Professional Documents
Culture Documents
Isk
Isk
Ms. #430
ing the exposure of workers to noise have to take representative measurements of the various circumstances of sound exposure and analyze the data in terms of usual daily exposure before deciding whether the results conform to the action levels of 85 or 90 dB(A). One of the approaches used to achieve this aim is multiple grab sampling, carried out throughout a working day or among a group of workers whose exposure is assumed to be homogeneous. In this case the value of a usual daily soumd exposure level is unlcnown but can be estimated from a series of samples, the confidence interval of which is then calculated. The estimation of a confidence interval based on values expressed in decibels runs into difficulties associated with the logarithmic nature of the measurement unit of the sound levels, which are overshadowed when the confidence interval is computed using an approximate method and applied to samples of small size and high standard deviation. This is the case with the method proposed in the standard, 640
ISO 9612,(') annex D: the mean equivalent continuous A-weighted sound pressure level, L`,q , is estimated using data collected by multiple grab sampling carried out with an identical duration T, within the framework of a lognormal hypothesis applied to the sound exposure values. Its confidence interval is estimated using an approximate method, which can be used even if the sample size is less than 5 or if the standard deviation of the measured values in dB(A) is higher than 6 dB(A). In the field of exposure to chemical substances the estimation of confidence intervals applicable to lognormally distributed data has been widely studied and has allowed, in particular, the comparison of values of confidence intervals calculated according to the exact method and according to approximate methods.(2 3 ) , These studies confirm that in case of lognormally distributed data the performance of approximate methods is poor for small sample sizes or high values of standard deviation. It should be noted that this drawback is specific to the calculation of the confidence interval of
September/October 2003
the mean sound exposure level. Other indexes such as the tolerance limit or the proportion of observations above a given value can be directly deduced from the distribution of sound pressure levels, and the usual estimation procedures(4) can be applied. 5 The original feature of this work is to use Land's publications' -7) to provide exact confidence intervals applicable in the field of occupational noise exposure. Although the decibel unit provides a convenient scale for representing noise data, it has been demonstrated,', that it is preferable to consider as base variable the sound exposure, expressed in Pa 2 .h or Pa'.s as defined in several international standards., 19.101 Sound exposure values are analvzed within the framework of a lognormal model and are assumed to be independent. Within the context of the ISO standard,") the authors propose a method of.analyzing samples of equivalent continuous A-weighted sound pressure levels LAeS T collected with the same duration T. An estimate of the mean A-weighted sound pressure level with a confidence interval expressed in dB(A) is provided.
9/
/
I
"V
.9
0.23, to obtain the corvith k = 1/p2T and a = log(10)/10 ~ responding properties of Ei. It should be noted that according to Occupational Safety and Health Administration (OSHA) noise ex0.14, but the formalism posure regulations, a = log(10)/16.61 remains unchanged. The values of E; are then independent and identically distributed in a lognormal distribution such that log(E,) - N(ap, Therefore the mean log(k), a2of2L. log(k), a2 o'2) or E, - LN(ap, 1 sound exposure related to the duration T is
(Xq,V
E = (EA-) = kea
{D/2w,k]
(7)
The measurement
of noise exposure in a worldng environment is based on the instantaneous A-weighted sound pressure p,(t) and on the equivalent continuous A-weighted sound pressure level defined by:
LAeqT= 10 log[
1
Hence, according to Equation 6, the mean sound exposure level value in dB(A) is
LAeq,T Ilog(kL.E)
pL. + 2-or2
(8)
T" ,
f 2PTA-dt]
,
(1)
, where T = [t 1 t2] and pa = 2 X 10 5 Pa. The A-weighted sound exposure over the duration T is defined in ISO 1999010 by:
EAT
pA(t) dt
(2)
(3)
or
EAT = p2T10
'LAqTJ/
(4)
with unit Pa 2 .h or Pa2 .s depending on the unit of time choser When the exposure measurement is performed by multi ple grab sampling, over a work shift for one employee or a group) of employees, the duration T must be identical for each measuLrement. The collected data is a set of n values LA, 9T expressedI in dB(A), noted below Li. The sample mean of these n values iis L and their sample stanidard deviation is SL. These data, expres sed . in Pa7.h or Pa2 .s using Equation 4, are noted E 1 The L, vali ies, expressed in dB(A), are assumed to be independent and identic ally distributed in a normal distribution with mean p.Land varia nce a', that is L, - N(p., o7L) This assumption conforms to usual knowledge on indust rial noise exposure and is applied to noise levels, in dB(A). WVI ien their corresponding exposure values Ei, in Pa .h, are used, tlieir distribution is assumed to be lognormal. These hypotheses are lata used to estimate the mean sound exposure level. Initial d and results are both expressed in the usual acoustic unit, the dB(A).
