Skip to main content

Special Correlations: Point Biserial and Phi Coefficient

Introduction​

Pearson's product moment correlation (r) is the workhorse of psychological statistics — but it requires continuous variables that ideally follow a normal distribution. What happens when one or both of your variables are dichotomous (taking only two values, like male/female, yes/no, pass/fail)?

This is where special types of correlation come in. Some of these are simply Pearson's r applied to dichotomously-coded data — they produce mathematically equivalent results but are given distinct names to reflect the nature of the data. Others are fundamentally different in their computational approach.

This unit introduces the first two such special correlations: Point Biserial Correlation (r_pb) and Phi Coefficient (φ) — both are Pearson's correlations adapted for dichotomous data. The distinction between these methods and when to apply each is critical for psychological research involving group comparisons, binary outcomes, and categorical predictors.

Source PDFs:

Special Types of Correlation: Overview​

When Pearson's standard assumptions are not met, alternative correlation methods are needed. These fall into three categories:

CategoryMethodsWhen Used
Pearson-type special correlationsPoint Biserial (r_pb), Phi (φ)One or both variables are truly dichotomous
Non-Pearson correlationsBiserial (r_b), Tetrachoric (r_tet)Dichotomous variables that have an underlying continuous distribution
Rank-order correlationsSpearman's rho (r_s), Kendall's tau (τ)Ordinal/rank data, or when Pearson's assumptions fail

Correlation vs. Association: An Important Distinction​

Following Howell (2002), there is a useful conceptual distinction:

  • Measures of correlation: Variables have an inherent order — higher scores represent more (or less) of something. Examples: number of friends, hopelessness scores, time taken to complete a task.
  • Measures of association: Variables are categorical/nominal with no inherent order. Examples: gender (male/female) × property ownership (own/don't own).

Despite this distinction, correlation formulas like phi coefficient are often used for dichotomous categorical variables in practice.

Dichotomous Variables: What Are They?​

A dichotomous variable can be divided into exactly two sharply distinguished, mutually exclusive categories. No underlying continuum is assumed.

True dichotomies (genuinely binary):

  • Gender: male / female
  • Diagnosis: diagnosed / not diagnosed with an illness
  • Location: rural / urban
  • Nationality: Indian / American
  • Group: experimental / control

Artificial dichotomies (underlying continuous variable, forcibly split):

  • Mood: classified as "low mood" or "normal mood" (actually a continuum)
  • Pass/fail: derived from a continuous marks scale
  • High anxiety / low anxiety: split at median from a scale

This distinction matters because:

  • True dichotomies → use Point Biserial or Phi Coefficient
  • Artificial dichotomies → use Biserial or Tetrachoric Correlation

Point Biserial Correlation (r_pb)​

Definition​

Point Biserial Correlation is the Pearson's product moment correlation computed between:

  • One truly dichotomous variable (coded as 0 and 1)
  • One continuous variable

The algebraic result is identical to Pearson's r: rpb=rr_{pb} = r

So r_pb is not a different formula — it IS Pearson's r, just applied to data where one variable is dichotomous. The name "point biserial" signals the nature of the data.

Coding the Dichotomous Variable​

When coding the dichotomous variable:

  • Assign any two distinct values (typically 0 and 1)
  • The specific values chosen (0 & 1, or 5 & 11) do not change the magnitude of r_pb
  • Only the sign changes depending on which group is coded as 0 vs. 1

💡 Interpretation of sign: The group coded as 1 has a higher mean on the continuous variable when r_pb is positive. When r_pb is negative, the group coded as 0 has the higher mean.

Calculating r_pb​

The formula is exactly Pearson's r:

rpb=CovXYSX⋅SYr_{pb} = \frac{Cov_{XY}}{S_X \cdot S_Y}

Where X is the dichotomous variable (0/1 coded) and Y is the continuous variable.

Worked Example: Gender and Marks​

Research question: Is gender related to examination marks?

Data: 20 subjects (9 male, 11 female) with marks in final exam.

