 

 

{15,

 

i: s...

... .5. r

‘13:?
s,

{to}? :3

5“. 3.. . ,. . ’ r
, 3mm.

 

 

THESIS

5W3?“

This is to certify that the
thesis entitled

COMPARING QUESTIONNAIRE BASED MEASURES OF
GREAT LAKES SPORT FISH CONSUMPTION FOR
PREDICTION OF HUMAN SERUM POLYCHLORINATED
BIPHENYL (PCB) LEVELS

presented by
Andrew Mullard

has been accepted towards fulfillment
of the. requirements for the

Master’s degree in EpidemiMy

 

 

Major Professor’s Signature

jth 7 J L‘D()(‘f
( L

 

Date

MSU is an Afﬁrmative Action/Equal Opportunity Institution

-v «r-o r-o- i 4» Wwvm—r—w—r-OM—d

'

0 r" ﬁ-vv—wv—wv

 

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE

DATE DUE

DATE DUE

 

I

b igﬁp'fb 9"“
w‘t—é 0_ L196

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 cJCIRC/DateDuepSS-p. 1 5

 

COMPARING QUESTIONNAIRE BASED MEASURES OF GREAT LAKES SPORT
FISH CONSUMPTION FOR PREDICTION OF HUMAN SERUM
POLYCHLORINATED BIPHENYL (PCB) LEVELS
By

Andrew Mullard

A THESIS

Submitted to
Michigan State University
in partial fulﬁllment of the requirements
for the degree of

MASTER OF SCIENCE
Department of Epidemiology

2004

ABSTRACT
COMPARING QUESTIONNAIRE BASED MEASURES OF GREAT LAKES SPORT
FISH CONSUMPTION FOR PREDICTION OF HUMAN SERUM
POLYCHLORINATED BIPHENYL (PCB) LEVELS
By

Andrew Mullard

Human consumption of Great Lakes Sport ﬁsh has been associated with increased
PCB exposure and adverse health outcomes. This thesis evaluates the criterion-related
validity of four simple questionnaire based measures of sport ﬁsh consumption for
prediction of human serum total PCB level.

The study population was a cohort of Michigan sport ﬁsh consumers (n=129).
Measures of consumption included: 1) lifetime years (YEARS); 2) number of meals in
past year (MEALS); 3) lifetime meals (MXY 1); and 4) modiﬁed lifetime meals (MXY2).
Multivariable regression was used to construct adjusted regression equations. Predictors
were compared using a test for comparing non-nested models.

All adjusted predictors were signiﬁcantly associated with serum PCBS (YEARS
[R-Square delta=4%], MEALS [5%], MXYl [7%], MXY2 [9%]). Only MXY2 showed
a statistically signiﬁcant increase in predictive power over YEARS and MEALS.
However, the observed trend for predictive power was consistent with theory.

The modiﬁed index for lifetime sport ﬁsh meals (MXY2) is simple to attain and
maximizes predictive power among the evaluated predictors. However, no test for model
reliability was performed due to insufﬁcient sample size. This study was performed on a
population of high sport ﬁsh consumers; therefore these ﬁndings may not apply to other

populations.

ACKNOWLEDGMENTS

I wish to thank my thesis advisor, Nigel Paneth, for his guidance and mentoring
throughout the course of my work and research training. Also, I wish to extend my
thanks to the other members on my committee including, John Hesse and Joseph
Gardiner, for their contributions and expertise.

I extend my thanks to the former Fisheaters Project Director, Julia Wirth, for her
contributions and encouragement. I also acknowledge all the students who worked with
dedication on the Fisheaters Study.

I also acknowledge Henry Anderson, and Lawrence Hanrahan, for their mentoring
and valued contributions to this area of research.

To my loving family, and especially toward mother and father, I express deep and
lasting gratitude for their support and patience. I also hold dearly my loving ﬁancée, and

thank her for sticking with me through thick and thin.

iii

TABLE OF CONTENTS

LIST OF TABLES ............................................................................................................. vi
LIST OF FIGURES .......................................................................................................... vii
INTRODUCTION .............................................................................................................. 1
BACKGROUND .................................................................................. . .............................. 3
Chemistry, Production, Industrial Use and Environmental Fate .................................... 3
Risk Assessment ............................................................................................................. 5
Hazard Control ................................................................................................................ 6
PCB Exposures in the Great Lakes Basin ....................................................................... 7
Pharmacodynamics and Body Burden ............................................................................ 9
Exposure Assessment .................................................................................................... 11
Sample Selection ........................................................................................................... 17
Selected Great Lakes Cohorts ....................................................................................... 18
Studies Reporting on Exposure-response Relationships .............................................. 20
Rationale and Aims ....................................................................................................... 24
Objectives ..................................................................................................................... 26
MATERIALS & METHODS ........................................................................................... 27
Study Population ........................................................................................................... 27
Questionnaire-based Measures of Sport Fish Consumption ......................................... 27
Biosample Collection and Analysis .............................................................................. 29
Derived measures for PCB exposure ............................................................................ 31
Descriptive Statistics ..................................................................................................... 32
Correlation Matrix ........................................................................................................ 32
Multiple Regression Modeling ..................................................................................... 32
Comparing Non-nested Multiple Regression Models ................................................... 35
Predicted Mean Serum PCB Levels at Selected Exposure Levels ............................... 37
RESULTS ......................................................................................................................... 38
Characteristics of the Study Data .................................................................................. 38
Correlation Matrix ........................................................................................................ 41
Multiple Regression Models ......................................................................................... 42
Model A: Model PCBs on Lifetime Years Eating Sport Fish (YEARS) ...................... 43
Model B: Model PCBS on Sport Fish Meals in Previous 12 Months (MEALS) .......... 46
Model C: Model PCBS on Lifetime Sport Fish Meals (MXY 1) ................................... 49
Model D: Model PCBS on Modiﬁed Lifetime Sport Fish Meals (MXY2) ................... 51
Comparing Predictors ................................................................................................... 54
Predicted Mean Serum PCB Levels with Oversampling .............................................. 56
DISCUSSION ................................................................................................................... 58
Applied Epidemiology .................................................................................................. 58
Findings ......................................................................................................................... 58
Strengths ....................................................................................................................... 63
Limitations .................................................................................................................... 63
Implications for Future Research .................................................................................. 64
Conclusions ................................................................................................................... 66
APPENDICES .................................................................................................................. 67
APPENDIX A: SOURCE QUESTION FOR YEARS ................................................. 68

iv

APPENDIX B: SOURCE QUESTION FOR MEALS ................................................. 69

APPENDIX C: INFLUENCE PLOT FOR MODEL A ................................................ 70
APPENDIX D: INFLUENCE PLOT FOR MODEL B ................................................ 71
APPENDIX E: INFLUENCE PLOT FOR MODEL C ................................................ 72
APPENDIX F: INFLUENCE PLOT FOR MODEL D ................................................ 73
REFERENCES ................................................................................................................. 74

LIST OF TABLES

Table 1.Demographic and sport ﬁsh related characteristics of 129 subjects in the
F isheaters Cohort-A, 1997-2000 ....................................................................................... 39

Table 2. Serum PCB level (ppb) for the Fisheaters Cohort-A, 1997-2000 ....................... 40

Table 3. Pearson correlations (prob > |r| under H0: r=0) between serum PCB level,
control variables, and measures of sport ﬁsh consumption in study data (n=129). .......... 42

Table 4. ANOVA and parameter estimates for regression of loglO serum total PCB level
on control variables: gender, age, and BMI. ..................................................................... 43

Table 5. ANOVA and parameter estimates for regression of loglO serum total PCB level
on gender, age, BMI and YEARS ..................................................................................... 45

Table 6. ANOVA and parameter estimates for regression of loglO serum total PCB level
on gender, age, BMI and MEALS. ................................................................................... 48

Table 7. ANOVA and parameter estimates for regression of loglO serum total PCB level
on gender, age, BMI and MXYl. ..................................................................................... 51

Table 8. AN OVA and parameter estimates for regression of loglO serum total PCB level
on gender, age, BMI and MXY 2. ..................................................................................... 54

Table 9. Partial regression mean square ratios with 95% conﬁdence intervals for
comparing non-nested models (n=128, alpha=0.05). ....................................................... 56

Table 10. Predicted adjusted mean serum PCB levels (ppb) with 95% conﬁdence
intervals and percent above full sample mean with no oversampling, oversampling top
50%, and oversampling top 20% for four questionnaire-based measures of sport ﬁsh
consumption (n=128). ....................................................................................................... 57

vi

LIST OF FIGURES

Figure 1. Distribution of serum PCB level in the Fisheaters Cohort-A (n=129). ............. 41

Figure 2. Partial residual plot of SERUMPCB on YEARS, with spline regression line,
and 90% conﬁdence intervals, aﬁer correcting for GENDER, AGE and BMI. ............... 44

Figure 3. Partial residual plot of SERUMPCB on MEALS, with spline regression line,
and 90% conﬁdence intervals, after correcting for GENDER, AGE and BMI. ............... 47

Figure 4. Partial residual plot of SERUMPCB on MXYl, with spline regression line, and
90% conﬁdence intervals, after correcting for GENDER, AGE and BMI. ...................... 50

Figure 5. Partial residual plot of SERUMPCB on MXY2, with spline regression line and
90% conﬁdence intervals, after correcting for GENDER, AGE and BMI. ...................... 52

Figure 6. Overlay of partial residual plots, and regression lines, of SERUMPCB on
standardized: YEARS, MEALS, MXYl and MXY 2 after controlling for GENDER, AGE
and BMI. ........................................................................................................................... 55

Figure 7. Diagram of a theoretically plausible relationship between self-reported sport
ﬁsh consumption (E1), biosample based measures (E2), relevant dose (D), and response
(R). .................................................................................................................................... 62

vii

CHAPTER I

INTRODUCTION

Polychlorinated Bipheny1s(PCBs) are a group of synthetic compounds with
unusual properties that made them desirable in many commercial and industrial
applications. Between 1929 and 1977, PCBs were produced in massive quantities. Large
quantities of PCBS were released into the environment leading to exposures in wildlife
and humans. By the mid-1960’s concerns arose over PCB pollution and hazards. PCBS
were found to persist in the environment, to be toxic to humans exposed to high doses,
and to be toxic to some wildlife even when exposed to very low doses. In 1977, these
discoveries led to the outright ban of PCBS resulting in major reductions in wildlife and
human exposures. However, the redistribution of PCBS already in the environment
continues to pose a diminishing but signiﬁcant hazard to wildlife and humans.

Since the ban, an extensive body of PCB toxicology and epidemiology research
has been produced. A number of excellent reviews are available that explicate the
contribution of this research toward PCB risk assessment. The toxicology research,
animal models, observational studies of wildlife, and other sources have proven
indispensable in contributing to an evidence based approach to PCB risk assessment.
However, only epidemiologic investigations can provide direct information about the
PCB exposure-response relationship in humans.

Many epidemiologic studies have been conducted that report associations
between PCB exposure and health effects such as cancer, and adverse reproductive and
hormonal related outcomes. These studies have been inﬂuential in the classiﬁcation of

PCBS as carcinogens, neurobehavioral toxicants, and suspect endocrine disruptors.

However, controversy surrounds these designations because of limitations with existing
studies. Although the current weight of epidemiologic evidence suggests an association
between exposure to PCBS and some adverse health outcomes in humans, strong
evidence for causal relationships has not been established.

