Lost sites, lost signals: How non-random abandonment of sampling sites biases trends in biodiversity

Long-term ecological monitoring is essential for tracking biodiversity trends and guiding conservation policy. However, these programs often experience changes in site coverage over time, through both the loss of existing sites and the addition of new ones. When this loss is non-random—corresponding to a “Missing Not At Random” (MNAR) mechanism, such as preferential abandonment of low-diversity or degraded sites—it can introduce substantial bias in population trend estimates. Although well documented in theory and simulations, this issue remains poorly supported by empirical evidence. Using 20 years of data from a national grassland bird monitoring program in France, we assessed whether site loss biased estimates of temporal change in hay meadow cover, grassland bird abundance, and species richness. Sites retained long-term exhibited higher initial values for all three variables, whereas early-abandoned sites started with lower hay meadow cover and biodiversity and experienced stronger declines, particularly in abundance. Between the first monitoring year and 2024, bird abundance declined by 19% at long-term sites but by more than 40% at early-abandoned ones. Our results reveal that well-preserved hay meadow sites were preferentially retained, while degraded sites were more likely to be abandoned. This selective process leads to systematic underestimation of national biodiversity trends when analyses rely only on long-term sites. To address this, we developed a simple correction method based on resampling of all initially monitored sites. This study provides one of the first empirical demonstrations and quantifications of MNAR bias in biodiversity monitoring. It highlights the need to explicitly account for site loss and selection effects when interpreting long-term trends and designing conservation strategies. • Non-random site loss biases biodiversity trends in long-term monitoring. • Grassland bird abundance fell 19% in long-term versus >40% in early-abandoned sites. • Well-preserved sites were retained over time while degraded sites rapidly abandoned. • First empirical evidence of MNAR bias in biodiversity monitoring. • Simple resampling corrects biased trend estimates effectively.