By the Wavefield Research team · Published Sep 15, 2026
Cross tabulation splits survey results by the groups you care about: a crosstab shows every answer as counts and column percentages for men, women, age bands, or customer segments, side by side. Reading one is easy. Reading one correctly — with letters, bases, and weights — is what this guide covers, worked example included.
The parts the existing guides skip are the parts that matter in a real cross tabulation analysis: what the significance letters actually test, why chi-square is the wrong default for survey work, what weighting does to your bases, and the honest line where Excel stops being enough.
| Total | Men (A) | Women (B) | |
|---|---|---|---|
| Yes | 52% | 61% B | 44% |
| No | 33% | 27% | 39% A |
| Not sure | 15% | 12% | 17% |
| Base (unweighted) | 400 | 190 | 210 |
| Effective base | 362 | 171 | 188 |
Row by row. Yes, 52% overall — but the Total column is the least interesting number on the table; it is what the topline already said. The crosstab exists for the split: 61% of men versus 44% of women. The “B” beside 61% is the finding — it says the difference cleared a statistical test, not just an eyeball. No, 39% among women carries an “A” for the same reason in the other direction. And Not sure shows why the letters earn their keep: women's 17% is higher than men's 12%, but no letter appears — on these bases, that gap is indistinguishable from sampling noise, and an honest table says so by staying silent.
The two base rows are not boilerplate. The unweighted base is how many humans answered; the effective base is what the sample is worth for inference after weighting — smaller, always, and it is the number the significance tests actually run on. A crosstab without base rows is a chart, not evidence.
Every banner column gets a letter. A letter in a cell means: this cell's percentage is significantly higher than the same row of the lettered column, under a pairwise two-proportion z-test on the two columns' effective bases. That sentence is the entire mechanic, and almost no guide states it. Three conventions follow from it. Letters only run within a banner group — Men versus Women, not Men versus 18–34, because cross-group comparisons compare overlapping people. The Total column is never tested — every subgroup is part of Total, so the test would compare a group to a blend containing itself. And columns under about 30 effective base are marked and excluded from testing: below that floor the test has no power, and a letter would be a costume.
The confidence level is a dial, not a law of nature. Banner books conventionally test at 95%; exploratory work sometimes runs 80 or 90 to surface leads worth a follow-up study, and high-stakes work runs 99. What matters is stating the level on the table and holding it constant — a report that quietly mixes levels is shopping for letters. Wavefield's crosstabs run the full range from 80 through 99 with the level printed on every table.
And the chi-square question, because every statistics textbook reaches for it here: chi-square tests whether two variables are related at all — one p-value for the whole contingency table. It answers “is gender related to switching?” but not “which gender, on which answer, by how much?” — and the second question is the one a marketing or campaign decision needs. That is why survey research settled on pairwise letters instead: same table, sharper question.
Real survey samples get weighted — raked to census targets so the percentages generalize. The reporting convention for a weighted crosstab: percentages weighted, counts unweighted. The percentage is the estimate, so it takes the correction; the count is an audit trail of real respondents, so it does not. A table showing weighted counts is quietly inventing people.
The deeper effect is on the tests. Weighting always costs precision: unequal weights inflate the variance, and the Kish effective base — the sample size your weighted data is actually worth — is what significance must be computed on. In the worked table above, 400 respondents carry an effective base of 362; a heavily weighted sample can lose a third or more. Run the letters on raw bases and a weighted study will flag differences it has no right to flag. This single mechanic — tests on effective bases, not headcounts — is the most common defect in homegrown crosstab setups, and it is invisible until a client's statistician checks. The raking mechanics behind the weights are covered in survey weighting.
Excel gets you further than the software vendors admit. A pivot table with the question in rows, the segment in columns, and values as a percentage of column total is a legitimate crosstab, free, on your desk in two minutes. For an unweighted, one-off, two-variable question — is satisfaction different by region, roughly? — it is the right tool, and a free online crosstab generator does the same job without the pivot-table detour.
The line where it stops: significance, weights, and scale. Pivot tables test nothing, weight nothing, and produce one table at a time. A study with thirty questions, raked weights, and a client who will ask “is that difference real?” needs crosstab software — the category's job is banner books (every question run against every banner column, letters computed on effective bases, exported as one workbook) rather than tables one at a time. That layer is where the market gets expensive: dedicated crosstab platforms price as enterprise seats. On Wavefield the same layer — weighted crosstabs, letters at a chosen confidence level, Excel banner books, SPSS export — is included in survey analysis software from $99 per study, because it is computed, not billed by the table. The step-by-step from raw responses to a defensible report is in how to analyze survey results.
A crosstab answers the question every topline hides: who. The overall result says 52% would switch providers; the cross tabulation shows that number is 61% among men and 44% among women — a difference that changes what the finding means. Crosstabs are used to compare answers across demographics, customer segments, arms of a test, or waves of a tracker, and to test whether those differences are statistically real rather than sampling noise.
Pivot tables: put the survey question in rows, the grouping variable in columns, show values as a percentage of column total. That genuinely produces a crosstab — counts and column percentages. What Excel does not produce is any of the inferential layer: no significance testing between columns, no support for weighted data on effective bases, and no way to generate a banner book across thirty questions without rebuilding thirty pivots. For a one-off two-variable check, Excel is enough; for a weighted study you plan to defend, it is not.
Each banner column gets a letter — A, B, C. A letter inside a cell means that cell's percentage is statistically significantly higher than the same row in the lettered column, at the confidence level of the test. In the worked example, “61% B” means men's 61% is significantly higher than women's 44%; the absence of a letter on “17%” means women's higher Not-sure share did not clear the significance bar. Letters run within a banner group only, and the Total column is never tested.
Same table, different rooms. Statisticians say contingency table and reach for chi-square tests of independence; market researchers say crosstab or banner table and reach for pairwise column tests with significance letters. The structure — categorical variables crossed, counts in cells — is identical. The reporting convention differs because the questions differ: chi-square asks whether the variables are related at all, while the letters ask which specific groups differ, which is what a decision usually needs.
Related: how to analyze survey results · survey weighting · sample size calculator · survey analysis software
Field the study and the crosstabs arrive weighted, lettered at your chosen confidence level, and exportable as an Excel banner book. From $99 per study.
Start a study