In the Great Lakes Basin, consumption of PCB contaminated sport ﬁsh is the
primary route for human exposure to PCBS. The population of Great Lakes sport ﬁsh
consumers, because of its increased risk for exposure to PCBS, has been of special
interest in researching the PCB exposure-response relationship. This exposure route is of
similar concern for other regions internationally. For example, PCB contamination in the
Baltic Sea ﬁshery poses similar exposures to ﬁsh-consuming populations in northern
Europe.

This thesis presents the background on PCB risk assessment and limits the scope
of the literature review to epidemiologic studies reporting PCB exposure-response
relationships in populations of Great Lakes sport ﬁsh consumers. The emphasis is on
methods of exposure assessment, but the main ﬁndings for study outcomes are also
considered. The rationale for this thesis is two-fold. First, the comparability of existing
research is hindered by the lack of minimum standards for reporting exposures. Second,
no evaluation of methods of exposure assessment has been performed that may offer
meaningful reductions in exposure measurement error. The main objective is to evaluate
a set of predictors by testing hypotheses comparing the predictive power of four simple,
but representative, methods for exposure assessment. Finally, recommendations are
made and the strengths and limitations of questionnaire and biosample based approaches

are discussed.

CHAPTER II

BACKGROUND

Chemistry, Production, Industrial Use and Environmental Fate

Polychlorinated Biphenyls (PCBS) are a class of synthetic chemicals that possess
a unique chemistry. Their qualities have made them valued and widely used in industry,
but also pose special concern as environmental pollutants.

Chemistry. PCBS are a class of synthetic organic chemicals, composed of a 12-
carbon biphenyl ring with 1 to 10 chlorine substitutions. Depending on the number and
position of chlorine substitution, up to 209 individual congeners are possible (1). PCBS
vary from hydrophobic/lipophilic oily liquids to waxy solids, and are chemically stable,
non-ﬂammable, have high boiling points, and electrical insulating properties (2).

Production. An estimated 1.5 billion pounds of PCBs were produced in the
United States between 1929 and 1979 (3). Most PCBS were produced and sold
commercially as Aroclor® mixtures with signature levels of individual congeners.
Aroclor® names reﬂect the number of carbon atoms and the percent of chlorine in the
mixture (e.g., ‘Aroclor® 1242’ has 12 carbon atoms and is 42% chlorine by weight).

Industrial use. PCBS had hundreds of practical uses in industrial and commercial
applications (e. g., in electrical and hydraulic equipment, as heat exchange ﬂuids, as
plasticizers in paints, plastics and rubber products, in pigments, dyes and carbonless copy
paper, and many other applications) (2, 3).

Environmental fate. PCBS are also currently released to the environment from
landﬁlls containing PCB waste materials and products, incineration of municipal refuse

and sewage sludge, and improper disposal of PCB materials, such as waste transformer

ﬂuid, to open areas (4). The EPA's Toxic Chemical Release Inventory estimates that in
the US. between 1987 and 1993, over 74,000 pounds of PCBs were released into land
and water (4).

Current evidence suggests that the major source of PCB release to the
environment is an environmental cycling process of PCBS previously introduced into the
environment; this cycling process involves volatilization from ground surfaces (water,
soil) into the atmosphere with subsequent removal ﬁom the atmosphere via wet/dry
deposition and then revolatilization (5).

Volatization is recognized as a major mode of transport for PCB redistribution (4-
7). The International Joint Commission estimates that atmospheric deposition accounts
for the majority of PCB inputs to the upper Great Lakes. According to the Mass
Balancing of Toxic Chemicals in the Great Lakes: The Role of Atmospheric Depositions,
a 1988 EPA report, the atmosphere accounts for 90% of PCB deposition into Lake
Superior, 58% in Lake Michigan and 63% in Lake Huron (4). The ﬁgures for the Lower
Great Lakes are much smaller, but may still be signiﬁcant. The majority of inputs of
PCBS to Lakes Erie and Ontario is from other sources, including the upper lakes. Recent
information continues to support the relative role of the atmosphere for Lakes Superior
and Huron. The relative role of atmospheric deposition of toxics in Lakes Michigan may
be somewhat less and is under active investigation (4).

PCBS that ﬁnd their way into watersheds become sequestered in sediments where
they persist indeﬁnitely and slowly release into ecosystems. Because of their lipophilic
properties, PCB congeners biologically magnify by moving from external media into

biological compartments of living organisms. Similarly, these same PCBS magnify up

food chains, when one organism feeds on another, and can result in a million-fold
increase in the PCB levels in apex species (8, 9). In 1966, Jensen discovered PCBS as a
contaminant in wildlife (10). Shortly thereafter, elevated PCBS levels were reported

among apex species, such as predatory ﬁsh, piscivorous birds and humans (11-16).

Risk Assessment

In 1968, Japanese investigators reported on an outbreak called “Yusho” or oil
disease (17). The source of this poisoning was linked to accidental contamination of rice
oil with heat exchange ﬂuids containing high levels of PCBS, polychlorinated
dibenzoﬁirans and quarterphenyls (17). The identiﬁcation of these chemicals as one of
the causative agents for this outbreak spurred investigation of PCBS.

Subsequently, the relationship between PCB exposure and adverse health
outcomes have been the focus of considerable research. Toxicological research including
investigation of wildlife and animal models has linked exposure to PCB contaminated
sport ﬁsh to adverse effects including cancers and dysfunctions of immune, reproductive,
nervous, and endocrine systems (18).

A weight of evidence based approach indicates that signiﬁcant adverse health
outcomes are also associated with human exposure to PCBS through consumption of
large amounts of contaminated sport ﬁsh (2, 19-21). These studies report associations
between PCB exposure and a variety of adverse health outcomes including: disruption of
reproductive ﬁmction, neurobehavioral deﬁcits, developmental deﬁcits, self-reported
liver disease, self-reported diabetes, cancer risk, and effects on the thyroid and immune

systems (19).

Hazard Control

In the 1970’s, rising concern among scientists and the public over PCB toxicity,
water contamination, and risk of human and animal exposure, prompted aggressive
hazard control.

Ban on production and use. In 1977, the EPA banned the manufacture of PCBS,
and regulated their removal from industrial use, under the Toxic Substances Control Act
(22). PCBS are now regulated under many federal and state laws. The EPA has broad
authority to regulate virtually all aspects of the manufacture, distribution, use, and
disposal of PCBS (3, 22).

Fish consumption advisories. Since the mid 1970s, all Great Lakes States have
issued annual sport ﬁsh consumption advisories that aim to limit human exposure to
PCBS by this route (23-26). While PCBS are a common contaminant that triggers
advisories, several other contaminants do so as well. These include mercury,
polychlorinated dibenzodioxin, polychlorinated dibenzofurans, and persistent pesticides
such as Chlordane (26, 27). To reduce risk of exposure, sport ﬁsh consumers are urged
to decrease the amount consumed and to switch to less contaminated species (26). A
long term goal stated in the Great Lakes Strategy 2002 is that all Great Lakes ﬁsh should
be safe to eat without restriction (28).

Time trends. Environmental PCB levels have declined greatly since their ban.
However the rate of reduction has decreased, and some authors are reporting having
reached a near steady state in environmental PCB levels. In any case, current levels
continue to pose a sufﬁciently likely hazard to warrant continuance of sport ﬁsh

consumption advisories (3, 26).

PCB Exposures in the Great Lakes Basin

Residents of the Great Lakes Basin are exposed to PCBS through multiple routes
including background, occupational and dietary exposures. The distribution of PCB
levels in the entire population is positively skewed. A positively skewed distribution
describes an asymmetrical distribution in which most members of the population exhibit
very low levels of exposure near, but not at, zero. The remaining and much smaller
portion of the population exhibit higher, more positive, levels of exposure that taper off
toward higher positive values and form what is called the positive tail of the distribution.
The sources of exposure that contribute toward the formation of this distribution include
background exposures, occupational exposures, and dietary exposures (2).

Background exposures. The general Great Lakes population experiences low
level background exposures to PCBS from contaminated air, water, soil and food (2).
Higher PCB exposure levels have been reported among individuals living in
contaminated areas (29, 30).

Occupational exposures. Prior to their ban, occupational exposures to
commercial PCB mixtures were more common but since the implementation of hazard
control, and near elimination of use in industry, occupational exposures have decreased
greatly. However, utility workers, ﬁreﬁghters and a handful of other occupations are still
at signiﬁcant risk for exposure (31, 32). Sweeney et a1. (33) and Schantz et al. (34) report
on a special exposure source originating from PCB-lined silos. Risk for exposure from
this source occurred between 1941 until identiﬁcation and remediation in the early 19705.

Dietary exposures. Dietary intake studies of PCBS indicate that >80% of cases of
human exposure to chlorinated organic compounds occur through eating contaminated

food including meat and dairy products but primarily from eating sport ﬁsh (35-37).

Interestingly, a recent investigation comparing contaminant levels in farmed salmon to
wild salmon, in a large intemational sample of farm, market and wild collected ﬁsh
tissue, report that even the least contaminated farmed salmon had signiﬁcantly higher
contaminant loads than wild salmon (38). The authors ﬁndings suggest that farmed
salmon may be an important new, but as yet unconﬁrmed, source for dietary exposures to
PCBs and other contaminants (3 8).

Exposure from eating sport ﬁsh. Among the Great Lakes population, ﬁsh
consumption has been reported as the primary route for exposure to PCBs (35-37).
Several studies have shown that individuals who consume large amounts of PCB
contaminated Great Lakes sport ﬁsh have PCB body burdens that average three-fold
higher than the general population (37, 39-46).

Sport ﬁsh consumption advisories have been reported to have resulted in
reductions in the PCB exposure level for the overall population. He (47) and Tee et a1.
(48) have evaluated changes in sport ﬁsh consumption and serum levels of PCBs among

-individuals in Michigan during three time periods in 1974-75, 1979-82, and 1989-91.
The researchers found that over time, there was a decline both in the number of sport ﬁsh
meals consumed (median = 66, 54, and 31 per year, respectively), and the total amount of
sport ﬁsh consumed (median = 40, 38, and 16 pounds per year, respectively) (48).

General population. In the Great Lakes Basin, the general population
experiences nearly ubiquitous PCB exposures that originate from background sources and
meat and dairy products. Only special subpopulations tend to experience substantially
higher exposures ﬁ'om dietary or occupational exposures. For the entire population, most

members experience very low levels of exposure to PCBs, a small proportion experience

medium levels of exposure, and only an even smaller fraction experience high exposure
levels.

Special populations. Special subpopulations have been identiﬁed as either
especially vulnerable to exposure or at increased risk for exposure. Factors that identify
these populations include: gender, age, ethnicity, lifestyle, and geography.

Vulnerable populations include couples attempting to reproduce, fetuses, the very
young and the very old. These populations, and especially reproductive age females,
typically receive special consideration in advisories and outreach programs. Even so, a
signiﬁcant proportion of women of reproductive age fail to heed such warnings (23, 49-
52).

