STANDARD ERROR OF CORRELATION 1

A Brief Note on the Standard Error of the Pearson Correlation

Timo Gnambs

Leibniz Institute for Educational Trajectories, Germany

Author Note

Timo Gnambs https://orcid.org/0000-0002-6984-1276

I have no conflicts of interest to disclose. The study was not preregistered. The simulation

syntax and statistical results are provided at https://osf.io/m6srf.

Correspondence concerning this article should be addressed to Timo Gnambs, Leibniz

Institute for Educational Trajectories, Wilhelmsplatz 3, 96047 Bamberg, Germany, E-mail:

[email protected].

Version: 2023-07-14
STANDARD ERROR OF CORRELATION 2

Abstract

The product-moment correlation is a central statistic in psychological research including

meta-analysis. Unfortunately, it has a rather complex sampling distribution which leads to

sample correlations that are biased indicators of the respective population correlations.

Moreover, there seems to be some uncertainty on how to properly calculate the standard error

of these correlations. Because no simple analytical solution exists, several approximations

have been previously introduced. This note aims to briefly summarize 10 different ways to

calculate the standard error of the Pearson correlation. Moreover, a simulation study on the

accuracy of these estimators compared their relative percentage biases for different population

correlations and sample sizes. The results showed that all estimators were largely unbiased for

sample sizes of at least 40. For smaller samples, a simple approximation by Bonett (2008) led

to the least biased results. Based on these results, it is recommended to use the expression

(1 − 𝑟 2 )⁄√𝑁 − 3 for the calculation of the standard error of the Pearson correlation.

Keywords: correlation; sampling distribution; standard error

STANDARD ERROR OF CORRELATION 3

A Brief Note on the Standard Error of the Pearson Correlation

The product-moment correlation (Pearson, 1896) is an essential statistic to quantify the

linear association between two variables. It is not only an elementary component of

descriptive and exploratory research but is also used as a measure of effect size to facilitate

interpretations and comparisons of results between studies. Therefore, it also plays a pivotal

role in research syntheses including quantitative meta-analyses (e.g., Cheung, 2015; Hafdahl,

2008). Although it is well known among statisticians that the sample correlation r represents a

biased estimator of the population correlation ρ (e.g., De Winter et al., 2016; Hedges, 1989;

Olkin & Pratt, 1958; Shieh, 2010; Zimmerman et al., 2003), applied researchers seldom adopt

unbiased estimators of ρ because the bias in r is often considered increasingly negligible with

larger sample sizes (Shieh, 2010). Moreover, there seems to be some uncertainty regarding

the calculation of the standard error of r that is often used in meta-analytic investigations as

precision weights and, thus, determines the contribution of each correlation coefficient to the

pooled estimate (e.g., Bonett, 2008; Hafdahl, 2008). Because ρ follows a rather complex

sampling distribution, the standard error is rarely calculated from the exact distribution.

Rather, various large-sample approximations have been suggested in the literature (e.g.,

Bonett, 2008; Hedges, 1989; Hotelling, 1953). However, if these approximations represent

biased estimators of standard errors, their use in meta-analyses might involuntarily distort

effect size estimations and interpretations. Therefore, this brief note reviews different

approaches on how to calculate the standard error of r and demonstrates their biases for

different values of r and sample sizes.

The Sampling Distribution of the Pearson Correlation

Let X = {x1, x2, …, xN} and Y = {y1, y2, …, yN} represent two bivariate normally distributed

variables with population means {μX, μY}, variances {σ𝑋2 , σ2𝑌 }, and correlation ρ that are

measured in a sample of N elements. Then, the Pearson product-moment correlation is given

by
STANDARD ERROR OF CORRELATION 4

∑𝑁
𝑖=1(𝑥𝑖 − μ
̂ 𝑋 )(𝑦𝑖 − μ̂𝑌 )
𝑟𝑋𝑌 =
(1)
√∑𝑁 ̂ 𝑋 )2 ∑𝑁
𝑖=1(𝑥𝑖 − μ ̂ 𝑌 )2
𝑖=1(𝑦𝑖 − μ

with μ̂𝑋 and μ̂𝑌 as the sample means of X and Y. The density probability distribution of r has

