MODULE III

PARAMETRIC AND
NON-PARAMETRIC TESTS
Detailed Study Notes

Based on: Statistical Methods for Psychology, 7th Edition — David C. Howell

Topics Covered
PARAMETRIC TESTS

1. Independent Samples t Test

2. Paired t Test (Matched Samples)

3. One-Way ANOVA

4. Two-Way ANOVA (Factorial)

5. Repeated Measures ANOVA

NON-PARAMETRIC TESTS

6. Median (as Nonparametric Measure)

7. Chi-Square Test

8. Sign Test

9. Wilcoxon Rank-Sum Test

10. Kruskal–Wallis H Test

 PART A: PARAMETRIC TESTS
Parametric tests make specific assumptions about the population distribution — primarily normality — and
involve estimation of population parameters. They are generally more powerful than nonparametric tests
when assumptions are met.

1. Independent Samples t Test

1.1 What Is It?

The independent samples t test is used to determine whether the means of two unrelated (independent)
groups are significantly different from each other. It is one of the most widely used statistical procedures in
psychology and the behavioral sciences.

Example: Comparing the mean exam scores of students taught by Method A vs. Method B. Each student
belongs to only one group.

1.2 Assumptions
• Independence: Observations in each group are independent of each other and of the other group.
• Normality: Scores in each population are normally distributed. With n ≥ 30 per group, the Central
Limit Theorem makes this assumption less critical.
• Homogeneity of Variance (Homoscedasticity): The variance in the two populations is equal (σ21 =
σ22). If violated, use Welch's correction.
• Interval/Ratio scale: The dependent variable must be measured on a continuous scale.

1.3 Formula
When population variances are assumed equal, variances are pooled:

Pooled Variance: s2p = [(n1 - 1)s21 + (n2 - 1)s22] / (n1 + n2 - 2)

t = (X■1 - X■2) / √[ s2p(1/n1 + 1/n2) ]

Degrees of freedom: df = n1 + n2 - 2

Where: X■1 and X■2 are the sample means; s21 and s22 are the sample variances; n1 and n2 are the
sample sizes.

1.4 When Variances Are Unequal — Welch's t Test

When Levene's test is significant (p < .05), homogeneity of variance is violated. Use Welch's corrected
formula:

t' = (X■1 - X■2) / √(s21/n1 + s22/n2)

The degrees of freedom are adjusted downward using the Welch-Satterthwaite equation. SPSS
automatically provides both versions.
1.5 Decision Rule
• State H0: µ1 = µ2 (no difference between population means)
• State H1: µ1 ≠ µ2 (two-tailed) or µ1 > µ2 / µ1 < µ2 (one-tailed)
• Choose significance level α (typically 0.05)
• Calculate t-statistic and compare with critical value t(α/2, df) from t-table
• If |t_obtained| > t_critical, reject H0

1.6 Effect Size — Cohen's d

Statistical significance alone does not tell us the practical importance of a result. Cohen's d expresses the
difference in standard deviation units:

d■ = (X■1 - X■2) / sp

Where sp = √s2p is the pooled standard deviation.

d Value Interpretation

0.20 Small effect

0.50 Medium effect

0.80 Large effect

1.7 Confidence Interval on µ1 – µ2

CI■■ = (X■1 - X■2) ± t(0.025, df) × s_X■1-X■2

If the CI does not include 0, we reject H0 at the corresponding α level.

1.8 Worked Example

Adams et al. (1996) studied sexual arousal in homophobic vs. non-homophobic men. Homophobic group
(n=35, X■=24.0, s2=148.87); Non-homophobic group (n=29, X■=16.5, s2=139.16).

s2p = [34(148.87) + 28(139.16)] / (35+29-2) = 144.48

t = (24.0 - 16.5) / √(144.48/35 + 144.48/29) = 7.5 / 3.018 = 2.48

df = 62, t_critical(0.025, 62) ≈ 2.003

Since t_obtained (2.48) > t_critical (2.003), we reject H0. The homophobic group showed significantly
more arousal (p < .05). Cohen's d = 0.62 (medium-large effect).

