Unit V: Non-parametric Statistics — Meaning

Non-parametric statistics are a set of statistical methods used to analyze data without assuming a specific
distribution (such as normal distribution) for the population from which the samples are drawn.

●​ Do not require the assumption of normality.​

Used for ordinal data, nominal data, or small sample sizes.​
Based on rankings or frequency counts rather than raw scores.​
Often called distribution-free tests.

When to Use Non-parametric Tests?

Data are not normally distributed.​
Data are ordinal or categorical.​
Sample sizes are too small to validate parametric assumptions.​
The measurement scale is nominal or ordinal.​

Advantages:
Flexible with fewer assumptions.​
Applicable to a wider range of data types.​
Robust against outliers and skewed data.​

Disadvantages:
Generally less powerful than parametric tests (i.e., require larger samples to detect effects).​
Provide less detailed information (e.g., do not estimate means or variances).​

Examples of Non-parametric Tests:

Chi-square test​
Mann-Whitney U test
Wilcoxon signed-rank test​
Kruskal-Wallis test​
Spearman’s rank correlation​

Aspect Parametric Tests Non-Parametric Tests

Definition Tests that assume the data follow a Tests that do not assume any specific
specific distribution, usually population distribution.
normal.

Data Type Require interval or ratio scale Suitable for nominal or ordinal data, and
data (continuous). also continuous data not meeting
parametric assumptions.

Assumption Require assumptions like Fewer assumptions; do not require normal

s normality, homogeneity of distribution or equal variances.
variance, and independence.

Sample Size Typically require larger sample Can be used with small sample sizes.
sizes for accuracy.

Test Use parameters like mean and Use ranks or frequency counts rather than
Statistics standard deviation. means.

Examples t-test, ANOVA, Pearson Mann-Whitney U, Wilcoxon signed-rank,

correlation, regression. Kruskal-Wallis, Spearman correlation,
Chi-square test.
Power Generally more powerful when Less powerful; may require larger samples
assumptions are met. to detect effects.

Results Provide estimates about population Provide information about median or

parameters (mean, variance). distribution-free comparisons.

Application When data meet assumptions and When data violate parametric assumptions
scale of measurement is or are ordinal/nominal.
appropriate.

Assumptions of Non-Parametric Tests

1.​ No Assumption of Normality:- Non-parametric tests do not require the data to follow a normal distribution.​

2.​ Level of Measurement:-Can be used with nominal, ordinal, or interval/ratio data (especially when assumptions
for parametric tests are violated).​

3.​ Independence- Observations should be independent of each other (except in some paired designs like Wilcoxon
signed-rank test).​

4.​ Random Sampling- Samples should be drawn randomly from the population.​

5.​ Shape of Distribution- Many non-parametric tests do not assume any specific shape for the distribution of
data.​

6.​ Homogeneity of Variance:- Usually not required; non-parametric tests are robust against heterogeneity of
variance.​

7.​ Sample Size:- Can be applied to small or large samples, but very small samples might reduce test power.

Characteristics of Non-Parametric Tests

1.​ Distribution-Free:- They do not require the population to follow a normal distribution or any specific
distribution.​

2.​ Use with Ordinal or Nominal Data:- Suitable for ordinal, nominal, or ranked data, as well as interval/ratio data
that violate parametric assumptions.​

3.​ Based on Ranks or Signs:- Often use ranking or categorization of data rather than raw scores.​

4.​ Robust to Outliers:- Less sensitive to outliers or extreme values in the data.​

5.​ Less Restrictive Assumptions:- Require fewer assumptions about the data, making them more flexible.​

6.​ Suitable for Small Sample Sizes:- Can be used effectively when sample sizes are too small to satisfy parametric
assumptions.​

7.​ Generally Less Powerful:= Usually have less statistical power compared to parametric tests when assumptions
for parametric tests are met.​

8.​ Simple Computation:- Often simpler to compute, especially before the widespread use of computers.​
Spearman’s Rank Difference Correlation
Spearman’s Rank Difference Correlation is a non-parametric measure that assesses the strength and direction of the
monotonic relationship between two variables based on the ranks of their values rather than the actual data. It uses the
differences between the ranks of paired observations to calculate the correlation coefficient.

