Tools › Environmental › Health and Safety › Exposure Sampling Statistics › Method

Exposure Sampling Statistics: Method

Open the Exposure Sampling Statistics tool

Exposure Sampling Statistics: Method, Sources, Assumptions and Limits

What the tool does

The tool computes the statistics an industrial hygienist runs on a set of personal exposure measurements and sets each beside the occupational exposure limit: the arithmetic and geometric mean, the geometric standard deviation, a goodness-of-fit test of the lognormal and the normal distribution, an upper percentile, the upper tolerance limit, the fraction of exposures above the limit with its confidence limits, and the estimate and exact confidence limits of the mean. Results below a detection limit are handled by regression on order statistics or by a substitution rule. A second part gives the one-day lower and upper confidence limits of the NIOSH Occupational Exposure Sampling Strategy Manual for a full-period sample or for consecutive samples. The page describes the data. Whether an exposure is acceptable, and what controls follow, is the judgment of the industrial hygienist.

The lognormal model

The NIOSH manual (section 4.4 and Technical Appendix I) treats the daily exposure averages of a worker or of a group with similar exposure as lognormally distributed. With y = ln x, the mean ȳ and the standard deviation s of the logs (n − 1 in the denominator) give the geometric mean GM = exp(ȳ) and the geometric standard deviation GSD = exp(s). The manual works in base-10 logarithms; the natural logarithms used here give the same GM, GSD and probabilities. The true arithmetic mean of a lognormal distribution is GM exp(σ²/2) (the manual's Technical Appendix B).

Goodness of fit. The Shapiro–Wilk W test (Royston's algorithm, the site's Data Summary code) is applied to the logs for the lognormal distribution and to the values for the normal distribution. A p-value above 0.05 does not reject the distribution; with fewer than about ten results the test can reject almost nothing, so a pass is weak evidence. The probability-plot correlation is reported as a second, descriptive measure. When some results are below detection the test uses the detected values only, as EPA's ProUCL does.

Upper percentile and upper tolerance limit

The estimate of the 100p-th percentile is exp(ȳ + zp s), with zp the standard normal quantile (1.645 for the 95th percentile). Because ȳ and s are themselves estimates, the percentile is uncertain. The upper tolerance limit exp(ȳ + K s) is the upper confidence limit of that percentile: with the stated confidence, at least 100p % of exposures lie below it (EPA Unified Guidance, equations [17.16] and [17.17]; ProUCL Technical Guide, equations (3-6) and (3-7)). The factor K depends on n, p and the confidence level and is K = t′ / √n, where t′ is the quantile of the noncentral t distribution with n − 1 degrees of freedom and noncentrality zp √n. The tool computes the noncentral t distribution by numerical integration of the normal distribution function over the chi density and solves for K; it does not use an approximation or a table. The normal model uses the same K on the values themselves.

Fraction above the limit

The estimated fraction of exposures above the limit L is 1 − Φ(z) with z = (ln L − ȳ) / s. This is the manual's "probability of noncompliance" Pn of section 4.4, which the manual reads as the proportion of days on which the worker would be above the limit if conditions stay as they were. The manual gives no confidence limit for it and notes that the true value is about as likely to be above the estimate as below. The tool adds one-sided confidence limits from the same noncentral t distribution: √n z follows a noncentral t distribution whose noncentrality is √n times the true standardized distance of the limit. The upper confidence limit of the fraction equals 1 − p exactly when the upper tolerance limit for the 100p-th percentile equals L, so the two statistics always agree.

Mean: estimate and Land's exact limit

The arithmetic mean of the results estimates the true mean without any assumption. Under the lognormal model the minimum-variance unbiased estimate is GM ψn(s²/2), where ψn is Finney's series ψn(t) = 1 + (n − 1) t / n + (n − 1)³ t² / (2! n² (n + 1)) + (n − 1)5 t³ / (3! n³ (n + 1)(n + 3)) + …

The confidence limit of a lognormal mean is not exp of a t limit on the logs (that is a limit on the geometric mean). Land's exact limit is exp(ȳ + s²/2 + s H / √(n − 1)) (EPA Unified Guidance, equations [21.5] and [21.6]; ProUCL Technical Guide, equation (2-36)). The factor H depends on s, n and the confidence level and is usually interpolated from Land's tables. The tool computes it. For a trial value ξ0 of ln(mean) = μ + σ²/2, subtracting ξ0 from every log makes the hypothesis a statement about one natural parameter of the normal family, so the most powerful unbiased test conditions on the sum of squares; the quantity w = Σz / √(n Σz²) then has a density on (−1, 1) proportional to exp(−½ √(n Σz²) w) (1 − w²)(n − 3)/2. The confidence limit is the ξ0 at which the observed w falls at the chosen tail probability, found by numerical integration and bisection. Land's limit needs at least three results. The normal model gives the ordinary Student t limits x̄ ± t s / √n.

