Sample statistics for data upon which a number of published Harker diagrams are based are reviewed in detail. A convincing Harker diagram—i.e. one in which most of the data points lie reasonably close to simple curves—seems to require that the variance of silica be at least as large as the sum of all other variances. For this situation it can be shown a priori that strong negative correlation is to be expected between silica and some or all other oxides making substantial contributions to the total variance of the array. The data suggest that the larger the relative contribution to total variance made by any other such oxide, the stronger its negative correlation with silica.