In references 5-7 Land provided a general framework to compute exact confidence intervals for any linear combination T = jt + (a/
2)c2,
where
,Lt
and
Cr
iable X, and a is a constant. In the case of noise exposure, a = log(10)/10 ~0.23 in ISO 1999 or a = log(10)/16.61 - 0.14 in the OSHA regulations. For a detailed description of the method, readers should refer to the previously mentioned articles. Several types of exact confidence intervals can be considered in noise exposure assessment: (1) exact one-sided lower or upper 1 - a confidence intervals; (2) exact two-sided 1 - c confidence intervals, obtained by combining the one-sided lower and upper 1 - a/2 confidence intervals; and (3) exact two-sided symmetrical 1 - cx confidence intervals (L + 8). In the case of the two-sided equal-tailed 1 - a confidence interval, both the upper and lower confidence level are equal to 1 - o/2, whereas with the two-sided symmetrical 1 - (xconfidence
Al HA Journal (64) September/October 2003 641
interval, the upper (1 - uj) and the lower (1 - aX) confidence levels are different, with au + oxl= a. As Land showed,(7) all these intervals can be expressed as
L+ 2S2L + C
n-L u'SL0
(10)
(11)
for the lower bound, where C. and C, are appropriate critical values, termed Land's critical values. Theses values depend on the type of interval, the sample size n, the critical level ot, and the product aSL, but not on L. Because these methods are cumbersome, Land published tables of selected one-sided and two-sided intervals.( 7 ,1 2 ) In the context of noise exposure, these tables correspond to values of SL equal to 0.43, 0.87, 2.17, 4.34, 8.69, 21.71 and 43.43, which make them of little use because the standard deviation of dB(A) values is usually within the range 0.1 to 6 dB(A). The present authors have computed tables suitable for use in the context of noise exposure. They were produced using the procedure given by Land.(') The procedure was transposed in MATLAB (The MathWorks, Inc., Natick, Mass.). Various confidence limits were computed for certain values of n and SL suitable for noise exposure assessment purposes. Table I gives values to compute one-sided upper and lower 1 - a confidence intervals for a = .05 and .025. It can be used to determine the limits of equal-tailed confidence intervals at level 1- 2a = 90% and 95%. Table II gives symmetrical confidence interval limits at level 1 -2a = 95%. In the case of 0.95 upper and lower one-sided confidence intervals (as well as 0.90 equal-tailed confidence intervals), the procedure given by Hewett et al.('3) can be followed to compute approximations of Land's critical values C. or C, easily, for all sample sizes between 3 and 1000 and all values of SL between 0.05 and 15.0, which encompasses a relevant range in the context of noise assessment. Applying the equations given in Appendix A of reference 13 using sy = aSL with a = log(10)/10 0.23, the authors verified that the difference in the values computed using their procedure and Hewett's method does not exceed 0.1 dB(A) in the range of Table I. In practice, the complete procedure must be applied as follows. (1) Calculate the sample (simple arithmetic) mean L and the sample standard deviation SL in dB(A).
with a = 0.23.
