Figure 157

The type of display used by Mynatt et al. (1977) to study confirmation bias. Subjects had to direct a particle that was fired from the upper left part of the screen, by selecting the direction of its path. The relative shading of the objects indicates the two levels of brightness at which objects were presented.

additional triples along with their reason for suggesting them. So, a subject might write 6-8-10, giving "numbers ascending by twos" as the rationale, to which the experimenter would answer that "yes, this triple is also an instance of the rule". After the subject has generated a number of such triples (e.g., 20-22-24, 4547-49) and has received the positive feedback that they are all instances of the rule, they can declare what they think the rule to be (e.g., "numbers ascending by twos"). However, this is not the experimenter's rule, and after they are informed of this they must continue generating triples and proposing other rules until they guess correctly. The task allows subjects to generate an infinite variety of hypotheses and tests. Unfortunately, most of these alternatives are not the rule that is required. Wason (1960) found that subjects tended to gather evidence that confirmed their hypothesised rules, rather than generating examples that would falsify their hypotheses (like 33-31-32). He called this tendency a confirmation bias, subjects sought confirmation for hypotheses rather than resorting to falsification. Mahony (1976) also found that scientists fared no better on the task than other groups (indeed, clergymen proved to be better at abandoning their hypotheses).

Confirmation bias appears to be quite prevalent. Mynatt, Doherty, and Tweney (1977) found similar results in a simulation world that was closer to real scientific testing. In this computer world, subjects fired particles at circles and triangles that were presented at two brightness levels (low and high). The world had other features but all of these were irrelevant to the task (see Figure 15.7). Subjects were not told that the lower-brightness shapes had a 4.2 cm invisible boundary around them that deflected particles. At the beginning of the experiment, they were shown arrangements of shapes which suggested the initial hypothesis that "triangles deflect particles". They were then presented with pairs of screens, where one screen contained similar features to those that deflected particles and the other screen contained novel features. Subjects were divided into three groups that were instructed to adopt either a confirmatory strategy, a disconfirmatory strategy, or no particular strategy (i.e., a control). Again, as in the 2-4-6 task, subjects tended to confirm their hypotheses by picking the confirming screen 71% of the time. Furthermore, the strategy instructions did not deflect subjects from this confirmation bias. Mynatt, Doherty, and Tweney (1978) found similar results using an interactive version of this simulation world. They also found that subjects tended to ignore falsificatory evidence when it occurred. More recently, Garavan, Doherty, and Mynatt (1997) have proposed that attempts to falsify in such complex environments are just too difficult for people to do, although they also proposed that a new classification for types of hypothesis tests needed to be developed. More generally, one of the things that should be remembered about falsification and confirmation is that either may be appropriate at different times. Chalmers (1982) has pointed out that established theories should be falsifiable, but that it will often be more beneficial to a scientist to seek confirmatory evidence during the development of a new theory.

The failure of instructions to undo people's confirmation bias led subsequent researchers to try a number of different manipulations to encourage disconfirmation. Tweney et al. (1980) used a manipulation that increased subjects' disconfirmatory responses by using either confirmatory instructions (it was pointed out that given a hypothetical triple 3-3-3 and the rule "three equal numbers", this rule could be tested with triples like 8-8-8 to confirm the hypothesis) or disconfirmatory instructions (given a hypothetical triple 33-3 and the rule "three equal numbers", subjects were told that if triples like 5-7-9 were correct then the rule would be wrong). However, even though subjects changed their strategy, Tweney et al. did not find any improvement in their success on the problem. Gorman and Gorman (1984) observed greater success, along with an increase in the use of disconfirmation on the 2-4-6 task, when they instructed subjects to use a disconfirmatory strategy and did not give subjects feedback as to the correctness of their hypotheses (see Gorman, 1992, for other related studies).

All these studies indicate that confirmation bias is difficult to modify by instruction. However, some researchers have argued that the studies fail to prove that subjects have the intention to confirm their hypotheses (see Evans, 1983; Klayman & Ha, 1987; Poletiek, 1996; Wetherick, 1962). They point out that subjects are led to induce a hypothesis that is a specific version of the experimenter's rule (e.g., "numbers ascending by twos" is more specific than "any ascending sequence"). As the subject's rule is a restricted version of the experimenter's rule, any triple that fits the subject's hypothesis will also fit the experimenter's. Thus, they fail to produce triples that do not fit their rule, but do conform to the experimenter's rule. In other situations, an attempt to test a hypothesis with a triple could result in a falsification (which could be intended by subjects). However, in the 2-4-6 problem falsification can only be achieved by explicitly trying negative tests of the hypothesis.