been derived by Hotelling (1951, 1953) based on prior work in Fisher (1915, 1921) and

follows a rather curious shape,

(𝑁 − 2) ∙ Γ(𝑁 − 1) ∙ (1 − ρ2 )(𝑁−1)/2 ∙ (1 − 𝑟 2 )(𝑁−4)/2

𝑓(𝑟|ρ, 𝑁) =
√2𝜋 ∙ Γ(𝑁 − 1⁄2) ∙ (1 − ρ𝑟)𝑁−3⁄2 (2)
1 1 2∙𝑁−1 ρ∙𝑟+1
∙ 2𝐹1 ( , ; ; )
2 2 2 2
involving the gamma function Γ(x) = (x - 1)! and Gauss’s hypergeometric function

(𝑎)𝑖 (𝑏)𝑖 𝑥 𝑖 Γ(𝑚+1)

2𝐹1 (𝑎, 𝑏; 𝑐; 𝑥) = ∑=∞
𝑖=0 with (𝑚)𝑖 = Γ(𝑚−𝑖+1). For demonstration, Figure 1 visualizes these
(𝑐)𝑖 𝑖!

distributions for two correlations, ρ = .20 and ρ = .80. Whereas the former represents a small

to medium effect that is typically observed in different areas of psychology (see Bosco et al.,

2015; Gignac & Szodorai, 2016; Lovakov & Agadullina, 2021), the latter reflects a rather

large effect that is often limited to specific domains such as competence research (e.g.,

Gnambs & Lockl, 2023). The distributions in Figure 1 highlight that the sample correlations r

follow a rather asymmetrical shape. In the present case, the modes of these distributions are

larger than the respective ρ, thus, resulting in a negative skew. Generally, the skew is more

pronounced for larger |ρ| because correlations are bounded at -1 and +1. In contrast, smaller

correlations often exhibit approximately normal distributions. Moreover, the asymmetry of

the sampling distribution is strongly affected by the sample size. It is stronger for small

sample sizes, while larger samples result in more symmetric distributions. In the presented

examples, already a sample size of N = 50 leads to an approximately normal sample

distribution, even for ρ = .80.

The estimator of the population correlation

Because of its skewed sampling distribution, the sample correlation r is a biased estimator

of the population correlation ρ (e.g., Hedges, 1989; Olkin & Pratt, 1958). As pointed out by

Zimmerman and colleagues (2013), this bias can reach up to .03 or .04 in many applied
STANDARD ERROR OF CORRELATION 5

situations. Therefore, Olkin and Pratt (1958) derived an estimator of ρ from (2) that corrects

for the bias in r as

1 1 2∙𝑁−2 𝑟 ∙ (1 − 𝑟 2 )
ρ̂ = 𝑟 ∙ 2𝐹1 ( , ; ; 1 − 𝑟 2) ≈ 𝑟 + (3)
2 2 2 2 ∙ (𝑁 − 4)
with ≈ indicating an approximation that is accurate within ± .001 for N ≥ 18. Monte Carlo

simulations confirmed that the approximate estimator in (3) is largely unbiased for different

sample sizes and population correlations, whereas r tends to underestimate ρ, particularly for

medium- to large-sized correlations and small sample sizes (Shieh, 2010). However, r tends to

exhibit a higher precision for |ρ| < .60 as reflected by the mean squared error. Therefore, in

these cases, the sample correlation r might, despite its bias, serve as a meaningful estimator of

the strength of the association between two variables.

Estimators of the sampling variance

The sampling variance of r was derived by Hotelling (1951, 1953) using the moments of r

about its mean as

1
σ2𝑟 = λ2 − λ12 with λℎ = ∫−1(𝑟 − ρ)ℎ ∙ 𝑓(𝑟|ρ, 𝑁 − 1) 𝑑𝑟.
(4)

Consequently, the standard error of r is σ𝑟 = √σ2𝑟 . Because (4) requires integrating over the

hypergeometric function, no simple analytic solution exists. Therefore, various

approximations have been suggested in the literature that often take the general form

(1 − 𝑟 2 )2
σ2𝑟 ≈ ∙ 𝑂(𝑛) (5)
𝑑𝑓
with df representing the degrees of freedom and O(n) a series of terms of the order n that

approximates the integral over the hypergeometric function in (4). For n → ∞, the estimator in

(5) approaches (4). An overview of 10 thus derived approximations is given in Table 1.