2. Paired t Test (Matched Samples / Repeated Measures)

2.1 What Is It?

The paired t test (also called the matched-sample, dependent-sample, or related-samples t test) is used
when the same subjects are measured twice (e.g., before and after treatment) or when subjects are paired
by matching. The key idea is that knowing one score tells us something about the other — the scores are
correlated.

Example: Measuring weight of anorexic girls before and after family therapy. Each girl contributes two
scores.

2.2 Assumptions
• The differences between paired scores are approximately normally distributed.
• The pairs of observations are independent of other pairs.
• The data are measured on an interval or ratio scale.
Note: Unlike the independent t test, we do NOT need equal variances across the two measurement occasions.

2.3 Logic — Difference Scores

The paired t test reduces to a one-sample t test on the difference scores. For each participant, compute:

D = X_after - X_before (or X_condition2 - X_condition1)

If the treatment has no effect, the average of all D values (D■) should be close to zero. We test H0: µD =
0.

2.4 Formula
t = D■ / (s_D / √N)

Degrees of freedom: df = N - 1

Where: D■ = mean of the difference scores; s_D = standard deviation of the difference scores; N =
number of pairs.

2.5 Worked Example — Anorexia Treatment (Everitt, 1994)

17 girls were weighed before and after family therapy. The difference scores (After - Before) had: D■ =
7.26 lbs, s_D = 7.16, N = 17.

t = 7.26 / (7.16 / √17) = 7.26 / 1.737 = 4.18

df = 16, t_critical(0.025, 16) = 2.12

Since 4.18 > 2.12, we reject H0. The girls gained a statistically significant amount of weight (p < .05).
CI■■ = 3.57 to 10.95 lbs.

Effect size: Cohen's d = 7.26/5.02 = 1.45 (very large effect — nearly 1.5 standard deviations gain in
weight).

2.6 When to Use Paired vs. Independent t Test

Feature Independent t Paired t

Groups Two separate groups Same subjects measured twice (or matched pairs)

Key question Do the groups differ? Did the treatment/time change scores?

Advantage of pairing — Removes between-subjects variability; more powerful

df n1+n2-2 N-1 (N = number of pairs)
3. One-Way Analysis of Variance (ANOVA)

3.1 What Is It?

One-way ANOVA is an extension of the independent t test to situations where there are three or more
groups. It tests whether the means of three or more independent groups are all equal. The term 'one-way'
refers to one independent variable (factor) with multiple levels.

Example: Comparing memory recall across five different encoding conditions (Counting, Rhyming,
Adjective, Imagery, Intentional) — Eysenck (1974).

Why not just run multiple t tests? Running multiple t tests inflates the Type I error rate. If we run 10 tests at
α = .05, the familywise error rate rises to ~40%. ANOVA controls the overall error rate at α.

3.2 Assumptions
• Independence of observations within and between groups.
• Normality: Scores are normally distributed within each group.
• Homogeneity of variance: σ21 = σ22 = ... = σ2k (test with Levene's test).
• Interval/ratio scale measurement.

3.3 The Structural Model

Every score can be decomposed into:

Xij = µ + τj + εij

Where: µ = grand mean of the entire population; τj = treatment effect for group j (= µj - µ); εij = random
error for person i in group j.

H0 states that all τj = 0, i.e., all group means equal the grand mean.

3.4 Partitioning of Variance

ANOVA works by partitioning total variability (SS_Total) into:

SS_Total = SS_Between + SS_Within

SS_Between = Σnj(X■j - X■..)² (variation DUE TO treatment)

SS_Within = ΣΣ(Xij - X■j)² (variation NOT due to treatment, 'error')

df_Between = k - 1 (k = number of groups)

df_Within = N - k (N = total observations)

MS_Between = SS_Between / (k-1)

MS_Within = SS_Within / (N-k)

F = MS_Between / MS_Within

3.5 Interpreting the F-Ratio

The F-ratio compares between-group variance to within-group variance. If H0 is true, both MS estimates
estimate the same σ2e, so F ≈ 1.0. If H0 is false, the between-group MS is inflated by treatment effects, so
F > 1.

Reject H0 if F_obtained > F_critical(df_between, df_within, α).

