Evaluate Alternate Methods of Disease Grouping

The constitution of disease groups is subject to varying definition, sometimes based on pathologic characteristics, sometimes based on clinical manifestations or disease course, and increasingly based on molecular analyses. Designing epidemiologic research requires making decisions about what constitutes a meaningful disease entity for study, and that question depends in part on the exposure of interest. If our interest is in the effect of heat waves on mortality, for example, all cause mortality may be the appropriate entity to consider if the mechanism is one in which persons whose health is compromised for any reason are more susceptible. On the other hand, we may believe that only a subset of diseases will mediate an effect of heat on mortality, perhaps heart or respiratory disease only if we believe that the isolated effect is one of cardiopulmonary stress, and heat would not be expected to increase risk of death due to injury or infection. The rationale for considering and choosing the optimal disease entity is a focus on what set of conditions may reasonably be expected to be influenced by the exposure under study. As we zero in on heart disease deaths, omitting deaths that are not thought to be plausibly related to heat, the adverse effects of high temperature should become increasingly clear. When we study all deaths from a wide range of factors not likely to be related to the heat wave, such as cancer, motor vehicle injury, and infant deaths from congenital anomalies, the inclusion of those entities not truly related to heat will generate a correct measure of the impact of heat on total mortality but the strength of association (relative risk) will be reduced even though in principle the risk difference would not be affected.

Selection of the proper disease entity depends on our understanding of the eti-ologic process. There is almost always some degree of uncertainty regarding the etiologic mechanism that might link a particular exposure and disease outcome, even for extensively researched topics. Furthermore, different hypothesized eti-ologic mechanisms often lead to different expectations regarding the disease subset that is likely to be causally related to the exposure under study. The optimal grouping of diseases to evaluate as the study outcome measure is the complete aggregation that is plausibly related to the exposure, including cases that could be affected and excluding cases that are not potentially affected. Exclusion of relevant cases leads to loss of power or precision, and inclusion of cases of disease that are not potentially related to the exposure of interest in the analysis biases the measure of the relationship between exposure and disease as a form of misclassification. The additional cases of disease that are not plausibly related to the exposure represent false positives relative to the exposure-disease relation ship under study, and have the same impact as any other source of false positives on the results.

If benzene exposure truly caused only one form of leukemia, acute myeloid leukemia, as some have argued (Wong, 1995), then studies of benzene and leukemia that include other forms, such as chronic myeloid leukemia and acute lymphocytic leukemia would be expected to yield weaker ratio measures of association. That weaker measure would accurately reflect the impact of benzene on total leukemia, but would reflect a smaller magnitude than would be found for acute myeloid leukemia alone. Those cases of other forms of leukemia would act analogously to false positive cases of acute myeloid leukemia, diluting the measured association. Under the hypothesis of an effect limited to acute myeloid leukemia, the exposure pattern of cases of other types of leukemia would be identical to those of persons free of disease. On the other hand, if multiple types of leukemia are in fact affected by benzene, as suggested in a recent review (Savitz & Andrews, 1997) and a report from a large cohort study (Hayes et al., 1997), then restricting an already rare disease, leukemia, to the subset of acute myeloid leukemia, is wasteful. Relative to studying all leukemias, there would be a substantial loss of precision, and may not be a gain in specificity of association with benzene exposure.

Often we are faced with uncertainty and reasonable arguments that would support more than one approach to disease grouping. Rather than arbitrarily adopting one strategy, the best approach may be to examine the results under several scenarios and consider what impact misclassification would be likely to have had under those alternative assumptions. If, in fact, there is a causal association with at least some subset of disease, then the analysis that is restricted to that subset will show a stronger exposure-disease association than analyses that are more inclusive. If there is reasonable doubt about whether etiologically distinctive subsets of disease may be present, there is an incentive to present and evaluate results for those subsets. Should the subsets all yield similar measures of effect, then one might infer that nothing was gained and the exposure has similar consequences for all the subgroups of disease. On the other hand, generating data for disease subsets is the only means for discovering that some subsets are affected by the exposure whereas others are not.

For example, we hypothesized that among all cases of preterm delivery, distinctive clinical presentations may correspond to different etiologic mechanisms: some occur following spontaneous onset of labor, some following spontaneous rupture of the chorioamniotic membranes, and some result from medical interventions in response to health complications of the mother or fetus that require early delivery, such as severe pre-eclampsia or fetal distress (Savitz et al., 1991). If this is a valid basis for dividing cases to study etiology, then associations with subsets of cases will be stronger than for the aggregation of all preterm delivery cases. At present, the empirical evidence regarding such heterogeneity is mixed, with some risk factors distinctive by subtype whereas other potential causes of preterm birth appear to be associated with two or all three subgroups (Lang et al., 1996; Berkowitz et al., 1998).