Originally, Pearson (1896) proposed df = N ⸱ (1 + ρ2) resulting in (6). However, he soon

noticed that this referred to a rather special case where the variances of X and Y are known

(see Pearson & Filon, 1898) and therefore revised the expression to df = N as in (7). But, the

expression in (8) with df = N – 1 provides an even more accurate estimator of σ𝑟 by assuming

STANDARD ERROR OF CORRELATION 6

estimated means in (1) because (7) implies that the means of the two variables in (1) are

known, that is, μ̂𝑋 = μ𝑋 and μ̂𝑌 = μ𝑌 (Olkin & Pratt, 1985). Moreover, simulation studies led

to Bonett’s (2008) suggestion of df = N – 3 in (9). The formulas in (7) to (9) are frequently

used in applied research because they are easy to calculate and give good approximations of

(4). However, they are biased to some degree because they ignore O(n) for the approximation

of the integrated hypergeometric function.

Several authors tried to give analytic solutions for O(n) and different n. Soper (1913)

derived O(1) for df = N and df = N – 1 resulting in (10) and (11). The latter was later extended

by Hotelling (1953) to O(2) as in (12). Moreover, Ghosh (1966) independently presented an

approximation of O(6) resulting in (13). The formulas in (10) to (13) are expected to provide

better estimators of (4) because of their closer approximation of the sampling distribution of r.

However, they are hardly used anymore today because researchers trying to improve the

accuracy of the estimated standard errors as compared to (7) to (9) can easily do so by directly

evaluating the integral in (4) using modern optimization routines implemented in standard

statistical software.

Hedges (1989) presented an alternative approach in the context of meta-analytic studies

based on the assumption σ2 (ρ𝑖 ) = σ2 (ρ) + σ2 (𝑒𝑖 ) derived from classical test theory. Here,

the variance σ2 (ρ𝑖 ) in study i is decomposed into the variance σ2 (ρ) of the distribution from

which the study-specific population correlations ρ𝑖 are sampled and sampling error σ2 (𝑒𝑖 ).

̂2𝑖 can be
Consequently, an estimator of the standard error σ(𝑒𝑖 ) = σ(ρ𝑖 ) − σ(ρ) = ρ̂2𝑖 − ρ

obtained by using unbiased estimates of ρ as given in (4) and the multiple correlation ρ2 as

derived by Olkin and Pratt (1958). This results in the approximate and exact expressions of

the unbiased standard error given in (14) and (15), respectively.

Finally, it might be helpful to illuminate a potential misconception arising from the fact

that the bivariate correlation can also be expressed in terms of the standardized linear
STANDARD ERROR OF CORRELATION 7

(𝑏−β)
regression coefficient. The studentized regression coefficient 𝑡 = with b and β as the
σ𝑏

estimated and true regression weights, respectively, and σb as the standard error of b follows a

𝑏 𝑟
t distribution with df = N - 2. Under the null hypothesis of β = 0, this reduces to 𝑡 = σ = σ =
𝑏 𝑏

−1
1−𝑅2
𝑟 ∙ (√ 𝑁−2 ) with R2 as the coefficient of determination (see Pugh & Winslow, 1966, for the

detailed derivation). Because for a single predictor R2 = r2, it might be tempting to mistake the

1−𝑟 2
term √ 𝑁−2 for the standard error of r. However, this is only true for the special case when ρ =

0, whereas this expression leads to increasingly biased estimators of σr for larger |ρ|.

Comparison of Estimators of the Standard Error

As highlighted in Table 1 different approximations have been suggested for the estimation

of the standard error of r. However, little is known about which of these estimators might

yield noteworthy benefits in substantive research. Therefore, the accuracy and efficiency of

these estimators were compared for different population correlations and sample sizes.