3.6 The ANOVA Summary Table

Source SS df MS F

Between Groups (Treatment) SS_B k-1 SS_B/(k-1) MS_B/MS_W

Within Groups (Error) SS_W N-k SS_W/(N-k)

Total SS_T N-1

3.7 Post-Hoc Tests

A significant F-ratio tells us that at least one pair of means differs, but NOT which ones. Post-hoc tests
identify which specific pairs differ while controlling the familywise error rate:

• Tukey's HSD (Honestly Significant Difference): Most commonly used; controls familywise α well.
• Bonferroni correction: Divide α by the number of comparisons.
• Scheffé's test: Most conservative; appropriate for complex comparisons.
• Fisher's LSD: Least conservative; use only with limited planned comparisons.

3.8 Effect Size — η² (Eta-squared)

η² = SS_Between / SS_Total

Interpretation: proportion of total variance accounted for by group membership. η² = 0.01 (small), 0.06
(medium), 0.14 (large).

3.9 Worked Example — Eysenck's Recall Study

Five encoding conditions: Counting (n=10, X■=7.00), Rhyming (n=10, X■=6.90), Adjective (n=10,
X■=11.00), Imagery (n=10, X■=13.40), Intentional (n=10, X■=12.00). Grand mean = 10.06.

SS_Between = 10[(7.00-10.06)² + (6.90-10.06)² + (11.00-10.06)² +

(13.40-10.06)² + (12.00-10.06)²]

= 10[9.36 + 9.99 + 0.88 + 11.16 + 3.76] = 351.52

SS_Within = sum of all within-group variances × (n-1) ≈ 667.65

F = (351.52/4) / (667.65/45) = 87.88 / 14.84 = 5.93, F_crit(4,45) ≈ 2.58

F_obtained (5.93) > F_critical (2.58), so we reject H0. Not all conditions produce equal recall. Post-hoc
tests would identify that deeper-processing conditions (Adjective, Imagery, Intentional) produce
significantly better recall than shallower conditions (Counting, Rhyming).
4. Two-Way ANOVA (Factorial Analysis of Variance)

4.1 What Is It?

Two-way ANOVA extends one-way ANOVA to two independent variables (factors) simultaneously. Every
level of Factor A is combined with every level of Factor B — hence the name 'factorial design'. The design
allows us to examine:

• Main effect of Factor A: The effect of A ignoring B.

• Main effect of Factor B: The effect of B ignoring A.
• Interaction (A × B): Whether the effect of A depends on the level of B (or vice versa).
Example: Eysenck studied both Age (Young vs. Old) and Recall Condition (5 levels), yielding a 2×5
factorial design.

4.2 Advantages of Factorial Designs

• Economy: Two factors tested simultaneously, fewer participants needed than two separate one-way
ANOVAs.
• Broader generalization: Conclusions about A apply across all levels of B.
• Interaction detection: Can reveal that the effect of one variable depends on the level of another —
often the most interesting finding.

4.3 Partitioning of Variance

SS_Total = SS_A + SS_B + SS_A×B + SS_Within

Each effect is tested by its own F-ratio:

F_A = MS_A / MS_Within

F_B = MS_B / MS_Within

F_A×B = MS_A×B / MS_Within

Degrees of freedom: df_A = a-1, df_B = b-1, df_A×B = (a-1)(b-1), df_W =

N-ab

4.4 The ANOVA Summary Table for Two-Way

Source df MS F

Factor A (e.g., Age) a-1 SS_A / (a-1) MS_A / MS_W

Factor B (e.g., Condition) b-1 SS_B / (b-1) MS_B / MS_W

A × B Interaction (a-1)(b-1) SS_AB/(a-1)(b-1) MS_AB/MS_W

Within (Error) N - ab SS_W/(N-ab) —

Total N-1

4.5 Understanding Interactions

An interaction occurs when the effect of one factor is not the same across levels of the other factor. For
example, if deeper-processing conditions benefit younger participants far more than older ones, there is an
Age × Condition interaction.

Interaction is best visualized with a profile (line) plot. Non-parallel lines indicate an interaction; parallel lines
indicate no interaction.

Note: Interactions are often among the most theoretically important findings in research. Always examine the
interaction before interpreting main effects — a significant interaction complicates the interpretation of main
effects.

4.6 Simple Effects

When an interaction is significant, we examine simple effects: the effect of one factor at each level of the
other. For example, we compare recall conditions separately for Young and Old participants. Simple
effects are themselves one-way ANOVAs conducted within subgroups.

