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

Area of Science:

• Ecology
• Environmental Science
• Statistical Modeling

Background:

• Biodiversity monitoring relies on sample data, which can be unrepresentative, leading to biased inferences.
• Unrepresentative samples occur when sampled locations differ from nonsampled ones in key variables.
• Auxiliary variables, common causes of sample inclusion and the variable of interest, can help adjust samples.

Purpose of the Study:

• To evaluate the effectiveness of six survey sample adjustment methods for unrepresentative biodiversity data.
• To estimate mean occupancy and trends for Calluna vulgaris in Great Britain using citizen science data.
• To assess the accuracy of adjusted estimates compared to unadjusted ones.

Main Methods:

• Applied six adjustment techniques: subsampling, quasirandomization, poststratification, superpopulation modeling, doubly robust procedure, and multilevel regression and poststratification.
• Utilized a large, unrepresentative citizen science dataset for Calluna vulgaris occupancy in Great Britain.
• Estimated mean occupancy for 1987-1999 and 2010-2019, and the trend between these periods.

Main Results:

• Most adjustment methods resulted in more accurate estimates of mean occupancy and trends compared to unadjusted data.
• Standard uncertainty intervals for adjusted estimates generally did not encompass the true values.
• The effectiveness of adjustments depended on the careful selection of auxiliary variables.

Conclusions:

• Sample adjustment techniques can significantly reduce bias in biodiversity monitoring from unrepresentative datasets.
• Complete unbiased inference is unattainable without knowledge of all relevant auxiliary variables.
• Acknowledging and reporting potential residual bias is crucial when using adjusted samples.