The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory.

the total number of papers may exceed 10,000. Nevertheless, cost consideration forced us to consider mostly published papers and technical reports in English. 4 Formula 4 in Seifert (1991) is in error—a multiplier of n, of cell size, is missing in the numerator. 5 Unfortunately, the technique of meta-analysis cannot be applied, at present time, to such effects because the distribution of dis based on a sampling of people, whereas the statistics of techniques such as ARIMA are based on the distribution of a sampling of observations in the time domain regardless of the size of the people sample involved (i.e., there is no way to compare a sample of 100 points in time with a sample of 100 people). That is, a sample of 100 points in time has the same degrees of freedom if it were based on an observation of 1 person or of 1,000 people. 258 KLUGER AND DENISI From the papers we reviewed, only 131 (5%) met the criteria for inclusion. We were concerned that, given the small percentage of usable papers, our conclusions might not fairly represent the larger body of relevant literature. Therefore, we analyzed all the major reasons to reject a paper from the meta-analysis, even though the decision to exclude a paper came at the first identification of a missing inclusion criterion. This analysis showed the presence of review articles, interventions of natural feedback removal, and papers that merely discuss feedback, which in turn suggests that the included studies represent 1015% of the empirical FI literature. However, this analysis also showed that approximately 37% of the papers we considered manipulated feedback without a control group and that 16% reported confounded treatments, that is, roughly two thirds of the empirical FI literature cannot shed light on the question of FI effects on performance—a fact that requires attention from future FI researchers. Of the usable 131 papers (see references with asterisks), 607 effect sizes were extracted. These effects were based on 12,652 participants and 23,663 observations (reflecting multiple observations per participant). The average sample size per effect was 39 participants. The distribution of the effect sizes is presented in Figure 1. The weighted mean (weighted by sample size) of this distribution is 0.41, suggesting that, on average, FI has a moderate positive effect on performance. However, over 38% of the effects were negative (see Figure 1). The weighted variance of this distribution is 0.97, whereas the estimate of the sampling error variance is only 0.09. A potential problem in meta-analyses is a violation of the assumption of independence. Such a violation occurs either when multiple observations are taken from the same study (Rosenthal, 1984) or when several papers are authored by the same person (Wolf, 1986). In the present investigation, there were 91 effects derived from the laboratory experiments reported by Mikulincer (e.g., 1988a, 1988b). This raises the possibility that the average effect size is biased, because his studies manipulated extreme negative FIs and used similar tasks. In fact, the weighted average d in Mikulincer's studies was —0.39; whereas in the remainder of the