Biserial and Tetrachoric Correlations: Non-Pearson Methods

Introduction

In the previous section, we examined Point Biserial (r_pb) and Phi (φ) coefficients — both of which are Pearson's r applied to truly dichotomous variables. However, many variables in psychology that appear dichotomous in measurement actually have a continuous underlying distribution — they have simply been artificially split into two categories.

For example:

• Mood may be measured as "happy vs. sad" (two categories), but mood is actually a continuous spectrum
• Pass/fail is derived from a continuous marks distribution
• Attitude (positive vs. negative) is usually a continuous latent trait

When this is the case, using Point Biserial or Phi would underestimate the true correlation because we're losing information by treating a continuous variable as binary. The appropriate methods are Biserial Correlation and Tetrachoric Correlation — which estimate what the correlation would be if we had measured the underlying continuous variables directly.

---

Source PDFs:

• 📄 Block-2/Unit-2.pdf — Pages 34–36
• 📚 MPC-006 Statistics in Psychology, Block 2: Correlation and Regression

---

The Central Concept: Artificial Dichotomy

An artificial dichotomy (also called a forced or imposed dichotomy) occurs when a fundamentally continuous variable is split into two categories for measurement convenience.

Examples of Artificial Dichotomies

Variable	How Measured	Underlying Continuity
Mood	Happy (1) / Sad (0)	Continuous mood spectrum
Anxiety	High (1) / Low (0)	Continuous anxiety scale
Academic performance	Pass (1) / Fail (0)	Continuous marks distribution
Attitude	Positive (1) / Negative (0)	Continuous attitude continuum
Depression severity	Clinical (1) / Non-clinical (0)	Continuous symptom score

When both variables are artificially dichotomised, the true association between the underlying continuous variables is greater than what phi or point biserial would reveal.

---

Biserial Correlation (r_b)

What It Is

Biserial Correlation estimates the correlation between two continuous variables when one of them has been measured dichotomously (artificially split into two categories).

The formula corrects for the fact that dichotomising a continuous variable loses information, and therefore estimates what the Pearson's r would be if the artificially dichotomised variable had been measured on its true continuous scale.

Formula

$r_b = \frac{\bar{Y}_1 - \bar{Y}_0}{S_Y} \cdot \frac{P_0 \cdot P_1}{h}$

Where:

• $\bar{Y}_1$ = mean of Y scores for cases with X = 1
• $\bar{Y}_0$ = mean of Y scores for cases with X = 0
• $S_Y$ = standard deviation of Y (the continuous variable)
• $P_0$ = proportion of cases with X = 0
• $P_1$ = proportion of cases with X = 1
• $h$ = the ordinate (height) of the standard normal distribution at the point that divides the distribution at the proportion P₀ / P₁

The ordinate h is found from the standard normal distribution table — it is the y-value of the normal curve at the z-value that separates the two groups.

Relationship with Point Biserial

The biserial correlation and point biserial are related:

$r_b = \frac{r_{pb}\sqrt{P_0 \cdot P_1}}{h}$

This is useful because once you compute r_pb (simpler), you can easily derive r_b.

Important Properties

Property	Value
Range	Theoretically −1.00 to +1.00 (but can exceed ±1 in practice with extreme splits)
Interpretation	Like Pearson's r — direction and strength
When r_pb is known	Use the conversion formula above

Example: Mood and Hope

Suppose we measure:

• Y (continuous): Hope scores on BHS (Beck Hopelessness Scale)
• X (artificially dichotomous): Mood classified as low (0) or normal (1)

We assume mood is normally distributed in the population, but we've only measured it as a binary variable.

Using the biserial formula (or converting from r_pb), we get an estimate of the correlation between hope and mood as if mood had been measured on its full continuous scale — a more accurate reflection of the true relationship.

💡 Key Insight: Biserial correlation is always larger in absolute value than the corresponding point biserial. This is because dichotomising a continuous variable reduces variance and thus attenuates (lowers) the correlation. Biserial corrects for this attenuation.

