy and practice, reduce differing interpretations of data and provide a clearer picture of the uncertainties associated with our understanding of reality.

occAssess: An R package for assessing potential biases in species occurrence data.

Species occurrence records from a variety of sources are increasingly aggregated into heterogeneous databases and made available to ecologists for immediate analytical use. However, these data are typically biased, i.e. they are not a probability sample of the target population of interest, meaning that the information they provide may not be an accurate reflection of reality. It is therefore crucial that species occurrence data are properly scrutinised before they are used for research. In this article, we introduce occAssess, an R package that enables straightforward screening of species occurrence data for potential biases. The package contains a number of discrete functions, each of which returns a measure of the potential for bias in one or more of the taxonomic, temporal, spatial, and environmental dimensions. Users can opt to provide a set of time periods into which the data will be split; in this case separate outputs will be provided for each period, making the package particularly useful for assessing the suitability of a dataset for estimating temporal trends in species' distributions. The outputs are provided visually (as ggplot2 objects) and do not include a formal recommendation as to whether data are of sufficient quality for any given inferential use. Instead, they should be used as ancillary information and viewed in the context of the question that is being asked, and the methods that are being used to answer it. We demonstrate the utility of occAssess by applying it to data on two key pollinator taxa in South America: leaf-nosed bats (Phyllostomidae) and hoverflies (Syrphidae). In this worked example, we briefly assess the degree to which various aspects of data coverage appear to have changed over time. We then discuss additional applications of the package, highlight its limitations, and point to future development opportunities.