A pencil and a kneaded eraser resting on graph paper on a library table, with a lamp lighting the near half of the page

Chi-square asks one question: is the gap between the offspring you counted and the offspring your cross predicted bigger than chance would comfortably produce? You compute a single number from the counts, compare it to a table value set by the number of outcome classes, and read off a verdict about the deviation, not about the truth of your hypothesis.

Everything below is one number carried from a set of counts to a sentence in English. Two crosses are worked in full, one with one degree of freedom and one with three, because the table value changes and that is where most of the lost marks live.

What is the chi-square test actually comparing?

Observed counts against expected counts, class by class. The expected counts come from the genetics, not from the data. A self-cross of two heterozygotes predicts a 3:1 phenotypic ratio in the F2 generation, and a self-cross of a plant heterozygous for two traits predicts 9 round-yellow : 3 round-green : 3 wrinkled-yellow : 1 wrinkled-green. OpenStax Concepts of Biology states both, and attaches the condition that gets dropped in class: these are the ratios expected assuming the crosses were performed with a large enough sample size.

That condition is the reason the test exists. No real plot of 160 plants comes out at exactly 120 and 40. The question is never whether you got the ratio. It is whether missing it by this much, with this many offspring, is ordinary.

The statistic on the College Board’s equations and formulas sheet is the sum, across every class, of the squared difference between observed and expected divided by expected. Biology LibreTexts writes the same thing for a genetic cross: take (observed minus expected) squared over expected for each category, then add the categories together. Squaring removes the signs so surpluses and shortfalls cannot cancel, and dividing by the expected count is what makes a shortfall of 8 out of an expected 40 count for more than a surplus of 8 out of an expected 120.

A monohybrid cross, worked from start to finish

Self-cross a heterozygous plant, Aa by Aa, and score 160 offspring. You count 128 with the dominant phenotype and 32 with the recessive one. Is that consistent with the 3:1 the Punnett square predicts?

Step 1. Expected counts, from the ratio and the total. The 3:1 splits 160 into four parts, so each part is 160 / 4 = 40.

  • Dominant: 3 x 40 = 120
  • Recessive: 1 x 40 = 40

Keep these as counts. Converting to percentages here is the single most common way this problem is lost, because the test is sensitive to sample size and a percentage has thrown the sample size away.

Step 2. One term per class.

  • Dominant: (128 - 120)^2 / 120 = 64 / 120 = 0.53
  • Recessive: (32 - 40)^2 / 40 = 64 / 40 = 1.60

Note the asymmetry. Both classes are off by exactly 8, and the small class contributes three times as much to the total.

Step 3. Add the terms. 0.53 + 1.60 = 2.13.

Step 4. Degrees of freedom. The College Board sheet defines them as the number of distinct possible outcomes minus one. There are two phenotype classes here, so the degrees of freedom are 2 - 1 = 1.

Step 5. Read the table. On the chi-square table printed on the AP Biology equations and formulas sheet, the row for p = 0.05 gives 3.84 at one degree of freedom, and the p = 0.01 row gives 6.63.

Step 6. Compare and say it in words. 2.13 is less than 3.84, so you fail to reject the null hypothesis. The 8-offspring gap in each class is the size of gap you would expect to see fairly often from sampling alone in a plot this size, and the data are consistent with a 3:1 ratio.

What does “fail to reject” actually mean?

It means the data gave you no reason to abandon your prediction. It does not mean the prediction is proved.

The null hypothesis in a genetics cross is the statement that there is no real deviation between what you counted and what the cross predicts, and that whatever gap exists is sampling noise. Biology LibreTexts puts it that a statistical test starts from the assumption that there are no significant deviations between the measured results and the predicted ones. A chi-square below the critical value leaves that assumption standing. It does not confirm it, and it never rules out a different model that predicts a similar ratio.

There is a practical consequence students meet on their own data. A small sample makes a test forgiving, because the expected counts are small and the denominators are small along with them. Failing to reject after counting 20 offspring is much weaker evidence than failing to reject after counting 400. If your result comes out comfortable, check how much work your sample size is doing before you treat it as support.

What changes in a dihybrid cross?

Two things: the number of classes, and therefore the table value.

Self-cross a plant heterozygous for both traits, AaBb by AaBb, and score 160 offspring. The dihybrid cross predicts 9:3:3:1, which splits 160 into sixteen parts of 160 / 16 = 10.

  • 9 x 10 = 90 round-yellow
  • 3 x 10 = 30 round-green
  • 3 x 10 = 30 wrinkled-yellow
  • 1 x 10 = 10 wrinkled-green

Suppose you count 110, 18, 21 and 11, which does total 160. Term by term:

  • (110 - 90)^2 / 90 = 400 / 90 = 4.44
  • (18 - 30)^2 / 30 = 144 / 30 = 4.80
  • (21 - 30)^2 / 30 = 81 / 30 = 2.70
  • (11 - 10)^2 / 10 = 1 / 10 = 0.10

Sum: 4.44 + 4.80 + 2.70 + 0.10 = 12.04.

Four phenotype classes means 4 - 1 = 3 degrees of freedom, and the p = 0.05 row of the College Board table gives 7.81 at three degrees of freedom. Some tables print 7.82 for the same entry, including the Bio-OER genetics chapter, which is the same value rounded differently and will not change a verdict at this precision.

12.04 is greater than 7.81, so you reject the null hypothesis. The deviation is larger than chance comfortably explains, and something in the model is wrong. In a dihybrid cross the usual suspect is independent assortment itself: the two parental combinations are in surplus and the two recombinant ones are short, which is the signature of genes sitting close together on the same chromosome rather than assorting independently. Rejecting tells you the prediction failed. Working out why is the biology, and the genotype and phenotype page is where the vocabulary for that sits.

Where are the marks usually lost?

Four places, in order of frequency.

Using the total as the degrees of freedom, or the number of individuals, rather than the classes minus one. Degrees of freedom never depend on how many offspring you counted.

Computing expected values from the observed data instead of from the predicted ratio. The expectation has to come from the hypothesis, or the test is comparing the data to itself.

Dividing by the observed count instead of the expected one. The expected value is always the denominator.

Reporting the statistic and stopping. The number 12.04 is not an answer. The answer is a sentence that names the comparison, the degrees of freedom, the critical value and what it means for the cross. Practising that sentence is part of preparing for the AP Biology free-response questions.

If you have a set of counts in front of you and cannot tell which line went wrong, photograph the problem and Cyllo works the same chain in order, expected counts first, one term per class, degrees of freedom, table value, verdict, so you can find the line where your version diverged and ask about that line specifically. Photo solving and chat need a subscription; the genetics quizzes and flashcards do not. The same test turns up again on the Hardy-Weinberg equation, and there is a set of Hardy-Weinberg practice problems worked line by line to go with it. The AP Biology study guide collects the equations sheet unit by unit.

AP is a trademark registered by the College Board, which is not affiliated with, and does not endorse, this site.

Sources