White & Smith, 2007, Testing measures of animal social association

Ref: White, D. J., & Smith, V. A. (2007). Testing measures of animal social association by computer simulation. Behaviour, 144, 1447-1468.

This paper demonstrates a unique and valuable use for agent-based modeling in the behavioral sciences. There are a number of different measures of social association used in animal behavior. Which method is used may depend on convenience related to the particular species.

They created a simulation of virtual animals, 10 of each sex, who moved with varying levels of social association. They then captured the true underlying distributions of social association according to each measure. In addition, they had several virtual "observer" agents who acted as field biologists, collecting data and restricted by location and speed. They then compared the different methods both objectively in terms of how well they reflect real tendencies, as well as how accurate an otherwise unbiased observer would be in using them.

The Simulation
The animals: 10 male and 10 female agents, occupying a 100x100 grid. Each time step, they moved with probability = .75. If they moved, they could move either randomly or socially. They moved socially with probability s, which was varied between 0 and 0.3. If the moved randomly, they moved to one of the adjacent 8 squares with equal probability. If they moved socially, each direction was weighted by the number of agents in the quadrant bisected by that direction, within 15 spaces in all directions.

The observers: Observers could sample data only every 5 time steps. Some took data in blocks of time (60 samples in 300 time steps) after which certain sampling rules were rest for the beginning of the next block.
Group membership: Began in random location and recorded all organisms within the group nearest the observer's current location. The closest organism was the target individual for the group. Group membership was defined as all individuals within 5 spaces of the target, as well as any individuals within 5 spaces of another group member. For the next sampled data, the observers moved to a new location and sampled again.
Nearest neighbor: Systematically selected each organism in a random order and noted the individual that was closest to it. If there were no other organisms within 30 spaces of the selected organism, the observer moved on to the next organism in the order. At each sampling time point, the observer took data on the next organism in the order. When the end of the list was reached, a new random order was generated. Thus, near-equal amounts of data was collected on all organisms.
Neighborhood: Taken by adlib scanning and focal sampling. Scanning: Observers noted all the organisms within 5 spaces of a target. One observer rnadomly selected individuals that had at least one organism in its neighborhood, and then did not re-sample that target until the collection block was over. The second observer did the same thing, but could re-use individuals for other associations within a block. The focal observer systematically chose each organism from a random order. At the end of each data sampling point, the focal observer noted all the other organisms within 5 spaces of the target. Used the same organism throughout the block.

The underlying distributions used all the data for all the organisms.

Results:
Overall, group membership was pretty accurate, though it had a tendency to report false positives in assortment between the sexes. This was apparent in the underlying distribution, and more pronounced in the observers' data. In large groups, associations between individuals compound rapidly. This bias is not a sampling error, but it inherent in the measure when taking large groups into account.
The nearest-neighbor method produced a number of inaccuracies. The underlying distribution was insensitive to differences in affiliative strength across classes, and was actually worse than the observers' sample at detecting sex differences. When one class was assorted and the other dispersed, there was a tendency not to record data for the dispersed class, and when there was, that neighbor was more likely to be of the dispersed class (since the other class was assorting). The critical variable to be considered is the giving up distance (here = 30 spaces).
For neighborhood methods, focal sampling was less sensitive to detecting effects within and between classes than scan sampling. Reducing the amount of time spent sampling each individual, while not increasing the overall amount of data collected, increased accuracy.

This paper demonstrates that there are real differences between how well certain measures capture real differences in behavior, and these differences can be compounded by observer limitations. It doesn't take into account human biases, which can also contribute to errors. This is a really creative use of agent-based modeling (ABM) to do behavioral research.