At Six Samples, the Statistical Framework You Pick Decides Whether the Industrial Hygiene Report Survives an OSHA Contest
← Back to Blog Industrial Hygiene

At Six Samples, the Statistical Framework You Pick Decides Whether the Industrial Hygiene Report Survives an OSHA Contest

Pick Bayesian or Frequentist for a small-business exposure assessment and the same $4,050 budget yields a different defensible decision. A practitioner walk-through.

by iSi Environmental

Two industrial hygienists walk into the same plant. Same six samples, same lognormal distribution, same operator, same OEL. One walks out with an exposure assessment that holds up at a contest hearing. The other walks out with six numbers in a spreadsheet that cannot answer the only question that matters: is the exposure controlled, and how do you know?

The difference is not the data. It is the statistical framework each one chose before the first pump turned on.

That choice — Frequentist versus Bayesian decision analysis — is the conversation small-manufacturer safety managers almost never have with their consultant, and it is the conversation that decides whether the $4,050 they just spent on an industrial hygiene engagement bought them a defensible decision or just bought them paper. This post walks through what each framework does, where the regulatory acceptance lines fall, and how to spend a small-business sampling budget so the report lands in the compliance binder ready for the inspector instead of ready for a redraft.

What Does OSHA Actually Use to Evaluate an Exposure Assessment?

OSHA’s enforcement vocabulary is Frequentist. The OSHA Technical Manual, Section II, Chapter 1, lays out the agency’s evaluation approach in classical confidence-interval terms: an exposure is judged compliant when the upper confidence limit on the measured concentration (UCL₁,₉₅%) sits below the permissible exposure limit, and is judged out of compliance when the lower confidence limit (LCL₁,₉₅%) exceeds it (https://www.osha.gov/otm/section-2-health-hazards/chapter-1). The substance-specific standards under 29 CFR Subpart Z carry the same logic forward: methods of monitoring “must have an accuracy, to a 95 percent confidence level, of not less than plus or minus 25 percent” of the relevant PEL (see e.g., 29 CFR 1910.1050 Appendix D — https://www.osha.gov/laws-regs/regulations/standardnumber/1910/1910.1050AppD).

That language matters because it is the math an OSHA compliance officer will read when reviewing an employer’s monitoring program. The OSHA Sampling and Analytical Methods technical document repeats the same Frequentist framing — a coverage factor of 2 for two-sided 95% confidence intervals, computed against the target concentration, which is typically the PEL (https://www.osha.gov/sites/default/files/sampling-and-analytical-methods1.pdf). The underlying sampling strategy textbook OSHA’s approach descends from is NIOSH’s 1977 Occupational Exposure Sampling Strategy Manual, Publication 77-173, by Leidel, Busch, and Lynch (https://www.cdc.gov/niosh/docs/77-173/default.html). That manual is still cited, still in the Navy Industrial Hygiene Field Operations Manual, and still the procedural template most state plan and federal inspectors carry in the back of their heads (https://www.med.navy.mil/Portals/62/Documents/NMFA/NMCPHC/root/Industrial%20Hygiene/IHFOM_CH3_220909.pdf).

So if the inspector showed up tomorrow, the framework on the desk would be Frequentist.

So Why Did AIHA Move to Bayesian Decision Analysis?

Because the Frequentist framework breaks down on the data sets small businesses actually generate.

The American Industrial Hygiene Association’s A Strategy for Assessing and Managing Occupational Exposures, 4th edition (Mulhausen, Damiano, et al.), is the field’s standard reference for how to structure exposure assessment defensibly (https://www.aiha.org/education/marketplace/strategy-book-4th-edition). Its central concept is the Similar Exposure Group, or SEG — workers grouped by task, process, or job function whose exposures share a common distribution. Decisions are made on the SEG’s exposure profile, characterized as a 95th-percentile estimate against the OEL.

The classical Frequentist test for that profile is the Upper Tolerance Limit at 95% confidence and 95% coverage (UTL95/95). It is mathematically sound, but it has a small-sample problem the industry has known about for decades. The UTL95/95 for a lognormal exposure distribution — and industrial hygiene exposures are nearly always lognormal, with geometric standard deviations commonly between 1.5 and 3.0 per NIOSH 77-173 — produces tolerance intervals that span more than an order of magnitude when n is less than six. With three samples, the confidence interval on the GSD itself routinely exceeds 10. The math is technically valid; the decision it yields is operationally useless.

AIHA’s response, formalized by Hewett, Logan, Mulhausen, Ramachandran, and Banerjee in 2006, was to relax the coverage requirement from 95% to 70% — the UTL95/70 — and to express the decision as a posterior probability distribution over exposure decision categories rather than a binary pass/fail (https://pubmed.ncbi.nlm.nih.gov/16998991/). The categories, as used in the free IHDA-AIHA application, run from Category 1 (under 10% of the OEL, “no action”) through Category 4 (“moderate certainty unacceptable, implement controls”) (https://www.aiha.org/public-resources/consumer-resources/apps-and-tools-resource-center/aiha-risk-assessment-tools/ihda-aiha). The acceptance criterion AIHA publishes is straightforward: an SEG is acceptable when the probability it sits in Category 4 is less than 5%, which corresponds to UTL95/70 below the OEL.

