However, a confounding factor is that for many of our pairs, the two copies may occur in contexts that are practically indistinguishable. Therefore, we bin the pairs by how different their absolute deviations are, and consider whether helpfulness ratios differ at least for pairs with very different deviations. More formally, for i, j ∈ {0, 0.5, · · · , 3.5} where i < j, we write ij (conversely, i≺j) when the helpfulness ratio of reviews with absolute deviation i is significantly larger (conversely, smaller) than that for reviews with absolute deviation j. Here, significantly larger or smaller means that the Mantel-Haenszel test for whether the helpfulness odds ratio is equal to 1 returns a 95% confidence interval that does not contain 1. The Mantel-Haenszel test [2] measures the strength of association between two groups, giving more weight to groups with more data. (Experiments with an alternate empirical sampling test were consistent.) We disallow j = 4 since there are only relevant 24 pairs which would have to be distributed among 8 (i, j) bins.