Folklore and many anecdotal stories have relayed how some individuals have claimed to be able to ``foretell'' the future, or have experienced premonitions of events before they actually occurred. While much of this information is likely due to misinterpretation, misrepresentation or other flaws of human perception, memory and reasoning, there are experimental findings which suggest that precognition may occur (see Wiseman and Morris  for an overview of ways we can be deceived, or can deceive ourselves into interpreting a normal incident as being paranormal).
Honorton and Ferrari  conducted a meta-analysis of 309 precognition studies conducted between 1935 and 1987. These studies all used a ``forced-choice'' methodology, in which the subject is aware of the possible target choices, and is asked to choose one of them as his answer (as opposed to ``free-response'' methodologies, such as ganzfeld studies). In all of these studies, the subject made their choice as to the target identity prior to the target identity actually being randomly generated. Thus the subjects' responses were to targets which did not exist at the time of their response. These studies are thought by some to be methodologically superior to other ESP studies as there is little possibility of the subject ``cheating'', or receiving any subtle cues about the target identity, as the target does not exist when their response is made.
The studies included in this meta-analysis were conducted by 62 different senior investigators, and included nearly two million individual trials contributed by over 50,000 subjects. While the mean effect size per trial is small ( es = .02), it is sufficiently consistent for the overall effect from these studies to be highly significant (combined ). Using eight different measures of study quality, no systematic relationship was found between study outcome and study quality. A ``fail-safe N'' estimate would require 14,268 unreported, null studies to reduce the significance of the database to chance levels. Given the wide diversity of study methods and procedures found in this database, it is not surprising that the study outcomes were extremely heterogeneous. The authors eliminated outliers by discarding those studies with z scores falling within the top and bottom 10 percent of the distribution, leaving 248 studies. It should be noted that the elimination of outlier studies to obtain homogeneity is a common practice, and in other, non-parapsychological reviews ``it is sometimes necessary to discard as many as 45% of the studies to achieve a homogeneous effect size distribution'' (p. 1507) . The resulting mean trial effect size was .012, and the combined z still highly significant ( ). While it was found that study quality improved significantly over the 55 year period during which these studies were conducted (correlation coefficient r[246 degrees of freedom] = .282, , study effect sizes did not significantly co-vary with the year of publication. Study effect sizes are homogenous across the 57 investigators contributing to the trimmed database. The rest of the analyses conducted were all performed upon this smaller database.
The authors identified four ``moderating'' variables that appeared to relate systematically to study outcome. The first variable involved the subject population. It was found that studies using subjects who were selected on the basis of good ESP performance in previous experimental sessions obtained significantly better ESP effects than those studies using unselected subjects (a t test with 246 degrees of freedom [df] giving t = 3.16, p = 0.001). Another variable which covaried with study effect size was whether the subjects were tested individually or in groups, with individual testing studies obtaining significantly higher outcomes than those using group testing methods ( ).
A further moderating variable involved the type of feedback subjects received about the accuracy of their responses. There were four feedback categories, including no feedback, delayed feedback (usually via mail), feedback given after a sequence of responses (often after 25 responses), and feedback given after each response. Of the 104 studies which supplied the necessary information, there was a linear and significant correlation between the precognition effect and feedback level ( , p = 0.009), with effect sizes increasing with level of feedback. A related finding involves the time interval between the subject's responses and the target selection. This finding is confounded by the feedback level, as time duration between the response and target generation may co-vary with feedback level (i.e., when feedback was given after every response, the time interval between response and target selection would have to be shorter than was necessarily the case when feedback was given after a sequence of calls, or a month after the responses had been made). There were seven different time interval categories, varying from a millisecond to months. There was found to be a significant decline in precognition effect sizes as the time interval between response and target selection increased ( ). The significant temporal decline/study effect size relationship is due entirely to those studies which used unselected subjects, with the studies that tested selected subjects showing a small, non-significant increase in precognition scoring as the time interval increased (the difference between these groups was not significant).
It should be noted that there was no significant difference in quality between studies using selected and unselected subjects. Also, studies which tested subjects individually did show significantly higher study quality than those utilising group testing procedures ( ). A correlation between feedback level and research quality was positive, but not significant ( .
In summarising the precognition findings, Honorton and Ferrari concluded ``the forced-choice precognition experiments confirm the existence of a small but highly significant precognition effect.'' (p. 300). Furthermore, they concluded that the most important outcome of the meta-analysis was the identification of moderating variables, which not only provides guidelines for future research, but may also help expand our understanding of the phenomena.