There is one manipulation that has been surprising successful at improving subjects' performance on the 2-4-6 task; this involved indicating to subjects that there were two distinct rules to be considered. Tweney et al. (1980) told subjects that the experimenter had two rules in mind; one of these rules generated DAX triples and the other generated MED triples. They were also told that 2-4-6 was a DAX triple. The DAX rule was intended to be what has been called the experimenter's rule (i.e., "any ascending sequence"), whereas the MED rule characterised the rule that generated any other triple. Thus, instead of being told that their triples were right or wrong, subjects were told that they were DAX or MED. This manipulation led to the striking result that the majority of subjects generated the correct rule on their first attempt. Furthermore, this success occurred even though subjects continued to make "confirmatory" tests of their hypothesis. In this version of the task subjects succeed because they do not have to disconfirm the DAX hypothesis, rather they can alternatively test the MED hypothesis in a confirmatory fashion. Wharton, Cheng, and Wickens (1993) have ruled out a number of other explanations for the DAX-MED effects, such as the information quantity in this version of the task (e.g., the greater number of tests performed before the rule is announced)

and the possible influence of positive labels. Their results supported the proposal that the effect follows from being given two complementary goals to test involving the DAX and MED hypotheses. However, Vallée-Tourangeau, Austin, and Rankin (1995) have shown that these results are not specifically due to the complementarity of the hypotheses but rather arise from the breadth of hypotheses that subjects generate based on the specific testing strategy they adopt.

Problem-space accounts of scientific discovery

Traditionally, for the most part, hypothesis testing and discovery have not been formulated in problemspace terms (see Chapter 14). Kulkarni and Simon (1988, 1990) adopted a historical perspective by simulating Hans Kreb's discovery of the urea cycle in biochemistry in a system called KEKDA. One of the key phenomena modelled in this system is how surprising new results can lead to new hypotheses and theories. Klahr and Dunbar (1988; Klahr, Fay, & Dunbar, 1993) have also looked at discovery in different task situations. In one task, they asked subjects to discover the function of a mystery button (labelled "RPT") for controlling a toy vehicle called Big Trak. Subjects tested the function of the button by including it in brief sets of instructions they had to write to make the toy move. Subjects, therefore, propose hypotheses about the function and then experimentally test them by seeing whether they work. Initially, subjects adopt a positive test strategy reasoning that "If Big Trak does X then my hypothesis is correct", although they are often forced to revise their theories by negative evidence.

Using problem-space theory, Klahr and Dunbar characterised scientific discovery as a dual-space search; one space contains the experimental possibilities in the situation and the other contains a space of possible hypotheses. These ideas are very similar to the ideas proposed by Ohlsson (1992) to explain insight problem solving (see Chapter 14). In searching the hypothesis space the initial state is some knowledge of the domain and the goal state is a hypothesis that can account for that knowledge in a more concise, universal form. Hypothesis generation in this space may be the result of a variety of mechanisms (e.g., memory search, analogical mapping, or remindings). Search in the experiment space is directed towards experiments that will discriminate between rival hypotheses and yield interpretable outcomes. On the basis of protocol analysis, Klahr and Dunbar distinguished two groups of subjects, theorists who preferred to search the space of hypotheses and experimenters who preferred to search the space of experiments (see Van Joolingen & DeJong, 1997, for a further elaboration of this approach). Gorman (1992) has suggested that much previous research has concentrated on the experiment space, ignoring the hypothesis space. The latter is important, as is illustrated in the DAX-MED study which shows how subjects' representation of hypothesis goals can be very important to their subsequent success on a task. Finally, it should be pointed out that all these studies were based on individual problem solving, when most scientific discoveries are collaborative efforts between groups of individuals. Unfortunately, the cognitive processes in discovery appear to change when more than one individual is involved (see Dunbar, 1997; Okada & Simon, 1997).

Business Correspondence

Business Correspondence

24 chapters on preparing to write the letter and finding the proper viewpoint how to open the letter, present the proposition convincingly, make an effective close how to acquire a forceful style and inject originality how to adapt selling appeal to different prospects and get orders by letter proved principles and practical schemes illustrated by extracts from 217 actual letter.

Get My Free Ebook


Post a comment