Populations at risk for elevated exposure include: sports anglers, subsistence
anglers, and their family members and acquaintances who eat signiﬁcant quantities of
contaminated sport ﬁsh. Men on average annually consume more sport ﬁsh than do
women (37, 47, 49). Similarly, elevated levels of sport ﬁsh consumption are often
observed in minority groups such as Native Americans and Asian Americans. Elevated
sport ﬁsh consumption among minority populations is largely attributable to cultural
practices, but may in part be explained by a decreased awareness of ﬁsh consumption

advisories (53, 54).

Pharmacodynamics and Body Burden

Pharmacodynamics describes the uptake, biotransformation, detoxiﬁcation,
elimination and accumulation of PCB exposures. Ingested PCBs rapidly absorb into the
blood stream, with an observable spike in serum PCB level, and are then absorbed by

liver, muscle, and other soft tissues before ﬁnally compartrnentalizing in fatty tissues (55-

57). PCBs sequestered in fatty tissue will approach an equilibrium state with all other
tissues in proportion to the respective tissue/blood ratios and the body burden (56).

In humans, PCB half lives for all 209 congeners vary immensely from months to
decades (58-60), but for the main congeners associated with sport ﬁsh, half lives range
ﬁ'om 2 to 6 years (60-65). Therefore, current PCB body burden is expected to be at least
moderately associated with past exposure. For example, He (47) concluded that sport
ﬁsh consumers who decreased their sport ﬁsh consumption (in a cohort whose ﬁsh
consumption habits and serum PCB levels were measured at three time periods in 1974-
75, 1979-82, and 1989-91) did not experience a parallel decline in serum PCB levels.

Other important mediators of PCB body burden are metabolic and compartmental
factors (29, 30, 66, 67). Metabolic factors include detoxiﬁcation systems and enzyme
induction. Compartrnental factors include compartment ratios and weight change.
Compartment ratios describe relative compartment volumes of blood and soft tissues to
fat tissues. Compartment ratio is directly associated with PCB half lives, since PCBs in
blood and soft tissue compartments are more readily excreted than PCBS sequestered in
fat tissue (68). Increasing Body Mass Index (BMI) is associated with a decreasing blood
and soft tissue to fat tissue compartment ratio that dilutes the level of PCBs in the blood,
decreases the rate of excretion, and consequently increases half lives (68). Increased
BMI has been associated with increased PCB body burden (29, 62, 69-74). Decreasing
BMI is associated with an increasing compartment ratio that concentrates the level of
PCBS in the blood, increases the rate of excretion, and consequently decreases half lives

(75,76)

10

Breast feeding is especially important because it can result in a transfer of a
signiﬁcant proportion of the mothers PCB body burden directly to the infant (77).
Lactation mobilizes and excretes fatty tissue stored lipids and PCBS in breast milk (56).
Schecter et a1. (78) estimated, in a case study, that breast feeding for two and half years
reduced the mother’s total level lipid-adjusted serum non-coplanar PCB levels by 78%
from 285 to 63 ng/g.

Bioaccumlation and body burden. Bioaccumulation describes the net result of
pharmacodynamics acting upon an individual’s cumulative exposure (9). Body burden is
a point-in-time measure of bioaccumulation. Pharrnacodynamic factors are powerful
mediators of PCB body burden that introduce considerable variation between a
cumulative exposure and an observed point in time body burden (9). These factors
warrant special consideration when investigating associations between PCB exposures,

body burdens and responses of interest.

Exposure Assessment

In environmental epidemiology, direct assessment of a dose-response relationship
is typically not feasible, since a direct measure of the biologically effective dose is rarely
available (79). Investigators therefore investigate exposure-response relationships by
basing inferences upon the best available proxy measures. Measures of the
characteristics of exposure can be based upon laboratory measures of biological samples
or derived from self-reported information (80).

Characteristics of exposure. Fundamental characteristics of exposure include
rate, duration, and timing. Theoretically, the product of rate and duration produce the

cumulative exposure, while timing describes when the exposure occurred (80).

ll

Information on timing is also often important when there are critical periods for exposure
(81). Timing of exposure is of particular concern for the investigation of reproductive
and developmental outcomes, especially since the biological mechanism involves soft
tissues that experience post-ingestion spikes (42, 82, 83).

The importance of appropriately characterizing an exposure is crucial to detecting
the true relationship between an exposure and response (81). A theoretically optimal
measure would ascertain rate, duration and timing to calculate an index of cumulative
exposure (80). For many exposure-related outcomes, it may be insufﬁcient to base
exposure assessment upon a single characteristic, such as rate or duration. When timing
is crucial, even an index of cumulative measures may be insufﬁcient. An exposure
measure that fails on theoretical grounds is very likely to introduce serious exposure
misclassiﬁcation (80).

Biosamples. Biosamples are the putative “Gold Standard” for PCB exposure
assessment. PCB levels are most commonly measured from venous blood, but cord
blood, fatty tissue, human milk, and neonatal meconium are also sampled (84, 85).
Meconium analysis is a new and potentially sensitive tool for the detection of fetal
exposure to environmental toxins (86-88).

Biosamples do face limitations however; an important issue is whether or not a
particular biosample measure actually characterizes exposure in a way that best
approximates the hypothetical biologically effective dose. Biosamples provide an index
of body burden, but typically do not measure cumulative exposure and offer no direct

information about the duration or timing of exposure. A biosample measure, that does

12

not measure the biologically effective dose inevitably introduces measurement error
resulting in exposure misclassiﬁcation.

However, even serum PCB levels are not always an accurate predictor of past
PCB exposure. In most other exposure studies, serum PCB levels have been used to
estimate body burden. However, as these other studies have reported, PCB levels in
blood can vary daily according to the food consumed and the state of the individual.
Pregnant women may have more PCBS in their blood due to the mobilization of fat stores
during pregnancy. Consequently, in interpreting relationships between ﬁsh consumption
and outcomes, the uncertainty of the actual PCB intake must be realized.

Serum PCB levels can be signiﬁcantly affected by serum lipid content (total
cholesterol, free cholesterol, triglycerides, and phospholipids) because of partitioning
between plasma and serum lipids. Non-fasting levels also tend to be higher. Therefore,
lipid-adjusted serum PCB levels are a better biomarker for body burden. Non-fasting
and/or non-lipid adjusted serum PCB levels introduce bias and/or measurement error
resulting in an additional source of exposure misclassiﬁcation.

A further, largely unacknowledged, limitation is encountered in reporting
exposure levels for individual congeners. The majority of the epidemiological studies
report exposure simply as summed or total PCBS, but a large body of data clearly
demonstrates different effects following exposure to different classes of PCB congeners
(89). In this light, summed or total PCBs is a conceptual construct and must be
recognized as a proxy measure for any biologically plausible disease causing agent.
Regardless of precision, the effect of reporting exposure as sumed or total PCBs, is a

loss of sensitivity that, like measurement error, dilutes strength of association (90).

13

DeVoto et al. (91) and Gladen et a1. (92) report that congeners tend to be well correlated
with each other (Pearson > 0.80). However such correlations still introduce a
considerable loss of sensitivity that contributes to exposure misclassiﬁcation.

Still other limitations strain sample selection and resource allocation. Biosample
collection is invasive, inconvenient and costly. Fear of needles, travel to clinics, waiting
rooms, and procrastination all decrease and potentially bias participation. Laboratory
analyses of collected biosamples are technically difﬁcult and costly (~$250 per sample
for full PCB proﬁle).

In summary, the most important advantage of biosamples is the ability to provide
an index of body burden for individual PCB congeners. Disadvantages of biosamples
include a lack of information about the duration and timing of exposure, a loss of
sensitivity when reporting individual congeners as summed or total PCBs, and a decrease
in participation rates. Moreover, the collection and analysis is costly. Sometimes a
researcher may decide that the disadvantages of biosample collection outweigh their
advantages, forgo biosample measures, and rely solely on questionnaire-based measures
of exposure.

Questionnaires. Even though questionnaire-based measures offer several
attractive advantages, they are generally regarded as inferior to biosamples. Advantages
of questionnaire-based measures are that they are relatively noninvasive, convenient and
cheap. Questionnaires are ﬂexible; they can be used in mass mailings, random digit
dialing, or face-to-face interviews. Because they are relatively noninvasive and

convenient, their use produces higher participation rates.

14

Self-reported sport ﬁsh consumption can serve as a proxy measure for PCB dose.
The rate of sport ﬁsh consumption is exposure per unit time (e. g., the number of
cigarettes per day). In this thesis, measures of sport ﬁsh consumption based upon the
number of meals in the past 12 months are considered rates of exposure where year is the
unit of time. However, a theoretical limitation of rate based-measures of exposure is that
it captures no information about the duration of exposure.

Duration of sport ﬁsh consumption is nearly always based on number of years.
Information about PCB half lives suggests that historical ﬁsh consumption should exert a
diminishing effect on current body burden (48). However, such reductions may be
attenuated by historically higher sport ﬁsh contaminant levels and rates of consumption
(61). In fact, some studies have reported that historic ﬁsh consumption predicts current
PCB levels better than recent consumption (48, 72). Hanrahan et al. (44) reported that
total number of years of eating sport ﬁsh was the best predictor of PCB body burden
when compared to rate-based measures.

Cumulative sport ﬁsh consumption is a calculated index based on the product of
exposure rate and duration. Several studies have used such measures of exposure (93-
96). Schwartz et al. (97), report a moderate correlation (r=0.21-0.29) between cumulative
ﬁsh consumption with PCB levels in maternal serum and milk. A theoretical limitation
of this measure of exposure is that it does not capture information about the timing of
exposure.

An index of cumulative PCB exposure (ng) is provided by the product of
exposure rate (meals/year), duration (years), meals size (grams), and is weighted for the

average ﬁsh tissue PCB level (ng/ g) by species consumed (50, 98, 99). Dar et al. (40)

15

used a similar weighting technique, but limited questions about ﬁsh consumption to the
12 months prior to pregnancy, and found a much higher correlation between ﬁsh
consumption and serum PCB levels than did studies that included years of consumption.

An ostensible limitation of questionnaire-based measures is that they all tend to
have inadequate criterion-related validity with serum PCB level (100). Problems with
obtaining accurate retrospective dietary information may be one factor contributing to the
relatively low associations between PCB body burdens and estimated ﬁsh consumption
based upon questionnaires (ﬁsh consumption diaries reduce the error but are limited to
use in prospective studies) (42, 101).

An alternative explanation that clearly limits the strength of association is that
questionnaire-based measures estimate cumulative exposures, not body burden. A
signiﬁcant proportion of the variance in serum PCB levels not explained by
questionnaire-based measures of exposure may be attributable to pharrnacodynamic
factors. This suggests that serum PCB level may not be an entirely meaningful criterion
for judging the accuracy, potential validity, or true value of questionnaire-based
measures.

Complementary measures of exposure. Consideration of the strengths and
limitations of biosample and questionnaire-based measures of exposure suggests that both
methods offer complementary information that help produce more meaningful and
accurate characterizations of exposure (102, 103). Although biologically-effective dose
(e. g., the level of particular PCB congeners in a body tissue during a speciﬁc time

window) is the ideal measure, hazard control is based on external exposures (e.g., number

16

of ﬁsh meals eaten by species per unit time), and therefore a researcher may want to

measure both exposure and dose.

Sample Selection

Sample selection for study subjects is a fundamental consideration for any study
design. Characteristics of the study sample determine statistical power (81, 104). In the
case of epidemiological investigations of exposure-response relationships, the researcher
strives to maximize statistical power while staying within budget and time constraints.

