5.1 Lognormal Distributions and Exposure Data Statistics

Lognormal Distributions and Exposure Data Statistics

Occupational hygiene exposure assessment relies heavily on applied biostatistics to characterize workplace air concentrations, validate exposure acceptability, and demonstrate regulatory compliance. Unlike classical industrial quality control data—which often exhibit symmetric, bell-shaped Gaussian behavior—workplace chemical, physical, and aerosol exposure measurements almost universally follow lognormal distributions.

A thorough understanding of lognormal mathematics, parameter estimation, upper percentile calculations, and goodness-of-fit validation is essential for any practicing Certified Industrial Hygienist (CIH).

---

1. Normal vs. Lognormal Distributions in Occupational Hygiene

A continuous random variable X is lognormally distributed if its natural logarithm, Y = ln(X), is normally distributed with mean µy and variance σy². That is, if Y ~ N(µy, σy²), then X = exp(Y) ~ Lognormal(µy, σy²).

Why Exposure Data Are Lognormal

Workplace environmental concentrations exhibit lognormality due to fundamental physical and statistical principles:

1. Physical Boundary at Zero (x > 0): Chemical concentrations and sound pressures cannot assume negative values. A normal distribution assigns finite non-zero probability to negative values whenever the mean is within two to three standard deviations of zero, which is physically impossible for chemical concentrations.
2. Multiplicative Generation and Dispersion Mechanisms: According to the Central Limit Theorem, when a process is governed by the addition of numerous independent random variables, the resulting sum tends toward a normal distribution. However, airborne contaminant concentrations in workplaces are determined by the product of multiple independent, fluctuating physical factors:
   • Source emission rate (mass volatilized per unit time)
   • General and local exhaust ventilation rates (air changes per hour, room capture velocity)
   • Worker proximity, position, and orientation relative to the contaminant source
   • Cyclic process duration and production throughput
   • Ambient barometric pressure, room temperature, and cross-draft turbulence

Because ln(A · B · C · D) = ln(A) + ln(B) + ln(C) + ln(D), taking the logarithm transforms a multiplicative physical process into an additive one, resulting in a normal distribution in log space and a lognormal distribution in linear concentration space.

Distributional Comparison Matrix

---

2. Geometric Mean (GM) and Geometric Standard Deviation (GSD)

When analyzing a dataset of n occupational exposure samples {x1, x2, …, xn}, the data are first transformed to the natural log scale: yi = ln(xi).

The Geometric Mean (GM)

The Geometric Mean (GM) is the antilogarithm of the arithmetic mean of the log-transformed data. It represents the median (50th percentile) of the lognormal exposure distribution:

$\bar{y} = \frac{1}{n} \sum_{i=1}^{n} \ln(x_i) = \frac{\ln(x_1) + \ln(x_2) + \dots + \ln(x_n)}{n}$

$\mathbf{\text{GM} = \exp(\bar{y}) = \exp\left( \frac{1}{n} \sum_{i=1}^{n} \ln(x_i) \right) = \sqrt[n]{x_1 \cdot x_2 \dots x_n}}$

• Physical Meaning: Exactly 50% of the daily exposure shifts in the sampled population are expected to fall below the GM, and 50% will exceed it. The GM has the same units as the original concentration measurements (e.g., ppm or mg/m³).

The Geometric Standard Deviation (GSD)

The Geometric Standard Deviation (GSD) is the antilogarithm of the sample standard deviation of the log-transformed data (sy):

$s_y = \sqrt{\frac{\sum_{i=1}^{n} (\ln(x_i) - \bar{y})^2}{n - 1}}$

$\mathbf{\text{GSD} = \exp(s_y)}$

• Physical Meaning: Unlike the arithmetic standard deviation, the GSD is dimensionless and strictly greater than or equal to 1.0 (GSD ≥ 1.0). A GSD = 1.0 indicates zero variance (sy = 0), where all measured exposure concentrations are identical.

Multiplicative Coverage Intervals

In a lognormal distribution, standard deviation intervals operate multiplicatively around the Geometric Mean:

• 68.27% of exposures lie within: [ GM/GSD, GM · GSD ] = [exp(ȳ - sy), exp(ȳ + sy)]
• 95.45% of exposures lie within: [ GM/GSD², GM · GSD² ] = [exp(ȳ - 2sy), exp(ȳ + 2sy)]
• 99.73% of exposures lie within: [ GM/GSD³, GM · GSD³ ] = [exp(ȳ - 3sy), exp(ȳ + 3sy)]

Interpreting GSD in Similar Exposure Groups (SEGs)

The American Industrial Hygiene Association (AIHA) utilizes the GSD as a primary diagnostic indicator of exposure variability within a defined Similar Exposure Group (SEG):

---

3. Arithmetic Mean (AM) vs. Geometric Mean (GM)

In occupational hygiene toxicology and epidemiology, understanding the divergence between the Arithmetic Mean and Geometric Mean is critical:

• Geometric Mean (extGM): Represents the median (50th percentile) exposure. It is the most robust measure of typical daily exposure because it is resistant to extreme high-end outliers.
• Arithmetic Mean (extAM): Represents the average mass of chemical inhaled over time. Chronic toxicological effects, cumulative internal body burden, and occupational disease risks are proportional to the true Arithmetic Mean, not the Geometric Mean.