(3) Estimate the standard deviation for L:
SL 2 a2 S4
2n-1
(16)
t(n - 1, 1 - o/2)9
(17)
where t(n - 1, 1 - ot/2) is the 1 - o/2 quantile of the Student's t distribution with n - 1 degrees of freedom. This method is very simple to apply using standard software packages but hides the extent of the bias produced by the approximation: the approximate confidence intervals are narrower than the exact confidence intervals, as has been shown in references 2 and 6, for example. Figure 1 indicates the range in dB(A) of the bias of the upper bound of the two-sided symmetrical 0.95 confidence interval for various values of sample sizes and standard deviations. The exact limit is always higher than the approximate limit. When the sample size is greater than 20 and the standard deviation is lower than 6 dB(A), the difference between the exact and the approximate limit is lower than 0.5 dB(A); but for n = 5 and SL ' 3 dB(A), the difference is greater than 1.2 dB(A).
EXAMPLE
2L
(12)
(3) If SL and n correspond to an entry in Table I, read the 8(a, n, SL) value in that table. For example, the two-sided 0.90 confidence interval is
[L + 8(0.05, n, SL), L + 8(0.95, n, SL)]
(13)
In other cases, calculate Land's critical value C. or C, using s, aSL with a = log(10)/10 0.23 (or a = log(10)/16.61 0.14 in the case of the OSHA regulations) in the Hewett's equations.(' 3 ) Then calculate the confidence interval as [L+ C, SL
L + C.
SV
(14)
642
September/October 2003
TABLE 1.Calculation of Exact One-Sided Upper or Lower 95% and 97.5% Confidence Intervals
Standard Deviation of Observed Values S, in dB(A) n 5
a
0.5 -0.57 -0.45 0.51 0.69 -0.49 -0.39 0.44 0.57 -0.44 -0.35 0.39 0.50 -0.40 -0.32 0.35 0.44 -0.37 -0.30 0.32 0.41 -0.34 -0.28 0.30 0.38 -0.31 -0.25 0.27 0.33 -0.28 -0.23 0.24 0.30 -0.26 -0.21 0.23 0.28 -0.24 -0.20 0.21 0.26 -0.23 -0.19 0.20 0.24 -0.20 -0.17 0.18 0.21 -0.18 -0.15 0.16 0.19
1 -1.07 -0.85 1.13 1.54 -0.93 -0.75 0.95 1.25 -0.84 -0.68 0.83 1.07 -0.77 -0.63 0.75 0.97 -0.71 -0.58 0.68 0.87 -0.67 -0.55 0.64 0.80 -0.60 -0.49 0.56 0.70 -0.55 -0.45 0.51 0.63 -0.51 -0.42 0.47 0.58 -0.48 -0.40 0.44 0.54 -0.45 -0.37 0.41 0.50 -0.40 -0.33 0.36 0.44 -0.36 -0.30 0.33 0.40
1.5 -1.53 -1.24 1.88 2.65 -1.35 -1.10 1.55 2.09 -1.21 -1.00 1.34 1.77 -1.12 -0.92 1.20 1.56 -1.04 -0.86 1.09 1.40 -0.98 -0.81 1.01 1.29 -0.88 -0.73 0.89 1.12 -0.81 -0.67 0.80 1.00 -0.75 -0.63 0.74 0.91 -0.71 -0.59 0.69 0.85 -0.67 -0.56 0.64 0.79 -0.59 -0.50 0.56 0.69 -0.54 -0.45 0.51 0.62
2 -1.97 1.61 2.82 4.10 -1.74 -1.43 2.28 3.14 -1.58 -1.31 1.95 2.61 -1.46 -1.21 1.73 2.27 -1.36 -1.13 1.57 2.03 -1.28 -1.07 1.44 1.85 -1.16 -0.97 1.26 1.60 -1.07 -0.90 1.13 1.42 -1.00 -0.84 1.04 1.29 -0.94 -0.79 0.96 1.19 -0.89 -0.75 0.90 1.11 -0.79 -0.67 0.79 0.97 -0.72 -0.61 0.71 0.86
