Combination Rules in Information Integration

人们所做的决定往往显示出特殊的偏见，尤其是在将信息结合在一起时。假设你是一起谋杀案的陪审员。控方向你提供两项证据：一名目击证人声称在犯罪现场看到被告，另一名专家将凶器上的指纹与被告相匹配。这两项证据都有力地指向被告，但都不完美。令人惊讶的是，我们发现，通过一个更严格的实验室实验，被告有罪的印象并不会随着第二件证据而增加。如果实验中的参与者认为基于一条证据有75%的可能性有罪，他们仍然相信基于两条证据有75%的可能性有罪。然而，我们并不总是错误地组合信息——我们用来组合信息的规则是可塑的。如果我们将谋杀案重新定义为一个感知任务，就会出现这样的例子。假设我们关心的是某人是否发出了“ba”或“pa”的声音，而不是确定有罪或无罪。如果不评估两个证人，假设我们有一段嘴动的视频和一段音频。这和谋杀案的例子是一样的问题，但这里人们正确地组合了信息。如果信息以“认知”的方式呈现给人们，作为单独的文本片段，他们就会错误地组合起来。如果信息是“感性地”呈现的，即作为同一图像的不同部分，他们就会正确地组合起来。其他研究人员也研究了人们在不同情况下组合信息的方式，但他们在不同的任务和不同的刺激下使用了不同的操作方法。在我们的试点工作中，我们使用同一组刺激，使谋杀案和声音识别任务尽可能相似，但仍然发现人们对信息的组合非常不同。我们的总体目标是通过使用一组共同的刺激和操作来探索各种任务，从而对驱动人们正确或错误地组合信息的原因形成统一的理解。除了我们的总体目标之外，我们对这项研究还有一些具体的目标。我们的第一个目标是调查为什么改变谋杀审判和声音识别任务之间的极端呈现会影响我们如何组合信息。这些操作应该告诉我们，这种差异是由于有更好的感知系统而不是认知系统，还是由于有更多的感知任务经验，还是人们对信息的来源有特定的假设。我们的第二个目标是了解人们如何将有关投资风险和可能获得多少潜在回报的信息结合起来。虽然投资者经常不得不在投资之间做出选择，但研究人员不知道人们在这项任务中是否正确地结合了风险和回报。一旦我们确定了人们使用的规则类型，我们就会看到谋杀审判/声音识别任务的操作是否有同样的效果——人们在两个任务中是否有共同的选择组合规则的机制。尽管人们经常不能正确地整合信息，但考虑到他们对正确整合规则的不确定性，他们的策略实际上可能是最优的。从统计学上讲，如果不确定正确的规则，接近真实答案的最佳策略是对候选规则的估计取平均值。我们的第三个目标是测试人们在面对整合规则的不确定性时是否真的表现得最优。

The decisions that people make often show peculiar biases, especially when combining information together. Let's say that you are a juror in a murder trial. The prosecution presents you with two pieces of evidence: an eyewitness who claims to have seen the defendant at the scene of the crime and an expert who matched the fingerprints on the murder weapon to the defendant. Both pieces of evidence strongly implicate the defendant, but neither is perfect. Surprisingly we've found, using a more controlled laboratory task, that the impression of the guilt of the defendant does not increase with the second piece of evidence. If a participant in the experiment thought that there was a 75% chance of guilt based on one pieces of evidence, they still believed that there was a 75% chance of guilt based on two pieces of evidence.However, we do not always combine information incorrectly -- the rule we use to combine information is malleable. An example of this comes if we reframe the murder case as a perceptual task. Instead of determining guilt or innocence, let's say we are concerned with whether someone made a "ba" or "pa" sound. And if instead of evaluating two witnesses, let's say that we are presented with a video of a mouth moving and an audio clip. This is the same problem as the murder case example, but here people combine information correctly. If the information is presented to people "cognitively", as separate pieces of text, they combine it incorrectly. If the information is presented "perceptually", as different parts of the same image, they combine it correctly.Other researchers have looked at when people combine information differently, but have used different manipulations in different tasks with different stimuli. In our pilot work, we made the murder case and sound identification task as similar as possible using the same set of stimuli and still found that people combined information very differently. Our overall goal is to develop a unified understanding of in what drives people to combine information correctly or incorrectly by exploring a variety of tasks using a common set of stimuli and manipulations.In addition to our overall goal, we have some specific aims for this research. Our first aim is to investigate why changing the presentation between the extremes of the murder trial and sound identification tasks affects how we combine information. These manipulations should tell us whether the difference is due to having a better perceptual rather than cognitive system, whether it is due to having more experience with the perceptual task, or whether people have particular assumptions about where the information comes from.Our second aim is to understand how people combine information about how risky an investment is and how much potential reward could be gained. Though investors often have to choose between investments, researchers do not know whether people combine risk and reward correctly in this task. Once we have determined what type of rule people are using, we see if the murder trial/sound identification task manipulations have the same effect here - whether people have common mechanisms for choosing combination rules in the two tasks.Although people often do not integrate information correctly, their strategy might in fact be optimal given their uncertainty about the correct integration rule. Statistically, if one is unsure of the correct rule, the best strategy to get close to the true answer is to average estimates from candidate rules. Our third aim is to test whether people are in fact behaving optimally in the face of uncertainty about integration rules.