---

Tetrachoric Correlation (r_tet)

What It Is

Tetrachoric Correlation estimates the correlation between two continuous underlying variables when both have been measured dichotomously (artificially split).

It answers the question: "If we had measured both variables on their true continuous scales, what would the Pearson's r be?"

Formula

The tetrachoric correlation is based on the angle between two vectors in multivariate normal space:

$r_{tet} = \cos\left(\frac{180°}{1 + \sqrt{\frac{ad}{bc}}}\right)$

Where a, b, c, d are the cell frequencies in a 2×2 contingency table:

	X = 0	X = 1
Y = 0	a	b
Y = 1	c	d

Worked Example: Attitude Towards Women and Liberalisation

Research question: Are attitudes towards women's rights and attitudes towards economic liberalisation related?

Both variables measured as: 0 = negative attitude, 1 = positive attitude.

Contingency table (n = 200):

	Attitude towards Women (X=0 negative)	Attitude towards Women (X=1 positive)	Row Sum
Attitude towards Liberalisation (Y=0 negative)	68 (a)	32 (b)	100
Attitude towards Liberalisation (Y=1 positive)	30 (c)	70 (d)	100
Column Sum	98	102	200

Calculation:

$r_{tet} = \cos\left(\frac{180°}{1 + \sqrt{\frac{ad}{bc}}}\right) = \cos\left(\frac{180°}{1 + \sqrt{\frac{68 \times 70}{30 \times 32}}}\right)$

$= \cos\left(\frac{180°}{1 + \sqrt{\frac{4760}{960}}}\right) = \cos\left(\frac{180°}{1 + \sqrt{4.958}}\right) = \cos\left(\frac{180°}{1 + 2.227}\right)$

$= \cos\left(\frac{180°}{3.227}\right) = \cos(55.78°) = 0.562$

Interpretation: There is a moderate positive correlation (r_tet = +0.562) between the underlying continuous latent attitudes towards women and liberalisation. People who tend to be more liberal on women's issues also tend to be more liberal on economic issues.

---

Biserial vs. Tetrachoric: A Comparison

Feature	Biserial (r_b)	Tetrachoric (r_tet)
Data type	One continuous + one artificial dichotomy	Both artificial dichotomies
Assumption	One underlying normal distribution	Both variables underlying normal distributions
What it estimates	Pearson's r for continuous × continuous	Pearson's r for continuous × continuous
Range	Can exceed ±1 in extreme cases	−1.00 to +1.00
Computation	Based on means and proportions	Based on 2×2 contingency table cell frequencies
Relationship to Pearson-type	Related to r_pb via conversion	Related to φ but estimates true underlying r

---

Choosing the Right Correlation: Complete Decision Tree

---

Practical Applications in Psychology

Biserial Correlation

1. Psychometric item analysis: Dichotomous items (correct/incorrect) correlated with ability scores when ability is assumed continuous — though r_pb is often used in practice for its simpler computation, r_b provides the theoretically correct estimate
2. Medical decision thresholds: A clinical "case/non-case" classification derived from a continuous severity measure, correlated with another continuous outcome
3. Developmental psychology: Age-based dichotomies (e.g., "reached developmental milestone: yes/no") correlated with cognitive test scores

Tetrachoric Correlation

1. Factor analysis of binary data: Tetrachoric correlations form the input correlation matrix when factor-analysing dichotomous survey items — crucial for attitude and personality scale development
2. Twin and genetic studies: Concordance data (both twins affected: yes/no) used with tetrachoric correlation to estimate genetic contributions
3. Educational measurement: Correlating pass/fail on two tests when both are cut-offs from continuous ability distributions

Indian Research Context 🇮🇳

• Attitude research at Indian universities: Tetrachoric correlation is used to compute factor structures when attitude survey items are dichotomous (agree/disagree format), helping establish the dimensionality of attitudes in Indian populations
• NIMHANS binary diagnostic outcomes: When two clinical diagnoses (yes/no) are thought to reflect underlying continuous dimensional pathology, tetrachoric correlation provides a more appropriate association measure than phi
• Agricultural extension and health research: Studies in rural India often measure outcomes as binary (adopted practice: yes/no; recovered: yes/no) — biserial or tetrachoric correlations correct for the artificial dichotomisation of what are fundamentally continuous outcomes

