Conditional Probability Tables and Marginal Checks

Reading a Two-Way Probability Table

A two-way probability table organizes the joint distribution of two discrete random variables. Its interior entries describe probabilities of pairs of values, while its row and column totals give the corresponding marginal probabilities. Conditional probabilities come from dividing joint entries by the appropriate marginal total. The central practical question is therefore: which variable is being conditioned on, and which table total belongs to that variable?

Let \(X\) take values \(x_1,\ldots,x_r\) and \(Y\) take values \(y_1,\ldots,y_s\). Write \(p_{ij}=P(X=x_i,Y=y_j)\). In the tables below, rows index values of \(Y\), and columns index values of \(X\). Thus the row total is \(P(Y=y_j)\), and the column total is \(P(X=x_i)\).

The arrangement matters. With \(Y\) in the rows and \(X\) in the columns, conditioning on \(Y=y_j\) means dividing entries in row \(j\) by that row's total. Conditioning on \(X=x_i\) means dividing entries in column \(i\) by that column's total. In either case, the probabilities for the variable being described must sum to \(1\), provided the conditioning event has positive probability.

From Joint Entries to Conditional PMFs

For any pair of values, the conditional probability is the joint probability divided by the probability of the conditioning event, when that event has positive probability. In a two-way table, this gives the following formulas.

The first formula divides each cell in a fixed row by that row's total. The second divides each cell in a fixed column by that column's total. A zero total cannot be used as a denominator: if \(P(Y=y_j)=0\), the conditional pmf of \(X\) given \(Y=y_j\) is not defined by this formula.

Why the Normalization Checks Work

Conditional rows and columns sum to one because their denominators are exactly the sums of the joint entries being divided. This is a useful verification, not just a convenient pattern to memorize.

Proof. Each conditional probability is nonnegative because each joint entry is nonnegative and the denominator is positive. For the row conditional, summing over all values of \(X\) gives

For the column conditional, summing over all values of \(Y\) gives

Nonnegativity and a total of \(1\) establish the pmf property in each case. If a row or column total is zero, the corresponding conditional probabilities are not defined, so that case is excluded. \(\square\)

The same formulas can be rearranged to recover a joint entry from a conditional probability and its marginal probability. This provides a second way to check a table or to build one from conditional information.

Proof. By the definition of conditional probability, when \(P(Y=y_j)>0\),

Multiplying both sides by \(P(Y=y_j)\) gives the first identity. The definition with \(X=x_i\) as the conditioning event gives

and multiplication by \(P(X=x_i)\) gives the second identity. \(\square\)

Zero Marginals and Careful Interpretation

A marginal total of zero means every joint entry in that row or column is zero, since entries are nonnegative. It does not make the conditional probabilities in that row or column equal to zero: those conditionals are undefined, because their denominator is zero. A conditional distribution must be normalized over an event that can occur with positive probability.

Common Mistake: Normalizing in the Wrong Direction

A frequent error is to normalize by the marginal of the variable whose conditional distribution is being reported, rather than the variable after the conditioning bar. For example, \(P(X=x_i\mid Y=y_j)\) requires division by \(P(Y=y_j)\), not by \(P(X=x_i)\). In the first worked example, \(P(X=0\mid Y=a)=0.15/0.40=0.375\). Dividing \(0.15\) by the \(X=0\) column total \(0.50\) instead gives \(P(Y=a\mid X=0)=0.30\), a different conditional probability answering a different question.

A useful check is to say the conditioning statement in words before calculating: “given \(Y=a\)” fixes the row, so use that row total; “given \(X=0\)” fixes the column, so use that column total. Then check that the probabilities for the possible values of the other variable sum to \(1\). This check can catch a wrong denominator or arithmetic error, but a row or column summing to \(1\) alone does not verify that the original joint table is valid. Also check that all joint entries are nonnegative and that the grand total is \(1\).