Math AI HL: Chi-Squared Test of Independence

The chi-squared test for independence is a powerful statistical tool for uncovering relationships between categorical variables. Beyond just confirming if a pattern exists, it provides a structured, quantitative method to move from a hunch about association—like wondering if musical taste is linked to age—to a defensible, evidence-based conclusion. Mastering this test is essential for the IB Math AI HL curriculum because it forms the backbone of analyzing real-world data where outcomes are counted, not measured.

Contingency Tables: Organizing Observed Data

Every chi-squared test begins with organizing raw counts into a contingency table (also called a two-way table). This table displays the frequency distribution of two categorical variables, showing how the categories of one variable are distributed across the categories of the other. The variables are often referred to as the row variable and the column variable.

For example, imagine a study investigating the relationship between preferred music genre (Rock, Pop, Classical) and school year (Year 12, Year 13). The contingency table would have three rows for the genres and two columns for the years. Each cell contains the observed frequency ($O$), which is the actual count of individuals falling into that specific combination of categories. The margins of the table (the "Total" row and column) contain the row totals and column totals, which are crucial for the next step. Constructing this table correctly is the first and most critical step in the analysis.

Calculating Expected Frequencies Under Independence

The core logic of the chi-squared test compares what we actually observed ($O$) with what we would expect to observe if there were no association between the two variables—that is, if they were independent. The expected frequency ($E$) for a given cell is calculated based on the assumption of independence, using the row and column totals.

The formula for the expected frequency of the cell in row $i$ and column $j$ is: $$E_{ij}=\frac{({\text{Row Total}}_i)\times ({\text{Column Total}}_j)}{\text{Grand Total}}$$

This calculation is performed for every cell in the contingency table. Intuitively, it distributes the total counts proportionally. If 40% of all respondents are in Year 13, then, assuming independence, we would expect roughly 40% of Rock fans, 40% of Pop fans, and 40% of Classical fans to also be in Year 13. A large discrepancy between many of the observed ($O$) and expected ($E$) values suggests the assumption of independence may be false.

Formulating Hypotheses and the Chi-Squared Test Statistic

Before any calculation, you must formally state the hypotheses. The null hypothesis ($H_0$) always states that the two variables are independent (i.e., no association). The alternative hypothesis ($H_1$) states that the two variables are not independent (i.e., there is an association).

• $H_0$: [Variable A] is independent of [Variable B].
• $H_1$: [Variable A] is not independent of [Variable B].

To test $H_0$, we compute the chi-squared test statistic (${\chi }^2$). This single number quantifies the overall discrepancy between the observed and expected frequencies across the entire table. The formula is: $${\chi }^2=\sum \frac{(O-E{)}^2}{E}$$ where the summation $\sum$ is over all cells in the contingency table.

You calculate $\frac{(O-E{)}^2}{E}$ for each cell and then sum these values. The squaring of $(O-E)$ ensures that positive and negative differences do not cancel each other out, and division by $E$ standardizes the differences (a difference of 5 is more meaningful when $E=10$ than when $E=1000$). A ${\chi }^2$ value of zero would mean the observed data perfectly match the expected data under independence. Larger values indicate greater evidence against the null hypothesis.

Degrees of Freedom, Critical Values, and Interpretation

The calculated ${\chi }^2$ statistic must be evaluated to determine if it is large enough to be considered statistically significant. This evaluation depends on the degrees of freedom (df) for the test. For a contingency table with $r$ rows and $c$ columns, the degrees of freedom are calculated as: $$\text{df}=(r-1)(c-1)$$ Degrees of freedom essentially reflect the number of cells in the table that are free to vary once the row and column totals are fixed.

You then compare your calculated ${\chi }^2$ statistic to a critical value from the chi-squared distribution table (using the appropriate df and significance level, typically $\alpha =0.05$). Alternatively, and more commonly with technology, you find the p-value associated with your ${\chi }^2$ and df.

• If ${\chi }_{\text{calculated}}^2>{\chi }_{\text{critical}}^2$ or if p-value < $\alpha$, you reject $H_0$. You conclude there is sufficient statistical evidence at the $\alpha$ level to suggest an association between the variables.
• If ${\chi }_{\text{calculated}}^2\le {\chi }_{\text{critical}}^2$ or if p-value ≥ $\alpha$, you do not reject $H_0$. You conclude there is insufficient evidence to suggest an association.

The final, crucial step is interpretation in context. Never just state "reject $H_0$." Instead, say: "The test provides significant evidence (p < 0.05) that there is an association between preferred music genre and school year." You should then refer back to the contingency table to describe the nature of the association (e.g., "Year 13 students showed a higher than expected preference for Rock music").

Conditions for a Valid Chi-Squared Test

The chi-squared test relies on approximations that are only valid under certain conditions. Violating these conditions compromises the test's reliability.

1. Data are Counts: The data in the contingency table must be frequencies or counts of individuals, not percentages, proportions, or measurements.
2. Independence of Observations: Each individual or case counted contributes to only one cell in the table. This is usually guaranteed by the study design (e.g., each student is surveyed only once).
3. Sample Size: A common rule of thumb is that all expected frequencies ($E$) should be at least 5. Some stricter guidelines require that at least 80% of cells have $E\ge 5$ and no cell has $E<1$. If this condition is not met, the test results may not be valid. For a 2x2 table, some statisticians use Yates' continuity correction to improve the approximation, though this is often handled by technology.

Common Pitfalls

• Misinterpreting Failure to Reject $H_0$: Concluding "the variables are independent" is too strong. You can only state there is "insufficient evidence of an association." The variables might still be related in a way your sample did not detect.
• Ignoring the Conditions for Use: Applying the test to percentages or when expected counts are too small invalidates the results. Always check that your data are counts and that $E\ge 5$ for all cells before proceeding.
• Confusing Test Statistics: The chi-squared test for independence (two variables) and the chi-squared goodness of fit test (one variable against a distribution) use the same ${\chi }^2$ formula but have different hypotheses and degrees of freedom calculations. Ensure you know which test you are performing.
• Overlooking Context in Interpretation: A statistically significant result tells you an association exists, but not its strength or practical importance. A very large sample can detect a trivial association as "significant." Always combine the statistical conclusion with a descriptive analysis of the table.

Summary

• The chi-squared test for independence analyzes the relationship between two categorical variables organized in a contingency table.
• It compares observed frequencies ($O$) with expected frequencies ($E$) calculated under the null hypothesis of independence, using the formula $E_{ij}=\frac{({\text{Row Total}}_i)\times ({\text{Column Total}}_j)}{\text{Grand Total}}$.
• The chi-squared test statistic ${\chi }^2=\sum \frac{(O-E{)}^2}{E}$ summarizes the total discrepancy. The degrees of freedom are $\text{df}=(r-1)(c-1)$.
• Compare the ${\chi }^2$ statistic to a critical value or, preferably, use the p-value to decide whether to reject $H_0$ and conclude there is evidence of an association.
• The test is valid only when data are counts, observations are independent, and all expected frequencies are at least 5.