Methods

The comparison varied the population correlations between ρ = .00 and .90 (in increments

of .10) and sample sizes N from 10 to 50 (in increments of 10) and 100. For each condition,

six types of standard errors σ

̂𝑟 for r were examined either following (7) in Pearson and Filon

(1898), (8) in Soper (1913), (9) in Bonett (2008), (4) using the integral by Hotelling (1953)

with ρ̂ = 𝑟, or (15) in Hedges (1989). Moreover, also the standard error of a regression

coefficient was derived to demonstrate its inadequacy for the correlation coefficient. The

approximations in (10) to (13) were not considered because they are rarely used in practice

and, more importantly, are superseded by the direct evaluation of the integral in (4). The

performance of these estimators was compared using the relative percentage bias (Hoogland

̂𝑟 , σ𝑟 ) = 𝐸(σ
& Boomsma, 1998) which is given as %𝐵𝑖𝑎𝑠(σ ̂𝑟 − σ𝑟 )⁄σ𝑟 ∙ 100. Values of

%Bias less than 5% were considered negligible. The true standard error σr was calculated
STANDARD ERROR OF CORRELATION 8

following (4) using the population correlation ρ from the data-generating process. Moreover,

the efficiency of the estimators was studied using the root mean squared error,

̂𝑟 − σ𝑟 )2 ]. Because the RMSE can be decomposed into the squared bias

̂𝑟 , σ𝑟 ) = 𝐸[(σ
𝑅𝑀𝑆𝐸(σ

and the variance of σ

̂𝑟 , it represents a trade-off between the two components. Thus, an

estimator might be more biased but at the same time also more efficient if it has a smaller

variance.1 The Bias and RMSE were computed by numerical integration using adaptive
1
Simpson’s quadrature with respect to r, that is, 𝐵𝑖𝑎𝑠(σ
̂𝑟 , σ𝑟 ) = ∫−1(σ
̂𝑟 − σ𝑟 ) ∙ 𝑓(𝑟|ρ, 𝑁) 𝑑𝑟

1
and 𝑅𝑀𝑆𝐸 2 (σ ̂𝑟 − σ𝑟 )2 ∙ 𝑓(𝑟|ρ, 𝑁) 𝑑𝑟.
̂𝑟 , σ𝑟 ) = ∫−1(σ

Results and Discussion

The relative bias of the different estimators of σr for different sample sizes is summarized

in Figure 2. These results show little differences between the compared estimators for sample

sizes of N = 40 or larger. Except for the standard error of the regression coefficient, the

approximations of the standard error of the correlation yielded largely unbiased estimates. In

contrast, at a small sample size such as N = 10 or 20 more pronounced differences were

observed. Estimators using N (Pearson & Filon, 1898) or N - 1 (Soper, 1913) as degrees of

freedom resulted in negative relative biases that increased for larger population correlations.

In contrast, the estimator by Bonett (2008) with N – 3 degrees of freedom resulted in unbiased

estimates of the standard error across a wide range of correlations; only for very large

correlations a slight negative bias was observed. Similarly, an evaluation of the integral in (4)

led to comparably unbiased estimates for sample sizes of at least 20. However, in extremely

small samples it was less precise than the approximation by Bonett (2008), presumably

because Equation (4) uses the estimate of the population correlation multiple times which is

estimated rather imprecisely in small samples. Interestingly, the approximation by Hedges

1
The simulation only considered correlations ρ ≥ 0 because the bias for negative correlations is simply the
opposite of the bias for positive correlations, that is, 𝐵𝑖𝑎𝑠(σ ̂𝑟 , −σ𝑟 ) = −𝐵𝑖𝑎𝑠(σ ̂𝑟 , σ𝑟 ), while the root mean
squared error is identical in both cases, that is, 𝑅𝑀𝑆𝐸(σ ̂𝑟 , −σ𝑟 ) = 𝑅𝑀𝑆𝐸(σ ̂𝑟 , σ𝑟 ).
STANDARD ERROR OF CORRELATION 9

(1989) was only unbiased for correlations up to about .60; larger correlations resulted in a

slightly negative bias. These results also emphasize that - independent of the sample size - the

standard error of the regression coefficient yielded unbiased results only for population

correlations close to 0. For larger correlations, the relative percentage bias increased up to

125%. Thus, mistakenly using this standard error as an indicator of precision might result in

severely distorted meta-analytic summaries of product-moment correlations.