2. Equation:

Where:

3. Steps to Calculate:

1.​ Assign ranks to the values of both variables separately.​

Calculate the difference did_i between the paired ranks.​
Square the rank differences​
Sum all squared differences:​
Substitute values in the formula to calculate ​

5. Advantages:
Does not require assumptions of normality or linearity.​
Suitable for ordinal data or continuous data that is not normally distributed.​
Robust to outliers as it uses ranks instead of raw data.​
Can detect monotonic relationships (either increasing or decreasing).​

6. Disadvantages:
Less powerful than Pearson’s correlation when data is truly interval and linear.​
Cannot capture non-monotonic relationships.​
Ties in ranks complicate calculation and interpretation (though correction methods exist).​
Sensitive to sample size; very small samples may give misleading results.​

7. Limitations:
Not suitable for nominal data (only ordinal or higher).​
Assumes a monotonic relationship; cannot detect more complex relationships.​
Presence of many ties can reduce accuracy.​
Cannot specify causality—only association.​

8. Alternatives:
Alternative Method When to Use

Pearson’s r When data is continuous, normally distributed, and relationship is linear

Kendall’s Tau (τ) For smaller samples or many tied ranks

Point-biserial When one variable is dichotomous

Phi Coefficient (φ) When both variables are binary

9. Scenario Example:

Imagine a researcher wants to study if students’ ranks in mathematics predict their ranks in physics. Since the interest is in
ranks and the data may not be normally distributed, Spearman’s rank difference correlation is appropriate. After ranking
both sets of scores, the researcher calculates the rank differences and applies the formula to find the correlation, which
indicates how closely related the performances are across subjects.

Biserial Correlation
Biserial correlation is a statistical measure used to determine the strength and direction of the association between:

●​ One continuous (interval/ratio) variable, and​

One artificially dichotomous variable (a continuous variable that has been split into two categories).​
It estimates the correlation assuming the dichotomous variable represents an underlying continuous distribution
split by a cutoff.

2. Types of Biserial Correlation:

●​ Point-biserial correlation: Used when the dichotomous variable is natural (truly categorical, e.g., male/female).​
Biserial correlation: Used when the dichotomous variable is artificial (created by splitting a continuous variable
into two groups, e.g., pass/fail based on a cutoff score).​

3. Formula for Biserial Correlation

4. Assumptions:

●​ The dichotomous variable is artificially created by cutting a continuous variable at some point.​
The underlying continuous variable is normally distributed.​
Both groups have equal variances (homoscedasticity).​
The relationship is linear.​

5. Advantages:

●​ More accurate than point-biserial when the dichotomy is artificial.​

Reflects the true correlation between the continuous variable and its underlying latent distribution.​
Useful in psychological testing and educational measurement (e.g., pass/fail split).

6. Disadvantages:

●​ Requires the assumption of an underlying continuous normal distribution which may not always hold.​
Calculation is more complex because of the need to estimate yy from the normal distribution.​
If the dichotomy is natural, point-biserial correlation is more appropriate.​
Sensitive to the cutoff point chosen to dichotomize the continuous variable.​

7. Example Scenario:

Suppose an educational researcher has a continuous variable: students’ intelligence scores. They create two groups by
dividing students into "high IQ" and "low IQ" based on a cutoff score of 100.

●​ The researcher wants to see how this artificial grouping correlates with another continuous variable such as
academic performance.​
Using biserial correlation, the researcher can estimate the strength of association between the underlying
continuous IQ distribution and academic performance more accurately than using point-biserial correlation.

