How is Cramer's V computed from a contingency table's chi-square statistic, and what does its scale mean?

## The problem V solves The chi-square statistic on a contingency table measures how far the observed cell counts sit from what proportionally identical rows would give. It is useful but unbounded and it scales with the sample: double every count in a table and the chi-square statistic doubles, even though the pattern of association is unchanged. That makes it useless for answering how strong, and useless for comparing two tables. Cramer's V is the standardisation that fixes both problems. ## The formula For a table with r rows, c columns and n total observations: `V = sqrt( chi2 / (n * (min(r, c) - 1)) )` Three pieces matter. Dividing by n removes the sample-size scaling, so V stays put when you duplicate the dataset. Dividing by min(r, c) - 1 removes the table-shape scaling: the largest chi-square value a table of that shape can attain is n x (min(r, c) - 1), reached when each level of the smaller dimension pins down the other variable exactly. Taking the square root brings the quantity back from a squared scale to something comparable to a correlation magnitude. Worked example: a 2x3 table of device type against subscription plan on 500 respondents with a chi-square statistic of 20. Here min(2, 3) - 1 = 1, so V = sqrt(20 / (500 x 1)) = sqrt(0.04) = 0.20. If the same pattern held on 5,000 respondents the chi-square statistic would be 200 and V would still be 0.20, which is the whole point. ## Phi as the 2x2 special case On a 2x2 table min(r, c) - 1 = 1, so V = sqrt(chi2 / n), which is the phi coefficient. Phi has a second, equivalent definition: code each binary variable 0/1 and compute Pearson's correlation between the two columns. That version carries a sign, which is meaningful for binary flags where the coding has a natural direction, for example email-opened against purchased: a positive phi says opening and purchasing co-occur more than proportionality would give. Cramer's V is the absolute value, generalised to tables bigger than 2x2 where a sign would be meaningless because there is no ordering of nominal levels to be positive or negative about. ## Reading the scale V = 0 means every row has the same conditional distribution as every other, that is, no association in this sample. V = 1 means the smaller dimension is perfectly determined: knowing the row tells you the column with certainty (or vice versa, whichever dimension is smaller). In between, V behaves like a correlation magnitude but with no direction, and its interpretation is not shape-free: conventional small/medium/large cutoffs are usually quoted separately for each table shape, so 0.2 on a 2x2 is not automatically the same evidence as 0.2 on a 6x6. ## What V does not tell you It does not say where the association lives. A 4x5 table with V = 0.3 might be driven by one unusual row, with the other three behaving like the marginal; only inspecting the conditional row proportions reveals that. It does not tell you direction, because nominal levels have no order. It says nothing about causation: a strong association between country and currency is definitional, not causal. And it is not invariant to how you build the table: merging two rare categories into an other bucket changes both the chi-square statistic and min(r, c), so it changes V, which means the category scheme is an analysis choice you must disclose. ## Small-sample inflation V is biased upward. Even with two completely unrelated columns the observed counts wobble, producing a positive chi-square statistic and therefore a positive V. The inflation grows with table size and shrinks with n, so a large sparse table on a small sample can report a respectable-looking V from noise alone. Sanity-check any V against the table's shape and sample size before treating it as a finding, and prefer a bias-corrected variant when comparing tables of different shapes.