The skewed distribution of PCBS in the Great Lakes population complicates
sample selection for exposure-response studies, since it is difﬁcult to sample a sufﬁcient
proportion of exposed subjects. The ideal solution would be to perform stratiﬁed
sampling based on actual serurrr total PCB level. Unfortunately, such a process would
require a highly inefﬁcient use of resources since biosample based screening of the
general population is impractical due to the expense and inconvenience of biosample
collection.

Targeting special populations. One successful approach has been to target
special populations such as charter boat captains, Native Americans, or Hmong
Americans (105) (Table 1). These studies may effectively capture a more highly exposed
subpopulation, but even such subpopulations exhibit positively skewed distributions for
PCB body burden and they may have limited generalizability.

Oversampling. Another possible strategy is oversampling with screening
surveys. Oversampling is a procedure designed to selectively sample a segment of a
population in larger proportion than its actual representation in the target population.

Many of these studies use sample allocation strategies designed to over select for

17

increased exposure. For example, screening surveys that ascertain estimators of exposure
such as meals in the previous 12 months are used to target individuals more likely to have

elevated serum PCB levels.

Selected Great Lakes Cohorts

Table 1 provides a review of selected cohorts for the epidemiologic investigation
of PCB exposure-response relationships among consumers of Great Lakes sport ﬁsh. All
seven cohorts targeted special populations at elevated risk for exposure to PCBS by
consumption of Great Lakes sport ﬁsh. Four of the studies employed some form of
oversampling based on a screening survey to measure sport ﬁsh consumption. All of
these sampling strategies boost the study sample mean serum PCB levels however the
distributions are still highly positively skewed suggesting the potential for improved

oversampling.

18

Table 1. Selected cohorts for epidemiologic investigations of PCB exposure-response

relationships among Great Lakes sport fish consumers.

 

 

Target
Population

Cohort Risk Factors
(Enrollment for PCB Geographic area
Period) Exposure sampled Sample Allocation
Michigan Anglers and Ten Lake Michigan Weighted lifetime PCB
Department of geography shoreline exposure based on sport
Public Health communities in the ﬁsh meals. Pounds of
Cohort (MDPHC), State of Michigan sport-caught Great Lakes
1982 (16) ﬁsh consumed annually
Lake Michigan Geography Western side of the Pounds of Lake
Maternal Infant State of Michigan in Michigan ﬁsh
Cohort study close proximity to consumption in past 6
(LMMIC), 1983 Lake Michigan years
(106)
New York Anglers Anglers Sixteen State of n.r.
Cohort, 1991 (107) New York counties

in close proximity

to Lake Ontario
Oswego Cohort, Geography Oswego county on Weighted lifetime PCB
1991-1994 (108) southeastern shore exposure based on Lake

of Lake Ontario Ontario sport ﬁsh meals
Great Lakes Currently active Statewide in Stratiﬁed sample of
Charter Boat Great Lakes Wisconsin, Great Lakes sport ﬁsh
Captains Cohort, charter boat Michigan, Indiana, consumers
1993-1995 (109) captains Illinois and Ohio
Fisheater Anglers and Ten State of n.r.
Reproductive geography Michigan counties
Health Screening containing or
Survey Cohort, adjacent to Areas of
1993-1995 (49) Concern
Fisheater Family Anglers and Ten Michigan n.r
Health Project geography counties containing
Cohort-A, or adjacent to Areas
1997-2000 of Concern

 

n.r., none reported.

19

Studies Reporting on Exposure-response Relationships

This section reviews the methodologies and key ﬁndings of 23 studies that have

reported on PCB exposure-response relationships for subjects in the selected cohorts

(Tables 1 & 2). Sample selection in these studies employs a variety of strategies to

capture populations with elevated exposure to PCBS. The majority of the studies target

populations at increased risk for exposure.

Table 2. Selected epidemiology studies reporting on PCB exposure-response

relationships among Great Lakes sport fish consumers.

 

 

 

 

Self-report Sport Fish Biosample
Consumption PCB Level

Author Year Measure Findings Specimen Findings
Jacobson et al. 1984 Cumulative + Cord serum —
Fein et al. 1984 Cumulative + Serum +
Jacobson et al. 1985 Cumulative + Cord serum +
Jacobson et al. 1990 n.r. Cord serum +
Dar et al. 1992 Cumulative — Serum —
Mendola et al. 1995 Multiple — n.r.
Lonky et al. 1996 Cumulative +/— n.r. .
Jacobson et a1. 1996 n.r. Composite +/—
Buck et al. 1997 Years — n.r.
Mendola et al. 1997 Multiple + n.r.
Buck et al. 1999 Multiple — n.r.
Courval et al. 1999 Cumulative +/— n.r.
Schantz et a1. 1999 n.r. . Serum —
Buck et al. 2000 Multiple +/— n.r. .
Darvill et al. 2000 Cumulative — Cord serum +
Stewart et a1. 2000 Cumulative + Cord serum +
McGuinness et al. 2001 Multiple +/— n.r. .
Persky et al. 2001 Multiple +/— Serum +/—
Schantz et al. 2001 n.r. Serum +
Karmaus et al. 2002 n.r. Serum +
Stewart et al. 2003 n.r. Cord serum +
Weisskopt et al. 2003 n.r. Serum +
Karmaus et al. 2004 n.r. Serum +

 

n.r., none reported. + Positive ﬁndings. — Negative ﬁndings.

multiple responses.

+/— Both positive and negative ﬁndings on

Findings from the 23 studies in Table 2 have been very inﬂuential. Eighteen of
the studies report at least one positive ﬁnding of either a questionnaire or biosample
based exposure-response relationship.

Limitations. Several authors have suggested concerns over the validity of
ﬁndings drawn from this body of research (101, 110-116). A summary of these
limitations include: 1) weak associations; 2) potential bias; 3) lack of speciﬁcity; and 4)
inconsistent ﬁndings. Together these limitations bring into question the validity of
arguments that make causal inferences about PCB exposures and adverse health
outcomes.

These studies are vulnerable to non-differential selection bias, differential
selection bias, and recall bias. Non-differential selection bias occurs when the study
sample is not randomly sampled ﬁom the target population. Stein et al. (117), in a cross
sectional survey called the Fisheater Reproductive Health Screening Study, report that
among men, nonresponders had ﬁshed fewer days in the past year (12% reported no
ﬁshing, compared to 4. 3% of responders). Almost one half of nonresponders reported
no ﬁsh consumption in the past year, compared to one quarter of responders. Non-
differential selection may affect the external validity of study ﬁndings, but not the
internal validity.

Differential selection bias is potentially much more problematic because it can
result in spurious associations between the exposure measure and response, or the study’s
internal validity. This form of bias occurs when there is interaction between selection
bias for exposure and a selection bias for the response. It is likely to be directly

associated with the inconvenience and invasiveness of a study design since motivation

21

becomes an important factor inﬂuencing participation. Stein et al. (117) found no
evidence for differential selection bias in survey data from the F isheater Reproductive
Health Screening Study. Biosamples are more likely to be vulnerable to differential
selection bias, because of the increased difﬁculty and inconvenience of participation, but
as yet there are no publications on its actual impact. Also, Larsen et al. (118) report
ﬁndings that subfertile men were more likely to participate in their reproductive health
study. Considered together, these two reports should serve as a warning that differential
selection bias may occur.

Recall bias is another serious limitation that can result in spurious associations
between an exposure and a response. Recall bias occurs when error in ascertaining
exposure status is not random but varies by response status. Biosamples plainly avoid
this limitation, but self-report based measures are vulnerable to recall bias and may result
in spurious associations.

Confounding is a study design limitation that can introduce spurious associations
between the putative exposure and response. Both questionnaire and biosample-based
measures of exposure are vulnerable to confounding. Confounding occurs when the
effects of the exposure factor and the confounder factor are not separated. An exposure
that has no causal relationship with a response, but is associated with an agent that is
causally related to the response, will be associated with the response, distorting the
apparent effect of the exposure on risk for exhibiting the response (119). Sport ﬁsh is an
exposure source not only for PCBS but also a complex mixture of other bioaccumulative
pollutants (e.g., methylrnercury, dioxins, furans, Lindane isomers, Chlordane isomers,

DDE, Polybrominated Diphenyl Ethers, and toxic metabolites such as hydroxlated and

22

methylsulfone PCBs, and other identiﬁed and unidentiﬁed contaminants) (13, 15, 104,
120). Given the complex mixture of contaminants in sport ﬁsh, PCBs exposures from
consuming sport ﬁsh are almost certainly associated with other contaminants.

Another potential confounder is disease-induced weight change (e.g., weight loss
associated with cancer). Recent weight loss is associated with increased serum PCB
level, probably caused by mobilization of lipid-soluble PCB from body fat (121-125).
Breast feeding also mobilizes organochlorines from body stores, and the body burden of
organchlorines usually declines during breast-feeding (71). If the response under
investigation is associated with weight change, than an increased serum PCB levels may
be associated with the outcome, without being causally related to it.

Sport ﬁsh consumption may also be confounded with known risk factors for a
response such as demographic and lifestyle factors. For example, a few studies have
indicated that smoking may be positively related to organochlorines level in plasma (126-
129)

Another limitation includes the inconsistent use of methods for measuring
exposure (Table 2). Of the 23 studies, 7 report both questionnaire and biosample-based
measures of exposure, 8 report biosample based measures of exposure only, and the
remaining 8 report questionnaire-based measures of exposure only. Of the studies
reporting questionnaire-based measure of exposure, no consistent measure were used.
These inconsistencies in methods of exposure assessment suggest that suboptimal
characterizations of exposure may have been used in at some of the studies. Studies

using suboptimal measure of exposure may be vulnerable to substantial exposure

23

misclassiﬁcation that dilutes the true strength of association between an ideal, or true,
characterization of exposure and the response (80).

Methodological inconsistencies for biosamples include: questions of
comparability of biosamples between laboratories (including PCB quantiﬁcation and
what set of congeners are reported), different protocols for imputation of values below
the levels of detection, and methods by which total PCBs are calculated (130). Together
these methodological limitations almost certainly impact observed exposure-response
relationships and hinder between-study comparability and metaanalysis for this body of
research.

Assessing the impact, and properly addressing, this set of limitations poses a
daunting challenge. Some of the limitations are inherent to this area of research and are
by in large intractable. But others are not. For example, it may be possible to decrease
measurement error by developing improved measures of exposure and response, the
effect of bias can be investigated, and comparability between studies can be improved
with standard measures (81). Presently, the extent to which the weight of evidence in
this body of research is affected by these limitations is unknown. However, it would be a

mistake to ignore these concerns.

Rationale and Aims

Investigation of PCB exposure-response relationships faces serious
methodological limitations that introduce exposure misclassiﬁcation and reduce statistical
power to discern exposure-response relationships. Evaluation of measures of exposure
may improve screening surveys, reduce measurement error and promote standardization

of measures of exposure.

24

Improve screening surveys. The effectiveness of sample selection methods that
utilize a screening survey to over-sample for subjects with increased serum PCB levels is
directly proportional to predictor criterion-related validity. Criterion-related validity, also
referred to as instrumental validity, is used to demonstrate the accuracy of a measure by
comparing it with another measure which has been demonstrated to be valid.
Identiﬁcation of measures with improved criterion-related validity can increase the
predictive power and effectiveness of screening surveys.

