DisplayTitle

Efficiency of Adaptive Cluster Sampling for Estimating Plant Disease Incidence

Conventional sampling designs such as simple random sampling (SRS) tend to be inefficient when assessing rare and highly clustered populations because most of the time is spent evaluating empty quadrats, leading to high error variances and high cost. In previous studies with rare plant and animal populations, adaptive cluster sampling, where sampling occurs preferentially in the neighborhood of quadrats in which the species of interest is detected during the sampling bout, has been shown to estimate population parameters with greater precision at an effort comparable to SRS. Here, we use computer simulations to evaluate the efficiency of adaptive cluster sampling for estimating low levels of disease incidence (0.1, 0.5, 1.0, and 5.0%) at various levels of aggregation of infected plants having variance-to-mean ratios (V/M) of ≈1, 3, 5, and 10. For each simulation, an initial sample size of 50, 100, and 150 quadrats was evaluated, and the condition to adapt neighborhood sampling (CA), i.e., the minimum number of infected plants per quadrat that triggers a switch from random sampling to sampling in neighboring quadrats, was varied from 1 to 4 (corresponding to 7.7 to 30.8% incidence of infected plants per quadrat). The simulations showed that cluster sampling was consistently more precise than SRS at a field-level disease incidence of 0.1 and 0.5%, especially when diseased plants were highly aggregated (V/M = 5 or 10) and when the most liberal condition to adapt (CA = 1) was used. One drawback of adaptive cluster sampling is that the final sample size is unknown at the beginning of the sampling bout because it depends on how often neighborhood sampling is triggered. In our simulations, the final sample size was close to the initial sample size for disease incidence up to 1.0%, especially when a more conservative condition to adapt (CA > 1) was used. For these conditions, the effect of disease aggregation was minor. In summary, both precision and the sample size required with adaptive cluster sampling responded similarly to disease incidence and aggregation, i.e., both were most favorable at the lowest disease incidence with the highest levels of clustering. However, whereas relative precision was optimized with the most liberal condition to adapt, the ratio of final to initial sample size was best for more conservative CA values, indicating a tradeoff. In our simulations, precision and final sample size were both simultaneously favorable for disease incidence of up to 1.0%, but only when infected plants were most aggregated (V/M = 10).