Population Relationships in Lognormal Distributions

For a true lognormal population with log-scale parameters µy and σy:

$\mathbf{\text{AM} = \exp\left( \mu_y + \frac{1}{2}\sigma_y^2 \right) = \text{GM} \cdot \exp\left( \frac{1}{2}\ln^2(\text{GSD}) \right)}$

$\text{Variance}(X) = \exp(2\mu_y + \sigma_y^2) \left( \exp(\sigma_y^2) - 1 \right) = \text{AM}^2 \cdot \left( \exp(s_y^2) - 1 \right)$

$\text{Mode}(X) = \exp(\mu_y - \sigma_y^2) = \frac{\text{GM}}{\exp(\sigma_y^2)}$

Because 1/2σy² > 0 whenever variability exists (sy > 0), the Arithmetic Mean is always strictly greater than the Geometric Mean (extAM > extGM). As exposure variability (extGSD) increases, the arithmetic mean is pulled substantially higher into the upper tail.

Sample Estimation Note: While the naive sample arithmetic mean x̄ = 1/nΣ xi is an unbiased estimator of the population mean, for small lognormal datasets (n < 30), the Minimum Variance Unbiased Estimator (MVUE) or parametric estimator AM = GM · exp(0.5 sy²) provides a more efficient point estimate of the true long-term arithmetic mean.

---

4. Estimating Exposure Percentiles (The 95th Percentile, X95)

Under the AIHA Exposure Assessment Strategy, exposure acceptability is judged by evaluating the 95th percentile (X95) of the exposure distribution relative to the Occupational Exposure Limit (OEL). The 95th percentile represents an exposure level that is exceeded on only 5% of work shifts (or 1 out of every 20 workdays).

General Formula for Lognormal Percentiles

For any desired percentile p (where 0 < p < 1), the corresponding standard normal Z-score (Zp) is applied in log space:

$\ln(X_p) = \bar{y} + Z_p \cdot s_y$

Exponentiating both sides yields the master percentile equation:

$\mathbf{X_p = \exp(\bar{y} + Z_p \cdot s_y) = \text{GM} \cdot \text{GSD}^{Z_p}}$

Standard Normal Z-Scores for Common IH Percentiles

Probability of Exceedance (fe)

Conversely, to calculate the probability that an individual random work shift will exceed the Occupational Exposure Limit (OEL), calculate the standardized log-scale Z-score:

$Z = \frac{\ln(\text{OEL}) - \bar{y}}{s_y} = \frac{\ln(\text{OEL} / \text{GM})}{\ln(\text{GSD})}$

The exceedance fraction (fe = P(X > OEL)) is determined from the standard cumulative normal distribution table:

$f_e = 1 - \Phi(Z) = \Phi(-Z)$

---

5. Cumulative Lognormal Probability Plots & Goodness-of-Fit

Before applying parametric lognormal equations (X95 = GM · GSD(1.645)), the industrial hygienist must confirm that the sample dataset adequately fits a lognormal distribution.

Constructing Cumulative Lognormal Probability Plots

A lognormal probability plot graphs log-transformed exposure concentrations against their empirical cumulative probabilities (or probit / normal quantiles):

1. Sort the n exposure measurements in ascending order: x((1)) ≤ x((2)) ≤ s ≤ x((n)).
2. Assign an empirical cumulative plotting position (Pi) to each ith ordered data point using an established ranking formula:
   • Hazen's Formula: Pi = (i - 0.5)/n × 100%
   • Blom's Formula: Pi = (i - 0.375)/(n + 0.25) × 100%
   • Chegodayev's Formula: Pi = (i - 0.3)/(n + 0.4) × 100%
3. Plot ln(x((i))) on the linear vertical axis against the normal score Zi = Φ⁻¹(Pi) (or cumulative percent on probability paper) on the horizontal axis.

Graphical Diagnostics

• Straight-Line Fit: If the plotted points fall along a straight line (R² > 0.90), lognormality is validated.
• Estimating Parameters Graphically:
  • The concentration at the 50th percentile line (Z = 0) equals the Geometric Mean (GM).
  • The ratio of the concentration at the 84.13th percentile (Z = +1) to the 50th percentile equals the Geometric Standard Deviation (GSD): GSD = X(84.13) / X50.
• S-Shaped Curve: Indicates a bimodal mixture of two distinct exposure groups incorrectly combined into one SEG.
• Down-Curving Upper Tail: Indicates truncation, such as the effective operation of an engineering capture hood capping maximum exposures.

