Piecewise continuous sampling: a method for minimizing bias and sampling effort for estimated metrics of animal behavior

Capturing qualitative features of animal behavior requires recording occurrences of behavior over time. Continuous sampling is best for capturing brief behaviors, but can be very time consuming. Instantaneous sampling can reduce the amount of labor required, but can miss short-duration behaviors. We therefore synthesized these techniques by continuously sampling during randomly scattered time intervals; a technique we call piecewise continuous sampling. To optimize and test the efficacy of this technique, we collected a continuous behavioral dataset of harvester ant workers, and then we developed a protocol to estimate the amount of sampling time necessary to reconstruct the proportion of time animals spend in different behavioral states. This protocol finds the sample size needed for the variance of the sample to converge on the variation of the population. We then divided this estimated time into equal-duration intervals that were randomly distributed across the entire continuous dataset. Finally, we calculated both time-dependent and time-independent error from this sample. We found that 4 to 16 sampling intervals minimize both types of error simultaneously. This finding was robust to differences in underlying behavior and was validated with simulations, implying that this method could be used for many types of organisms.

Methods
In order to create a continuously-sampled dataset to compare sampling methods against, we manually coded the behavior of nine Pogonomyrmex californicus ants continuously over a three-hour timespan. Six were from a small colony (~30 workers, 2 queens) and three were from a larger one (~110 workers, 2 queens), though both colonies were still considered small as colonies in this species typically reach a size of 2,000–4,500 workers in the field (Johnson 2000). The nest was partitioned into a foraging arena and a brood chamber with a total surface area of 242 cm2. The workers that were followed were selected based on the task they were doing at the beginning of the video, so as to capture a range of repertoires; brood care (interacting with brood), food processing (interacting with seeds or artificial diet), or resting (immobile). Two from the small colony, and one from the larger, were selected for each task group. Switches between activities were manually coded using the program Cowlog (Version 3.0.2; Hänninen & Pastell 2009) and results were visualized with the ggplot2 package in R. We categorized behaviors into 9 discrete tasks and 3 activity levels (Table S1). Our lab-reared colonies were started from newly mated P. californicus foundresses that were collected in 2017 from Pine Valley, California (lat 32°49′20″N, long 116°31′43″W, 1136 m elevation).