The paper opens with a short review of previous studies that have uses sports data to evaluate decision making. They conclude that “In the case of many, if not most, of the studies showing a departure from simple rationality or maximization principles, questions arise as to whether the departure arises because agents incorrectly process the available information or because the objective functions for agents are more complex than simple models incorporate”.
Then the authors proceed to start by modelling the coach/manager as a points maximizer while acknowledging that “our results align very closely with Romer’s and appear to offer and expand confirmation that NFL managers are not simple maximizers of expected points”. However, it is difficult to get away from the implicit assumption (or maybe it is a value judgement) that the decision makers should be maximizers of expected points. This is despite the fact that Goff and Locke draw attention to Romer’s point that the fundamental objective for coaches is to win.
I wonder if there is an analogy with golf. Should players seek to minimise the distance from the ball to the hole with each shot? Is a player that lays-up stupid or fearful? If we implicitly assume that the player should seek to minimise the distance to the hole on each shot, and the player takes a three-wood rather than a driver on the tee, then we presume it is incorrect information processing or risk-aversion that led to the decision.
The Goff and Locke paper does indeed dig deeper. It uses more of the information from each game but not all of it. Telling the reader that neither themselves nor Romer use all the game data acts like a flashing red light. It prompts one to wonder why that is the case. Why would one not use the data from the last 3 minutes? To reinforce the point, they produce estimates that suggest that a “team that trails by 7 points is 3.2% more likely to go for a fourth down” and the fourth quarter “raises the likelihood of going for it by 9.7%”. Different people might interpret these finding differently.
Overall, I think the paper suggests that “the objective functions for agents are more complex than simple models incorporate”.