Some diseases are aggregations of subgroups, in a sense demanding consideration of subtypes of a naturally heterogeneous entity. Brain cancer is defined solely by the anatomic location of the tumor, with a wide range of histologic types with varying prognosis and quite possibly varying etiology. In a rather sophisticated examination of the issue of magnetic field exposure and brain cancer in a Canadian case-control study, Villeneuve et al. (2002) hypothesized that the exposure acts as a tumor promoter and would thus show the strongest association for the most aggressive subtypes of brain cancer. Subsets of brain cancer were examined empirically (Table 9.2) and there was clear heterogeneity in pat-

Table 9.2. The Risk of Brain Cancer According to the Highest Average Level of Occupational Magnetic Field Exposure Ever Received by Histological Type. Canadian National Enhance Cancer Surveillance System, Male Participants, 1994-1997

HIGHEST AVERAGE OCCUPATIONAL

Table 9.2. The Risk of Brain Cancer According to the Highest Average Level of Occupational Magnetic Field Exposure Ever Received by Histological Type. Canadian National Enhance Cancer Surveillance System, Male Participants, 1994-1997

HIGHEST AVERAGE OCCUPATIONAL

EXPOSURE MAGNETIC FIELDS EVER RECEIVED

CASES

CONTROLS

ODDS RATIO*

95% CI

ODDS RATIOt

95% CI

All Brain Cancers

<0.3 ¡T$

410

420

1.0

1.0

>0.3 ¡T

133

123

1.11

0.84-1.48

1.12

0.83-1.51

>0.6 ¡T

42

29

1.38

0.79-2.42

1.33

0.75-2.36

Astrocytomas

<0.3 ¡ T

163

160

1.0

1.0

>0.3 ¡T

51

54

0.93

0.60-1.44

0.93

0.59-1.47

>0.6 ¡T

12

16

0.61

0.26-1.49

0.59

0.24-1.45

Glioblastoma

Multiforme

<0.3 ¡T

143

156

1.0

1.0

>0.3 ¡T

55

42

1.50

0.91-2.46

1.48

0.89-2.47

>0.6 ¡T

18

6

5.50

1.22-24.8

5.36

1.16-24.78

Other

<0.3 ¡T

92

94

1.0

1.0

>0.3 ¡T

23

21

1.11

0.59-2.10

1.10

0.58-2.09

>0.6 ¡T

9

7

1.50

0.53-4.21

1.58

0.56-4.50

♦Unadjusted odds ratio obtained from the conditional logistic model.

tThe odds ratio was adjusted for occupational exposure to ionizing radiation and vinyl chloride. ¿Referent group. Villeneuve et al., 2002.

♦Unadjusted odds ratio obtained from the conditional logistic model.

tThe odds ratio was adjusted for occupational exposure to ionizing radiation and vinyl chloride. ¿Referent group. Villeneuve et al., 2002.

terns of association across tumor groupings. A modest association was found for brain cancer in the aggregate (relative risks of 1.3-1.4 in the highest exposure category), with markedly stronger associations for the more aggressive subtype, glioblastoma multiforme, with relative risks over 5.0. Whether this pattern reflects a causal effect or not, the heterogeneity in risk across subtypes provides informative suggestions and helps to focus additional research that addresses the same hypothesis or actually refine the hypothesis about whether and how magnetic fields might affect brain cancer.

As in many suggested approaches to epidemiologic data analysis, there is no analysis that can discern the underlying truth. Hypotheses are proposed, results are generated, and then interpretations are made, with greater information provided when informative disease subsets can be isolated, and considered. Several caveats to this approach must be noted however. Alternative grouping schemes need to have a logical basis in order for the results to be interpretable. A plausible theoretical foundation is needed for each approach to grouping that is then examined in order for the association to have any broader meaning and to advance understanding of disease etiology. To note that an arbitrarily chosen subset of cases, such as those who came to the clinic on Tuesdays, shows a stronger relationship to disease than cases in the aggregate, is of little help in evaluating misclassification and understanding the causal process. Through random processes, there will always be disease subsets more and less strongly related to exposure, but to be worthy of evaluation, finding such heterogeneity or even the absence of heterogeneity that might have been expected under some plausible hypothesis should advance knowledge. In fact, random error becomes a much greater problem for case subgroups than for the disease group in the aggregate, simply due to a diminution of the numbers of cases in the analysis. Arbitrary, excessive splitting of cases for analysis has the danger of generating false leads based solely on random error. Nonetheless, except when imprecision is extreme, it would often be preferable to have a less precise result for the subgroup of cases that is truly affected by the exposure than to have a more precise result for a broader aggregation of cases, some of which are affected by the exposure and some of which are not.

10 Ways To Fight Off Cancer

10 Ways To Fight Off Cancer

Learning About 10 Ways Fight Off Cancer Can Have Amazing Benefits For Your Life The Best Tips On How To Keep This Killer At Bay Discovering that you or a loved one has cancer can be utterly terrifying. All the same, once you comprehend the causes of cancer and learn how to reverse those causes, you or your loved one may have more than a fighting chance of beating out cancer.

Get My Free Ebook


Post a comment