In plain English: AIHA’s framework asks “what is the probability the 95th-percentile exposure is over the limit, given everything we know,” and treats that posterior as the decision. The Frequentist framework asks “can we reject, at 95% confidence, the hypothesis that the mean is over the limit,” and treats the rejection (or failure to reject) as the decision. They are different questions, and on a six-sample data set with high variability, they produce different answers.

Where the Frameworks Diverge in Practice

The cleanest way to see the divergence is to put both engines on the same data set. AIHA provides the free tools for that: IHSTAT is the Excel workbook that runs classical descriptive statistics and UCL95 / UTL95 calculations; IHSTAT_Bayes is the Bayesian module in the same workbook; IHDA-AIHA is the standalone desktop application that combines goodness-of-fit testing, censored-data handling, and the full Bayesian Decision Chart output (https://www.aiha.org/public-resources/consumer-resources/apps-and-tools-resource-center/aiha-risk-assessment-tools). For practitioners who want a hierarchical Bayesian approach with multi-censoring point handling, the IRSST-funded Expostats web application (Lavoué, Joseph, et al.) runs the same class of models in R behind a Shiny front end (https://www.expostats.ca/site/en/info.html; foundational paper at https://pubmed.ncbi.nlm.nih.gov/30551169/).

Run six samples through both engines. Two patterns recur:

  • When the SEG is clearly controlled (geometric mean well below the OEL, modest variability), both frameworks agree. The Frequentist UCL95 sits below the OEL; the Bayesian posterior puts the SEG in Category 2 with high probability. Either report is defensible.
  • When the SEG is borderline (geometric mean around 30–60% of the OEL, GSD above 2), the frameworks diverge. The Frequentist UCL95 will often cross the OEL — inconclusive at 95% confidence. The Bayesian posterior will give a probability distribution: maybe 60% probability of Category 2, 30% of Category 3, 10% of Category 4. The Frequentist test cannot reject overexposure; the Bayesian framework gives a categorical decision the safety manager can act on, plus the probability attached.

The borderline case is the small-business case. It is where the framework choice changes the deliverable.

What Does This Mean at Six Samples and a $4,050 Budget?

iSi’s median industrial hygiene engagement is $4,050. That figure comes from internal proposal data across more than 1,200 engagements, and it funds approximately six samples per single-SEG, single-analyte assessment — planning time, on-site sampling, lab analysis at AIHA-accredited rates, statistical analysis, and a written report.

At that budget, the Frequentist play is to run UCL95 and report pass, fail, or inconclusive. On a lognormal data set with a high GSD, “inconclusive” is the most likely output. The client has an exposure assessment on paper and no decision in hand.

The Bayesian play, at the same budget and the same six samples, builds a documented prior before sampling — sourced from comparable-SEG history, from a validated exposure model such as ART, IH-MOD, or Stoffenmanager, or from a structured subjective assessment with named inputs — and then computes the posterior distribution across the AIHA decision categories. The report says, for example, “P(Category 2) = 78%, P(Category 3) = 19%, P(Category 4) = 3%.” The decision is Category 2: well-controlled, periodic re-monitoring on a five-year cadence. That decision is defensible because the prior is documented, the posterior is computed in a peer-reviewed published framework, and the conclusion meets AIHA’s UTL95/70 acceptance criterion.

Same data. Same money. Different decision-yield.

The cost stack behind the dollar anchor: IH planning, on-site sampling, and write-up typically run $2,500–$5,000; lab analysis at AIHA-LAP-accredited facilities runs $30–$120 per sample, so six samples typically add $180–$720; decision-analysis report production sits inside the IH engagement, not as a separate line. Multi-SEG, multi-analyte programs covering an entire production area scale to $8,000–$15,000 depending on analyte count and shift coverage. Compare those numbers to OSHA’s penalty schedule under 29 CFR 1903.15 — $16,550 per serious citation and up to $165,514 per willful citation under the 2025-2026 inflation adjustment (https://www.osha.gov/penalties) — and the cost ratio is between 5x and 40x in favor of doing the assessment with the framework that actually produces a decision.

Where the Critics of Bayesian Decision Analysis Are Right

A practitioner who advocates for Bayesian decision analysis without acknowledging its weaknesses is not a serious practitioner. Four critiques are worth taking on directly.

The subjective-prior problem. A Bayesian posterior is only as good as the prior it started from. Informal “professional judgment” priors carry the industrial hygienist’s cognitive biases — anchoring on memorable past exposures, availability bias from recent jobs, confirmation bias toward the answer the client wants. The regulatory science literature documents this carefully (https://pmc.ncbi.nlm.nih.gov/articles/PMC7265656/). The mitigation is procedural: use a structured prior elicitation tool, not freehand professional judgment. Structured Subjective Assessment, exposure models such as those developed by Ramachandran and Banerjee (https://pmc.ncbi.nlm.nih.gov/articles/PMC4665102/), and COSHH Essentials-derived priors all reduce — though do not eliminate — the dependence on individual judgment. The prior source has to be named in the report. If the prior cannot be defended in writing, the posterior cannot be defended either.