4.7 Worked Example — Eysenck Age × Condition Study

2 × 5 factorial: Age (Young, Old) × Condition (5 levels), n = 10 per cell.

Results: Main effect of Age is significant — younger participants recall more. Main effect of Condition is
significant — deeper processing leads to better recall. Age × Condition interaction is significant — the
advantage of deeper processing is larger for younger participants.

This illustrates the value of factorial designs: the interaction tells us something neither main effect alone
could reveal.

5. Repeated Measures ANOVA

5.1 What Is It?

Repeated measures ANOVA (also called within-subjects ANOVA) is used when the same participants are
measured under multiple conditions or at multiple time points. It is the extension of the paired t test to three
or more conditions.

Example: Measuring a patient's anxiety at 4 different time points during therapy (before, after 1 month,
after 3 months, after 6 months).

5.2 Advantages over Between-Subjects ANOVA

• Greater statistical power: Individual differences are removed from the error term, reducing noise.
• Fewer participants needed: Each participant serves as their own control.
• Controls for individual differences: Systematic variation between people is partitioned out.

5.3 Partitioning of Variance

In repeated measures ANOVA, the total variance is partitioned as:

SS_Total = SS_Between-subjects + SS_Within-subjects

 SS_Within = SS_Treatment + SS_Error (residual)

F = MS_Treatment / MS_Error

SS_Between-subjects represents individual differences across participants and is removed from the error
term — this is why repeated measures has greater power.

5.4 Assumptions
• Normality: The differences between any pair of conditions are normally distributed.
• Sphericity (Mauchly's test): The variances of the differences between all pairs of conditions must be
equal. Violation inflates Type I error.
Note: If Mauchly's test is significant, apply a correction: Greenhouse-Geisser (conservative) or Huynh-Feldt
(less conservative) epsilon corrections adjust the degrees of freedom.

5.5 Mixed Designs

A common and important design combines repeated measures on one factor with independent groups on
another. For example: compare therapy outcomes (between-subjects: therapy type) measured at multiple
time points (within-subjects). This is called a split-plot or mixed design.

• Between-subjects factor: Participants belong to only one level (e.g., Control vs. Treatment).
• Within-subjects factor: All participants measured at each level (e.g., Time 1, Time 2, Time 3).
• Interaction (Group × Time): Tests whether the treatment groups change differently over time —
typically the key test of interest.
 PART B: NON-PARAMETRIC TESTS
Non-parametric (distribution-free) tests do not assume a specific population distribution (such as normality)
and do not require estimation of population parameters. They are used when parametric assumptions are
violated, sample sizes are small, or data are ordinal in nature.

When to Choose Non-Parametric Tests

• Data are ordinal (ranked) rather than interval/ratio.
• Data are severely skewed or contain extreme outliers.
• Sample sizes are very small and normality cannot be assumed.
• The dependent variable is better described by the median than the mean.
Note: The main cost of non-parametric tests is lower power compared to their parametric counterparts when
assumptions ARE met. When assumptions are violated, non-parametric tests may actually have greater
power.

6. The Median as a Measure in Non-Parametric Analysis

6.1 Definition
The median is the middle value when data are arranged in ascending order. For N observations, the
median is the value at position (N+1)/2 for odd N, or the average of positions N/2 and N/2+1 for even N.

Median location = (N + 1) / 2

6.2 Why the Median Matters in Non-Parametric Statistics

Many non-parametric tests are specifically tests of medians rather than means. They are sensitive to
differences in central tendency in terms of median values. The median has several important properties:

• Resistant to outliers: Extreme values do not distort the median, unlike the mean.
• Appropriate for skewed distributions: When data are heavily skewed, the median better represents
the 'typical' score.
• Appropriate for ordinal data: When the scale does not support arithmetic operations, the median is
more meaningful than the mean.

6.3 Comparison: Mean vs. Median

Property Mean Median

Sensitive to outliers Yes — heavily influenced No — resistant

Best for Symmetric distributions Skewed distributions or ordinal data

Used in Parametric tests (t, ANOVA) Non-parametric tests

Formula Σx/n Middle value of sorted data

7. Chi-Square (χ²) Test

7.1 What Is It?

The chi-square test is a non-parametric test used with categorical (nominal or ordinal) data. It tests
whether the observed frequency distribution differs from an expected one, or whether two categorical
variables are independent.

