Using heuristics on smaller sample sizes is inherently less efficient. You can get reasonable polling data on 50 million voters using 10,000 samples. That’s 1 sample per 5,000 voters. Drop the election down to a local election with 10,000 voters and the same ratio provides 2 samples telling you almost nothing.

Thus, if 5 insurance companies send patients to the same doctor you simply have more overhead.