Formal Goodness-of-Fit Hypothesis Tests

---

6. Worked Step-by-Step Calculation Examples

Worked Example 4.1.1: Calculating GM, GSD, AM, and 95th Percentile

Problem: An industrial hygienist collects six (n = 6) full-shift personal air samples for toluene vapor in a paint spray booth. The laboratory reports the following 8-hour TWA concentrations (in ppm): $\{12.0, \; 18.0, \; 24.0, \; 35.0, \; 48.0, \; 75.0\}$

1. Calculate the sample Geometric Mean (GM) and Geometric Standard Deviation (GSD).
2. Calculate the estimated population Arithmetic Mean (AM).
3. Calculate the estimated 95th percentile (X95) and determine if it complies with the ACGIH TLV-TWA of 20 ppm.

Solution Steps:

1. Transform data to natural logarithms (yi = ln(xi)):

   • ln(12.0) = 2.4849
   • ln(18.0) = 2.8904
   • ln(24.0) = 3.1781
   • ln(35.0) = 3.5553
   • ln(48.0) = 3.8712
   • ln(75.0) = 4.3175

2. Calculate log-scale mean (ȳ) and Geometric Mean (GM): $\sum y_i = 2.4849 + 2.8904 + 3.1781 + 3.5553 + 3.8712 + 4.3175 = 20.2974$ $\bar{y} = \frac{20.2974}{6} = 3.3829$ $\mathbf{\text{GM} = \exp(3.3829) = 29.46\text{ ppm}}$

3. Calculate log-scale standard deviation (sy) and GSD:

   • Sum of squared deviations: Σ (yi - ȳ)² $(-0.8980)^2 + (-0.4925)^2 + (-0.2048)^2 + (0.1724)^2 + (0.4883)^2 + (0.9346)^2$ $= 0.8064 + 0.2426 + 0.0419 + 0.0297 + 0.2384 + 0.8735 = 2.2325$
   • Sample variance in log space: sy² = 2.2325/(6 - 1) = 2.2325/5 = 0.4465
   • Sample standard deviation: sy = √(0.4465) = 0.6682 $\mathbf{\text{GSD} = \exp(0.6682) = 1.95}$ (Note: GSD = 1.95 indicates a well-defined, homogeneous SEG since 1.5 ≤ GSD ≤ 2.0.)

4. Calculate estimated population Arithmetic Mean (AM): $\widehat{\text{AM}} = \text{GM} \cdot \exp(0.5 \cdot s_y^2) = 29.46 \cdot \exp(0.5 \times 0.4465) = 29.46 \cdot \exp(0.22325)$ $\widehat{\text{AM}} = 29.46 \times 1.2501 = \mathbf{36.83\text{ ppm}}$ (Notice that AM = 36.83 ppm > GM = 29.46 ppm, reflecting the skewness of peak exposures).

5. Calculate the 95th Percentile (X95): $\ln(X_{95}) = \bar{y} + 1.645 \cdot s_y = 3.3829 + (1.645 \times 0.6682) = 3.3829 + 1.0992 = 4.4821$ $X_{95} = \exp(4.4821) = \mathbf{88.42\text{ ppm}}$ Alternatively: X95 = GM · GSD(1.645) = 29.46 · (1.95)(1.645) = 29.46 · 3.0014 = 88.42 ppm.

Compliance Evaluation: The estimated 95th percentile (88.42 ppm) substantially exceeds the ACGIH TLV-TWA of 20 ppm. The exposure profile is unacceptable, requiring immediate engineering controls.

---

Worked Example 4.1.2: Calculating Exceedance Fraction from Baseline Parameters

Problem: Historical baseline monitoring of an abrasive blasting operation indicates that personal respirable crystalline silica exposures follow a lognormal distribution with a Geometric Mean GM = 0.022 mg/m³ and a Geometric Standard Deviation GSD = 2.10. Calculate the statistical probability (fe) that a worker's daily shift exposure exceeds the OSHA Permissible Exposure Limit (PEL) of 0.050 mg/m³.

Solution Steps:

1. Convert parameters to log space:

   • ȳ = ln(GM) = ln(0.022) = -3.8167
   • sy = ln(GSD) = ln(2.10) = 0.7419
   • ln(OEL) = ln(0.050) = -2.9957

2. Calculate standard normal score (Z): $Z = \frac{\ln(\text{OEL}) - \bar{y}}{s_y} = \frac{-2.9957 - (-3.8167)}{0.7419} = \frac{0.8210}{0.7419} = +1.107$

3. Determine exceedance probability (fe):

   • From standard normal cumulative tables: Φ(1.107) = 0.8658
   • Exceedance fraction: fe = 1 - Φ(1.107) = 1 - 0.8658 = 0.1342 = 13.42%

Result: There is a 13.4% probability that any given work shift will exceed the OSHA PEL, confirming that exposure controls are currently inadequate under AIHA criteria (which require exceedance < 5%).