Chapter 15: The Chi-Square Statistic: Tests for Goodness of Fit and Independence

Chapter 15: The Chi-Square Statistic: Tests for Goodness of Fit and Independence

Chapter Overview

  • This chapter covers the chi-square statistic and its applications in analyzing data through tests for goodness of fit and tests of independence.

Learning Outcomes

  • Explain when a chi-square test is appropriate.
  • Test hypotheses about the shape of a distribution using the chi-square goodness of fit.
  • Test hypotheses about the relationship between variables using the chi-square test of independence.
  • Evaluate the effect size using the phi coefficient or Cramér’s V.

Tools Required

  • Proportions (for math review, reference Appendix A).
  • Frequency distributions (refer to Chapter 2).

15-1: Introduction to Chi-Square: The Test for Goodness of Fit

  • Statistical tests previously discussed aim to test hypotheses regarding population parameters; these tests are categorized as parametric tests.
Characteristics of Parametric Tests
  • Share several fundamental assumptions:
    • Normal distribution in the population.
    • Homogeneity of variance across the population.
    • Requirement of a numerical score for each individual.
    • Data must stem from an interval or ratio scale.
Nonparametric Tests
  • If research data does not meet the requirements for parametric tests, nonparametric tests serve as an alternative:
    • Do not specify hypotheses based on a specific population parameter.
    • Fewer assumptions about population distribution (termed as "distribution-free" tests).
    • Chi-square tests are two types of nonparametric tests.
Classification and Measurement Scale
  • Nonparametric tests typically classify participants into categories using nominal or ordinal scales.
  • Data for nonparametric tests may include frequency counts (e.g., number of individuals from different political affiliations).
  • Nonparametric methods are useful when interval or ratio measures cannot be achieved.
  • Situations arise where it is impossible to obtain precise scores; categorical classification is then employed.
Selection of Statistical Procedure
  • Choice of statistical analyses rests primarily on the measurement level:
    • Chi-square tests, t-tests, or ANOVA may be selected for hypothesis testing concerning relationships.
    • Evaluation of relationship strength may also involve the use of $r^2$.

The Chi-Square Test for Goodness of Fit

  • The chi-square test evaluates research questions regarding proportions or relative frequencies within a distribution.
  • It utilizes sample data to test hypotheses regarding the shape or proportions of the population’s distribution.
  • The test assesses how well the sample data proportions align with the expected proportions postulated for the population.
Null Hypothesis for the Goodness-of-Fit Test
  • The null hypothesis (H₀) defines the proportion or percentage distribution of the population across each category.
  • Main rationale for the null hypothesis includes:
    • No preference among categories (implying equal proportions).
    • No observed difference in the specific population relative to another known population’s proportions.
  • The alternative hypothesis (H₁) posits that the population distribution deviates from the specified proportions in H₀, often concisely presented as "Not H₀".
Data Requirements for Goodness-of-Fit Test
  • Sample mean or SS calculation is not necessary:
    • Individuals classified based on categories (e.g., grades, frequency of exercise).
    • Each measurement category has documented observed frequencies.
    • Observations must be mutually exclusive (an individual belongs to one category).
Expected Frequencies
  • The goodness-of-fit test contrasts observed frequencies against expected frequencies based on the null hypothesis.
  • Expected frequencies are computed to align with the null hypothesis and are dependent on sample size (n).
  • Represent an idealized prediction of sample distribution.
The Chi-Square Statistic
  • Denoted by $\chi^2$, it represents the chi-square statistic:
    • $f_o$ is the set of observed frequencies.
    • $f_e$ is the set of expected frequencies.
Chi-Square Value Implications
  • The chi-square value signifies the discrepancy magnitude between observed data and expected frequencies:
    • A smaller chi-square value suggests a closer alignment to the null hypothesis, indicating a good fit.

Chi-Square Distribution and Degrees of Freedom

Null Hypothesis Discrimination Criteria
  • The null hypothesis should be upheld when there is a small discrepancy between observed and expected values.
  • Rejection occurs with substantial discrepancies.
Characteristics of Chi-Square Distribution
  • The chi-square distribution encompasses values for all possible random samples when H₀ is accurate.
  • All chi-square values are $
    geq 0$, indicating that when H₀ holds, chi-square values should remain low.
  • The distribution exhibits positive skewness and is a family of distributions influenced by degrees of freedom (df).
Degrees of Freedom for Goodness-of-Fit
  • Formula for degrees of freedom:
    • df=C−1df = C - 1
    • Where C represents the number of categories.
    • df is unaffected by sample size (n).

Locating the Critical Region for a Chi-Square Test

  • Steps to find the critical region include:
    • Setting significance level (alpha).
    • Using a chi-square distribution table listed in Appendix B, identifying critical chi-square values based on:
    • Degrees of freedom (df) in the first column.
    • Significance level (alpha) in the top row.
Critical Value Representation
  • The critical values of chi-square are found within the table body, facilitating decision making regarding the null hypothesis.
Reporting Results of Chi-Square
  • The results should detail significant differences among categories and incorporate summary elements:
    • Presenting results as $\chi^2$ with degrees of freedom, sample size (n), and the test statistic.
    • Example: $\chi^2(3, n = 50) = 9.36, p < .05$.
    • Include observed frequencies for each category when applicable.

Goodness of Fit versus Single-Sample t Test

  • Comparison of nonparametric chi-square tests against parametric t tests:
    • The chi-square test for goodness of fit operates independently of population distribution assumptions.
    • Conversely, the t test presumes a normal population, necessitating numerical scores, and evaluates hypotheses concerning population means.
Test Similarities
  • Both tests use a single sample to infer conditions regarding the single population.
  • The measurement level dictates the choice of testing method:
    • Numerical scores (interval/ratio) imply usage of t test.
    • Non-numerical classifications (ordinal or nominal) necessitate use of chi-square tests based on proportions or percentages.