  • Males coded as X = 0; Females coded as X = 1

Summary statistics from calculation:

StatisticValue
Mean (sex variable)0.55
Mean (marks)64.8
Mean marks (males, coded 0)60.88
Mean marks (females, coded 1)68.0
S_sex0.497
S_marks9.17
Cov_XY1.76

rpb=CovXYSX⋅SY=1.760.497×9.17=1.764.557=+0.386r_{pb} = \frac{Cov_{XY}}{S_X \cdot S_Y} = \frac{1.76}{0.497 \times 9.17} = \frac{1.76}{4.557} = +0.386

Interpretation:

  • r_pb = +0.386 → Moderate positive correlation
  • Since females are coded as 1 and r_pb is positive → females scored higher on average (68.0 vs 60.88)
  • Coefficient of determination: r_pb² = 0.386² = 0.149 → 14.9% of variance in marks is accounted for by gender

Significance Testing of r_pb​

Since r_pb IS Pearson's r, significance testing follows the same procedure — using the t-distribution with df = n − 2:

t=rpbn−21−rpb2t = \frac{r_{pb}\sqrt{n-2}}{\sqrt{1 - r_{pb}^2}}

Example (gender and marks data): t=0.38620−21−0.3862=0.386×4.2430.851=1.6380.922=1.775t = \frac{0.386\sqrt{20-2}}{\sqrt{1-0.386^2}} = \frac{0.386 \times 4.243}{\sqrt{0.851}} = \frac{1.638}{0.922} = 1.775

df=20−2=18df = 20 - 2 = 18

Decision: At α = .05, two-tailed, critical t at df = 18 is approximately 2.101. Since 1.775 < 2.101, we fail to reject H₀.

Conclusion: The correlation between gender and marks is not statistically significant at the 0.05 level. (Note: this is largely due to the small sample of n = 20.)

Hypotheses: H0:ρ=0(no relationship in population)H_0: \rho = 0 \quad \text{(no relationship in population)} HA:ρ≠0(relationship exists in population)H_A: \rho \neq 0 \quad \text{(relationship exists in population)}

📌 With small samples, even a moderate correlation may not reach significance. Larger samples are needed to reliably detect the effect.

Phi Coefficient (φ)​

Definition​

The Phi Coefficient is the Pearson's product moment correlation computed when both variables are truly dichotomous (each coded 0/1).

Like r_pb, φ is algebraically identical to Pearson's r: φ=rXY(when both X and Y are dichotomous)\varphi = r_{XY} \quad \text{(when both X and Y are dichotomous)}

Example: Gender and Property Ownership​

Research question: Is gender related to property ownership?

  • X: Gender (0 = Male, 1 = Female)
  • Y: Ownership (0 = No ownership, 1 = Ownership)

Data: 12 subjects (dichotomous scores on both variables)

Summary calculations:

StatisticValue
Mean(X) = Mean(sex)0.50
S_X0.52
Mean(Y) = Mean(ownership)0.58
S_Y0.51
Cov_XY−0.13

φ=CovXYSX⋅SY=−0.130.52×0.51=−0.130.265=−0.465\varphi = \frac{Cov_{XY}}{S_X \cdot S_Y} = \frac{-0.13}{0.52 \times 0.51} = \frac{-0.13}{0.265} = -0.465

Interpretation:

  • φ = −0.465 → Moderate negative correlation
  • Males (coded 0) → more ownership; Females (coded 1) → less ownership
  • r² = φ² = 0.465² = 0.216 → 21.6% shared variance between gender and ownership

⚠️ The sign of φ depends entirely on which group is coded as 0 vs. 1. Once coding is fixed, the sign IS interpretable: here, "as we move from male to female (0→1), we move in the negative direction on ownership (1→0), meaning males have more ownership."

Significance Testing of Phi (φ): Chi-Square Method​

Unlike r_pb (which uses t-distribution), the significance of φ is tested using the Chi-Square (χ²) distribution:

χ2=nφ2\chi^2 = n\varphi^2

This Chi-Square has df = 1.

Example: χ2=nφ2=12×0.216=2.59\chi^2 = n\varphi^2 = 12 \times 0.216 = 2.59

Decision: At df = 1, critical χ² = 3.84 (α = .05). Since 2.59 < 3.84, we fail to reject H₀.

Conclusion: The relationship between gender and property ownership is not statistically significant — but this is primarily due to the small sample (n = 12). With a larger sample, the same φ of −0.465 would likely reach significance.

Relationship Between φ and χ²​

φ=χ2n\varphi = \sqrt{\frac{\chi^2}{n}}

This means you can compute φ from a Chi-Square test result — a common practice in psychological research using 2×2 contingency tables.

Comparing Point Biserial and Phi Coefficient​

FeaturePoint Biserial (r_pb)Phi Coefficient (φ)
Variable typesOne dichotomous + one continuousBoth dichotomous
FormulaPearson's rPearson's r
Range−1.00 to +1.00−1.00 to +1.00
Significance testt-distribution, df = n−2Chi-square, df = 1
% variancer_pb² × 100φ² × 100
ExampleGender × GPAGender × Pass/Fail
Sign interpretationGroup coded 1 has higher meanDepends on coding
IGNOU MAPC relevanceComparing groups on scale scores2×2 categorical relationships

Practical Applications in Psychology​

Point Biserial — Common Uses​

  1. Item-total correlation in psychometric scale development: correlating each dichotomous item (correct/incorrect) with the total test score — a key step in item analysis
  2. Group comparisons with continuous outcomes: clinical vs. control group on a symptom scale; trained vs. untrained on a performance measure
  3. Diagnostic validity studies: does a binary diagnosis (yes/no) correlate with severity score?

