How do people revise their beliefs once an item of evidence is discredited? Do they simply return to the belief states they had prior to acquiring that false information, or does the change propagate more widely through their belief network? For example, when the testimony of a key witness is shown to be fabricated, how does this affect jurors’ beliefs about the testimony of other witnesses, or even other forensic evidence?
Previous research in jury decision-making has looked at whether people can exclude inadmissible evidence (Kassin & Sommers, 1997), but has not addressed the more general question of how information that discredits one item of evidence affects other items of evidence. This is especially important given that there is no unambiguous normative answer - whether or not one should generalize the discredit of one item of evidence to other items depends on their causal relations. If both pieces of evidence come from the same source
(e.g., two statements from the same witness), then it seems appropriate that the discredit of one should lead to the discredit of the other. However, if they come from unrelated sources (e.g., a witness statement and a footprint match) then the discredit of one should not affect the status of the other. How do people adjust their beliefs in these situations, and what are the cognitive mechanisms that they use?
Scenario: House burglary, local suspect arrested
HYPOTHESIS: Suspect is guilty
EVIDENCE 1
Neighbour 1 says suspect often loiters in area
Does the discredit of item 2 affect the assessment of item 1?
?
EVIDENCE 2
Neighbour 2 says suspect was outside house on night of crime
Neighbour 2 is lying because he dislikes suspect
In exploring these questions two prominent models of jury belief updating are tested:
Belief adjustment model (Hogarth & Einhorn, 1992): online judgments formed by adjusting from a prior anchor; do not take account of causal relations between items of evidence.
Story model (Pennington & Hastie, 1992): holistic judgments based on causal relations between evidence.
Mock jurors were given simplified criminal cases, and asked to judge the probability that a suspect was guilty on the basis of several items of evidence. In
Experiment 1 two items of incriminating evidence were presented sequentially, followed by information that discredited the second item of evidence. The degree of relatedness between the items of evidence was manipulated: evidence either came from the same source (e.g., two statements from same neighbour), similar sources (e.g., two statements from different neighbours) or different sources (e.g., one statement and one blood test). Experiment 2 manipulated the ordering of the evidence: in the late condition the discrediting information was presented last (as in Experiment 1), in the early condition it was presented after the first item of evidence. We also removed the ‘similar’ condition, so there were just two levels of relatedness.
In each problem participants made four judgments of probability of guilt:
1. On the basis of the background story
2. Given evidence 1 (e.g., neighbour 1 says suspect often loiters in area)
3. Given evidence 2 (e.g., neighbour 2 says suspect was outside house)
4. Given discredit of evidence 2 (e.g., neighbour 2 is lying because he dislikes suspect)
NB In Experiment 2 the order of 3 and 4 was switched, and the discredit was of evidence 1
Experiment 1
In all three conditions the discrediting of evidence 2 affected the evaluation of evidence 1. Mean probability judgments at stage 4 were significantly lower than at stage 2:
Same: t (23) = 3.74, p <.05
Similar: t (23) = 2.71, p <.05
Different: t (23) = 3.13, p <.05
30
20
10
0
60
50
40
100
90
80
70
1 2
Evidence
3 4 same similar different
100
90
80
70
60
50
40
30
20
10
0
In Experiment 1 the discrediting information was over-generalized regardless of the degree of relatedness of the evidence. For example, when participants discovered that one witness was lying, they discounted the testimony of another witness, but they also discounted an unrelated piece of forensic evidence.
Experiment 2
LATE EARLY same different
100
90
80
70
60
50
40
30
20
10
0 same different
1 2
Evidence
3 4 1 2
Evidence
3 4
LATE ORDER
There was no difference between Same and Different conditions: in both cases mean judgments at stage 4 were lower than at stage 2 ( Same , t (41) = 3.41, p <.01; Different , t (41) = 2.89, p <.01), showing that participants over-generalized the discredit of one item to another, irrespective of relatedness. This replicated results from Experiment 1.
EARLY ORDER
There was a marked difference between Same and Different conditions. In the Same condition, mean judgments at stage 4 were lower than at stage 2, t (41) = 4.11, p <.001), whereas in the Different condition there was no difference, t (41) = 1.02, ns ).
This shows that participants in the early condition were sensitive to the relatedness between items of evidence, and generalized appropriately. For example, when participants discovered that one witness was lying, they discounted further testimony from that same witness, but did not discount an unrelated piece of forensic evidence.
In Experiment 2 the order in which evidence was presented affected the impact of the discredited evidence, and thus the final judgments of guilt. In the late condition discrediting information was over-generalized to unrelated items, as in Experiment 1.
In contrast, in the early condition discrediting information was only generalized to related items.
People made reasonable revisions of judgment when the discrediting information was presented early, but unreasonable revisions when it was presented late. This pattern of results is difficult to explain on existent models of belief revision. The Belief adjustment model can account for the late condition, but not the early one; the Story model can account for the early condition, but not the late one. Instead we propose a limited-capacity Bayesian network model inspired by recent formal work in forensic statistics (Beidermann & Taroni, 2006). This combines qualitative Bayesian reasoning (Pearl, 1988) with integrative encoding of evidence (Schul & Mayo, 1999), and can explain both the normative inferences shown in the early condition and the sub-normative inferences in the late condition. It is naturally extended to other contexts in which evidence is acquired and updated (e.g., medical and social reasoning), and coheres with recent theoretical work in cognitive science (Chater,
Tenenbaum & Yuille, 2006). Finally, these findings should have practical implications for how evidence is presented to decision makers in the court room.
Biedermann, A., & Taroni, F. (2006). Bayesian networks and probabilistic reasoning about scientific evidence when there is a lack of data. Forensic Science International, 157, 163-7.
Chater, N., Tenenbaum, J. B., &Yuille, A. (2006). Special issue: Probabilistic models of cognition.
Trends in Cognitive Science , 10, 7, 287-344.
Hogarth, R. M., & Einhorn,H.J. (1992). Order effects in belief updating: The belief adjustment model.
Cognitive Psychology , 24,1-55.
Kassin, S. M., & Sommers, S. R. (1997). Inadmissible testimony, instructions to disregard, and the jury: substantive versus procedural considerations. Journal of Personality and Social
Psychology , 23, 1046-1054.
Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference.
Morgan Kaufman Publishers.
Pennington, N., & Hastie, R. (1992). Explaining the evidence: test of the story model for juror decision making. Journal of Personality and Social Psychology , 62, 189-206.
Schul, Y., & Mayo, R. (1999). Two sources are better than one: the effects of ignoring one message on using a different message from the same source. Journal of Experimental Social
Psychology , 35, 327-345.