Results below the detection limit

A result written with a less-than sign is known only to lie below its detection limit. Robust regression on order statistics (EPA Unified Guidance, section 15.4, equations [15.8] to [15.13]) handles one or several detection limits. For each distinct limit it estimates the fraction of the distribution above that limit from the counts of detected values and of all results below it; these fractions give a plotting position to every detected value and spread the results below each limit evenly over the probability beneath it. A least-squares line of ln(detected value) against the normal quantile of its plotting position is then used to read a value for each result below detection. Those values are combined with the measured ones and all statistics are computed on the combined set. At least three detected values are needed. The substitution rules (half the limit, the limit divided by √2, or the limit itself) are offered because they are widely used; they are crude, they distort the standard deviation, and EPA's ProUCL guide advises against them.

One day's samples: the NIOSH decision statistics

For a single workshift the manual attaches the uncertainty of the sampling and analytical method, its total coefficient of variation CVT, to the measured exposure. With x the measured exposure divided by the limit and z the one-sided normal quantile (1.645 at 95 %):

Full-period single sample (section 4.2.1): LCL = x − z CVT, UCL = x + z CVT.

Consecutive samples with durations Ti and results Xi: TWA = Σ Ti Xi / Σ Ti. For uniform exposure during the shift (section 4.2.2.1) the limits are TWA / L ∓ z CVT √(Σ Ti²) / Σ Ti. For nonuniform exposure (section 4.2.2.2) they are TWA / L ∓ z CVT √(Σ Ti² Xi²) / (L Σ Ti √(1 + CVT²)), a conservative form.

The manual classifies a day by these limits: an LCL above 1 is its "noncompliance exposure", a UCL at or below 1 its "compliance exposure", and anything between its "possible overexposure". The tool prints the limits and quotes the rule; it does not classify. The action level described in the manual's Technical Appendix L is one-half of the permissible exposure limit; it is an input here because substance-specific standards can differ.

What it assumes

The results are a random, representative sample of one stable exposure distribution: one worker or one group with similar exposure, no trend with time, no change of process or controls during the period, and independent results (not several samples from one shift treated as separate days). The lognormal statistics assume the lognormal shape; the normal statistics the normal shape. The one-day limits assume that sampling and analytical errors are normally distributed with a known CVT, and that the consecutive samples cover the whole averaging period of the limit. With results below detection, the regression assumes that detected and undetected results come from the same lognormal distribution.

When it misleads

With few results every limit is wide: for six results the 95 %, 95 % tolerance factor is 3.7, so the tolerance limit of a group with a GSD of 2.5 is about thirty times its geometric mean. A set that mixes two jobs or two operating conditions has an inflated GSD, and the tolerance limit and Land's limit then grow very large; EPA's guides warn that Land's limit becomes impractically high when the standard deviation of the logs exceeds about 1 to 1.5 and the sample is small. One unusually low value raises s and therefore raises the upper limits. A goodness-of-fit test that does not reject says little when n is small. After results below detection have been replaced, the confidence limits treat the replaced values as measurements and are too narrow, increasingly so as the share of such results grows. The upper tail of the fitted distribution is an extrapolation beyond the largest result. The one-day limits cover measurement error only: they say nothing about other days. Exposure limits that are ceilings or short-term limits need their own sampling strategy (the manual's section 4.3), which this tool does not cover.

Not included

The manual's chart-based test for grab samples (section 4.2.3), its ceiling-standard procedure (section 4.3), partial-period sampling, the Kaplan–Meier and maximum-likelihood estimates for censored data, between-worker and within-worker variance components, Bayesian decision charts, and the exposure rating categories of the American Industrial Hygiene Association (the tool states each statistic as a multiple of the limit and leaves the rating to the user).

Verification

The library reproduces the manual's printed examples of sections 4.2.1, 4.2.2.1, 4.2.2.2, 4.2.4, 4.3.2 and 4.4; the 246 tolerance factors of Table 17-3 of the EPA Unified Guidance; the 4,600 Land factors of its Tables 21-1 to 21-8 (88 % to the printed digit and 98.6 % within one unit of the last printed digit; 56 cells differ by more than 0.3 %, 26 of them by a dropped or added minus sign, and are taken as misprints in the document); the Unified Guidance examples 15-2 (regression on order statistics), 17-3 (tolerance limit) and 21-3 (Land's limit); and the values printed for Example 2-2 of the ProUCL Technical Guide (its Shapiro–Wilk statistic, 0.948, differs from the 0.950 of Royston's algorithm used here). Every statistic is also compared with an independent computation written separately with numerical quadrature. The Finney series was written from memory and is checked against the equivalent hypergeometric function and by simulation.