Occupational Epidemiology for CIH: From Workplace Questions to Useful Evidence

Your learning route

Define the health question
          ↓
Count cases with the right denominator
          ↓
Choose the study design
          ↓
Calculate and explain the comparison
          ↓
Check uncertainty, bias and causal reasoning
          ↓
Use exposure evidence to guide prevention

How to use it: On your first pass, read each example and memory aid. On your second pass, cover the calculations and work them yourself. All numerical workplace examples below are invented for teaching.

01 Define What You Are Trying to Learn

People, exposure, outcome and time, plus the case definition and the terms to learn early.

Start with people, exposure, outcome and time

Epidemiology examines patterns of health outcomes in populations and factors associated with those patterns. A useful occupational question specifies:

“Are workers getting sick?” is a starting concern. “Is the two-year incidence of defined asthma higher among exposed workers?” is a question that guides a study.

A case definition states the criteria used to count an outcome consistently. It can include clinical findings, test results and relevant person/place/time boundaries. Apply comparable outcome criteria to both groups. When testing whether a particular exposure is associated with disease, avoid defining disease status by that very exposure in a way that predetermines the answer. CDC: case definitions.

Distinctions to learn early

Memory aid: Define the people, the exposure, the outcome and the clock before calculating.

02 Count Disease: New Cases, Existing Cases or Person-Time

Prevalence, risk and incidence rate answer different questions, with a worked person-time example.

Three denominators that answer different questions

Period prevalence includes people who had the condition at any time in a defined interval. Point and period prevalence need clearly defined populations and time boundaries. Prevalence is a proportion, although older sources often call it a “prevalence rate.” CDC: disease-frequency measures.

Example: prevalence. On one survey date, 30 of 300 workers have asthma. Prevalence = 30/300 = 10%. You cannot tell from that snapshot when their asthma began.

Example: risk. A separate group of 100 workers is free of the outcome at baseline. Four develop it during complete two-year follow-up. Risk = 4/100 = 4% over two years.

Why person-time matters

Workers do not always contribute equal follow-up. They may develop the outcome, die, leave observation or reach study end. For a first-event analysis, count their time only while they are observed and still at risk of that first event. Study entry is set by the protocol; it need not coincide with first exposure.

Suppose the same 100 workers are followed for up to two years:

The 4% risk and 20.4 per 1,000 person-years rate are both valid here, but answer different questions. Do not label the rate as a percentage. More complex follow-up may require survival-analysis methods rather than a simple baseline denominator. CDC: person-time.

Why high prevalence does not necessarily mean high incidence

Prevalence depends on new cases and how long people remain cases. Longer survival or longer disease duration can increase prevalence even without an increase in new disease. In a stable population with an uncommon condition, prevalence ≈ incidence rate × average duration can be a useful approximation; it is not a universal formula.

Memory aid: Prevalence = present. Incidence = incoming cases. Rate = person-time.

03 Choose the Design by How Participants Enter the Study

Cohort, case-control and cross-sectional, plus the other labels worth recognizing.

The main designs at a glance

Watch for: a cohort needs follow-up and enough events; case-control controls must represent the source population's exposure experience; a cross-sectional study cannot show which came first.

A rare exposure often suits a cohort; a rare disease often suits case-control. These are selection guides, not guarantees. A cohort has no automatic advantage in statistical power. CDC: analytic study design.

Retrospective cohort versus case-control

Both may use old records. The distinction is how the study population is assembled:

• A historical employee roster is classified by earlier job exposure, and later disease records are examined → retrospective cohort.
• Workers with a disease and sampled controls are selected, then prior jobs are compared → case-control.

In a conventional case-control study, controls are not chosen merely because they are unexposed. Selecting them that way would distort the exposure comparison. CDC: cohort and case-control investigation.

Other labels worth recognizing

Sources: NIOSH CIB 69, NIOSH Practices in Occupational Risk Assessment, CDC-hosted ecological-methods guide (violence research; the method applies generally), CDC design guidance.

Memory aid: Disease-selected → case-control. Defined workforce followed over time → cohort. One snapshot → cross-sectional.

04 Compare Groups: One Table, Several Questions

Risk ratio, risk difference, attributable fraction and odds ratio from one 2 x 2 table.