The OSHA acceptance gap. OSHA’s enforcement vocabulary, as noted above, is still Frequentist. The agency has not formally adopted the AIHA decision categories as a citation criterion. In a contested hearing, an opposing expert can argue that the Bayesian output is “not OSHA’s framework.” The practical mitigation is to report both. The deliverable that survives a contest hearing leads with the AIHA decision category (because that is the operational decision the employer made) and also reports the Frequentist UCL95 / UTL95 numbers in the same document. Belt and suspenders. The marginal cost of running both engines on the same data set is essentially zero once the IH has the data in IHDA-AIHA or Expostats — both tools produce both outputs.

Prior dominance at small n. With n = 1 or n = 2, the Bayesian posterior is dominated by the prior. The data barely moves the answer. This is mathematically real. AIHA’s own decision rules in the 4th edition Strategy book police it internally: an SEG with fewer than six measurements and any single reading above the OEL defaults to Category 4 (“poorly controlled”) until additional sampling refutes that assignment. The default position protects workers; it does not protect employers who declined to invest in adequate sampling. A Bayesian decision at n < 3 should not be published without explicit flagging of prior dominance in the report, and a small-business sampling plan should size to n = 6 wherever the budget allows.

Recordkeeping burden. Bayesian analysis produces more documentation per decision — prior source, model choice, posterior distribution, decision chart — than a Frequentist UCL95 with a pass/fail conclusion. For a small business without an industrial hygienist on staff, that looks like more paper. The mitigation is operational: the IH engagement absorbs the documentation cost once, in the deliverable, and the small business holds a compliance binder it can read but does not have to construct.

What Should a Small Manufacturer Actually Do?

Six steps, in order. This is the sequence iSi runs on a baseline IH engagement.

Step 1 — Define the SEG before sampling. Group by task, by process, by job classification — not by job title alone. Two welders in different departments are usually different SEGs. Document the SEG basis in the proposal so the rationale is on the page before any sample is collected.

Step 2 — Document the prior explicitly. Source the prior from one of: historical comparable-SEG data, a validated exposure model (ART, IH-MOD, Stoffenmanager), or a structured subjective assessment with named inputs. Do not use “professional judgment” alone without a named structuring tool. The prior source goes in the report.

Step 3 — Collect six random samples per SEG, per analyte. Random sampling, not worst-case judgmental sampling. Worst-case introduces an upward bias that inflates both the Frequentist UCL and the Bayesian posterior. The bias is known and predictable, and an opposing expert will name it.

Step 4 — Run both engines on the same data. IHSTAT for the descriptive stats and the Frequentist UCL95 / UTL95; IHDA-AIHA or Expostats for the Bayesian posterior and decision chart. When the two outputs agree, the decision is doubly defensible. When they disagree, the disagreement is itself diagnostic — usually a misspecified prior or a non-lognormal distribution worth investigating.

Step 5 — Document the categorical decision and the re-sampling cadence. The AIHA framework matches re-sampling frequency to category. Category 1 (highly controlled) typically every five years; Category 2 (well controlled) every two to five years; Category 3 (controlled, borderline) every six to twelve months; Category 4 triggers controls plus re-sampling within ninety days.

Step 6 — Land the deliverable in the safety manager’s compliance binder. Proposal, SEG basis, sampling plan, lab data, both statistical outputs, decision categorization, re-sampling cadence. A safety manager should be able to hand the binder to an OSHA inspector and explain every step on the page.

For a small manufacturer without a credentialed industrial hygienist on staff, this is where iSi’s industrial hygiene consulting fits. iSi’s IH program is built around AIHA-grade exposure assessment using both statistical frameworks — SEG construction, baseline sampling at the n = 6 threshold, Bayesian decision analysis paired with Frequentist UCL95, written report, and a re-sampling cadence sized to the operation. The median engagement runs $4,050; comprehensive multi-SEG programs run $8,000–$15,000 depending on analyte count.

Compare that to one willful OSHA citation at $165,514 under 29 CFR 1903.15, or three serious citations at $16,550 each stacked across workers and analytes, and the framework-selection question stops being academic. A single willful OSHA violation costs up to $165,514. An iSi industrial hygiene assessment that catches it before the inspector does costs $4,050. That is a 41:1 return on a phone call.

The Bottom Line

OSHA enforces in Frequentist. AIHA decides in Bayesian. A defensible small-business industrial hygiene report does both on the same data set, documents the prior, lands the decision in one of AIHA’s four exposure categories, and sets a re-sampling cadence the safety manager can put in a calendar.

The statistical framework is not a technicality. It is the difference between an exposure assessment that closes a question and one that opens a contest hearing. At six samples and a $4,050 budget, that difference is decided before the first pump turns on.

If you have an industrial hygiene report sitting in a binder that was built on a single round of sampling or that reports only a Frequentist UCL with no documented prior — that report is as defensible as you have made it. To rebuild it on the framework that holds up: call iSi at (316) 264-7050 and ask for an AIHA-grade exposure assessment scoped to your similar exposure groups.

Sources

Tags: