Natural resource managers need accurate depictions of existing resources to make informed decisions. The classical approach to describing resources for a given area in a quantitative manner uses probabilistic sampling and design-based inference to estimate population parameters. While probabilistic designs are accepted as being necessary for design-based inference, many recent studies have adopted non-probabilistic designs that do not include elements of random selection or balance and have relied on models to justify inferences. While common, model-based inference alone assumes that a given model accurately depicts the relationship between response and predictors across all populations. Within complex systems, this assumption can be difficult to justify. Alternatively, models can be trained to a given population by adopting design-based principles such as balance and spread. Through simulation, we compare estimates of population totals and pixel-level values using linear and nonlinear model-based estimators for multiple sample designs that balance and spread sample units. The findings indicate that model-based estimators derived from samples spread and balanced across predictor variable space reduce the variability of population and unit-level estimators. Moreover, if samples achieve approximate balance over feature space, then model-based estimates of population totals approached simple expansion-based estimates of totals. Finally, in all comparisons made, improvements in estimation were achieved using model-based estimation over design-based estimation alone. Our simulations suggest that samples drawn from a probabilistic design, that are spread and balanced across predictor variable space, improve estimation accuracy.
- Model-based estimation
- Sample design