How To Calculate Point Biserial Correlation By Hand

Calculate the correlation between a continuous variable and a binary variable

Comprehensive Guide: How to Calculate Point-Biserial Correlation by Hand

The point-biserial correlation coefficient (r_pb) measures the relationship between a continuous variable and a binary variable. It’s particularly useful in educational research (e.g., correlating test scores with pass/fail outcomes) and psychological studies (e.g., correlating personality scores with gender).

When to Use Point-Biserial Correlation

• When one variable is continuous (interval/ratio data)
• When the other variable is naturally binary (dichotomous)
• When you want to understand the strength and direction of the relationship
• When you need to test the significance of this relationship

The Point-Biserial Correlation Formula

r_pb = (M₁ – M₀) / s_n × √[p(1-p)]

Where:

• M₁ = Mean of continuous variable for group coded as 1
• M₀ = Mean of continuous variable for group coded as 0
• s_n = Standard deviation of all continuous scores
• p = Proportion of cases in group 1

Step-by-Step Calculation Process

1. Organize Your Data:

   Create a table with three columns: Subject ID, Continuous Variable (X), and Binary Variable (Y coded as 0/1).

   Subject	Test Score (X)	Pass (Y)
   1	85	1
   2	72	0
   3	91	1
   4	68	0
   5	88	1

2. Calculate Group Means:

   Compute M₁ (mean for Y=1) and M₀ (mean for Y=0) separately.

   Important: Ensure your binary variable is properly coded with only two distinct values (typically 0 and 1).

3. Compute Overall Standard Deviation:

   Calculate s_n using all continuous variable scores regardless of group membership.

   s_n = √[Σ(X – M)² / N]
4. Determine Proportion p:

   Calculate p as the number of cases in group 1 divided by total number of cases.

5. Plug Values into Formula:

   Substitute all calculated values into the point-biserial correlation formula.

6. Interpret the Result:

   Use standard correlation interpretation guidelines:

   • ±0.00-0.10: Negligible
   • ±0.10-0.30: Weak
   • ±0.30-0.50: Moderate
   • ±0.50-1.00: Strong

Testing Significance of Point-Biserial Correlation

To determine if your correlation is statistically significant:

t = r_pb × √[(N-2)/(1-r_pb²)]

Compare your calculated t-value against critical values from the t-distribution table (NIST) with N-2 degrees of freedom.

Example Calculation

Let’s work through a complete example with 10 subjects:

Subject	Study Hours (X)	Passed Exam (Y)
1	15	1
2	8	0
3	20	1
4	5	0
5	18	1
6	12	0
7	22	1
8	9	0
9	16	1
10	7	0

1. M₁ (Passed) = (15 + 20 + 18 + 22 + 16)/5 = 91/5 = 18.2
2. M₀ (Failed) = (8 + 5 + 12 + 9 + 7)/5 = 41/5 = 8.2
3. Overall mean = (15+8+20+5+18+12+22+9+16+7)/10 = 132/10 = 13.2
4. s_n = 5.70 (calculated from all scores)
5. p = 5/10 = 0.5
6. r_pb = (18.2 – 8.2)/5.70 × √(0.5×0.5) = 10/5.70 × 0.5 = 0.877

This extremely high correlation (0.877) indicates a very strong relationship between study hours and exam outcomes in this sample.

Common Mistakes to Avoid

• Incorrect binary coding: Always verify your binary variable only contains two distinct values
• Unequal group sizes: Very small groups (especially n<5) can distort results
• Assuming causality: Correlation doesn’t imply causation
• Ignoring assumptions: Point-biserial assumes:
  • Continuous variable is normally distributed
  • Binary variable divides the continuous variable into two groups
  • Homogeneity of variance (equal variances in both groups)

Comparison with Other Correlation Measures

Correlation Type	Variable Types	When to Use	Range
Point-Biserial	Continuous + Binary	One naturally binary variable	-1 to +1
Biserial	Continuous + Artificial Binary	Binary variable created from underlying continuous variable	-1 to +1
Pearson’s r	Continuous + Continuous	Both variables continuous and normally distributed	-1 to +1
Spearman’s ρ	Ordinal + Ordinal	Non-normal distributions or ordinal data	-1 to +1
Phi Coefficient	Binary + Binary	Both variables are binary	-1 to +1

Advanced Considerations

For more sophisticated applications:

1. Confidence Intervals:

   Calculate 95% CIs using Fisher’s z transformation:

   z = 0.5 × ln[(1+r)/(1-r)]

   SE_z = 1/√(N-3)

2. Effect Size Interpretation:

   Cohen’s guidelines for r_pb:

   • Small: |0.10|
   • Medium: |0.24|
   • Large: |0.37|
3. Power Analysis:

   Use G*Power or similar software to determine required sample size for desired power (typically 0.80).

Real-World Applications

Point-biserial correlation is widely used in:

• Educational Research:
  • Correlating SAT scores with college admission (yes/no)
  • Examining relationship between study habits and passing exams
  • Evaluating whether tutorial attendance predicts course success
• Psychological Testing:
  • Validating test items (correct/incorrect responses vs total scores)
  • Examining gender differences in personality traits
  • Assessing whether clinical diagnosis correlates with symptom severity
• Medical Research:
  • Correlating biomarker levels with disease presence/absence
  • Examining relationship between treatment adherence and recovery
  • Assessing whether risk factors predict disease outcomes
• Market Research:
  • Correlating product usage with purchase decisions
  • Examining relationship between advertising exposure and brand preference
  • Assessing whether customer satisfaction predicts repeat purchases

Software Implementation

While this guide focuses on manual calculation, most statistical software can compute r_pb:

• SPSS:

  Use Analyze → Correlate → Bivariate (treat binary variable as continuous)

• R:

  cor.test(continuous_var, binary_var, method="pearson")

• Python:

  from scipy.stats import pointbiserialr
  r, p = pointbiserialr(continuous_data, binary_data)

• Excel:

  =CORREL(continuous_range, binary_range)

Historical Context

The point-biserial correlation was first described by Karl Pearson in 1900 as a special case of his product-moment correlation. It was further developed by:

• Charles Spearman (1904) in his work on intelligence testing
• Louis Thurstone (1931) in psychometric applications
• Lee Cronbach (1949) in educational measurement

For a deeper historical perspective, see the Pearson’s original papers at York University’s Psychology Classics archive.

Mathematical Derivation

Point-biserial correlation can be derived from the general Pearson correlation formula:

r = Cov(X,Y) / (s_x × s_y)

When Y is binary (coded 0/1):

• Cov(X,Y) = p(1-p)(M₁ – M₀)
• s_y = √[p(1-p)]
• Substituting these into the Pearson formula yields the point-biserial formula

Limitations and Alternatives

While useful, point-biserial correlation has limitations:

1. Assumption Violations:

   If the continuous variable isn’t normally distributed, consider:

   • Spearman’s rank correlation for ordinal data
   • Biserial correlation if the binary variable is artificial
2. Small Sample Issues:

   With N<30, results may be unstable. Consider:

   • Exact permutation tests
   • Bayesian approaches
3. Unequal Variances:

   If Levene’s test shows unequal variances, consider:

   • Welch’s correction
   • Nonparametric alternatives

Reporting Guidelines

When reporting point-biserial correlations in academic work, include:

1. The correlation coefficient (r_pb) with two decimal places
2. Degrees of freedom in parentheses
3. p-value (exact if possible, otherwise as p<.05 etc.)
4. Confidence intervals (preferably 95%)
5. Effect size interpretation
6. Sample size for each group
7. Assumption checks performed

Example APA-style reporting:

“Study hours were strongly correlated with exam outcomes, r_pb(8) = .88, p = .001, 95% CI [.56, .97], representing a large effect size according to Cohen’s (1988) criteria.”

Frequently Asked Questions

1. Can I use point-biserial if my binary variable has unequal group sizes?

   Yes, but extreme imbalances (e.g., 90% in one group) may reduce statistical power and inflate the correlation coefficient.

2. What’s the difference between point-biserial and biserial correlation?

   Point-biserial uses a naturally binary variable, while biserial assumes the binary variable was created from an underlying continuous variable (e.g., passing a test with a cutoff score).

3. How do I handle missing data?

   Use listwise deletion for small amounts of missing data (<5%). For more missing data, consider multiple imputation.

4. Can r_pb be negative?

   Yes. A negative value indicates that as the continuous variable increases, the likelihood of being in group 1 decreases.

5. What’s a good sample size for point-biserial correlation?

   Minimum N=30 for reasonable stability. For publishing, aim for N≥100 with at least 10-15 cases per group.

Further Learning Resources

For deeper understanding, consult these authoritative sources:

• Laerd Statistics Guide to Correlation – Comprehensive explanation with examples
• VassarStats Computational Tools – Free online calculators with detailed output
• NIH/NLM Statistics Review – Medical research applications
• Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Lawrence Erlbaum Associates. – The standard reference for effect size interpretation