8. Difference between Point-biserial and Biserial:

Aspect Point-biserial Biserial

Nature of Dichotomous Natural (true binary Artificial (dichotomized from continuous)

Var. categories)
 Assumptions Less strict Assumes underlying normal continuous
distribution

Calculation Simpler More complex (requires normal curve

ordinate)

Point-Biserial Correlation
Point-biserial correlation is a statistical measure used to calculate the strength and direction of the relationship between:

●​ One continuous (interval/ratio) variable, and​

One naturally dichotomous (binary) variable (e.g., male/female, yes/no).​
It is a special case of the Pearson correlation applied when one variable is binary.

2. Formula:

Where:

3. Assumptions:
The dichotomous variable is naturally occurring (not artificially created).​
The continuous variable is normally distributed within each group.​
Groups have homogeneity of variance (equal variances).​
The relationship is linear.​

4. Advantages:
Simple to compute and interpret.​
Measures association between binary and continuous variables effectively.​
Widely used in psychological, educational, and social research.​
Directly related to t-test for difference between two means.​

5. Disadvantages:
Assumes natural dichotomy; inappropriate if the binary variable is artificially created.​
Sensitive to unequal group sizes (unbalanced groups).​
Requires normality and equal variance assumptions.​
Does not measure causality, only association.​

6. Example Scenario:

A researcher studies whether there is a relationship between gender (male = 0, female = 1) and math test scores
(continuous variable).

M1M_1 = mean math score for females​

M0M_0 = mean math score for males​
pp = proportion of females​
qq = proportion of males​
SyS_y = standard deviation of math scores for all students​
The point-biserial correlation rpbr_{pb} indicates how strongly gender is associated with math scores.

7. Comparison with Biserial Correlation:

Aspect Point-Biserial Correlation Biserial Correlation

Dichotomous Variable Natural (true binary) Artificial (dichotomized continuous)

Nature
 Calculation Complexity Simpler More complex (requires normal curve
ordinate)

Assumptions Normality within groups, Same + assumes latent continuous

homogeneity underlying variable

Use Case Gender, yes/no, Pass/fail based on cutoff scores

presence/absence

Tetrachoric Correlation
Tetrachoric correlation is a measure of association between two dichotomous variables that are assumed to represent
underlying continuous latent variables which have been artificially dichotomized.In other words, it estimates the
correlation between two unobserved continuous variables from the observed binary data.

2. When to Use:

●​ Both variables are binary (dichotomous)

●​ The dichotomies are artificial (e.g., cutoff scores on continuous scales).​
You assume the original latent variables are normally distributed.

3. Difference from Phi Coefficient:

Aspect Phi Coefficient Tetrachoric Correlation

Variables Two observed binary variables Two latent continuous variables behind binaries

Assumptions No assumptions about latent traits Assumes underlying bivariate normal distribution

Use Direct association of binaries Estimation of correlation between latent traits

4. Formula and Computation:

●​ The formula is complex and usually computed via iterative methods or software.​
Based on the frequencies in a 2x2 contingency table, and the thresholds for dichotomizing the latent continuous
variables.​
Uses the bivariate normal distribution to estimate correlation.

5. Assumptions:
Each observed binary variable represents an underlying normally distributed continuous variable.​
The latent continuous variables are linearly related.​
The dichotomization thresholds are known or estimated.​
The sample size is adequate for stable estimation.​

6. Advantages:
Provides more accurate estimates of association for dichotomous variables from continuous origins.​
Useful in psychometrics when testing relationships between binary test items.​
Corrects downward bias of simpler measures like phi coefficient when artificial dichotomization exists.​

7. Disadvantages:
Computationally intensive (requires special software like R, SPSS with add-ons, or dedicated packages).​
Assumptions (normality of latent variables, linearity) may not always hold.​
Not suitable for naturally dichotomous variables (use point-biserial or phi instead).​
Sensitive to small sample sizes.​
8. Example Scenario:

A researcher studies the correlation between two dichotomous test items (pass/fail):

Item 1: Pass/fail based on a cutoff on a math ability scale.​

Item 2: Pass/fail based on a cutoff on a reading ability scale.​
Both are dichotomized versions of continuous abilities. Tetrachoric correlation estimates the true correlation between
math and reading abilities behind these binary outcomes.