7.2 Types of Chi-Square Tests

• Goodness-of-Fit (One-way): Tests whether observed frequencies fit a hypothesized distribution.
• Test of Independence (Two-way / Contingency table): Tests whether two categorical variables are
related.
• Test of Homogeneity: Tests whether several populations have the same distribution.

7.3 Assumptions
• Data are frequencies (counts), not proportions or means.
• Each observation falls into exactly one category (mutually exclusive).
• Expected frequencies ≥ 5 in each cell (if violated, use Fisher's Exact Test or combine categories).
• Observations are independent.

7.4 Formula
χ² = Σ [ (O - E)² / E ]

Where: O = observed frequency; E = expected frequency. The sum is over all categories/cells.

Degrees of freedom (goodness-of-fit): df = k - 1 (k = number of

categories)

Degrees of freedom (independence): df = (r-1)(c-1) (r rows, c columns)

7.5 Expected Frequencies — Test of Independence

E_ij = (Row_i total × Column_j total) / Grand total = (R_i × C_j) / N

7.6 Decision Rule

Reject H0 if χ²_obtained > χ²_critical(df, α). The chi-square distribution has only positive values and is
right-skewed (not symmetric like t).

7.7 Effect Size — Cramér's V

V = √[ χ² / (N × min(r-1, c-1)) ]

V ranges from 0 (no association) to 1 (perfect association). For 2×2 tables, V = φ (phi).

7.8 Worked Example

A study tested therapeutic touch: 150 trials. Expected: 75 correct, 75 incorrect. Observed: 53 correct, 97
incorrect (therapists did worse than chance).

χ² = (53-75)²/75 + (97-75)²/75 = 484/75 + 484/75 = 12.91

df = 1, χ²_critical(0.05, 1) = 3.84

χ²_obtained (12.91) >> χ²_critical (3.84). Reject H0. The proportion of correct responses differed
significantly from chance — therapists performed worse than expected, providing evidence against
therapeutic touch.

Note: Chi-square does not specify the direction of an effect, only that a difference exists. A significant result
with small expected frequencies should be interpreted cautiously.
8. The Sign Test

8.1 What Is It?

The sign test is the simplest non-parametric test for matched pairs or repeated measures. It ignores the
magnitude of differences and looks only at their direction (sign). It is less powerful than the Wilcoxon
signed-ranks test but makes even fewer assumptions.

Use case: When you have paired data and want to test whether one condition tends to produce higher
scores than the other, without assuming anything about the distribution of differences.

8.2 Assumptions
• The data consist of matched pairs or repeated measures on the same subjects.
• The differences are independent of each other.
• The measurement scale is at least ordinal (we need to know which score is larger).

8.3 Procedure
• Step 1: Compute difference D = X2 - X1 for each pair.
• Step 2: Note the sign of each difference (+ or −). Ignore ties (D = 0).
• Step 3: Count the number of positive signs (k) out of N non-zero differences.
• Step 4: Under H0, each sign is equally likely to be + or −, so p = 0.5. Test using the binomial
distribution.

8.4 Formula
Under H0 (p = 0.5), the probability of k or more positive signs out of N follows the binomial distribution:

P(X ≥ k) = Σ C(N,x) × (0.5)^x × (0.5)^(N-x) for x = k to N

For large N, use the normal approximation:

z = (k - N/2) / √(N/4) = (k - N/2) / (√N / 2)

8.5 Worked Example — Manning et al. (1990)

16 participants recalled words in two conditions (Glucose vs. Saccharin). Differences computed: 13
positive (Glucose better), 3 negative (Saccharin better), 0 ties.

P(X ≥ 13 | N=16, p=0.5) = P(13) + P(14) + P(15) + P(16)

= 0.0085 + 0.0018 + 0.0002 + 0.0000 = 0.0105

Two-tailed p = 2 × 0.0105 = 0.021. Since p < .05, reject H0. Glucose condition produces significantly better
recall.