The root mean squared error resulted in only marginal differences between the compared

estimators (see Figure 2). For samples of N = 50 or N = 100, the RMSE fell below .001 in all

conditions. Although it was slightly larger for smaller sample sizes, the RMSE did not indicate

pronouncedly different efficiencies of the studied estimators. Except for N = 10, the different

estimators resulted in comparable RMSEs. But again, the standard error of the regression

coefficient was less efficient at larger population correlations. Together, these results indicate

that the simple approximation by Bonett (2008) results in the least biased estimates of the

standard error for different values of ρ and sample sizes. More complex estimators that either

require the evaluation of the integral in (4) or rely on the hypergeometric function such as

(15), do not seem to provide noteworthy benefits for the accuracy of the standard error.

Conclusion

Despite its popularity in applied research, the distributional properties of the sample

correlation are often not well understood or simply neglected. Particularly, the calculation of

the standard error of the Pearson correlation remains challenging because of the complex

sampling distribution of r which does not give a simple analytical solution. Therefore, various

approximations are currently used in substantive research. Although some of the more

complex expressions in Table 1 have only historical value today because the integration of

complex functions became substantially easier with modern computers, it was unclear

whether the choice between the simpler approaches might matter. Therefore, the simulation

evaluated the accuracy of these estimators for different population correlations and sample
STANDARD ERROR OF CORRELATION 10

sizes. The respective findings suggest that the least biased estimator used the expression

(1 − 𝑟 2 )⁄√𝑁 − 3 (Bonett, 2008). Particularly, for larger population correlations in small

samples, this estimator should result in more precise standard errors. However, it needs to be

emphasized that differences between estimators become negligible as soon as sample sizes

increase. At typical sample sizes in psychology that often exceed 50 or 100 respondents the

choice of the estimator is unlikely to substantially matter. Thus, the estimator in (8) that is

currently often used in practice should be usually also acceptable in many applied situations.

However, further research is needed to identify specific conditions under which biased

estimators might yield non-negligible consequences such as in meta-analytic applications with

small samples.
STANDARD ERROR OF CORRELATION 11

Declarations

Author contributions

TG conceived the study, conducted the analyses, and wrote the manuscript.

Funding

No funding was received to assist with the preparation of this manuscript.

Conflicts of Interest

The author has no competing interests to declare that are relevant to the content of this

article.

Availability of data and materials

Not applicable.

Code availability

The computer code for the simulation is available at https://osf.io/m6srf.

STANDARD ERROR OF CORRELATION 12

References

Bosco, F. A., Aguinis, H., Singh, K., Field, J. G., & Pierce, C. A. (2015). Correlational effect

size benchmarks. Journal of Applied Psychology, 100(2), 431–449.

https://doi.org/10.1037/a0038047

Bonett, D. G. (2008). Meta-analytic interval estimation for bivariate correlations.

Psychological Methods, 13, 173-189. https://doi.org/10.1037/a0012868

Cheung, M. W.-L. (2015). Meta-Analysis: A Structural Equation Modeling Approach. Wiley.

De Winter, J. C., Gosling, S. D., & Potter, J. (2016). Comparing the Pearson and Spearman

correlation coefficients across distributions and sample sizes: A tutorial using simulations

and empirical data. Psychological Methods, 21(3), 273.

https://doi.org/10.1037/met0000079

Fisher, R. A. (1915). Frequency distribution of the values of the correlation coefficient in

samples from an indefinitely large population. Biometrika, 10, 507-521.

https://doi.org/10.2307/2331838

Fisher, R. A. (1921). On the ‘probable error’ of a coefficient of correlation deduced from a

small sample. Metron, 1, 1-32. https://hdl.handle.net/2440/15169

Ghosh, B. K. (1966). Asymptotic expansions for the moments of the distribution of

correlation coefficients. Biometrika, 53(1), 258-262. https://doi.org/10.2307/2334076

Gignac, G. E., & Szodorai, E. T. (2016). Effect size guidelines for individual differences

researchers. Personality and Individual Differences, 102, 74–78.