2.5 -2.40 -1.97 4.00 5.96 -2.13 -1.77 3.16 4.44 -1.94 -1.62 2.67 3.64 -1.80 -1.50 2.35 3.13 -1.69 -1.41 2.12 2.78 -1.59 -1.33 1.94 2.52 -1.44 -1.21 1.69 2.15 -1.33 -1.12 1.51 1.91 -1.25 -1.05 1.38 1.73 -1.17 -0.99 1.27 1.59 -1.11 -0.94 1.19 1.48 -1.00 -0.84 1.03 1.27 -0.91 -0.77 0.93 1.14
3 -2.82 -2.34 5.44 8.23 -2.53 -2.11 4.22 6.02 -2.31 -1.94 3.53 4.86 -2.15 -1.80 3.08 4.14 -2.02 -1.70 2.76 3.65 -1.91 -1.61 2.52 3.29 -1.74 -1.47 2.17 2.79 -1.61 -1.36 1.93 2.46 -1.50 -1.27 1.76 2.22 -1.42 -1.20 1.62 2.03 -1.35 -1.14 1.51 1.89 -1.21 -1.02 1.31 1.62 -1.11 -0.94 1.17 1.44
3.5 -3.25 -2.72 7.15 10.93 -2.93 -2.46 5.47 7.89 -2.69 -2.27 4.53 6.30 -2.51 -2.12 3.92 5.33 -2.36 -1.99 3.50 4.67 -2.23 -1.89 3.18 4.19 -2.04 -1.73 2.72 3.52 -1.89 -1.60 2.41 3.09 -1.77 -1.50 2.19 2.77 -1.67 -1.42 2.01 2.54 -1.59 -1.35 1.87 2.35 -1.43 -1.21 1.62 2.01 -1.31 -1.11 1.44 1.78
4 -3.70 -3.11 9.13 14.06 -3.34 -2.82 6.91 10.05 -3.08 -2.61 5.68 7.96 -2.88 -2.44 4.89 6.69 -2.71 -2.30 4.34 5.83 -2.57 -2.19 3.92 5.20 -2.36 -2.00 3.35 4.36 -2.19 -1.86 2.96 3.80 -2.05 -1.75 2.67 3.40 -1.94 -1.65 2.45 3.10 -1.85 -1.57 2.27 2.86 -1.67 -1.42 1.96 2.44 -1.53 -1.30 1.74 2.16
4.5 -4.16 -3.52 11.39 17.60 -3.77 -3.21 8.55 12.50 -3.49 -2.97 6.98 9.85 -3.26 -2.78 5.98 8.23 -3.08 -2.63 5.28 7.14 -2.93 -2.50 4.77 6.35 -2.69 -2.29 4.04 5.29 -2.50 -2.13 3.56 4.60 -2.35 -2.01 3.21 4.10 -2.23 -1.90 2.94 3.73 -2.12 -1.81 2.72 3.44 -1.91 -1.63 2.33 2.92 -1.76 -1.50 2.07 2.57
5 -4.64 -3.95 13.91 21.56 -4.22 -3.60 10.39 15.24 -3.92 -3.34 8.44 11.95 -3.67 -3.14 7.20 9.95 -3.47 -2.97 6.34 8.60 -3.31 -2.83 5.70 7.63 -3.04 -2.60 4.82 6.33 -2.83 -2.42 4.23 5.48 -2.66 -2.28 3.80 4.88 -2.53 -2.16 3.47 4.43 -2.41 -2.06 3.22 4.07 -2.18 -1.86 2.75 3.45 -2.01 -1.71 2.43 3.03
5.5 -5.14 -4.40 16.70 25.95 -4.70 -4.02 12.42 18.27 -4.36 -3.74 10.05 14.27 -4.10 -3.52 8.55 11.85 -3.88 -3.33 7.51 10.22 -3.70 -3.17 6.74 9.05 -3.41 -2.92 5.67 7.47 -3.18 -2.73 4.97 6.45 -3.00 -2.57 4.45 5.73 -2.84 -2.44 4.06 5.19 -2.71 -2.32 3.76 4.77 -2.46 -2.10 3.20 4.02 -2.27 -1.94 2.83 3.53
6 -5.67 -4.87 19.76 30.74 -5.19 -4.46 14.64 21.59 -4.83 -4.16 11.81 16.82 -4.55 -3.91 10.02 13.93 -4.31 -3.71 8.78 11.99 -4.12 -3.54 7.87 10.59 -3.80 -3.27 6.61 8.72 -3.55 -3.05 5.77 7.51 -3.35 -2.88 5.17 6.66 -3.18 -2.73 4.71 6.02 -3.04 -2.61 4.35 5.53 -2.75 -2.36 3.70 4.65 -2.54 -2.18 3.26 4.08
10
12
14
16
18
20
25
30
0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975 0.025 0.05 0.95 0.975
+ Note: The confidence interval is computed from L +-2S2 8 (a, n, Sj), where 8(a, n, Sj is the value given in the table for the corresponding value of n and S, and a, with a = .05 or .025 for lower confidence intervals and a = .95 or .975 for upper confidence intervals. This table provides exact equal-tailed 90 and 95% confidence
intervals.