Phi Coefficient — Common Uses​

  1. 2×2 contingency table analysis in epidemiology and clinical psychology
  2. Correlation between two binary personality traits: introverted/extraverted × anxious/calm
  3. Survey research: agreement (yes/no) on two binary attitude items
  4. Genetic and diagnostic studies: presence/absence of two risk factors

Indian Research Context 🇮🇳​

  • NEET/entrance exam research: Point biserial correlation is the standard method for item analysis in competitive exam question banks in India — each MCQ item (correct = 1, incorrect = 0) is correlated with the total score to assess item discrimination
  • NIMHANS clinical studies: Phi coefficient used to assess co-occurrence of two binary diagnoses (e.g., depression and anxiety disorder) in clinical samples
  • NCERT assessments: Educational psychologists at NCERT use r_pb to evaluate the effectiveness of individual test items in standardised achievement tests

Memory Aids 🧠​

"PDCP" — When to Use Which Special Correlation​

  • Pearson: both variables continuous
  • Dichotomous + Continuous → Point Biserial
  • Categorical + Categorical (both dichotomous) → Phi Coefficient
  • Pseudo-dichotomous (underlying continuum) → Biserial or Tetrachoric

Point Biserial Sign Rule: "1 is Higher"​

If r_pb is positive, the group coded 1 has the higher mean on the continuous variable.

Phi and Chi-Square: "Square Root of Chi Over n"​

φ=χ2n\varphi = \sqrt{\frac{\chi^2}{n}} This formula links phi to the Chi-Square test — both assess independence of two categorical variables.

Significance Tests: "r_pb uses t; φ uses chi"​

  • Point biserial → t-test (like Pearson's r)
  • Phi → Chi-square test (like contingency table analysis)

External Resources 🔗​

Wikipedia Articles​

Educational Videos​

Research Articles​

Online Tools​

Self-Assessment Questions ✅​

Recall Level​

  1. What is the key difference between Point Biserial Correlation and Phi Coefficient?
  2. Name three examples of truly dichotomous variables in psychology.
  3. What formula is used to test the significance of the Phi Coefficient?

Understanding Level​

  1. A researcher codes "pass" as 1 and "fail" as 0, and finds r_pb = +0.55 between exam outcome and study hours. What does this mean? If the researcher had coded pass = 0, fail = 1, what would r_pb be?
  2. Why does the choice of 0/1 coding affect the sign but not the magnitude of Point Biserial Correlation?
  3. A chi-square test on a 2×2 table yields χ² = 8.40 with n = 120. Compute the phi coefficient and interpret its meaning.

Application Level​

  1. A clinical psychologist wants to know if a diagnosis of ADHD (yes=1, no=0) predicts attention test scores (continuous scale). Which correlation method should she use, and why? Write out the null hypothesis and describe the significance testing procedure.
  2. In an Indian school study, 50 students' pass/fail status (dichotomous) is correlated with their attendance category (regular=1, irregular=0). n = 50, φ = +0.42. (a) What is the percentage of shared variance? (b) Test significance using χ². (c) What does this finding imply for school administrators?

Critical Thinking Level​

  1. A psychologist uses Point Biserial Correlation to relate gender (male=0, female=1) to a clinical anxiety scale in n = 200 adults, obtaining r_pb = +0.25, p < .01. A newspaper headline reads "Women Are 25% More Anxious Than Men." What is wrong with this interpretation?
  2. Discuss the conditions under which the Phi Coefficient would be preferred over Pearson's r for two genuinely dichotomous variables. Would there ever be a reason to use Pearson's r instead? Explain.

Summary​

When one or both variables in a correlation analysis are truly dichotomous, standard Pearson's r is modified — but not fundamentally altered. Point Biserial Correlation (r_pb) is used when one variable is dichotomous and the other is continuous; it IS Pearson's r and uses the t-distribution for significance testing. Phi Coefficient (φ) is used when both variables are dichotomous; it IS also Pearson's r, but its significance is tested using Chi-Square with df = 1. Both are Pearson-type correlations, both share the −1 to +1 range, and both yield a coefficient of determination (r²) for percentage of shared variance. The sign in both cases depends on how the 0/1 coding is assigned and must be interpreted accordingly.

➡️ Next: Biserial and Tetrachoric Correlations — Non-Pearson Methods for Underlying Continuous Distributions