https://doi.org/10.1016/j.paid.2016.06.069

Gnambs, T., & Lockl, K. (2023). Bidirectional effects between reading and mathematics

development across secondary school. Zeitschrift für Erziehungswissenschaft, 26, 345-371.

https://doi.org/10.1007/s11618-022-01108-w
STANDARD ERROR OF CORRELATION 13

Hafdahl, A. R. (2008). Combining heterogeneous correlation matrices: Simulation analysis of

fixed-effects methods. Journal of Educational and Behavioral Statistics, 33(4), 507-533.

https://doi.org/10.3102/1076998607309472

Hedges, L.V. (1989). An unbiased correction for sampling error in validity generalization

studies. Journal of Applied Psychology, 74(3), 469-477. https://doi.org/10.1037/0021-

9010.74.3.469

Hoogland, J. J., & Boomsma, A. (1998). Robustness studies in covariance structure modeling:

An overview and a meta-analysis. Sociological Methods & Research, 26, 329–367.

https://doi.org/10.1177/0049124198026003003

Hotelling, H. (1951). The impact of R. A. Fisher on statistics. Journal of the American

Statistical Association, 46, 35– 46. https://doi.org/10.1080/01621459.1951.10500765

Hotelling, H. (1953). New light on the correlation coefficient and its transforms. Journal of

the Royal Statistical Society Series B. Methodological, 15(2), 193-232.

https://doi.org/10.1111/j.2517-6161.1953.tb00135.x

Lovakov, A., & Agadullina, E. R. (2021). Empirically derived guidelines for effect size

interpretation in social psychology. European Journal of Social Psychology, 51(3), 485-

504. https://doi.org/10.1002/ejsp.2752

Olkin, I., & Pratt, J. W. (1958). Unbiased estimator of certain correlation coefficients. Annals

of Mathematical Statistics, 29, 201-211. https://doi.org/10.1214/aoms/1177706717

Pearson, K. (1896). Mathematical contributions to the theory of evolution. III. Regression,

heredity, and panmixia. Philosophical Transactions A, 373, 253-318.

https://doi.org/10.1098/rsta.1896.0007

Pearson, K., & Filon, L. N. G. (1898). Mathematical contributions to the theory of evolution.

IV. On the probable errors of frequency constants and on the influence of random selection

on variation and correlation. Proceedings of the Royal Society of London, 62, 173-176.

https://doi.org/10.1098/rspl.1897.0091
STANDARD ERROR OF CORRELATION 14

Pugh, E. M., & Winslow, G. H. (1966). The Analysis of Physical Measurement. Addison-

Wesley.

Shieh, G. (2010). Estimation of the simple correlation coefficient. Behavior Research

Methods, 42, 906 –917. https://doi.org/10.3758/BRM.42.4.906

Soper, H. E. (1913). On the probable error of the correlation coefficient to a second

approximation. Biometrika, 9(12), 91-115. https://doi.org/10.2307/2331802

Zimmerman, D. W., Zumbo, B. D., & Williams, R. H. (2003). Bias in estimation and

hypothesis testing of correlation. Psicológica: Revista de Metodología y Psicología

Experimental, 24, 133-158.