Alternatively, using chi-square: χ² = (13-8)²/8 + (3-8)²/8 = 6.25. With df=1, χ²_critical = 3.84 → reject H0.

8.6 Limitations
 • The sign test is less powerful than the Wilcoxon signed-ranks test because it discards magnitude
information.
• Use it when: differences are unreliable in magnitude, or the measurement scale only gives direction.
• When more information is available, prefer the Wilcoxon test.

9. Wilcoxon Rank-Sum Test (Mann-Whitney U Test)

9.1 What Is It?

The Wilcoxon rank-sum test is the non-parametric equivalent of the independent samples t test. It tests
whether two independent groups were drawn from identical populations, and is particularly sensitive to
differences in central tendency (medians). It works by replacing raw scores with ranks.

The Mann-Whitney U test is mathematically equivalent and gives the same result.

9.2 Assumptions
• Two independent random samples.
• The dependent variable is at least ordinal.
• The two distributions have the same shape (for the test to be specifically about medians).
• No assumption of normality required.

9.3 Logic
Rank all N scores (from both groups combined) from lowest to highest. If H0 is true (groups are from the
same population), low and high ranks should be evenly spread between groups. The test statistic W_S is
the sum of ranks in the smaller group.

9.4 Procedure
• Step 1: Combine all scores from both groups and rank them 1 to N.
• Step 2: Handle ties by assigning the average of the tied ranks.
• Step 3: Sum the ranks for the smaller group (or either group if sizes equal): W_S.
• Step 4: Also compute W'_S = 2W - W_S, where 2W = n1(n1+n2+1).
• Step 5: Compare the smaller of W_S and W'_S to the critical value from Wilcoxon tables.

9.5 Formula
W_S = sum of ranks in the smaller group

W'_S = n1(n1 + n2 + 1) - W_S

For large samples (n1, n2 > 25), use the normal approximation:

Mean of W_S: µ_W = n1(n1 + n2 + 1) / 2

SE of W_S: σ_W = √[ n1 × n2 × (n1+n2+1) / 12 ]

z = (W_S - µ_W) / σ_W

9.6 Worked Example — Stressful Life Events
Cardiac patients (n=6): scores 32, 8, 7, 29, 5, 60. Orthopedic patients (n=5): 0, 1, 2, 2, 3.

Combined ranking (lowest to highest): 0(1), 1(2), 2(3.5), 2(3.5), 3(5), 5(6), 7(8), 8(9), 29(10), 32(11)

W_S (orthopedic, n=5) = 1 + 2 + 3.5 + 3.5 + 5 = 21

From Wilcoxon tables, n1=5, n2=6, α=0.05 two-tailed: critical value = 18

Since W_S = 21 > 18 (critical), we fail to reject H0. The cardiac patients did not show significantly more
stressful life events than orthopedic patients at α = .05.

9.7 Wilcoxon Matched-Pairs Signed-Ranks Test

This is the non-parametric equivalent of the paired t test. It ranks the absolute differences and then
assigns signs, thus using more information than the sign test.

• Step 1: Compute D = X2 - X1 for each pair. Discard zero differences.

• Step 2: Rank the absolute values |D| from smallest to largest. Assign average ranks for ties.
• Step 3: Reattach the sign of each D to its rank.
• Step 4: T+ = sum of positive ranks. T- = sum of negative ranks.
• Step 5: Test statistic T = min(T+, T-). Reject H0 if T ≤ T_critical.
For large N: z = (T - N(N+1)/4) / √[ N(N+1)(2N+1)/24 ]
10. Kruskal–Wallis H Test

10.1 What Is It?

The Kruskal–Wallis test is the non-parametric analogue of one-way ANOVA. It tests whether three or more
independent groups were drawn from identical populations, and is particularly sensitive to differences in
central tendency. It uses ranks rather than raw scores.

Use case: When you have three or more groups and the normality assumption of ANOVA is violated, or
the data are ordinal.

10.2 Assumptions
• Three or more independent random samples.
• The dependent variable is at least ordinal.
• No assumption of normality.
• Each group should have at least 5 observations for the chi-square approximation to be valid.

10.3 Procedure
• Step 1: Combine all scores from all groups and rank them 1 to N (assigning average ranks for ties).
• Step 2: Compute the sum of ranks R_j for each group j.
• Step 3: Compute the H statistic.
• Step 4: Compare H to the chi-square distribution with df = k - 1.