The approximate two-sided svmmetrical 95% confidence interval computed according to the procedure suggested in ISO 9612 is L + 3.52 dB(A).
DISCUSSION
eidel et al. 15, and Rovester et al.14) show how confidence inLtervals can be used to enforce criterion compliance. Each
bound is interpreted with respect to its one-sided level of confidence to perform a one-sided test. Thus, equal-tailed confidence limits are recommended, especially because equal-tailed confidence intervals are not symmetrical, as can be seen in Tables I and II. Equal-tailed confidence intervals are not symmetrical except for small values of SL and large values of n; as a rule of thumb, an equal-tailed confidence interval can be considered as symmetrical within a 0.1 dB(A) range when Sf-/n is less than 0.15, which may be uncommon in practical noise exposure survevs. When n is small AIHA Journal (64) September/October 2003 643
2.5 4.08 3.25 2.77 2.46 2.23 2.05 1.80 1.62 1.48 1.38 1.29 1.14 1.02
3 5.47 4.27 3.59 3.15 2.84 2.60 2.27 2.03 1.86 1.73 1.62 1.41 1.27
3.5 7.16 5.49 4.56 3.97 3.55 3.24 2.80 2.50 2.28 2.11 1.97 1.72 1.55
4 9.14 6.92 5.69 4.91 4.37 3.96 3.40 3.02 2.75 2.54 2.37 2.06 1.85
4.5 11.39 8.55 6.99 5.99 5.30 4.79 4.08 3.61 3.27 3.01 2.80 2.43 2.17
5 13.91 10.39 8.44 7.20 6.35 5.71 4.84 4.26 3.85 3.53 3.28 2.84 2.53
5.5 16.70 12.42 10.05 8.55 7.51 6.74 5.69 4.99 4.49 4.11 3.81 3.28 2.92
6 19.76 14.64 11.81 10.02 8.78 7.87 6.61 5.78 5.19 4.74 4.39 3.77 3.35
+ a (n, Sj), where 8(n, Sj)is the value given in the table for the corresponding values of n and S,.
or S, is large, the upper bound of the symmetrical 95% confidence interval is very close to the bound of the one-sided upper 95% confidence interval. Thus, when using the upper bound of the 1 - 2c symmetrical confidence interval to compare it with a level led down in the regulations, the surveyor thinks that he has used a confidence level close to 1 -of whereas in fact this can be close to 1- 2c. This makes the use of symmetrical confidence intervals questionable as a biased result which is unfavorable in terms of employee protection is obtained. Therefore, in the context of noise exposure, the use of equaltailed confidence intervals is preferable. However, the computation of these intervals requires a specific computer program that is not included in standard statistical packages. In practical situations, tables similar to Table I can be used. Another alternative is to use Hewett's procedure, which is both easy to use and gives reasonably accurate values applicable to noise assessment. In contrast, the approximate method suggested in ISO 9612 should not be used, because it underestimates the width of the
symmetrical confidence interval. When the sample size is lower than 20 or the sample standard deviation greater than 3 dB(A), the bias reaches an unacceptable level.