Reduce measurement error. Poor criterion-related validity reduces precision,
increases exposure rrrisclassiﬁcation, and dilutes the observed strength of association
between exposure and response. Exposure measurement error decreases the power of a
study or increases the sample size required to detect a signiﬁcant health risk. If error is
random and does not vary by response status (i.e., there is no recall bias), the sample size
needed to detect a statistically signiﬁcant effect is inversely related to the squared
correlation (R-Square; i.e., a criterion for criterion-related validity) between the imperfect
exposure measure and the perfect exposure measure (new; nme/R-Square) (80).
Therefore, even small improvements in criterion-related validity, if cost effective, are
meaningful. For example, if the study exposure measure has a criterion-related validity
of an R-square of 0.1, then the sample size (nwm) necessary to detect a signiﬁcant health
effect would be approximately 10 times higher than the sample size (um...) necessary if
the perfectly measured dose was used (new; ﬁrm/0.1 = 10*n,me).

Promote standard measures. Inconsistent measures of exposure limit

comparability and metaanalysis. Studies that adopt a minimum standard for measuring

25

and reporting exposure promote comparability and thereby increase the likelihood of
discerning real exposure-response relationships in metaanalysis.

In conclusion, existing epidemiologic investigations of PCB exposure—response
relationships utilize suboptimal methods for sample selection and inconsistent measures
of exposure. Evaluation of self-report based measures of sport ﬁsh consumption for their
criterion-related validity to serum PCB level may distinguish strengths and limitations of
individual predictors, promote adoption of a minimum set of standards for future research
in this area, and contribute toward less controversial research permitting strong causal

inference.

Objectives

This thesis investigates the predictive power of questionnaire-based measures of
sport ﬁsh consumption as predictors of serum total PCB level. The thesis has three main
objectives: 1) Evaluate individual questionnaire-based measures of sport ﬁsh
consumption as predictors of human serum PCB levels (Ho: 51: ﬁg: 33: 34:0), 2)
Compare individual measures for statistically signiﬁcant improvement in the predictive
power (Ho: MS1= MS; versus H]: MS1 #MS;), and 3) Predict the performance of

individual predictors to determine if differences are meaningful.

26

CHAPTER III
MATERIALS & METHODS

Study Population

This thesis uses data from the Fisheaters Family Health Project (FFHP) Cohort-A.
This cohort was created to investigate the relationship between exposure to
polychlorinated biphenyls (PCBs) through consumption of Great Lakes sport ﬁsh and the
reproductive health of males and females. This cohort is a sample of couples, one or both
of whom is a licensed Michigan angler, and resident in one often Michigan counties
abutting Lake Michigan, Lake Huron, or Lake Erie and within International Joint
Commission designated Areas Of Concern (AOC), because of their proximity to water
bodies contaminated with high levels of PCBs and other persistent and toxic critical
pollutants. The full cohort enrolled 305 subjects between July, 1997, to November, 2000.
However, the thesis data is limited to subjects with data available for serum PCB level
(n=141) and complete data for important demographic and exposure related information

(n=129).

Questionnaire-based Measures of Sport Fish Consumption

Cohort members performed a telephone questionnaire with separate male and
female sections for reproductive history, but shared common sections for exposure
assessment including the base questions of the predictors (APPENDICES A & B). The
four measures of sport ﬁsh consumption evaluated in this thesis include: 1) lifetime years
eating sport ﬁsh (YEARS); 2) sport ﬁsh meals in the previous 12 months (MEALS); 3)

lifetime sport ﬁsh meals (MXYl); and 4) modiﬁed lifetime sport ﬁsh meals (MXY2).

27

Lifetime Years Eating Sport Fish (YEARS). Lifetime years eating sport ﬁsh
(YEARS) characterizes the duration of exposure. This predictor was assessed as a single
question with the response broken into ﬁve year age groups (APPENDIX A).

Sport Fish Meals in Previous 12 Months (MEALS). Sport ﬁsh meals in
previous 12 months (MEALS) characterizes the rate of exposure. This predictor assessed
a single question with the response broken into four intervals, of three months each, that
distinguish between winter, spring, summer, and fall in order to distinguish seasonal
variation (APPENDIX B).

Lifetime Sport Fish Meals (MXYI). Lifetime sport ﬁsh meals (MXY 1) is an
index that measures cumulative exposure. MXYl combines rate and duration of
exposure to provide an index of the total number of lifetime sport ﬁsh meals. MXYl is
calculated as the product of YEARS times MEALS. This index assumes that ﬁsh
consumption patterns are relatively consistent across years that any ﬁsh is consumed. A
potential limitation of this index is that subjects who report having eaten sport ﬁsh during
their lifetime (YEARS > 0), but report eating zero sport ﬁsh meals during the previous 12
months (MEALS = 0), will be misclassiﬁed as never having eaten sport ﬁsh during their
lifetime, since the product of zero meals times any number of total years is always zero.
Consequently, lifetime sport ﬁsh consumers who did not eat sport ﬁsh in the previous 12
months, are misclassiﬁed as non-consumers, regardless of how long or what amount of
sport ﬁsh they may have eaten in previous years.

Modiﬁed Lifetime Sport Fish Meals (MXY2). Modiﬁed lifetime sport ﬁsh meals
(MXY 2) is also an index that measures cumulative exposure. This index improves on

MXYl by imputing rate of sport ﬁsh consumption when zero meals were consumed in

28

the previous 12 months. The modiﬁed index is calculated as the product of multiplying
non-zero responses for number of sport ﬁsh meals consumed in the previous 12 months
by the total number of lifetime years in which any sport ﬁsh was consumed. For subjects
reporting zero sport ﬁsh meals in the previous 12 months, zero is replace by an imputed
value calculated as the overall sample median value of sport ﬁsh meals in the previous 12
months (8 meals for thesis data) and multiplied by the total number of lifetime years in
which any sport ﬁsh was consumed. The status of true non-ﬁsh eaters is preserved since
the product of multiplying the median value for sport ﬁsh meals in the previous 12
months by zero total years still results in an index value of zero. This method treats a
response of zero to sport ﬁsh meals for the previous 12 months as missing information
since no information has been provided about a subject’s exposure frequency when he or
she may have eaten sport ﬁsh in prior years. This salvages exposure information
captured by lifetime years. MXY2 also assumes that ﬁsh consumption patterns are
relatively consistent across years that any ﬁsh is consumed. A theoretical limitation of

this measure of exposure is that it captures no information about the rate of exposure.

Biosample Collection and Analysis

Upon recruitment into the study, subjects were asked to report for a blood drawn
at one of several Quest Clinical Laboratories located in the study counties. Cohort
members who failed to have their blood drawn at a clinic were given the option of having
a ﬁeld phlebotomist visit their home to draw the blood sample. Subject inclusion in
either group occurred by self-selection with 103 subjects reporting to a clinic and the

remaining 26 requesting ﬁeld collection.

29

Protocols for both sample groups speciﬁed collection of blood samples into 10 ml
glass vacutainers without an anticoagulant, clotting at room temperature and decanting
into storage containers. Processing of samples between the two groups was identical.
However, sera processed at Quest Clinical Laboratories were decanted into polystyrene
storage containers, while those collected and processed by the ﬁeld phlebotomist were
decanted into glass storage containers. Serum-containing storage containers from both
groups were then stored at —20°C until analysis. Storage container type is the only known
difference between the two sample groups. The median storage time was 8 months with a
range of l to 32 months. All samples were analyzed within 3 months of each other.

Storage of serum in polystyrene containers was an unplanned violation of the
study protocol that raises at least two concerns. First, the polystyrene storage containers
used to store serum may contain PCBs, or other substances, that may contaminate the
serum samples and introduce experimental error. Second, container type is perfectly
confounded with clinic and ﬁeld sample groups. This limits the comparability of data
between the two groups. It also complicates evaluation of the potential effects of this
protocol violation, since the two groups are likely to vary on other factors besides type of
storage container. However, for the purposes of this thesis, the preceding concerns are
considered a limitation of the data. The risk of experimental error from this source is
plausible, but highly unlikely, and is not of sufﬁcient concern to prevent its use for this
thesis.

The analysis was performed, between October and November of 1999, at the
Michigan Department of Community Health (MDCH), Bureau of Laboratories Analytical

Chemistry Section. The MDCH laboratory adheres to strict internal and external quality

30

assurance and quality control practices and has over twenty years of experience in
developing and conducting chemical analyses on human specimens.

Sample preparation involves extraction of the serum sample using a liquid/liquid
method. The extract is then cleansed of interfering lipids using a micro ﬂorisil column,
and the PCBS are separated from the pesticides with a micro-silica gel 60 column
technique. The two PCB fractions (coplanar PCBs and other PCBs) are separated by an
automated carbon column/casium silicate-acid silica/alumina separation and cleanup
device. Capillary column gas chromatography (V arian model 3800) was used to separate
and identify polychlorinated biphenyl (PCB) congeners. Individual PCB congeners were
quantiﬁed using custom calibration mixtures. Separate calibration standard mixtures
were obtained commercially for the PCB congeners (composed of 83 individual PCB

congeners).

Derived measures for PCB exposure

The laboratory did not report values for individual congeners below the level of
detection (LOD) that varied by congener from 0.2 to 1.3 parts per billion (ppb). For
values below the LOD, the square root of the LOD was recorded when less than 40% of
the subjects sampled were reported below the level of detection on any individual
congener. When greater than 40% of the subjects sampled were reported below the level
of detection on an individual congener, the zero value was retained. No adjustment was
made for serum lipid levels. Total PCBs were determined by summing individual

congeners.

31

Descriptive Statistics

All statistical analyses were performed using SAS software versions 8.02.
Descriptive statistics are presented for selected demographic factors, questionnaire-based
measures of sport ﬁsh consumption, and serum PCB levels. Frequencies are presented
for categorical variables. Means, the minimum (Min), the 10th percentile (PlO), the
median (Med), the 90th percentile (P90), and the maximum (Max) are presented for

continuous variables.

Correlation Matrix

A Pearson nonparametric correlation matrix is presented to show unadjusted
associations between modeled variables. These associations aid in the interpretation of
relationship between variables in the multivariable regression models. For example, age
is likely to be associated with total years eating sport ﬁsh and act as an intervening
variable that limits the maximum number of total years that a subject may have eaten

ﬁsh.

Multiple Regression Modeling

SAS regression procedures were used to model serum PCB level (y) on control
variables (x,) and predictors ()9). Criteria for selection include partial sum of squares and
R-Square delta. Model diagnostics are assessed for violation of model assumptions.
Transformations of model variables are implemented where appropriate to remedy
violations of model assumptions. Standardization procedures are used to facilitate
comparisons between modeled predictors.

To interpret control and predictor variables, partial regression coefﬁcients are

provided. Partial regression coefﬁcients represent the expected increase in y per unit

32

increase in x,, with all other variables held constant and are estimated by the parameter
estimate bj (131). If there is potential confounding between predictor and control
variables, then the predictor partial regression coefﬁcient may differ considerably from
the simple linear-regression coefﬁcient obtained by modeling the predictor without
considering control variables (131).

