Why Contingency Tables Multiply Marginal Totals to Find Expected Frequencies

The binding constraint for calculating an expected frequency is the assumption of absolute statistical independence. For the mathematics to work, the probability of an observation falling into a specific row must be completely unaffected by which column it belongs to. If that condition holds, the expected count in any cell of a contingency table is simply the product of its marginal totals divided by the grand total.[1][6]

This calculation forms the engine of the chi-square test of independence, a statistical method developed by Karl Pearson in 1900 and used across disciplines from epidemiology to marketing. Before researchers can determine if a relationship exists between two categorical variables—such as a patient's treatment group and their recovery status—they must first establish exactly what the data would look like if no relationship existed at all.[1][2]

A contingency table, or cross-tabulation, organizes data into a grid. In a standard 2x2 table, there are 4 individual cells. The "marginal totals" are the sums of each row and column, appearing in the margins of the table, while the "grand total" is the overall sample size of the dataset.[3][4]

To find the expected frequency for a specific cell, researchers multiply that cell's row total by its column total, then divide by the grand total. For example, if a clinical trial includes 100 total patients, with 50 receiving a new drug and 40 experiencing a specific side effect, the expected frequency of patients both receiving the drug and experiencing the side effect under the assumption of independence is (50 × 40) / 100, or exactly 20 patients.[4][5]

The mechanics behind this formula rely on the fundamental multiplication rule of probability. When two events are independent, the probability of both occurring simultaneously is the product of their individual probabilities.[2][3]

The probability of an observation landing in a specific row is the row total divided by the grand total. Similarly, the probability of it landing in a specific column is the column total divided by the grand total.[3]

Multiplying these two marginal probabilities yields the joint probability for that specific cell. To convert that theoretical probability back into an expected count, it must be multiplied by the grand total.[1][6]

Mathematically, this is expressed as: (Row Total / Grand Total) × (Column Total / Grand Total) × Grand Total. When simplified, one of the grand totals in the denominator cancels out the grand total in the numerator, leaving the standard formula.[5][6]

This elegant algebraic simplification is why researchers do not need to calculate individual probabilities for every cell. As the National Institute of Standards and Technology documentation notes, "The expected frequency for each cell is computed as the product of the row and column marginal totals divided by the grand total."[2]

Once the expected frequencies are calculated for every cell in the table, they are compared against the actual observed frequencies. The difference between the observed and expected counts forms the basis of the chi-square statistic, which squares the differences and divides by the expected counts to standardize the variance.[1][2]

If the observed counts closely match the expected counts, the deviation is small, suggesting that the variables are indeed independent. If the observed counts diverge significantly from the expected baseline, it indicates a statistical relationship that is unlikely to have occurred by chance.[1][3]

However, this mathematical framework has strict limitations. The calculation assumes that the sample is randomly drawn and that the observations are mutually exclusive, meaning no single subject can be counted in more than one cell.[3][5]

Furthermore, the reliability of the subsequent chi-square test degrades if the expected frequencies are too small. A widely accepted rule of thumb dictates that the expected frequency in any given cell should be at least 5 to ensure the test statistic accurately approximates a chi-square distribution.[1][5]

When expected counts fall below this threshold, researchers must either combine categories to increase the marginal totals or abandon the chi-square test in favor of alternatives like Fisher's exact test, published in 1922, which calculates exact probabilities rather than relying on approximations.[2][3]

The dimensions of the table also dictate the degrees of freedom, calculated as the number of rows minus 1 multiplied by the number of columns minus 1. For a 3x4 table, this yields 6 degrees of freedom, which determines the exact shape of the chi-square distribution used to evaluate whether the deviations from the expected frequencies are statistically significant.[4][5]