EPID168 Final Exam Answer Guide, Fall 1999

University of North Carolina at Chapel Hill
School of Public Health
Department of Epidemiology
Fundamentals of Epidemiology (EPID 168)
Final Examination, Fall 1999

Answer Guide

Case-control studies are well-suited for studying risk factors for brain cancer because the disease is rare (hence difficult to study in a cohort design). Also, the case-control design facilitates examining many risk factors of current interest, a substantial advantage when so few risk factors have been identified. A retrospective cohort study can examine only exposures for which historical data are available.

A "population-based case-control study" is a case-control study for which the study base is a defined population. With a hospital-based case-control study, it is difficult to specify the study base, since which cases come to a given hospital is influenced by such factors as seriousness and treatability of the disease, type of hospital, and health care financing ability and arrangements. A representative sample from this same defined population yields a control group that permits valid estimation of odds ratios. In contrast, the validity of measures of association estimated using a control group selected from among hospitalized persons is always somewhat uncertain, since it is generally impossible to know how well such controls provide valid estimates of the study base.

B. The method of finding cases was passive surveillance.

D. Using incident cases allows the odds ratio to estimate the incidence density ratio or risk ratio. In contrast, the exposure distribution among prevalent cases will reflect differential survival in relation to exposures as well as differential incidence.

C. Selective recall (the disease status has an influence on the recall of exposures) is a form of information bias, not selection bias.

Since 26 of the cases reported using a private well, 85-26=59 cases did not. Sensitivity=0.77 means that the proxy respondents correctly classified as "exposed" 0.77x26 approx.=20 brain cancer cases. Specificity=0.75 means that the proxy respondents correctly classified as "unexposed" 0.75x59 approx.=44 brain cancer cases. The rest of the table can be completed by subtraction and addition. As a check on the arithmetic, the positive predictive value is 20/35 approx.=0.57.

Validation of proxy reports of use water from a private well
	Case's self report
Report of proxy	Yes	No	Total
Yes private well	20	15	35
No private well	6	44	50
Total	26	59	85

C. Manifestional criteria – histological criteria are observable characteristics of tumor cells in microscopic examination.

C. 10% – the proportion of cases who are exposed is 85/291 approx.=0.29, and the OR approx.=1.5. Substituting into the formula for PARP in a case-control study gives 0.29x(1.5–1)/1.5 approx.=0.097.

A. The study base consisted of those people who if they developed brain cancer could have been selected as a case.

E. The OR for the oldest group is less likely to be due to chance because the confidence interval does not include 1.0 (although not without problems, this response was the best).

C. The narrower confidence interval indicates that the estimate for men is slightly more precise.

D. 2.6 = (7x423)/(30x38) for men with above median tap water intake

	Exposure to chlorinated water
	40+ years	< 40 years
Cases	7	30
Controls	38	423

	Average population
	≥50,010	≤2,500
Cases	32	112
Controls	246	780

Crude OR = (32x780)/(112x246) = 0.91 versus 0.7 adjusted. The estimates differ because the OR in the table has been adjusted for age and sex (according to the footnote to Table 1).

Bradford Hill criteria are:

The associations observed for this association were of medium strength (1.7 for 20-39 years of exposure to chlorinated surface water, 2.5 for >=40 years). The authors measured lifetime exposure (through recall) so in spite of the prolonged induction and latent periods for brain cancer, the criterion of temporality is satisfied to some extent. Some of the exposure history in Table 3 must have occurred after the brain cancer had begun and is therefore not relevant. However, it seems unlikely that if the association were causal it could go in the opposite direction (i.e., brain cancer causes exposure to chlorinated water). There is little evidence to support the plausibility of the association nor of its being found for men but not for women. Studies of the association have not yielded consistent results. (The remaining criteria – coherence, experiment, and analogy – are not applicable to the information in the article.)

The "ecologic fallacy" is the inference from aggregate data that a relationship exists at the level of the individual. The flaw in this inference is that the prevalences of a characteristic (e.g., exposure to trihalomethanes in drinking water) and a condition (e.g., brain cancer) can both be elevated in a population even if the individuals who possess the characteristic are not those with the condition. In the study described in the question, people who developed brain cancer may not themselves have ingested large amounts of THM despite living in counties with high THM levels in the county water supplies. A related analytic problem is that the absence of individual-level data precludes individual-level control for potential confounders, such as farming occupation.

D (both A and C). "Induction period" refers to the time between exposure and the onset of the disease; "latent period" refers to the time from disease onset to diagnosis. For exposure to be causal in early stages of tumor development, the exposure must be present prior to the latent period. In principle, exposure prior to the sum of the longest possible induction period and the longest possible latent period would not be relevant, either.

B. Stratification of cases by brain cancer type would permit examination of the relationship for the individual subtypes.

B. "We selected cases and controls with at least 70 percent of their lifetime years with a known source of drinking water in order to …minimize misclassification of exposure …" (end of p 554).

(question was not asked)

Risk of brain cancer by number of years resided in a dwelling supplied with chlorinated surface water
	>=40 years	None	Total
Cases	13 + 7 = 20	92 + 78 = 170	190
Controls	81 + 60 = 141	875 + 400 = 1275	1416
Total	161	1445	1606

OR = (20x1275) / (170x141) approx.= 1.1

The presence of confounding is usually determined on the basis of existence of a meaningful difference between the crude and adjusted OR's, which there is not. Since the OR's for men (2.5) and women (0.7) are quite different, for an unambiguous indication of confounding the crude OR would have to be above 2.5 or below 0.7.

Yes, there is modification of the OR by gender, in that they differ meaningfully. Although the two confidence intervals overlap substantially, neither point estimate is contained within the confidence interval for the other gender's estimate, so besides giving opposite "messages", the two OR's are likely to differ in fact (not necessarily for biological reasons).

The ORs shown in Table 1 are (according to the footnote) controlled for farming occupation, so that should not be a source of confounding, except to the extent that the crudeness of the measure ("yes" versus "no") prevents the control from being fully effective.

Farming occupation and brain cancer risk
	Farming occupation
	Yes	No	Total
Cases	85	206	291
Controls	628	1355	1983
Total	713	1561	2274

OR = (85x1355) / (206x628) approx.= 0.89

The OR of 0.89 indicates no (or possibly a slight inverse) crude association between brain cancer risk and having had a farming occupation.

If 10% of "unexposed" cases in fact had had a farming occupation, then 0.10x206=21 cases should be reclassified as exposed; if 15% of "exposed" controls in fact had not had a farming occupation, then 0.15×628=94 controls should be reclassified as exposed. The resulting table and OR would be:

Farming occupation and brain cancer risk
	Farming occupation
	Yes	No	Total
Cases	85 + 21 = 106	206 – 21 = 185	291
Controls	628 – 94 = 534	1355 + 94 = 1449	1983
Total	640	1634	2274

OR = (106x1449) / (185x534) approx.= 1.6

Correcting for the misclassification produces a table with a moderate positive association between odds of farming occupation and brain cancer.

A. The odds ratios are different due to differential misclassification of exposure.

D. Sensitivity is a measure of validity (kappa is a measure of agreement that gives equal weight to both classifications; standard error measures variability of an estimate)

The crude OR = (85 x 1,355) / (206 x 628) = 0.89. This value is substantially different from the adjusted value of 1.5, indicating that confounding by age and sex are present.

Controls were matched by age and sex to cancer cases for five cancer sites. Thus, the control group is not a simple random sample from the study base, so that analyses must control for the matching variables.

12/20/1999, Wayne D. Rosamond and Victor J. Schoenbach