Abuse of statistical packages: the case of the general linear model

Since their introduction in the 1960s, packaged computer programs have freed researchers from much drudgery and painstaking computational labor. Before the distribution of these packages, many meaningful projects could not be attempted, simply because the data analysis phase would have been too labor intensive. However, the very ease of use of these packages can lead to their abuse. In particular, performing all possible t tests among numerous groups, calculating many correlations and circling the ones with P values less than 0.05, and performing stepwise regression all can easily lead to spurious statistical conclusions. The danger of these three practices is illustrated with three simple computer runs using random data. In each run statistically significant results were found for the random data.