The strategy for ﬁtting models is based upon forward selection. The procedure
uses four steps: 1) Specify the maximum model to be considered including candidate
control variables and the four predictors; 2) Model control variables; 3) Add predictor
variables to the control variables as separate models (Models A, B, C & D), and
individually examine predictor partial f-value with an a of .05 for signiﬁcance, 4) for
each model evaluate for goodness of ﬁt and transform variables as indicated.

Assumptions. Assumptions for multiple regression are: 1) Existence: For each
speciﬁc combination of values of the independent variables X1, X2,. . ., Xk , Y is a random
variable with a certain probability distribution having ﬁnite mean and variance. 2)
Independence: the Y observations are statistically independent of one another. 3)
Linearity: the mean value of Y for each speciﬁc combination of X 1, X2,. . ., Xk is a linear
function of X 1, X2,. . .., Xk. 4) Homoscedasticity: The variance of Y is the same for any
ﬁxed combination of X 1, X2,. . .., Xk. 5) Normality: for any ﬁxed combination of X 1,

X2,. . .., Xk, the variable Y is normally distributed (132).

Goodness of ﬁt. Residuals are important in model critique for detecting outliers,
evaluating goodness of ﬁt, and diagnosing departures from assumptions. A SAS macro
was used to produce an "inﬂuence plot" for regression model residuals, leverage and their

combined effect as inﬂuence. Residuals are calculated as the observed value minus the

33

predicted value for each observation (ea. = Y9. — Y, ). This plot shows Studentized

residuals versus hat values for leverage, with Cook’s Distance as the size of a bubble
symbol for inﬂuence. The value of Cook’s distance for each observation represents a
measure of the degree to which the predicted values change if the observation is left out
of the regression. Observations exhibiting unusually large values for the Cook’s distance
are suspect outliers and considered candidates for deletion. Horizontal reference lines
added to the plots delimit observations whose Studentized residuals are individually
signiﬁcant at p=0.05. A vertical reference line in the plot shows observations which are
of "high leverage" (133).

Transformations. Models with non-normally distributed residuals can usually be
corrected by transforming variables. Many of the full model variables are highly skewed.
Skewed data is transformed to approximate normality using rescaling and fractional
exponentiation (134). The exponent for transformation is selected based upon an
iterative procedure for convergence of the distribution mean and median (135).

Standardizations. Two standardizations procedures are used to ease comparisons
of predictors with different units. First, standardized regression coefﬁcients are used to
compare the predictive value of each measure of sport ﬁsh consumption ()9). This
method is useful for comparing estimators with different units of measurement (131). It
describes the predicted increase in standard deviation units of serum PCB level ()2) that
would be expected per standard deviation increase in each questionnaire-based measure
(xj). Second, for comparison of overlay plot and predicted means, exposure units are
standardized to a 0 to 1 scale by dividing each exposure measure value by its maximum

value.

34

Partial residual plots with ﬁtted splines. Partial residual plots are used to
visualize the relationship between serum PCB level (y) and the predictors (xj), while
adjusting for control variables. Partial-residual plots are constructed as follows: 1) A
multiple regression is performed for y on all control variables (i.e., gender, age, BMI) and
the residuals are saved. 2) A multiple regression is performed for x,- on all other control
variables. 3) The partial-residual plot is then constructed as a scatter plot of the residuals
in step 1 on the y axis versus the residuals in step 2 on the x,- axis (131).

To assess the plausibility of a linear relationship between y and x,-, smoothed
regression lines are ﬁtted to plotted data using spline routines with 90% bootstrap
conﬁdence intervals (136). This method allows great ﬂexibility in ﬁtted-line
curvelinearity. A linear model is considered plausible when an approximately straight

line is ﬁt, or a straight line can be ﬁt within the conﬁdence intervals.

Comparing Non-nested Multiple Regression Models

When comparing the performance of two non-nested regression ﬁrnctions for
predicting y the partial F-test does not apply. Graybill and Iyer (137), describe a
procedure for comparing predictors between non-nested models.

First the full model is described; superscripts A and B are used to represent the

predictor variables in the two comparison subsets. The full model includes all the
covariates and predictor variables that are in model A and model B, X ,‘ ,...., X f
and X,” ,....,X: , and is denoted by the collection Xl ,....,X,.

Then the criterion for selection is declared; this thesis uses the model partial
regression root mean square, RMS(X4 | X I , X 2 , X 3 ), of the predictor. The partial

regression root mean square is a measure of the magnitude of variance explained by a

35

predictor after adjusting for the control variables already in the model. Therefore, to
compare predictors, it is of interest to know how much bigger or smaller the partial MS;
is than the partial RMSA. The approach used in this thesis is to examine the
ratioRMS(Xf | X1”,Xf,Xf)/RMS(X;‘ | XIA,X2",X3‘), or for brevity RMSg/RMSA, and
make a practical decision based on these results.

The test for signiﬁcance computes Bonferroni two-sided conﬁdence intervals for
RMSB/RMSA with conﬁdence coefﬁcient greater than or equal to l-OL The procedure for
its construction is given below:

1) Let r and m denote the number of independent variables in models A and B,
respectively, and let n be the number of sample observations.

2) Regress y on the independent variables in model A and obtain the partial
regression sum of squares SSA for model A predictor Jr].

3) Regress y on the independent variables in model B and obtain the partial
regression sum of squares for SS3 for model B predictor x,-.

4) Compute a 1 — oz/2 two-sided conﬁdence interval for RMSA, which is given by
C[LA SMSA SUA] = 1 — a/2.

Where

 

 

SS(Xf|X1A,X2A,X§‘)

2
la / 4:n—-r—l

SSIX: |X,",X,",X;‘)

 

 

LA: UA:

2 and
ZI—a/4zn—r—I

5) Compute a 1 — 01/2 two-sided conﬁdence interval for RS513, which is given by
C[LB SAISB SUB] = I — (ll/2.

Where

36

 

 

SS(Xf|X,B,Xf,Xf) U _ SS(Xf|XIB.Xf.Xf)
L3 = 2 and B _ 2
Zl—a / 4:n—m—l Za/ 4:n-m—l

 

. 6) Compute the following conﬁdence statement for MSB/MSA:

C[LB/UA sRMSB/RMSA SUB/LA]=1— 01/2.

Predicted Mean Serum PCB Levels at Selected Exposure Levels

For each modeled predictor, the serum PCB level predicted means, with 95%
conﬁdence limits, are presented at three sample selection related exposure levels: 1) full
sample (midpoint=0.50); 2) oversampling top 50% (midpoint=0.75); and 3) oversampling

top 20% (midpoint=0.90).

37

CHAPTER IV

RESULTS

Characteristics of the Study Data

55.0% of study subjects were male and 45.0% were female (Table 1). The mean
age was 36.7 years, with a range of 22 to 53 years. The average body mass index (BMI)
was 26.9 kilograms per meters squared, with a range of 16 to 68. One subject (ID, 0009-
1) reported an extremely high BMI of 68. This subject was a 36 years old male with a
recorded weight of 462 pounds and height of 5 feet 9 inches. Therefore, while this is an
extremely high BMI it is nevertheless correctly calculated from plausible base
parameters. During modeling, this subject was deleted from models as a inﬂuential
outlier. For highest level of education attained, 3.1% received less than high school,
20.9% received GED/high school, 46.5% reported some college, 24.0% reported being

college graduates, and 5.4% reported Graduate school.

38

Table 1. Demographic and sport fish related characteristics of 129 subjects in the

Fisheaters Cohort-A, 1997-2000.

 

N % Mean Min P10 Med P90 Max

 

Gender (GENDER)
Female 58 45.0
Male 71 55.0

Ageinyears(AGE) 36.7 22 31 35 45 53
Body Mass Index (BMI) 26.9 16 21 26 34 68

Education (EDUCAT)
Less than high school 4 3.1
GED/high school 27 20.9
Some college 60 46.5
College graduate 31 24.0
Graduate school 7 5.4

Grams of sport ﬁsh eaten
per day in past 12 months 9.0 0 O 5 27 58

Sport ﬁsh meals in past 12
months (MEALS) 11.8 0 O 8 32 56

Lifetime years eating sport
ﬁsh (YEARS) 23.8 0 10 25 35 35

Lifetime years eating sport
ﬁsh among subjects
reporting zero meals in the

past 12 months (n=23) 16.8 0 5 15 25 35
Index of lifetime sport-

caught ﬁsh meals (MXYl) 306.7 0 O 160 910 1,960
Modiﬁed index of lifetime

sport-caught ﬁsh meals

(MXY2) 330.7 8 50 200 910 1,960

 

The average number of grams of sport ﬁsh eaten per day in the past 12 months
was 9.0, the median value was 5, and the range was 0 to 58 (Table 1). The average
number of meals eaten in the previous 12 months was 11.8, the median value was 8, and
the range was 0 to 56. The average number of lifetime years eating sport ﬁsh was 23.8
and the median value of 25, and the range 0 was 35. The average of the index for

lifetime meals was 306.7, the median value was 160, and the range was 0 to 1,960. The

39

average of the modiﬁed index for lifetime meals was 330.7, the median was 200, and the

range was 8 to 1,960.

Table 2. Serum PCB level (ppb) for the Fisheaters Cohort-A, 1997-2000.

 

 

N % Mean Min P10 Med P90 Max
Serum PCB level (ppb) 1.9 0.1 0.2 0.8 4.8 13.5
Serum PCB group (ppb)
<2 95 73.6
2 to <4 17 13.2
4 to <6 6 4.7
6 to <8 4 3.1
8 to <10 4 3.1
10 to <12 2 1.6
212 1 0.9

 

The mean serum PCB level was 1.9 ppb, with a median value of 0.8 ppb, and a
range of 0.1 to 13.5 (Table 2). Nearly three quarters of the sample (73.6%) exhibited
serum PCB levels in the lowest exposure level grouping, or less than 2 ppb. The
relationship between the mean, median and range relative to zero suggests a highly
positively skewed, or lognormal, distribution. To illustrate the distribution, a histogram
was constructed and ﬁtted with a lognorrnal curve (Figure 1). The resulting histogram
veriﬁed the highly positive skew of the distribution and indicated that a transformation

was required to meet the assumptions for multiple regression.

40

Histogram of Serum Total PCB Concentration in Fisheaters Cohort—A

wlth Fitted Lognormal Curve
(n-lZB)

 

 

 

 

 

 

 

 

 

 

 

40— l
| N 129
as— I Mean 1.9
l Median 0.8
|
30_ | Std Dev 2.6
l
25- |
P
. l
r
C 20‘
9
I1
I:
15‘ —
7
10“
5-
1 F—ﬁ r—4——F—n r—w r—1
0 l I l I I I I I I I I 17 I l I
0 I 2 3 4 S 8 7 8 5 10 ll 12 13 H 15

Serum_PCB_Concentratlon

Figure 1. Distribution of serum PCB level in the Fisheaters Cohort-A (n=129).

Correlation Matrix

Inspection of the Pearson correlation matrix reveals that AGE is strongly
correlated with SERUMPCB, and GENDER and BMI are less strongly associated with
SERUMPCB (Table 3). All of the measures of sport ﬁsh consumption, YEARS (F036),
MEALS (0.35), MXYl (0.37) and MXY2 (0.38) are all strongly correlated with
SERUMPCB.