Learning Check 1

  • Question: Expected frequencies in a chi-square test…
    • are always whole numbers.
    • can contain fractions or decimal values.
    • can contain both positive and negative values.
    • can contain fractions and negative numbers.

Learning Check 1 – Answer

  • Correct Option: can contain fractions or decimal values.

Learning Check 2

  • Determine True or False:
    • In a chi-square test, the observed frequencies are always whole numbers. (T/F)
    • A large value for chi-square will tend to retain the null hypothesis. (T/F)

Learning Check 2 – Answers

  • True: Observed frequencies are counts, hence no fractions.
  • False: Large chi-square values suggest considerable disparity from null hypothesis predictions.

15-3: The Chi-Square Test for Independence

  • Chi-square can also be employed to ascertain auxiliary relationships between two variables:
    • Each participant is measured against both variables, categorized into a matrix.
    • This design may be derived from either experimental or non-experimental frameworks.
  • Frequency data from the sample tests the evidence for the relationship between two variables across the population using a two-dimensional frequency distribution matrix.
Null Hypothesis for Test of Independence
  • States that the two variables are independent, indicating that no inherent relationship exists.
  • Two interpretations of this hypothesis:
    • Single sample where each individual measures two variables, indicating no relationship.
    • Multiple separate samples imply no differing distribution across variable proportions amongst the populations.
Observed and Expected Frequencies
  • In the sample population distribution, observed frequencies will be counted.
  • Expected frequencies derive from the null hypothesis and are computed by the distribution proportions governed by both variables’ rows and columns.
Computing Expected Frequencies
  • Utilizes a straightforward method for any cell in the frequency distribution matrix:
    • f<em>e=n⋅f</em>cfrf<em>e = \frac{n \cdot f</em>{c}}{f_{r}}
    • Where:
    • $f_c$ = column frequency total.
    • $f_r$ = row frequency total.
    • $n$ = total number of individuals in the sample.
Chi-Square Statistic and Degrees of Freedom for Independence
  • The chi-square statistic calculation remains consistent with the goodness-of-fit test:
  • Degrees of freedom for the test of independence determined by:
    • df=(R−1)(C−1)df = (R - 1)(C - 1)
    • Where R is the count of rows and C the number of columns.
Summary Steps for Chi-Square Test for Independence
  • Standard procedure encompasses four steps:
    1. State hypotheses and establish alpha level.
    2. Identify the critical region.
    3. Compute the test statistic.
    4. Reach a conclusion based on results.

15-4: Effect Size and Assumptions for Chi-Square Tests

  • A significant chi-square test outcome indicates a non-chance occurrence.
  • Significance tests hinge on both treatment effect magnitude and sample size.
  • Each hypothesis test result should be accompanied by an appropriate effect size measure.
Cohen’s w
  • Applicable for both chi-square tests regarding effect size:
    • The Pₒ values signify the observed proportions.
  • As per Cohen, effect size standards are as follows:
    • 0.10 indicates a small effect.
    • 0.30 indicates a medium effect.
    • 0.50 indicates a large effect.
  • Notably, Cohen’s w is independent of sample size; only sample proportions relative to the null hypothesis signal the computation of w.
  • w and chi-square possess algebraic relations.
Phi-Coefficient and Cramér’s V
  • The phi coefficient (Φ) assesses strength for 2 × 2 matrices:
    • Φ2\Phi^2 accounts for the proportion of variance.
  • For larger matrices, Cramér’s V, an adjustment to the phi coefficient, measures effect size:
    • df∗df^* is derived from the lower of (R - 1) or (C - 1).
Standards for Interpreting Cramér’s V
  • Interpretive Standards for Cramér’s V:
    • Small Effect | Medium Effect | Large Effect
    • For df* = 1: 0.10 | 0.30 | 0.50
    • For df* = 2: 0.07 | 0.21 | 0.35
    • For df* = 3: 0.06 | 0.17 | 0.29
Assumptions and Restrictions for Chi-Square Tests
  • Specific conditions must be adhered to in order to apply a chi-square test for goodness of fit or independence:
    • Independence of Observations: Observed frequencies must derive from different individuals.
    • Expected Frequencies' Size: Performing chi-square tests is inadvisable if any cell's expected frequency is below 5.

Learning Check 3

  • Question: A basic assumption for chi-square hypothesis test is .
    • The population distributions must be normal.
    • The scores necessitate an interval or ratio scale.
    • The observations must remain independent.
    • None of the above options are assumptions for chi-square.

Learning Check 3 – Answer

  • Correct Option: The observations must remain independent.

Learning Check 4

  • Determine True or False Statements:
    • The value of df for a chi-square test does not rely on sample size (n). (T/F)
    • A positive chi-square statistic indicates a positive correlation between the two variables. (T/F)

Learning Check 4 – Answers

  • True: The df value remains contingent solely on the count of rows and columns in the observation matrix.
  • False: Chi-square cannot yield negative numbers and thus cannot accurately reflect correlation type between variables.

Example of SPSS Output for Chi-Square Test for Independence

  • Crosstabulation Example:
    • VAR00002 * VAR00003 counts presented in a 2x2 matrix format, detailing observed values across categories.
  • Chi-Square Test Values:
    • Results indicate statistical significance in relation to null hypothesis with provided p-values and the expectation that at least 5 counts appear in every cell.

Conclusion

  • The chapter offers extensive insights into chi-square tests used within statistical analysis frameworks and critically engages students to understand application logistics including hypotheses, calculations, and interpretative reporting.