 Table 1

Approximations for the Standard Error of the Pearson Correlation

Source Standard error

1 − ρ2
Pearson (1896, p. 266) σ𝑟 = (6)
√𝑁 ∙ (1 + ρ2 )
1 − ρ2
Pearson & Filon (1898, p. 174) σ𝑟 = (7)
√𝑁
1 − ρ2
Soper (1913, p. 107) σ𝑟 = (8)
√𝑁 − 1
1 − ρ2
Bonett (2008, p. 174) σ𝑟 = (9)
√𝑁 − 3
2
1−ρ 1 + 5.5 ∙ ρ2
Soper (1913, p. 107) σ𝑟 = ∙ (1 + ) (10)
√𝑁 2∙𝑁
1 − ρ2 11 ∙ ρ2
Soper (1913, p. 107) σ𝑟 = ∙ (1 + ) (11)
√𝑁 − 1 4 ∙ (𝑁 − 1)
1 − ρ2 11 ∙ ρ2 −192 ∙ ρ2 + 479 ∙ ρ4
Hotelling (1953, p. 212) σ𝑟 = ∙ (1 + + ) (12)
√𝑁 − 1 4 ∙ (𝑁 − 1) 32 ∙ (𝑁 − 1)2
0.5
2 1 − ρ2 14 + 11 ∙ ρ2
Ghosh (1966, p. 260) σ𝑟 = ∙ (1 + +⋯) (13)
√𝑁 + 6 2 ∙ (𝑁 + 6)
σ𝑟 = √𝑞 2 − 𝑄 with
𝑞 = 𝑟 + 𝑟 ∙ (1 − 𝑟 2 )/[2 ∙ (𝑁 − 3)]
𝑁−3
𝑄 = 1−( )∙
𝑁−2
Hedges (1989, p. 474) 3 2 − 4 ∙ 𝑟 2 + 2 ∙ 𝑟 4 8 − 24 ∙ 𝑟 2 (14)
( +
𝑁 𝑁 ∙ (𝑁 + 2)
48
+ )
𝑁 ∙ (𝑁 + 2) ∙ (𝑁 + 4)
σ𝑟 = √𝑞 2 − 𝑄 with
1 1 𝑁−2
𝑞 = 𝑟 ∙ 2𝐹1 ( , ; ; 1 − 𝑟 2)
Hedges (1989, p. 477) 2 2 2 (15)
𝑁
(𝑁 − 3)(1 − 𝑟 2 ) 2𝐹1 (1,1; ; 1 − 𝑟 2 )
𝑄 =1− 2
𝑁−2
Note. ρ = population correlation, r = sample correlation, N = sample size, 2𝐹1 (𝑎, 𝑏, 𝑐, 𝑥) =
Gauss hypergeometric function.

2
Ghosh (1966) presented an approximation to the order of 6. Because this resulted in rather complex terms,
for ease of presentation (13) reports only the first ones.
3
Hedges’ (1989) approximation of q (Equation 14) seems to mistakenly use a value of 3 in the denominator
instead of 4, thus, adopting N as the degrees of freedom, whereas the degrees of freedom was N - 1 in the
remainder of the paper. Moreover, the approximation of Q given in Equation 19 seems to be incorrect as well.
STANDARD ERROR OF CORRELATION 16

Figure 1

Sampling Distributions for Small and Large Correlations at Different Sample Sizes
Figure 2

Relative Percentage Bias of Different Estimators of the Standard Error of the Pearson Correlation

Note. Dashed lines represent relative biases of 5%. Relative biases exceeding 20% are not presented.
STANDARD ERROR OF CORRELATION 18

Figure 3

Root Mean Squared Error of Different Estimators of the Standard Error of the Pearson Correlation
 Supplemental Material for
A Brief Note on the Standard Error of the Pearson Correlation

In addition to the relative bias discussed in the main text, Table S1 also summarizes the

̂𝑟 , σ𝑟 ) = 𝐸(σ
absolute bias, 𝐵𝑖𝑎𝑠(σ ̂𝑟 − σ𝑟 ). These results corroborate the basic pattern already

discussed for the relative bias, with more precise standard errors for the estimators in (4) and

(9) as compared to the other estimators, particularly at larger correlations.