MXYl and MXY2 are extremely highly correlated (0.99) suggesting that they
may be indistinguishable in the regression. MEALS, MXY 1 and MXY2 are all highly

co-linear (greater than 0.6). Therefore, these three variables should not be considered in

41

the same model, since they are likely to measure the same phenomenon; including them

may cause the model to appear to explain the observed data better than is really does.

Table 3. Pearson correlations (prob > |r| under Ho: r=0) between serum PCB level,

control variables, and measures of sport ﬁsh consumption in study data (n=129).

 

 

Serum
Total
PCB GENDERT AGE BMI YEARS MEALS MXY]
GENDER -.20* —
AGE .49*** -.27** —
BMI .19* -.22* .17 —
YEARS .36*** -.40*** .39*** .23** —
MEALS .35*** -.Ol .19* .01 .21* —
MXYl .37*** -.12 .27" .06 .42*** .93*** —
MXY2 .37*** -.13 .26" .05 .42*** .91*** .99***

 

‘I'Male, 0; Female, 1; *p<.05; **p<.01; ***p<.001

Multiple Regression Models

Candidate multiple regression models were developed to predict serum PCB
level, identify control variables, and compare predictors. The four models are presented
sequentially and each section is organized similarly and presents the model ANOVA and
a partial residual plot for serum PCB level versus each predictor.

The maximum model (X , ,...., X k ) to be considered was SERUMPCB on
candidate control variables GENDER, AGE, BMI, EDUCAT, and the four predictors
YEARS, MEALS, MXY], and MXY 2. Candidate control variables were identiﬁed by a
review of literature and theoretical considerations. All four models were modeled for
GENDER, AGE, BMI and EDUCAT as control variables. EDUCAT was included in the

initial full model but ﬁnally restricted on the grounds that it did not meaningfully impact

42

estimates for predictors, did not improve precision, and was not statistically signiﬁcant in
any of the models.

Gender and age are signiﬁcantly associated with SERUMPCB. BMI is not found
to be signiﬁcantly associated with SERUMPCB. Of the control variables age is the most

strongly associated with SERUMPCB (Table 4).

Table 4. AN OVA and parameter estimates for regression of loglO serum total PCB

level on control variables: gender, age, and BMI.

 

 

Sum of Mean
Source 0F Squares Square F Value Pr > F
Model 3 12.81618 4.27206 23.68 <.0001
Error 124 22.36991 0.18040
Corrected Total 127 35.18608
Root MSE 0.42474 R-Square 0.3642
Dependent Mean -0.05429 Adj R-Sq 0.3489

Coeff Var -782.39850

Parameter Estimates

Parameter Standard Standardized
Variable 0F Estimate Error t Value Pr > It] Type I SS Estimate
Intercept 1 -1.82193 0.32251 -5.65 <.0001 0.37722 0
gendert 1 -0.18904 0.07956 -2.38 0.0190 3.75943 -0.17948
age 1 0.04913 0.00708 6.94 <.0001 9.04616 0.52405
BMI 1 0.00195 0.00804 0.24 0.8090 0.01059 0.01805

 

Model A: Model PCBs on Lifetime Years Eating Sport Fish (YEARS)

A standard multiple regression was performed modeling SERUMPCB on YEARS
while controlling for GENDER, AGE and BMI. Results of evaluation of assumptions led
to transformation of the variables to reduce skewness, reduce the number of outliers, and
improve the normality, linearity, and homoscedasticity of the residuals. A logarithmic
transformation was used on SERUMPCB (t_SERUMPCB=Log10[SERUMPCB]). An
exponential transformation was used on YEARS (t_YEARS=[YEARS + 1]**2.8). With

the use of Cook’s D, visualized on the inﬂuence plot, one extreme outlier was identiﬁed.

43

 

The outlying observation was identiﬁed as the subject exhibiting the extreme BMI (ID,

0009-1). This subject was deleted from the analysis data as an extreme outlier (n=128).

The model was reﬁtted and the inﬂuence plot redrawn (APPENDIX C: MODEL A

GOODNESS OF FIT STATISTICS).

A partial residual plot was constructed relating t_SERUMPCB to t_YEARS

(Figure 2). A spline with 90% conﬁdence intervals was ﬁtted to this plot and is

consistent with a linear relationship between t_SERUMPCB and t_YEARS.

LogIO SERUMPCB Residuals

Parlid residual plot of Serum PCBs versus YEARS

Spline Fitted Curve with Bootstrap Confidence interval

 

  
 

 

 
 

(n-128)
A
A A A
A A
A A A A A
AA A A A ,,,,,, ’3
AA A A A A A #K’
A ..,..A-A 13$”
A A Elsi-Juli ————
A
AA
A A A [Sig ‘dltlA A. A
... A AA A A A
1 ”.., A A A A A
A A A 191A 5.2%15 A A
A A 8 A
A A
A

 

 

[TYIITIYIIIUIYYFTT—IV—VTIIITTTVI'Y‘IIFIUITI‘IIV'UTTTIIIIVIIIIITIIUVUII

'17000 '14000 '1 I 000 '0000 '5000 -2000 1000 4000 7000 10000 13000 15000
(YEABBOI)‘*2.8 Residuals

 

 

A A A r_y Predicted Value of r_y
——‘ Lover 90! Confidence Linit for I """ Upper 90! Confidence Lilli: for ll

 

 

 

Figure 2. Partial residual plot of SERUMPCB on YEARS, with spline regression

line, and 90% conﬁdence intervals, after correcting for GENDER, AGE and BMI.

44

 

The t test for t_YEARS added to the model last is statistically signiﬁcant (t-
value=3.01, Pr<0.0032) (Table 5). The R—Square delta indicates that t_YEARS explained
an additional 4% of the overall variation in t_SERUMPCB after controlling for gender,
age and BMI.

According to the parameter estimation column, the regression equation is given
by SERUMPCB = (-1.627 -O.105*GENDER +0.040"'AGE -0.001*BMI +0.0000171
*[YEARS+1]**2.8)**10.

The partial-regression coefﬁcient for t_YEARS=b4=O.0000171 *[YEARSH]
W2.8 represents the expected increase in log10 serum total PCB level per unit increase in
total years eating sport ﬁsh for subjects of the same gender, age and BMI (Table 5).

The standardized estimate for age decreases from 0.524 to 0.427 suggesting that
AGE and YEARS are confounded. The standardized estimate for GENDER decreases

from -O.179 to -0.100 suggesting that GENDER and YEARS are confounded.

Table 5. AN OVA and parameter estimates for regression of log10 serum total PCB

level on gender, age, BMI and YEARS.

 

 

Sum of Mean
Source DF Squares Square F Value Pr > F
Model 4 14.34842 3.58711 21.17 <.0001
Error 123 20.83766 0.16941
Corrected Total 127 35.18608
Root MSE 0.41160 R-Square 0.4078
Dependent Mean -0.05429 Adj R-Sq 0.3885

Coeff Var -758.19118

Parameter Estimates

Parameter Standard Standardized R-Square
Variable DF Estimate Error t Value Pr > |t| Type I SS Estimate Delta
Intercept 1 -1.62657 0.31921 -5.10 <.0001 0.37722 0
gender1 1 -0.10533 0.08197 -1.28 0.2012 3.75943 -0.10001
age 1 0.04006 0.00750 5.34 <.0001 9.04616 0.42723
BMI 1 -0.00110 0.00786 -0.14 0.8885 0.01059 -0.01023
t_YEARS 1 0.00001705 0.00000567 3.01 0.0032 1.53224 0.25776 0.0436

 

45

 

Model B: Model PCBs on Sport Fish Meals in Previous 12 Months (MEALS)

A standard multiple regression was performed modeling SERUMPCB on
MEALS while controlling for GENDER, AGE and BMI. A logarithmic transformation
was used on SERUMPCB (Log10[SERUMPCB]). An exponential transformation was
used on MEALS (t_MEALS=[MEALS + 1]**0.002). With the use of Cook’s D,
visualized on the inﬂuence plot, ID 0009-1 was again identiﬁed as an extremely
inﬂuential outlier and was deleted from the analysis data (n=128). The model was
reﬁtted and the inﬂuence plot redrawn (APPENDIX D: MODEL B GOODNESS OF FIT
STATISTICS).

A partial residual plot was constructed relating t_SERUMPCB to t_MEALS
(Figure 3). A spline with 90% conﬁdence intervals was ﬁtted to this plot and is

consistent with a linear relationship between t_SERUMPCB and t_MEALS.

46

Paﬁal residual plot of Serum PCBs versus MEALS

Spline Fitted Ctrve Iitl} Bootstrap Confidence Interval
n-l28

 

A I A A 4. . A A A A A

 

LoglO SERUMPCB Res i dual 3

 

 

 

 

jirrilﬁrrilIfTTIrVVlerTIITTVTIFYII[ITTYIV‘ITIITYWIU]

-0.005 -0.004 -0.003 -0.002 -0.00l 0.000 0.00l 0.002 0.003 0.004 0.005
(“801)"0J02 Reclduals

 

D D D r_y _ Predicted Value of r_y
_—" Lover 90! Confidence Limit for ll "“ Upper 90! Confidence Unit for l

 

 

 

Figure 3. Partial residual plot of SERUMPCB on MEALS, with spline regression

line, and 90% conﬁdence intervals, after correcting for GENDER, AGE and BMI.

The t test for t_MEALS added to the model last is statistically signiﬁcant (t-
value=3.27, Pr<0.0014) (Table 6). The R-Square delta indicates that t_MEALS
explained an additional 5% of the overall variation in t_SERUMPCB after controlling for
gender, age and BMI.

According to the parameter estimation column, the regression equation is given
by SERUMPCB = (~53 -0.18*GENDER +0.045*AGE +0.001*BMI+ 51.2*[MEALS+1]

**0.002)**10.

47

The partial-regression coefﬁcient for t_MEALS=b4=51.20703*[MEALS+1]

a""‘O.002 represents the expected increase in log10 serum total PCB level per unit increase

in MEALS for subjects of the same gender, age and BMI (Table 6).

The standardized estimate for age decreases from 0.524 to 0.475 suggesting that

AGE and MEALS are confounded. The standardized estimate for GENDER decreases

from -O.179 to -O.175 suggesting that GENDER and YEARS are not meaningfully

confounded.

Table 6. AN OVA and parameter estimates for regression of log10 serum total PCB

level on gender, age, BMI and MEALS.

 

 

Sum of Mean

Source 0F Squares Square
Model 4 14.60432 3.65108
Error 123 20.58176 0.16733
Corrected Total 127 35.18608

Root MSE 0.40906 R-Square 0.4151
Dependent Mean -0.05429 Adj R-Sq 0.3960
Coeff Var -753.52124

Parameter Estimates

Parameter Standard
Variable DF Estimate Error t Value Pr
Intercept 1 -53.04125 15.67136 -3.38
gender1 1 -0.18431 0.07664 -2.40
age 1 0.04451 0.00697 6.39
8M! 1 0.00115 0.00775 0.15
t_MEALS 1 51.20703 15.66452 3.27

F Value Pr > F

21.82

> ltl
0.0010
0.0177
<.0001
0.8820
0.0014

<.0001

Type I 88
0.37722
3.75943
9.04616
0.01059
1.78814

Standardized R-Square
Estimate Delta
0
-0.17500
0.47477
0.01067
0.23156 0.0509

 

48

 

Model C: Model PCBs on Lifetime Sport Fish Meals (MXY 1)

A standard multiple regression was performed modeling SERUMPCB on MXYl
while controlling for GENDER, AGE and BMI. A logarithmic transformation was used
on SERUMPCB (Log10[SERUMPCB]). An exponential transformation was used on
MXYl (t_MXY1=[MXY1 + 10]**0.05). With the use of Cook’s D, visualized on the
inﬂuence plot, one extreme outlier was identiﬁed (ID, 0009-1). This subject was deleted
from the sample on the basis of its extreme BMI reducing the modeled data to 128
subjects. The model was reﬁtted and the inﬂuence plot redrawn (APPENDIX E:
MODEL C GOODNESS OF FIT STATISTICS).