A two-year cohort example

All 400 workers are free of the outcome initially, and follow-up is complete.

First calculate the two risks

The denominator includes all initially at-risk people in that exposure group, not only the cases.

In some texts the risk difference is called attributable risk; this is not the attributable fraction AFₑ in the table above. RR below 1 means lower risk in the numerator group; the example below shows why. RR = 1 means no difference. Multiply AFₑ by 100 for a percentage. The term “attributable” requires a valid causal interpretation; an unadjusted association alone does not establish preventable cases. Five percentage points is an absolute difference; the relative increase from 5% to 10% is 100%. CDC: association measures, CDC: attributable proportion.

Odds are different from probability

At 10% risk, odds are 0.10/0.90 = 1/9. In the cohort table:

RR = 2 and OR = 2.11 are close, but not identical. When outcomes are common, the difference can be substantial: risks of 60% and 30% give RR = 2, but OR = 3.5.

How to read an odds ratio

Memory aid:

• Odds = yes ÷ no.
• OR = one group's odds ÷ the comparison group's odds.
• OR tells us the association; the confidence interval tells us its precision.

In the cohort table above, the exposed odds are 20 ÷ 180 = 0.111 and the unexposed odds are 10 ÷ 190 = 0.053, so OR = 0.111 ÷ 0.053 = 2.11.

An OR above 1 alone does not establish statistical significance or causation. Also, "not significant" does not prove there is no effect. Source: CDC: interpreting study results.

Why case-control studies usually use OR

Suppose a case-control study enrolls:

Step 1: odds = yes ÷ no. Here "yes" is previously exposed and "no" is previously unexposed, and we take the odds separately in each column.

Step 2: the odds of exposure in each group.

Step 3: OR = one group's odds ÷ the comparison group's odds.

Step 4: simplify. Dividing by a fraction is the same as multiplying by its flip, so the two fractions cross-multiply into one:

Both routes give the same answer: OR = 6. In the 2 × 2 table, this is the top-left cell times the bottom-right cell, divided by the top-right cell times the bottom-left cell.

Cases have six times the exposure odds of controls. The investigator chose how many cases and controls to enroll, so 30/(30+20) is not the population disease risk among exposed workers. For conventional cumulative case-control sampling, OR approximates RR when the outcome is uncommon in the source population. In risk-set sampling, OR estimates a rate ratio without requiring a rare disease; this is an optional advanced distinction. CDC published study illustrating risk-set sampling. CDC: OR interpretation, NIOSH CIB 69, NIOSH Practices in Occupational Risk Assessment.

Read labels, not just letters. Transposing the table preserves the same comparison. Reversing exposed and unexposed reference groups takes the reciprocal of the ratio.

Other ratios and population impact

• Incidence rate ratio: divide exposed cases/person-time by comparison cases/person-time. For example, 20/1,000 PY ÷ 10/1,000 PY = 2.
• Prevalence ratio: divide the two prevalence proportions in a cross-sectional study. Do not describe this as new-disease risk.
• Population attributable fraction: (population risk − unexposed risk) ÷ population risk, under appropriate causal/comparability assumptions. It differs from AFₑ because it considers the entire population, including how common exposure is. In the 400-worker example, population risk is 30/400 = 7.5%, giving (7.5 − 5) ÷ 7.5 = 33.3%.

Memory aid: Ratio asks “how many times?” Difference asks “how many extra?” Attributable fraction asks “what share?”

05 Compare Mortality Fairly: SMR, SIR and PMR

Observed against expected deaths or cases, and why PMR is not a risk.

SMR: observed deaths versus expected deaths

An older workforce naturally has more deaths than a younger one. Standardization accounts for differences such as age and calendar period before interpreting excess deaths.

For indirect standardization, apply reference death rates to the cohort's corresponding person-time strata and add the results:

If 15 deaths are observed:

This is 50% above expected, not a 150% increase. Real studies commonly use more detailed age, sex and calendar-period strata. The result depends on the reference rates and needs an uncertainty interval. NIOSH-hosted SMR methods study.

SIR—standardized incidence ratio—uses observed and expected incident cases, rather than deaths. Do not confuse this with an infection-specific metric that happens to use the same acronym. CDC: SIR and statistical investigation.

PMR: what share of deaths was from this cause?