9. Interpretation:
the value of tetrachoric correlation ranges from -1 to +1.​
A high positive value suggests a strong positive relationship between the latent continuous traits.​
Unlike phi, it corrects for the artificial dichotomization.​

10. Applications:
Psychometric test analysis (item analysis).​
Behavioral sciences when dealing with dichotomized continuous data.​
Structural equation modeling when latent variables are involved.

Phi Coefficient (ϕ)

Phi coefficient is a measure of association for two binary (dichotomous) variables. It quantifies the strength and
direction of the relationship between two variables that each have two categories (e.g., yes/no, male/female).

It is equivalent to the Pearson correlation coefficient when both variables are dichotomous.

2. When to Use:

●​ Both variables are dichotomous and observed (naturally binary).​

You want to measure the degree of association between these two binary variables.

3. Formula:

4. Range

Values range from -1 to +1.​

+1 indicates a perfect positive association.​
-1 indicates a perfect negative association.​
0 indicates no association.​
5. Assumptions:

Both variables are binary.​

Observations are independent.​
Large enough sample size for stable estimates.​

6. Advantages:

Simple and easy to calculate.​

Provides a symmetric measure of association between two dichotomous variables.​
Interpreted similarly to Pearson’s r.​
7. Disadvantages:

Sensitive to marginal totals; can be misleading if marginal frequencies are very uneven.​
Does not account for underlying latent continuous variables if the binary variables come from dichotomizing
continuous variables.​
Limited to 2x2 tables (only two dichotomous variables).​

8. Example:

9. Applications:

Behavioral sciences for binary variables (e.g., yes/no responses).​

Medical research to find association between presence/absence of conditions.​
Any 2x2 contingency data analysis.​

. Related Measures:
Use Tetrachoric correlation when variables are dichotomized continuous variables with latent normality assumptions.​
Use Point biserial correlation when one variable is dichotomous and the other is continuous.​

Chi-Square Test (χ²)

Chi-Square test is a non-parametric statistical test used to determine whether there is a significant association between
two categorical variables.

It tests if the observed frequencies in categories differ significantly from the expected frequencies under the null
hypothesis of independence.

2. Types of Chi-Square Tests:

Test Type Purpose Data Type

Chi-Square Test of Tests association between two categorical variables Two categorical variables
Independence

Chi-Square Goodness of Fit Tests if observed frequency distribution fits an One categorical variable
Test expected distribution

Chi-Square Test of Compares distribution of a categorical variable Two categorical variables

Homogeneity across different populations across groups

3-formula

4. Calculating Expected Frequency (for 2x2 table):

5. Assumptions
Observations are independent.​
Data are categorical (nominal or ordinal).​
Expected frequency in each cell should be 5 or more for validity. For smaller samples, use Fisher’s exact test or Yates
correction.​
Random sampling from the population.​

6. Degrees of Freedom (df):

df=(r−1)Where:rr = number of rows​
cc = number of columns​

7. Advantages:
Simple and widely used test for categorical data.
No assumptions about normality or distribution shape (non-parametric).​
Applicable for nominal and ordinal data.​
Tests independence without requiring interval data.​

8. Disadvantages:
Requires sufficiently large sample size for accuracy.​
Sensitive to small expected frequencies; can give misleading results if expected counts <5.​
Does not indicate strength or direction of association, only existence.​
Cannot be used for paired or matched data without modifications.​

9. Example: A researcher studies the association between Gender (Male/Female) and Preference for Tea or Coffee
among 100 people(chatgpt)
10. Applications:

●​ Testing independence between categorical variables in social sciences, psychology, medicine.​

Testing if sample distribution fits a theoretical distribution (Goodness of fit).​
Quality control and market research.​

Mann-Whitney U Test
The Mann-Whitney U Test (also called the Wilcoxon rank-sum test) is a non-parametric test used to compare
differences between two independent groups on an ordinal or continuous variable when the assumptions of the
parametric independent t-test are not met (e.g., non-normal distribution).

It tests whether the distribution of scores differs between two groups.

. When to Use:
Comparing two independent samples/groups.​
Data are ordinal or continuous but not normally distributed.​
Sample sizes can be small or unequal.
Data do not require interval scaling.​

3. Assumptions:
Observations are independent between groups.​
The dependent variable is at least ordinal.​
The shapes of the distributions of the two groups are similar (for comparing medians).​
Samples are randomly selected.​

4. Hypotheses:-Null hypothesis (H₀): The two populations have the same distribution (no difference in ranks).​
Alternative hypothesis (H₁): The two populations differ in distribution.​
6. Steps to Perform Mann-Whitney U Test:

1.​ Combine data from both groups and rank all values from lowest to highest.​
Assign tied ranks the average of their positions.​
Calculate the sum of ranks (R1R_1, R2R_2) for each group.​
Compute U statistics U1U_1 and U2U_2.​
Find the smaller U value.​
Compare U to critical values from Mann-Whitney U tables (based on n1n_1 and n2n_2) or calculate p-value (with
normal approximation for large samples).​
Decide to reject or fail to reject H0H_0.​

7. Advantages:
Does not require normal distribution.
Suitable for small samples and ordinal data.​
Can handle unequal sample sizes.​
Robust to outliers.​

8. Disadvantages:
Less powerful than parametric t-test if data are normally distributed.​
Tests differences in distributions, not specifically means.​
Interpretation focuses on ranks, not raw data.​
Assumes similar distribution shape to compare medians meaningfully.​

9. Example:

Suppose we want to compare stress scores (measured on an ordinal scale) between two independent groups:(chatgpt)

10. Applications:Psychology: Comparing anxiety levels between two therapy groups.​

Education: Comparing test scores between two teaching methods when normality is not met.​
Medical research: Comparing pain scores between two independent treatment groups.​

Wilcoxon Matched-Pairs Signed-Rank Test

The Wilcoxon Matched-Pairs Signed-Rank Test is a non-parametric test used to compare two related samples or
matched pairs to assess whether their population mean ranks differ. It is the non-parametric alternative to the paired
t-test when the data do not meet the assumptions of normality.

2. When to Use:
Comparing two related or paired samples (e.g., pre-test/post-test scores).​
Data are ordinal or continuous but not normally distributed.​
Sample size is small or data violate parametric assumptions.​

3. Assumptions:
The data are paired and come from the same subjects or matched pairs.​
Differences between pairs are independent.​
The dependent variable is at least ordinal.​
The distribution of differences is symmetric (less strict than normality).​

4. Hypotheses:

●​ Null hypothesis (H₀): The median difference between pairs is zero (no change).​
Alternative hypothesis (H₁): The median difference is not zero (there is change).​
5. Test Procedure:

Calculate the difference between each pair of observations (Post - Pre).​

Ignore pairs with zero difference

Rank the absolute differences from smallest to largest.​

Assign signs (+ or -) to ranks based on the sign of the original differences.​
Calculate the sum of positive ranks (T⁺) and sum of negative ranks (T⁻).​
The test statistic T is the smaller of T+T^+ and T−T^-.​
Compare T to critical values from Wilcoxon signed-rank tables or calculate p-value (for larger samples, use normal
approximation).​

. Formula for Test Statistic:

7. Advantages:
Does not require normal distribution.​
Suitable for small sample sizes and ordinal data.​
Uses magnitude and direction of differences.​
More powerful than the sign test as it uses rank information.​

8. Disadvantages:
Less powerful than the parametric paired t-test if normality holds.​
Assumes differences are symmetric in distribution.​
Cannot be used if many zero differences occur.​
Interpretation is based on ranks, not raw scores.​

9. Example:

Suppose a researcher measures anxiety levels before and after therapy in 6 patients(chatgpt)

10. Applications:

Psychology: Comparing pre- and post-treatment scores.​

Medicine: Comparing patient health measures before and after intervention.​
Education: Comparing student performance before and after a training session.​

Kruskal-Wallis H Test
The Kruskal-Wallis H Test is a non-parametric test used to compare three or more independent groups to determine
if there are statistically significant differences between their median ranks. It is the non-parametric alternative to the
one-way ANOVA when the assumptions of normality and/or homogeneity of variance are violated.

2. When to Use:

●​ Comparing three or more independent groups on an ordinal or continuous variable.​

Data are not normally distributed or sample sizes are small.​
The dependent variable is at least ordinal.​
Groups are independent (no overlap between participants).
3. Assumptions:

●​ Samples are independent.​

Observations within each group are independent.​
The dependent variable is ordinal or continuous but not necessarily normally distributed.
●​ The distributions of groups have the same shape (for median comparisons to be valid).

4. Hypotheses:

●​ Null hypothesis (H₀): The populations have the same median (no difference among groups).​
Alternative hypothesis (H₁): At least one group differs in median from others.

5. Test Procedure:

1.​ Combine all data from all groups and rank all observations together from lowest to highest, ignoring group
membership.​
Calculate the sum of ranks for each group.​
Compute the Kruskal-Wallis test statistic HH using the formula:​

2.​ Where:

6. AdvantagesDoes not require normal distribution of data.​

Can handle ordinal or non-normally distributed interval data.​
Can compare more than two independent groups.

Robust against outliers

7. Disadvantages:
Less powerful than parametric ANOVA if assumptions for ANOVA hold.​
If significant, does not specify which groups differ (post hoc tests needed).​
Assumes groups have similar shaped distributions.​
Based on ranks, so loses some information about magnitude.​

8. Example:

A researcher wants to compare stress levels among three groups of employees working in different departments (A, B, C).
Stress scores are ordinal and non-normal:

Group A Group B Group C

8 10 7

9 11 6

7 13 5

●​ ​
Combine all scores, rank them from 1 (lowest) to 9 (highest).
●​ Sum ranks for each group and calculate HH.
●​ Check if stress levels differ across departments.

9. Post Hoc Tests:

If HH is significant, follow up with pairwise comparisons such as:Dunn's Test​

Mann-Whitney U Test with Bonferroni correction​
10. Applications:

●​ Psychology: Comparing groups on mood scores across therapy types.​

Education: Comparing student satisfaction across different teaching methods.​
Medicine: Comparing treatment effects across multiple drugs with non-normal data.

Median Test
The Median Test is a non-parametric test used to determine if there are differences in the medians of two or more
independent groups. It tests whether the groups come from populations with the same median.

2. When to Use:
Comparing the medians of two or more independent groups.​
Data are ordinal or continuous but do not meet parametric test assumptions (e.g., non-normal distribution).​
Suitable for small samples or skewed data.

3. Assumptions:
Observations are independent within and between groups.​
The measurement scale is at least ordinal.​
Groups are independent.​

4. Hypotheses:

●​ Null hypothesis (H₀): The medians of all groups are equal.​

Alternative hypothesis (H₁): At least one group median differs.

5. Test Procedure:

1.​ Calculate the overall median of all observations combined.​

Classify each observation as either above or below (or equal to) the overall median.​
Create a contingency table showing the count of scores above and below the median for each group.​
Use a Chi-square test of independence on the contingency table to check for association between groups and
above/below median counts.​
If Chi-square is significant, reject H0H_0, indicating medians differ.

6. Advantages:

Simple to compute and interpret.

Does not require normality or equal variances.​
Applicable for ordinal data.​
Robust to outliers because it uses the median.
7. Disadvantages:

Less sensitive (less powerful) than other non-parametric tests like Kruskal-Wallis.​
Only compares medians, ignoring other distribution features.​
Reduces data to above/below median, losing information.

8. Example:
Researchers want to test if three diet plans differ in their median weight loss. They record weight loss for participants and:
Calculate the overall median weight loss.​
Count how many participants in each diet group lost more or less than the median.​
Perform a Chi-square test on these counts.​
9. Applications:

●​ Psychology: Comparing median anxiety levels across therapy groups.​

Education: Comparing median test scores among different teaching methods.
●​ Medical research: Comparing median recovery times across treatments.​