---

Memory Aids 🧠

"BINA" Rule for Biserial

• Biserial is used when one variable is a Binary disguise of a continuous variable
• Its estimate is Inflated compared to r_pb (which is attenuated by dichotomisation)
• Normal distribution assumed for the underlying variable
• Always larger (in absolute value) than r_pb

"TETRA" for Tetrachoric

• Two variables, Each Truly continuous but Recoded as binary → need Acorrection
• Uses cell frequencies a, b, c, d from 2×2 table
• Estimates the Pearson r that would exist if both were measured continuously

---

External Resources 🔗

Wikipedia Articles

• 📖 Biserial Correlation — Mathematical derivation and comparison with r_pb
• 📖 Tetrachoric Correlation Coefficient — History, formula, and use in factor analysis

Educational Resources

• 🎥 Tetrachoric Correlation in Factor Analysis (YouTube) — Application in scale development
• 📄 Tetrachoric and Polychoric Correlations in Item Analysis (Uebersax, 2000) — Comprehensive applied reference

Research Articles

• 📄 Tetrachoric vs. Pearson Correlation in Binary Data (Muthen, 1984) — Seminal comparison study
• 📄 When to Use Tetrachoric Correlation (Greer et al., 2006) — Practical guidelines for researchers

Online Tools

• 🔧 Tetrachoric Correlation Calculator (VassarStats) — Free online computation
• 🔧 Biserial Correlation Calculator — R package psych documentation
• 🔧 R psych package — polychoric/tetrachoric — Tetrachoric correlation in R

---

Self-Assessment Questions ✅

Recall Level

1. What is an "artificial dichotomy"? Give three examples from psychology.
2. What is the main assumption underlying biserial correlation that distinguishes it from point biserial?
3. When would you use tetrachoric rather than phi coefficient?

Understanding Level

4. A researcher measures depression as "clinically depressed" (1) vs. "not depressed" (0) by applying a cut-off to a continuous depression scale. She wants to correlate this with a continuous anxiety score. Should she use point biserial or biserial correlation? Why will these give different values?
5. Why is biserial correlation always larger in absolute value than the corresponding point biserial on the same data?
6. In what type of research is tetrachoric correlation particularly important, and why?

Application Level

7. A personality scale developer creates 20 yes/no items and wants to perform factor analysis. She needs a correlation matrix. Should she use Pearson's r, phi, or tetrachoric correlations as input? Justify your answer.
8. Given the contingency table below, calculate the tetrachoric correlation and interpret the result:

	High Stress (X=1)	Low Stress (X=0)
Poor sleep (Y=1)	45 (a)	15 (b)
Good sleep (Y=0)	20 (c)	70 (d)

Critical Thinking Level

9. A researcher computes both phi and tetrachoric correlation on the same 2×2 data. She finds φ = 0.35 and r_tet = 0.52. Why are these different, and which should she report if her research question is about the "true" relationship between the two underlying constructs?
10. Discuss the limitations of tetrachoric correlation. What assumptions must hold for it to give an accurate estimate of the true correlation?

---

Summary

Biserial correlation and Tetrachoric correlation are non-Pearson methods that correct for the artificial dichotomisation of variables that have underlying continuous distributions. Biserial (r_b) is used when one variable is continuous and the other is an artificial dichotomy; it is related to point biserial (r_pb) and always yields a larger absolute value. Tetrachoric correlation (r_tet) is used when both variables are artificial dichotomies; it estimates the Pearson's r that would exist if both were measured on their true continuous scales. Both are particularly important in factor analysis of binary data, attitude scale development, and clinical research where continuous constructs are force-categorised for practical reasons.

➡️ Next: Spearman's Rho and Kendall's Tau — Rank Order Correlations