Suppose 20 of 100 workforce deaths were due to a particular cause. In the reference population, that cause accounts for 10% of deaths:

The cause represents twice the share of deaths. This does not establish twice the mortality risk among living workers. If deaths from other causes fall, the share from this cause can rise without its own rate increasing. Adjusted PMRs require appropriate stratum-specific methods. CDC: proportionate mortality.

Memory aid: SMR counts deaths against expectation; PMR compares shares of deaths.

06 Ask Whether the Comparison Is Distorted

Selection and information bias, confounding, random error and the healthy-worker effect.

Recall bias and interviewer bias are information-bias mechanisms. Detection/surveillance bias can occur if one group is examined much more intensively. Larger samples improve precision but do not automatically fix systematic error. CDC: interpreting study results.

The healthy-worker effect, explained simply

People healthy enough to obtain employment can have lower mortality than the general population, which includes people unable to work. Separately, healthier people may remain in demanding jobs while ill workers leave or transfer. These are commonly taught as occupational selection effects:

• Healthy hire effect: who starts work.
• Healthy-worker survivor effect: who remains exposed over time.

Therefore, low overall mortality does not prove that a workplace is harmless. Internal worker comparisons can help, but do not automatically remove survivor bias. Detailed analysis can involve both selection and time-varying confounding. Healthy-worker survivor methods paper.

Confounding versus effect modification

A classical confounder relates to exposure and outcome and is not an intermediate step in the causal pathway being studied. For example, adjusting for a biological effect caused by the exposure can remove part of the very effect you intend to estimate. Matching needs appropriate analysis; it does not automatically eliminate confounding.

Exposure misclassification: assigning a worker to the wrong category can conceal or distort an association. Under common simple assumptions, nondifferential binary misclassification tends to dilute an effect, but “misclassification always biases toward the null” is false.

Memory aid: Selection = who is included; information = what is measured; confounding = another explanation mixed in.

07 Read the Uncertainty Before Accepting the Headline

Confidence intervals, Type I and II error, power, p-values, SD versus SE.

Effect size and precision answer different questions

The null is 1 for ratios (above 1 is higher, below 1 is lower) and 0 for differences. “Not statistically significant” does not demonstrate absence of an effect, especially when the interval is wide. A 95% CI is generated by a method that covers the true parameter in 95% of repeated samples under its assumptions; it is not the range containing 95% of workers. CDC: confidence intervals.

Type I error, Type II error and power

Power depends on effect size, sample size/events, variability and the chosen significance level. Repeatedly testing many outcomes can generate chance positives. Statistical significance cannot correct poor exposure assessment or establish practical importance. CDC: statistical considerations.

SD versus SE: a short bridge to biostatistics

Standard deviation (SD) describes spread among observations. Standard error (SE) describes uncertainty in an estimate. For an ordinary sample mean of independent observations:

If SD = 10 and n = 100, SE = 1. That does not mean individual values vary by only 1. Repeated measurements on the same worker are not automatically 100 independent workers. Confidence intervals require an appropriate method; do not apply mean formulas to RR or OR without the correct transformation/model. NIST: confidence limits for a mean.

08 Decide How Strongly the Evidence Supports Causation

Temporality, the Bradford Hill viewpoints, latency and lagging.

Exposure must precede the outcome. This is temporality. A diagnosis today may reflect exposure years earlier, so a clean air sample today does not reconstruct the relevant history.

The Bradford Hill viewpoints help organize evidence:

These are viewpoints, not a scorecard that must reach nine checks. A weak or absent exposure gradient can reflect latency, thresholds or exposure error; strong associations can still be biased. Hill AB, The environment and disease: association or causation? Proc R Soc Med, 1965.

Latency is the delay between relevant exposure and detectable disease, as operationally defined in a study. Exposure lagging excludes recent exposure from a cumulative metric when it is assumed too recent to affect the outcome; select and justify the lag rather than searching only for a significant result.

09 Understand Screening and Surveillance

Sensitivity, specificity, PPV and NPV, and lead-time bias.

Screening looks for possible disease in people who may not yet have recognized symptoms. A positive screen often requires further assessment. Surveillance repeatedly collects, interprets and communicates health information to guide action. A screening programme can supply surveillance data; collecting results without interpreting or acting on them is incomplete. CDC: surveillance functions.

