## Transcript

PSYCHOMETRIKA--VOL. 65, NO. 1, 23-28 MARCH 2000
SAMPLE SIZE REQUIREMENTS FOR ESTIMATING PEARSON, KENDALL AND SPEARMAN CORRELATIONS DOUGLAS G. B O N E T T IOWA STATE UNIVERSITY
THOMAS A. W R I G H T UNIVERSITY OF NEVADA, RENO Interval estimates of the Pearson, Kendall tau-a and Spearman correlations are reviewed and an improved standard error for the Spearman correlation is proposed. The sample size required to yield a confidence interval having the desired width is examined. A two-stage approximation to the sample ~ize requirement is shown to give accurate results. Key words: sample size, interval estimation, correlation, rank correlation.
1. Introduction Pearson, Kendall tau-a and Spearman correlations, which will all be denoted by the symbol 0, are used frequently in behavioral research. Although hypothesis testing is common, interval estimation may be more appropriate in applications where the magnitude of a correlation is of primary interest. Cohen (1988), Desu and Raghavarao (1990), Odeh and Fox (1991), and several intermediate level statistics texts, such as Cohen and Cohen (1975) and Zar (1984), give formulas that can be used to determine the sample size required to test a hypothesis regarding the value of a population Pearson correlation with desired power. To date, sample size formulas to determine the sample size required for interval estimation of Pearson, Kendall tau-a and Spearman correlations are not available. Recently, Looney (1996) produced a table of sample sizes needed to obtain a 95% confidence interval for the Pearson correlation. 2. Confidence Intervals To keep notation simple, let 0 denote both the estimator and the estimate of a population Pearson, Kendall tau-a or Spearman correlation. Define ( = tanh -1 0 and ~ = tanh -10. The population variance of ~ and its estimate are both denoted as trY. For the Pearson correlation, tr ? ~- l / ( n - 3) for bivariate normal random variables (Fisher, ~ 1925). For absolute values of Kendall correlations less than .8, tr 2 -~ .437/(n - 4) for any monotonic transformation of the bivariate normal random variables (Fieller, Hartley, & Pearson, 1957). We show that for absolute values of Spearman correlations less than .95, tr ? ~-- (1 + 0 2 / 2 ) / ( n - 3) for any monotonic transformation of the bivariate normal random variables. Assuming asymptotic normality of 0, a large-sample 100%(1-u) confidence interval (Hahn & Meeker, 1991, p. 238) for 0 may be defined as (1)
L at Requests for reprints should be sent to Douglas G. Bonett, Department of Statistics, Iowa State University, Ames, IA 50011. 0033-3123/2000-1/1997-0607-A $00.75/0 © 2000 The Psychometric Society
23
24
PSYCHOMETRIKA
where zet/2 is the point on the standard unit normal distribution exceeded with probability or/2. Note that [d(tanh ()/d(]2tr~ ----(sech 2 ()2~r~ - - ( 1 - tanh 2 ~ 2 ) # __ ( 1 - 02)2~r2¢ is the approximate variance of 0 obtained by the delta method (Stuart & Ord, 1994, p. 351) which simplifies to (1 - 02)2/(n - 3), .437(1 - 02)2/(n - 4), and (1 + 02/2)(1 - 02)2/(n - 3) for the Pearson, Kendall and Spearman correlations, respectively. If [01 is large and n is small, (1) may have a coverage probability that is quite different from 1 - or. A better confidence interval, originally proposed by Fisher (1925) for the Pearson correlation, is defined as Lower Limit:
Upper Limit:
[exp(2Ll) - 1] [exp(2Ll) + 1] [exp(2L2) - 1] [exp(2L2) + 1]
(2)
where L1 = .5[ln(1 + 0) - ln(1 - 0)1 L2 = .5[ln(1 + 0) - ln(1 - 0)] +
C(Zot/2) v't- __ b)1/2
c(zot/2) (n - b) l/2'
with c = 1, (.437) 1/2, and (1 + 02/2)1/2, b = 3, 4, and 3 for the Pearson, Kendall and Spearman correlations, respectively. David (1938) recommends the use of (2) for Pearson correlations only if n _> 25. The Fisher confidence interval for the Pearson correlation also assumes bivariate normality, and the effects of violating this assumption deserve careful consideration. Pearson (1929) concluded that "the normal bivariate surface can be mutilated or distorted to a remarkable degree" without affecting the sampling distribution of the Pearson correlation estimator. Subsequent simulations by Pearson (1931), Dunlap (1931), Rider (1932), and Gayen (1951) led to similar conclusions. However, Haldane (1949), Kowalski (1972), and Duncan and Layard (1973) have shown that the robust properties of the Pearson estimator apply only under independence and that marginal kurtosis can have a serious effect on the asymptotic sampling distribution of the Pearson estimator in the non-null case. If the assumption of bivariate normality cannot be justified, Kendall or Spearman correlations should be considered. The Kendall and Spearman correlations are attractive because (2) can be used to generalize from the sample to the population correlation for any monotonic transformation of bivariate normal variables, As noted previously, the approximate variance of ( for a Kendall correlation is accurate only for 101 < .8. Under bivariate normality, a Kendall correlation is equal to 2/zr times the inverse sine of the Pearson correlation so that a Kendall correlation of .8 corresponds to a Pearson correlation of about .95. Long and Cliff (1997) found that (2) works reasonably well for Kendall correlations if n > 10. Fieller et al. (1957) claim that or? _ 1.06/(n - 3) for absolute values of Spearman correlations less than .8. We claim that (1 + ~2/2)/(n - 3) is a more accurate estimate of a2. The results of a computer simulation are summarized in the Appendix and provide support for our claim. 3. Sample Size Determination The sample size required to obtain a 100(1 - 00% Fisher confidence interval with a desired width (Upper Limit minus Lower Limit) can be obtained by first solving for n in (1). This gives a first-stage sample size approximation, denoted as no, equal to
D O U G L A S G. B O N E T T A N D T H O M A S A. W R I G H T
no
= 4c2(1 -02)
2
(z~/2] 2 + b, \
w
25 (3)
/
where w is the desired width of the Fisher confidence interval (2) and 0 is a planning estimate of 0 obtained from previous research or expert opinion. Round (3) up to the nearest integer and set no = 10 if no < 10. Note that c 2 = 1 + 02/2 for the Spearman correlation. L e t Wo denote the width of the Fisher confidence interval (2) for a sample of size n o and set equal to 0. Let n denote the sample size that yields a Fisher confidence interval having the desired width. Assume Wo = k l (no - b) -1/2 in the neighborhood of no and w = k 2 ( n - b) - I / 2 in the neighborhood of n where kl and k2 are the constants of proportionality. For no close to n, assume kl = k2 and define Wo w
(n - b) 1/2
= ............
( n o . - b) 1/2"
(4)
Solving for n gives a second-stage approximation to the required sample size n = (no - b )
+ b,
(5)
which is rounded up to the nearest integer. A planning estimate of 0 is often obtained from a range of possible values based on expert opinion or confidence intervals from previous research. All other factors held constant, the sample size requirement is inversely proportional to 101. Given a range of possible values for 0, some researchers will want to compute (5) for both minimum and maximum values of 0 to obtain maximum and minimum sample size requirements. 4. Example The following example illustrates the computation of (5) for a Pearson correlation. For = .8, w = .2, and a = .05, use (3) to compute no = 4(1 - .82)2(1.96/.2)2 + 3 ~_ 52.8 and round up to 53. Setting n = no = 53, c = 1, b = 3, and 0 = 0 = .8 in (2) gives lower and upper Fisher interval limits of about .6758 and .880 with an interval width of Wo ~ .2042. Use (5) to compute (53 - 3)(.2042/.2) 2 + 3 ~_ 55.1 and round up to n = 56. If a sample of size n = 56 is taken from the population and the Pearson correlation estimate is close to the planning estimate of .8, the Fisher confidence interval width should be close to the desired width of .2. 5. Accuracy of the Two-stage Sample Size Approximation The sample size given by (5) approximates the correct sample size for the Fisher confidence interval. The correct sample size is defined as the smallest value of n which yields a Fisher confidence interval width that is less than or equal to the desired width. The accuracy of the twostage approximation is evaluated by comparing the value obtained by (5) with the correct sample size for eight values of 0, three values of w, and two values of o~. The correct sample size was obtained by systematically incrementing n by 1 until the width of the Fisher confidence interval attained the desired width. The results are summarized in Table 1. It can be seen from Table 1 that (5) gives a value of n that is exactly equal to the correct sample size or exceeds the correct sample size by a small amount. Although not shown in Table 1, (5) tends to overstate the correct sample size to a slightly greater degree if I01 > .9 and w > 2(1 -[01). If the correct sample size must be obtained (e.g., in a commercial software package), the value given by (5) can be systematically decreased by 1 and the width of (2) can be checked at each step. Given the accuracy of (5), only one or two checks will be required in most cases.
26
PSYCHOMETRIKA TABLE 1, Accuracy of Sample Size Approximation Pearson tO
Spearman
Kendall
Eq. 5
Correct n
Eq. 5
Correct n
Eq. 5
Correct n
.10 .10 .10 .10 .10 .10
.1 .1 .2 .2 .3 .3
.05 .01 .05 .01 .05 .01
1507 2601 378 650 168 288
1507 2601 378 650 168 288
1517 2614 382 653 169 290
1517 2614 382 653 169 290
661 1139 168 269 77 129
661 1139 168 269 77 129
.30 .30 .30 .30 .30 .30
.1 .1 .2 .2 .3 .3
.05 .01 .05 .01 .05 .01
1274 2198 320 550 143 245
1274 2198 320 550 143 244
1331 2297 334 574 149 255
1331 2297 334 574 149 255
560 963 143 243 65 110
560 963 143 243 65 110
.40 .40 .40 .40 .40 .40
.1 .1 .2 .2 .3 .3
,05 .01 .05 .01 .05 .01
1086 1874 273 469 123 209
1086 1874 273 469 123 209
1173 2024 295 507 132 226
1173 2024 295 507 132 226
448 822 122 208 57 94
448 822 122 208 57 94
.50 .50 .50 .50 .50 .50
.1 .1 .2 .2 .3 .3
.05 .01 .05 .01 .05 .01
867 1495 219 376 99 168
867 1495 219 376 99 168
975 1682 246 422 111 189
975 1682 246 422 111 189
382 656 99 167 46 76
382 656 99 167 46 76
.60 .60 .60 .60 .60 .60
.1 .1 .2 ,2 .3 .3
.05 .01 .05 .01 .05 .01
633 1091 161 276 74 125
633 1091 161 276 74 125
746 1287 189 325 86 146
746 1287 189 325 86 146
280 480 73 123 35 57
280 480 73 123 35 57
.70 .70 .70 .70 .70 .70
.1 .1 .2 .2 .3 .3
.05 .01 .05 .01 .05 .01
404 696 105 178 49 82
404 696 105 178 49 82
503 866 129 221 60 101
503 866 129 221 60 101
180 307 49 81 24 39
180 307 49 81 24 39
,80 .80 .80 .80 .80 .80
.1 .1 .2 .2 .3 .3
.05 .01 .05 .01 .05 .01
205 352 56 94 28 46
205 352 56 93 28 45
269 463 72 122 36 59
269 463 72 122 35 59
93 157 27 44 15 23
93 157 27 44 15 23
.90 .90 .90 .90 .90 .90
.I .1 .2 .2 .3 .3
.05 .01 .05 .01 .05 .01
63 106 21 34 13 21
62 105 20 33 12 20
87 147 28 46 18 28
86 147 27 45 16 25
30 49 12 18 8 11
30 49 11 17 8 11
27
DOUGLAS G. BONETT AND THOMAS A. WRIGHT TABLE2. Empirical Coverage Rates for Spearman Correlations (ct = .05) Variance Estimate 0
n
A
B
C
•1
20 50 100 200
95.4 95.7 95.6 95.3
95.8 95.5 95.3 95.0
94.8 95.1 95.2 95.0
.3
20 50 100 200
95.2 95.2 95.3 95.3
95.3 95.3 95.3 95.2
94.6 94.9 95.0 95.1
.5
20 50 100 200
94.8 95.0 94.7 94.8
95.6 95.6 95.4 95.5
94.2 95.0 94.9 95.1
.7
20 50 100 200
94.5 94.1 94.0 94.0
95.7 95.5 95.8 95.8
94.4 94.4 94.5 94.9
.8
20 50
95.8 95.5
94.1 94.1
95.6
94.1
200
93.8 93.2 93.0 93.1
95.6
94.4
20 50
92.4 92.3
95.1 95.5
92.6 93.6
100
.9
.95
100
92.1
95.6
94.1
200
92.1
95.6
94.2
20 50 100 200
89.5 90.3 90.5 91.0
94.0 94.5 94.6 95.5
90.8 92.3 93.0 93.7
Key: A = 1.06/(n - 3) (Fieller, et al., 1957) B = (1 + g 2 / 2 ) / ( n - 3) C = 1 / ( n - 2) + [~l/(6n + 4n 1/2) (Caruso& Cliff, 1997)
6. Conclusion Testing the null hypothesis that a population correlation is equal to zero m a y not always be interesting--a confidence interval may be more informative as suggested by Gardner and Altman (1986), Schmidt (1996), and many others. W h e n designing a study to estimate a Pearson, Kendall or Spearman correlation, the sample size required to obtain a Fisher confidence interval with the desired width will be a primary concern. An accurate sample size approximation can be obtained using (5). Appendix The accuracy o f (2) for a Spearman correlation is investigated using three different estimates for a.z: 1) 1.06/(n - 3), 2) (1 + 0 2 / 2 ) / ( n - 3), and 3) 1 / ( n - 2) + [~[/(6n + 4nl/2). The third estimate was recently proposed by Caruso and Cliff (1997)• All three variance estimates were determined empirically.
28
PSYCHOMETRIKA
A computer simulation (20,000 random samples per condition) of the empirical coverage of a 95% Fisher confidence interval (2) for Spearman correlations under bivariate normality was performed for 0 = [. 1 . 3 . 5 . 7 . 8 . 9 . 9 5 ] and n = [20 50 100 200]. Column A of Table 2 shows that the empirical coverage rate with the Fieller et al. (1957) variance estimate is liberal for 0 > .7. Column B of Table 2 shows that the empirical coverage rate with (1 + 0 2 / 2 ) / ( n - 3) is close to-95% for 0 < .9 and slightly liberal for 0 = .95 with small n. Column C of Table 2 shows that the empirical coverage rate with the Caruso and Cliff (1997) variance estimate is close to 95% for 0 < .7 and has liberal tendencies for 0 > .7 that are most pronounced with small n. The results of Table 2 hold for any monotonic transformation of bivariate normal random variables. References Caruso, J.C., & Cliff, N. (1997). Empirical size, coverage, and power of confidence intervals for Spearman's rho. Educational and Psychological Measurement, 57, 637-654. Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Hillsdale, NJ: Lawrence Erlbaum Associates. Cohen, J., & Cohen, P. (1975). Applied multiple regression~Correlation analysis for the behavioral sciences. Hillsdale, NJ: Lawrence Erlbaum Associates. David, F.N. (1938). Tables of the ordinates and probability integral of the distribution of the correlation coefficient in small samples. Cambridge: Cambridge University Press. Desu, M.M., & Raghavarao, D. (1990). Sample size methodology. Boston, MA: Academic Press. Duncan, G.T., & Layard, M.W.J. (1973). A Monte-Carlo study of asymptotically robust tests for correlation coefficients. Biometrika, 60, 551-558. Dunlap, H.E (1931). An empirical determination of the distribution of means, standard deviations, and correlation coefficients drawn from rectangular distributions. Annals of Mathematical Statistics, 2, 66-81. Fieller, E.C., Hartley, H.O., & Pearson, E.S. (1957). Tests for rank correlation coefficients. I. Biometrika, 44, 470-481. Fisher, R.A. (1925). Statistical methods for research workers. London: Hafner Press. Gardner, M.J., & Altman, D.G. (1986). Confidence intervals rather than p-values: Estimation rather than hypothesis testing. British Medical Journal, 292, 746-750. Gayen, A.K. (1951). The frequency distribution of the product moment correlation in random samples of any size drawn from non-normal universes. Biometrika, 38, 219-247. Hahn, G.J., & Meeker, W.Q. (1991). Statistical intervals: A guide for practitioners. New York, NY: Wiley. Haldane, J.B.S. (1949). A note on non-normal correlation. Biometrika, 36, 467468. Kowalski, C.J. (1972). On the effects of non-normality on the distribution of the sample product moment correlation coefficient. Applied Statistics, 21, 1-12. Long, J.D., & Cliff, N. (1997). Confidence intervals for Kendall's tau. British Journal of Mathematical and Statistical Psychology, 50, 31-41. Looney, S.W. (1996). Sample size determination for correlation coefficient inference: Practical problems and practical solutions. American Statistical Association 1996 Proceedings of the Section on Statistical Education, 240-245. Odeh, R.E., & Fox, M. (1991). Sample size choice (2nd ed.). New York, NY: Marcel Dekker. Pearson, E.S. (1929). Some notes on sampling tests with two variables. Biometrika, 21, 337-360. Pearson, E.S. (1931). The test of significance for the correlation coefficient. Journal of the American Statistical Association, 26, 128-134. Rider, P.R. (1932). The distribution of the correlation coefficient in small samples. Biometrika, 24, 382-403. Schmidt, E (1996). Statistical significance testing and cumulative knowledge in Psychology: Implications for training of researchers. Psychological Methods, 1, 115-119. Stuart, A., & Ord, J.K. (1994). Kendall's advanced theory of statistics, Vol. 1, Distribution theory. New York, NY: Halsted Press. Zar, J.H. (1984). Biostatistical analysis (2nd ed.). Englewood Cliffs, NJ: Prentice-Hall. Manuscript received 20 OCT 1997 Final version received 23 FEB 1999