Table S1

Bias for Estimators of the Standard Error of the Correlation

Population correlation ρ
Method Eq.
.00 .10 .20 .30 .40 .50 .60 .70 .80 .90
N = 10
True standard error (4) .333 .331 .323 .310 .292 .267 .235 .196 .146 .083
Pearson & Filon (1898) (7) -.052 -.052 -.051 -.050 -.048 -.046 -.042 -.037 -.030 -.019
Soper (1913) (8) -.037 -.037 -.037 -.036 -.035 -.034 -.032 -.029 -.024 -.015
Bonett (2008) (9) .003 .002 .002 .001 -.001 -.003 -.005 -.006 -.007 -.006
Hotelling (1953) (4) -.030 -.029 -.028 -.025 -.022 -.018 -.013 -.007 -.004 .003
Hedges (1989) (15) .013 .012 .009 .003 -.003 -.011 -.018 -.024 -.026 -.020
Regression -.001 .000 .003 .008 .015 .024 .035 .048 .061 .069
N = 20
True standard error (4) .229 .227 .221 .211 .197 .178 .154 .125 .091 .049
Pearson & Filon (1898) (7) -.017 -.017 -.017 -.017 -.016 -.015 -.014 -.012 -.009 -.005
Soper (1913) (8) -.012 -.012 -.012 -.012 -.011 -.011 -.010 -.009 -.007 -.004
Bonett (2008) (9) .001 .000 .000 .000 -.001 -.001 -.002 -.002 -.002 -.002
Hotelling (1953) (4) -.011 -.010 -.010 -.009 -.007 -.005 -.003 -.001 .000 .001
Hedges (1989) (15) .006 .006 .004 .002 -.001 -.003 -.006 -.008 -.008 -.006
Regression .000 .001 .003 .008 .014 .022 .031 .040 .049 .053
N = 30
True standard error (4) .186 .184 .179 .170 .158 .142 .123 .099 .071 .038
Pearson & Filon (1898) (7) -.009 -.009 -.009 -.009 -.0009 -.008 -.007 -.006 -.005 -.003
Soper (1913) (8) -.006 -.006 -.006 -.006 -.006 -.006 -.005 -.005 -.004 -.002
Bonett (2008) (9) .000 .000 .000 .000 -.001 -.001 -.001 -.001 -.001 -.001
Hotelling (1953) (4) -.006 -.006 -.006 -.005 -.005 -.004 -.003 -.002 .000 .001
Hedges (1989) (15) .003 .003 .002 .001 .000 -.002 -.003 -.004 -.004 -.003
 STANDARD ERROR OF CORRELATION 20

Population correlation ρ
Method Eq.
.00 .10 .20 .30 .40 .50 .60 .70 .80 .90
Regression .000 .001 .003 .007 .012 .019 .027 .035 .042 .044
N = 40
True standard error (4) 160 .159 .154 .147 .136 .122 .105 .085 .060 .032
Pearson & Filon (1898) (7) -.006 -.006 -.006 -.006 -.006 -.005 -.005 -.004 -.003 -.002
Soper (1913) (8) -.004 -.004 -.004 -.004 -.004 -.004 -.003 -.003 -.002 -.001
Bonett (2008) (9) .000 .000 .000 .000 .000 -.001 -.001 -.001 -.001 -.001
Hotelling (1953) (4) -.004 -.004 -.003 -.003 -.002 -.002 .001 .000 .000 .000
Hedges (1989) (15) .002 .002 .001 .001 .000 -.001 -.002 -.003 -.003 -.002
Regression .000 .001 .003 .006 .011 .017 .024 .031 .037 .038
N = 50
True standard error (4) .143 .142 .137 .131 .121 .109 .093 .075 .053 .028
Pearson & Filon (1898) (7) -.004 -.004 -.004 -.004 -.004 -.004 -.003 -.003 -.002 -.001
Soper (1913) (8) -.003 -.003 -.003 -.003 -.003 -.003 -.002 -.002 -.001 -.001
Bonett (2008) (9) .000 .000 .000 .000 .000 .000 .000 -.001 .000 .000
Hotelling (1953) (4) -.003 -.003 -.002 -.002 -.002 -.001 -.001 .000 .000 .000
Hedges (1989) (15) .001 .001 .001 .000 .000 -.001 -.001 -.002 -.002 -.001
Regression .000 .001 .003 .006 .010 .015 .021 .028 .033 .034
N = 100
True standard error (4) .101 .100 .097 .092 .085 .076 .065 .052 .037 .020
Pearson & Filon (1898) (7) -.002 -.002 -.001 -.001 -.001 -.001 -.001 -.001 -.001 .000
Soper (1913) (8) -.001 -.001 -.001 -.001 -.001 -.001 -.001 -.001 -.001 .000
Bonett (2008) (9) .000 .000 .000 .000 .000 .000 .000 000 000 .000
Hotelling (1953) (4) -.001 -.001 -.001 -.001 .000 .000 .000 .000 .000 .000
Hedges (1989) (15) .000 .000 .000 .000 .000 .000 -.001 -.001 -.001 .000
Regression .000 .000 .002 .004 .007 .011 .016 .020 .024 .024