03 · What You Need to Know
Calculating Sample Size for Correlation Estimation
What Is the Parameter You Want to Estimate?
The Pearson population correlation coefficient, represented by ρ, measures the strength and direction of a linear association between two numerical variables.
Its values range from −1 to +1. Positive values indicate positive linear association, negative values indicate negative linear association, and zero indicates the absence of population linear correlation.
The sample Pearson correlation, represented by r, estimates ρ.
Before planning sample size, researchers should distinguish Pearson correlation from Spearman rank correlation, Kendall's tau, partial correlation, and other measures of association. These estimands have different sampling properties.
A correlation also does not establish causation. Confounding, restricted ranges, nonlinear relationships, and measurement error can affect its interpretation.
Why Is a Confidence Interval Important for Correlation?
A sample correlation is a point estimate subject to sampling uncertainty.
Suppose a study reports r = 0.30. Without an uncertainty interval, readers cannot readily judge whether the population correlation might plausibly be near 0.10 or closer to 0.50.
These values may imply substantially different interpretations depending on the research question.
For estimation-focused studies, sample-size planning should therefore consider the confidence interval width needed to answer the research question, rather than relying exclusively on whether the correlation differs statistically from zero.
Why Is Fisher's z Transformation Used?
The sampling distribution of Pearson's r is not generally symmetric, particularly when the population correlation is far from zero.
Fisher's transformation maps correlations from the interval between −1 and +1 onto an unbounded scale. Under the conventional bivariate-normal approximation, the transformed sample correlation is approximately normally distributed.
How Do You Calculate Sample Size for a Target Width on Fisher's z Scale?
If the desired precision is specified as a confidence-interval half-width H on Fisher's z scale, the normal approximation gives a simple sample-size expression.
The calculation is convenient, but researchers often care about precision in the original correlation units. That requires an additional step.
How Do You Plan for a Confidence Interval Width on the Correlation Scale?
For an anticipated population correlation ρ0, first transform the anticipated value using Fisher's transformation.
Next, calculate the approximate lower and upper limits on the transformed scale for a proposed sample size. Finally, transform those limits back using the hyperbolic tangent function.
The resulting approximate interval endpoints are:
To identify the required sample for a desired width on the r scale, researchers can evaluate successive sample sizes until U − L is no greater than the specified target.
Software can perform this search directly. The method should be documented, including the anticipated correlation and the interval procedure.
Does the Anticipated Correlation Affect Required Sample Size?
Yes, when precision is defined on the original correlation scale.
Fisher's transformation is nonlinear. A fixed interval width on the transformed scale corresponds to different widths on the correlation scale depending on the anticipated correlation.
Near zero, the transformation changes relatively slowly. Near −1 or +1, the back-transformation compresses distances on the correlation scale.
Consequently, a target width on the r scale can require different sample sizes at different anticipated population correlations.
This does not mean that a strong anticipated correlation is inherently easier to estimate in every practical setting. The conclusion depends on the estimator, distributional assumptions, and chosen precision criterion.
What If the Expected Correlation Is Unknown?
Researchers may use comparable previous studies, meta-analytic evidence, or a plausible range of correlations to inform planning.
If reliable evidence is unavailable, evaluating several values may be preferable to selecting an arbitrary correlation.
For a fixed absolute width on the correlation scale under the Fisher approximation, values near zero often provide a demanding planning scenario compared with correlations closer to the boundaries.
However, researchers should not automatically assume ρ = 0 without considering the research objective and relevant evidence.
Using a range of plausible planning assumptions can show how the required sample changes across candidate correlations.
Is Estimating a Correlation Different From Testing Whether It Is Zero?
Yes. A test of the null hypothesis ρ = 0 asks whether the data provide sufficient evidence against zero population correlation.
A precision-based study asks how accurately the population correlation can be estimated.
A sample may have adequate power to reject a zero-correlation hypothesis for a specified alternative while producing a confidence interval too broad for meaningful interpretation.
Conversely, a precision-focused study may estimate a weak correlation with useful accuracy even when the interval includes zero.
The distinction between precision-based and power-based sample-size planning should be resolved before choosing a formula.
Can You Use the Fisher Formula for Spearman Correlation?
Not without qualification. Spearman's rank correlation measures monotonic association using ranks, whereas Pearson's correlation measures linear association.
Their sampling distributions and variance properties differ, particularly with ties, discrete outcomes, and nonstandard distributions.
Some approximations for rank correlations resemble Fisher-transformation methods, but their accuracy depends on assumptions and the intended confidence-interval procedure.
For Spearman correlation, researchers should use a planning method specifically justified for that estimator. Simulation may be appropriate when ties, skewed distributions, or other complications materially affect precision.
What If the Data Are Clustered or Repeated?
The conventional Fisher approximation assumes independent paired observations. It is not directly appropriate when several observations come from the same participant or when students are clustered within classrooms.
Repeated-measures correlations, within-person associations, and multilevel correlations target different quantities and require methods that reflect the dependence structure.
A dataset containing 500 observations from 50 participants does not necessarily provide the same information as 500 independent participant pairs.
Researchers should identify the level at which the association is defined and use a sample-size method corresponding to that estimand.
How Do Measurement Error and Restricted Range Affect Correlation Planning?
Measurement error can attenuate observed correlations under common classical measurement-error assumptions. Restricted variability in one or both variables may also change the observed association.
For example, a study restricted to very high-achieving students may produce a different correlation between study time and examination scores than a study involving the full student population.
Increasing sample size improves precision for the correlation in the sampled population, but it does not automatically correct measurement error or restore the association that would exist in a broader population.
These concerns should inform both the sampling design and the interpretation of the resulting coefficient.
How Should Missing Paired Observations Be Handled?
Correlation estimation generally requires usable information on both variables for each observational unit under a complete-case Pearson analysis.
If some participants provide only one measurement, the number of complete pairs may be smaller than the number recruited.
Researchers should therefore distinguish the required number of usable paired observations from the initial recruitment target.
More advanced missing-data methods may use partially observed information under appropriate assumptions, but their precision cannot necessarily be determined using the simple complete-pair formula.
Additional recruitment does not automatically remove bias caused by systematic missingness.
Should You Conduct a Sensitivity Analysis?
Yes, when the anticipated correlation, missing-data rate, or distributional assumptions are uncertain enough to affect the design.
A sample-size sensitivity analysis can compare required sample sizes under several anticipated correlations and precision targets.
For complex estimators, simulation can evaluate the expected confidence-interval width or the probability of achieving a target width.
As with other estimation studies, the final sample-size justification should explain why the selected precision is sufficient for the intended scientific interpretation.