10.4 Formula
H = [ 12 / (N(N+1)) ] × Σ [ R_j² / n_j ] - 3(N+1)

Where: N = total number of observations; n_j = number of observations in group j; R_j = sum of ranks in
group j; k = number of groups.

Degrees of freedom: df = k - 1

H is approximately distributed as chi-square when each n_j ≥ 5.

10.5 Correction for Ties

When there are many tied ranks, H is adjusted upward:

H_corrected = H / [1 - Σ(t_i³ - t_i) / (N³ - N)]

Where t_i is the number of observations in the i-th group of ties. This correction increases H slightly,
making rejection of H0 more likely.

10.6 Worked Example

Three drug conditions (Depressant, Stimulant, Control) were tested on problem-solving scores, n=7 per
group. After ranking all 21 scores:

R_Depressant = 35, R_Stimulant = 75.5, R_Control = 120.5

 H = [12/(21×22)] × [(35²/7) + (75.5²/7) + (120.5²/7)] - 3(22)

= [12/462] × [175 + 814.6 + 2072.3] - 66

= 0.026 × 3061.9 - 66 = 79.61 - 66 = 13.61

df = 2, χ²_critical(0.05, 2) = 5.99

Since H = 13.61 > 5.99, reject H0. The three drug conditions produce significantly different levels of
problem solving.

10.7 Post-Hoc Tests for Kruskal–Wallis

A significant Kruskal–Wallis test indicates that at least one group differs from the others. To identify which
groups differ, use:

• Dunn's post-hoc test: Compares all pairs of groups using a Bonferroni correction.
• Wilcoxon rank-sum tests with Bonferroni correction: Run pairwise Wilcoxon tests and adjust α.
Quick Reference: Summary of All Tests — Module III

PARAMETRIC TESTS
Test Purpose Data Type Key Statistic Assumptions

Independent t Compare 2 independent groups

Interval/Ratio t, df=n1+n2-2 Normality, Homogeneity of variance, Independence

Paired t Compare 2 related measures (same

Interval/Ratio
subjects) t, df=N-1 Normality of differences, Independence of pairs

One-Way ANOVA Compare 3+ independent groups

Interval/Ratio F, df=(k-1, N-k) Normality, Homogeneity of variance, Independence

Two-Way ANOVA Compare 2 factors + interaction

Interval/Ratio F for A, B, A×B Normality, Homoscedasticity, Independence

Repeated Measures ANOVA 3+ conditions on same subjects

Interval/Ratio F, sphericity Normality, Sphericity (use ε correction if violated)

NON-PARAMETRIC TESTS
Test Parametric Analogue Purpose Test Statistic Key Feature

Median Mean Measure of central tendency

— Resistant to outliers; used with skewed data

Chi-Square None (categorical) χ², df=(r-1)(c-1)

Test categorical associations Works on frequencies; tests independence or GOF

Sign Test Paired t test Compare 2 related conditions

Binomial
(direction
/ z only) Discards magnitude; uses only sign of differences

Wilcoxon Rank-Sum Independent t test Compare 2 independent groups

W_S / z Ranks all scores; sensitive to median differences

Kruskal–Wallis H One-Way ANOVA ~ χ²(k-1)

Compare 3+ independentHgroups Ranks all scores; extension of Wilcoxon rank-sum

Decision Guide: Choosing the Right Test

Situation Recommended Test

2 independent groups, normal data, interval scale Independent t Test

2 related groups / same subjects measured twice Paired t Test

3+ independent groups, normal data One-Way ANOVA

2 factors, independent groups Two-Way (Factorial) ANOVA

Same subjects across 3+ conditions Repeated Measures ANOVA

2 independent groups, non-normal / ordinal Wilcoxon Rank-Sum Test

2 related groups, non-normal / ordinal Wilcoxon Signed-Ranks or Sign Test

3+ independent groups, non-normal / ordinal Kruskal–Wallis H Test

Categorical data (frequencies) Chi-Square Test

Source: These notes are based on David C. Howell, Statistical Methods for Psychology, 7th Edition (Cengage, 2010). Adapted for
Module III curriculum.