Four results, four useful measures

Example: Of 1,000 workers, 100 have the condition. A test detects 90 of them and incorrectly flags 90 of the other 900. Sensitivity = 90%; specificity = 90%; PPV = 90/180 = 50%; NPV = 810/820 = 98.8%.

Lower prevalence also raises NPV. In practice, test performance can also change with the population and threshold. Sensitivity/specificity are not automatically the same as analytic detection limits. CDC: surveillance-test calculation table, CDC: predictive value and prevalence.

Memory aid: Sensitivity starts with diseased people; specificity starts with disease-free people; predictive value starts with a test result.

When evaluating a screening programme, also recognize lead-time bias: earlier diagnosis can lengthen measured survival from diagnosis without postponing death. A programme should demonstrate meaningful health benefit, not just more diagnoses. NCI: interpreting screening statistics.

10 Connect the Health Findings to the Right Exposure

Exposure metrics, SEGs, job-exposure matrices and what the hygienist does at each stage.

Choose a metric that fits the mechanism

Example: 2 mg/m³ for 3 working-years plus 0.5 mg/m³ for 4 working-years gives 8 mg/m³-years on this defined exposure scale. It is an exposure index, not a measured absorbed dose. State the work schedule and concentration assumptions.

SEG: a similar exposure group shares relevant task, agent and control patterns. JEM: a job-exposure matrix assigns exposure estimates to jobs, often by period or industry. A shared title does not establish identical exposure; workers may perform different tasks. NCI: exposure assessment methods.

The IH's work through the study

Medical information needs authorized handling and confidentiality. Employer reports should communicate prevention findings without unnecessary disclosure of identifiable health information. Regulatory medical-surveillance triggers vary by agent and jurisdiction; a rule such as “30 exposure days” is not a general epidemiologic study requirement.

Exposure reconstruction should preserve historical process changes and the basis for each estimate. Document how uncertain estimates were assigned so future reviewers can understand the reasoning.

11 Apply the Ideas to Three Short Workplace Cases

A process change, a long-latency cancer and a ventilation improvement, then six check questions.

Case A: symptoms after a process change

A factory receives 12 respiratory complaints after introducing a material. First define the symptoms/outcomes, dates, population and tasks; investigate the agent and controls. Counts alone lack a denominator. If a complete workforce is identifiable, compare new outcomes across exposure groups. Check whether increased awareness or medical testing contributed to the apparent rise. Exposure control can proceed while the investigation continues.

Case B: long-latency cancer

A small number of workers has an unusual cancer. A case series can alert investigators, but cannot quantify comparative risk. A case-control study may efficiently compare work histories; a historical cohort may be feasible if rosters and follow-up records are strong. A present-day sample alone cannot represent exposure decades earlier.

Case C: improved ventilation

Complaints decline after local exhaust ventilation is installed. This supports further evaluation, but production volume, workforce composition or reporting may also have changed. Check exposure reduction, use comparable outcome definitions, examine repeated observations and consider a suitable comparison department. A single before/after count does not isolate the intervention effect.

Check yourself

1. Forty existing cases among 400 workers on one date: which measure and value?
2. Four first cases in 196 person-years: what incidence rate per 1,000 PY?
3. Exposure risks of 10% and 5%: what RR, RD and AFₑ?
4. Fifteen observed deaths and ten expected: what SMR and excess percentage?
5. RR 2.0 with 95% CI 0.6–6.7: does this establish no effect?
6. Ninety true positives and ninety false positives: what PPV?

Answers: 1. Point prevalence, 10%. 2. 20.4/1,000 PY. 3. RR 2; RD 5 percentage points; AFₑ 50% under causal assumptions. 4. SMR 1.5 or 150; 50% above expected. 5. No: the estimate is imprecise and compatible with a range of effects. 6. 50%.

12 Final Review Card

One line per cue, and where this chapter ends.

Where this chapter ends

For the wider CIH biostatistics area, study arithmetic/geometric means, normal and lognormal distributions, variance and GSD, percentiles, sampling strategies, confidence versus tolerance limits, and selection of statistical tests in a separate chapter. This guide provides the epidemiologic reasoning that makes those tools useful; it does not claim to cover every item in the combined rubric.