CONCLUSION
measurements are taken using multiple grab sampling, it is now possible to estimate the confidence interval of the mean sound exposure level based on Land's exact procedure, assuming a normal distribution for the LAeq,T values of noise sample expressed in dB(A). The procedure is easy to use and suitable even for small sample sizes (n < 20) or large sample standard deviations (SL > 3 dB(A)). Furthermore, when a two-sided confidence interval is required, the equal-tailed confidence interval should be preferred to the symmetrical confidence interval. This observation strengthen the view that approximate symmetrical procedures such as the procedure suggested by ISO 9612 should not be applied.
When noise
REFERENCES
8 <: m 0
A
n=5
n= 10 n=30
~~~ 20/ n=
6 4-
0-
~
0)
0)~~~~~~~~ o
0)
- 2o^n O
1. International Organization for Standardization (ISO): Acoustics-Guidelinesfor the Measuirement and Assessment of Exposure to Noise in a Working Environment (ISO 9612). [Standard] Geneva: ISO, 1997. 2. Armstrong, B.G.: Confidence intervals for arithmetic means of lognormally distributed exposure. Am. Ind. Hyg. Assoc. J. 53:481-485 (1992). 3. Hewett, P.: Mean testing: II. Comparison of several alternative procedures. Appl. Occup. Environ. Hyg. 12:347-355 (1997). 4. Royster, L.H., E.H. Berger, and J.D. Royster: Noise surveys and data analysis. In E.H. Berger, W.D. Ward, J.C. Morill, and L.H. Royster, editors, Noise and Hearing Conservation Manual, pp. 97-176. Akron, Ohio: American Industrial Hygiene Association, 1986. 5. Land, C.E.: Confidence intervals for linear functions of the normal mean and variance. Ann. Math. Stat. 42:1187-1205 (1971). 6. Land, C.E.: An evaluation of approximate confidence interval estimation methods for lognormal means. Technometrics 14(1):145-158 (1972). 7. Land, C.E.: Standards confidence limits for linear functions of the normal mean and variance. J. Am. Stat. Assoc. 68:960-963 (1973). 8. Waclcernagel, H., C. Lajaunie, L. Thiery, and M. Grzebyk: Evaluation de l'exposition sonore en milieu professionnel: Application de methodes g6ostatistiques a l' stimation du Leq et constequences sur
644
September/October 2003
les strategies de mesurage (MAV NT-373/LT). lTechnical Report] Paris: INRS, 1998. 9. International Electrotechnical Conunission (IEC): Electroacoustics-Specificationsfor Personal Sonnd Exposure Meters (IEC 1252). [Standard] Geneva: IEC, 1990. 10. International Organization for Standardization (ISO): Acou7sticsDetermination of Occupational Noise Exposure and Estimation of Noise-Induced Hearing Impairment (ISO 1999). [Standard] Geneva: ISO, 1990. 11. Land, C.E.: Tables of confidence limits for linear functions of the normal mean and variance. In H. Harter and D. Owen, editors, Selected Tables in MathematicalStatistics, volume 3, pp. 385-419. Providence, Rl.: American Mathematical Society, 1975.
12. Land, C.E.: Hypothesis tests and interval estimates. In E. Crow and K. Shimizu, editors, Lognormal Distributions. Theory and Applications, pp. 87-112. New York: Marcel Decker (1988). 13. Hewett, P., and G.H. Ganser: Simple procedures for calculating confidence intervals around the sample mean and exceedance fraction derived from lognormaUly distributed data. Appl. Occup. En iiron.Eyg. 12:132-142 (1997). 14. Mohn, E.: Confidence estimation of measures of location in the log normal distribution. Biometrika 66:567-575 (1979). 15. Leidel, N.A., K.A. Busch, and J.A. Lynch: Occutpational Exposure Sampling Strategy Manual (DHEW [NIOSH] Publication no. 77173): Cincinnati, Ohio: National Institute for Occupational Safety and Health, 1977.
Al HA Journal (64)
September/October 2003
645
COPYRIGHT INFORMATION
TITLE: Confidence Intervals for the Mean of Sound Exposure Levels SOURCE: AIHA J 64 no5 S/O 2003 WN: 0324407528008 The magazine publisher is the copyright holder of this article and it is reproduced with permission. Further reproduction of this article in violation of the copyright is prohibited. To contact the publisher: http://www.aiha.org/