A partial residual plot was constructed relating t_SERUMPCB to t_MXYl
(Figure 4). A spline with 90% conﬁdence intervals was ﬁtted to this plot. The ﬁtted
spline was only approximately linear for the relationship between SERUMPCB and

MXY 1. However, a more effective transformation was not found.

49

Paﬁal residual plot of Serum PCBs versus WY1

Spline Fitted Curve with} Bootefrap Confidence Interval
n-120

 

LoglO SERUMPCB Residuals

 

 

 

 

(HXYIOI0)‘"0.05 Heeiduals

 

 

0 O 0 r_y Predicted Value of r_y
——" Lower 90! Confidence Limit for H ‘-"'" Upper 90! Confidence Unit for II

 

 

 

Figure 4. Partial residual plot of SERUMPCB on MXY 1, with spline regression line,

and 90% confidence intervals, after correcting for GENDER, AGE and BMI.

The t test for t_MXYl added to the model last is statistically signiﬁcant (t-
value=3.27, Pr<0.0014) (Table 7). The R-Square delta indicates that transformed MXY]
explained an additional 7% of the overall variation in t_SERUMPCB aﬁer controlling for
gender, age and BMI.

According to the parameter estimation column, the regression equation is given
by SERUMPCB = (-3.46 -0.161*GENDER +0.0425*AGE -0.000239*BMI + 1.50202

*[MXY1+10]**0.05)**10.

50

The partial-regression coefﬁcient for t_MXY1=b4=1.50202*[MXY1+10]**0.05
represents the expected increase in log10 serum total PCB level per unit increase in
MXY 1 for subjects of the same gender, age and BMI (Table 7).

The standardized estimate for age decreases from 0.524 to 0.454 suggesting that
AGE and MEALS are confounded. The standardized estimate for GENDER decreases

from -0.179 to -0.153 suggesting that GENDER and YEARS are confounded.

Table 7. AN OVA and parameter estimates for regression of log10 serum total PCB

level on gender, age, BMI and MXYl.

 

Sum of Mean
Source DF Squares Square F Value Pr > F
Model 4 15.24092 3.81023 23.50 <.0001
Error 123 19.94517 0.16216
Corrected Total 127 35.18608
Root MSE 0.40269 R-Square 0.4332
Dependent Mean -0.05429 Adj R-Sq 0.4147
Coeff Var -741.77651

Parameter Estimates

 

Parameter Standard Standardized R-Square
Variable DF Estimate Error t Value Pr > |t| Type I SS Estimate Delta
Intercept 1 -3.45682 0.52177 —6.63 <.0001 0.37722 0
gender1 1 -0.16140 0.07577 -2.13 0.0352 3.75943 -0.15324
age 1 0.04253 0.00693 6.14 <.0001 9.04616 0.45359
BMI 1 o0.00023899 0.00765 -0.03 0.9751 0.01059 -0.00221
t_MXY1 1 1.50202 0.38843 3.87 0.0002 2.42474 0.27722 0.0690

 

 

Model D: Model PCBs on Modified Lifetime Sport Fish Meals (MXY 2)

A standard multiple regression was performed modeling SERUMPCB on MXY2
while controlling for GENDER, AGE and BMI. A logarithmic transformation was used
on SERUMPCB (LoglO[SERUMPCB]). An exponential transformation was used on
MXY2 (t_MXY2=[MXY2+ 10]**0.125). With the use of Cook’s D, visualized on the

inﬂuence plot, ID 0009-1 was again identiﬁed as an extremely inﬂuential outlier and was

51

deleted from the analysis data (n=128). The model was reﬁtted and the inﬂuence plot
redrawn (APPENDIX F: GOODNESS OF FIT STATISTICS).

A partial residual plot was constructed relating transformed SERUMPCB to
_ MXY2 (Figure 5). A spline with 90% conﬁdence intervals was ﬁtted to this plot and is

consistent with a linear relationship between SERUMPCB and MXY2.

Partial residual plot of Serum PCBs versus MXY2

Spline Fitted Curve with Bootstrap Confidence Interval
(n-128)

 

L0910 SERUMPCB Residual s

 

 

-l“

 

 

-0.7 -0.6 -0.5 -0.4 -0.3 -0.2 -0.l 0.0 0.1 0.2 0.3 0.4 0.5 0.8 0.7

Residuals of (MXYZ+I0)”0.IZS Residuals

 

. . . r_y — Predicted Value of r__y
“" Lower 90! Confidence Limit for M ““ Upper 90% Confidence Limit for M

 

Figure 5. Partial residual plot of SERUMPCB on MXY2, with spline regression line
and 90% conﬁdence intervals, after correcting for GENDER, AGE and BMI.
The t test for t_MXY 2 added to the model last is statistically signiﬁcant (t-

value=4.64, Pr<.0001) (Table 8). The R-Square delta indicates that transformed MXY 2

52

explained an additional 9% of the overall variation in t_SERUMPCB after controlling for
gender, age and BMI.

According to the parameter estimation column, the regression equation is given
by SERUMPCB = (-2.90 -.140*GENDER +0.043*AGE -0.000726*BMI +0.684
*[MXY2+10]**0.125)**10.

The partial-regression coefﬁcient for t_MXY2=b4=1.50202*[MXY1+10]**0.05
represents the expected increase in log10 serum total PCB level per unit increase in
MXY 1 for subjects of the same gender, age and BMI (Table 8).

The standardized estimate for age decreases from 0.524 to 0.454 suggesting that
AGE and MEALS are confounded. The standardized estimate for GENDER decreases
from -O.179 to -O.133 suggesting that GENDER and YEARS are confounded (Tables 4 &

8).

53

Table 8. AN OVA and parameter estimates for regression of log10 serum total PCB

level on gender, age, BMI and MXY2.

 

 

Source
Model
Error

Corrected Total

Root MSE

Dependent Mean
Coeff Var

Variable
Intercept

gender1
age
BMI
t_MXY2

F
1
1
1
1
1

OF
4
123
127

0.39336
-0.05429
-724.59133

Parameter

Estimate
o2.89992
-0.13964

0.04323

-0.00072628

0.68379

Sum of
Squares

16.15437
19.03171
35.18608

R-Square
Adj R-Sq

Standard

Mean

Square
4.03859

0.15473

0.4591
0.4415

Parameter Estimates

Error t Value

0.37825
0.07445
0.00668
0.00747
0.14721

-7.67
-1088

6.47

-0.10

4.64

Pr > |t|
<.0001
0.0631
<.0001
0.9227
<.0001

F Value Pr > F
26.10 <.0001

Type I SS

0.37722
3.75943
9.04616
0.01059
3.33819

Standardized

Estimate
0
-0.13258
0.46104
-0.00672
0.32322

R-Square
Delta

0.0949

 

Comparing Predictors

An overlay plot of the four partial residual plots (Figures 2-5) was constructed

relating log10 serum PCB level residuals to standardized predictor residuals with

regression lines ﬁtted for all four predictors (Figure 6). Comparing standardized

regression coefﬁcients for t_SERUMPCB we observe the following estimates: t_YEARS,

0.25776; t_MEALS, 0.23156; t_MXYl, 0.27722; t_MXY2, 0.32322. These coefﬁcients

tell us that the predicted increase in standard deviation units of y that would be expected

per standard deviation increase in x, holding GENDER, AGE and BMI constant. Thus

the rank order for least to most increase in predictive power is: MEALS, YEARS,

MXY], MXY2.

54

 

Overlaid Partial Residual Plots of SERUMPCB versus: YEARS, MEALS, MXY1 and MXY2

with Regression Lines
(n-IZB)

 

 

LoglO SERUMPCB Residuals

 

 

 

 

Standardized Predictor Hes lduals

 

lPredictor: mnrsms Era-82111518 mum“ Hormel

 

Figure 6. Overlay of partial residual plots, and regression lines, of SERUMPCB on
standardized: YEARS, MEALS, MXY1 and MXY2 after controlling for GENDER,

AGE and BMI.

The portion of the variance in serum total PCB level explained by individually

modeled predictors was YEARS (R-Square delta=4%), MEALS (5%), MXY1 (7%), and

MXY2 (9%).

55

Table 9. Partial regression mean square ratios with 95% confidence intervals for

comparing non-nested models (n=128, alpha=0.05).

 

MSE/MS), YEARS MEALS MXY1
(CI 95%) (MS..) (MS...) (MSA)

 

MEALS (MSE) 1.08 (0.81, 1.43)
MXY1 (MSE) 1.26 (0.95, 1.67) 1.16 (0.88, 1.55)
MXY2 (MSE) 1.48 (1.11, 1.96) 1.37(1.03, 1.81) 1.17(0.88, 1.56)

 

Only two pairs of predictors, MXY2 versus YEARS and MXY2 versus MEALS,
have root mean square ratios signiﬁcantly greater than 1 (Table 9). MXY2 explains 1.48
times more variance than YEARS, and 1.37 times more variance than MEALS. MXY 2
explained 1.17 time more variance than MXY1 but is not signiﬁcantly better (Table 9).
A general trend was observed for increasing overall explanation of variance from

MEALS, YEARS, MXY1 to MXY2.

Predicted Mean Serum PCB Levels with Oversampling

Predicted mean SERUMPCB levels are presented at standardized exposure levels
(0.50, 0.75, 0.90) for the four predictors (Table 10). At 0.50, all four predictors share the
overall same predicted mean and supports the appropriateness of selected transformations
used in the models. At 0.75, average levels are: 1.18 ppb for YEARS, 1.16 ppb for
MEALS, 1.23 ppb for MXY1 and 1.37 ppb for MXY2. At 0.90, average levels are: 1.41
ppb for YEARS, 1.37 ppb for MEALS, 1.52 ppb for MXY1 and 1.80 ppb for MXY2.

The ratio of the 0.50 standardized exposure level mean to the higher exposure
level means gives the predicted percent increase. At 0.75, the percent increases are: 32%

for YEARS, 30% for MEALS, 38% for MXY1 and 53% for MXY2. At 0.90, the percent

56

increases are: 58% for YEARS, 53% for MEALS, 70% for MXY1 and 102% for MXY2

(Table 10).

Table 10. Predicted adjusted mean serum PCB levels (ppb) with 95% confidence
intervals and percent above full sample mean with no oversampling, oversampling
top 50%, and oversampling top 20% for four questionnaire-based measures of sport

fish consumption (n=128).

 

Predicted adjusted mean serum PCB level (ppb) with 95% CI. and
percent increase over full sample

 

Full sample Oversampling top 50% Oversampling top 20%
Predictor (midpoinF.50) (midpoint=.75) (midpoint-=90)

 

YEARS 0.89 (0.72, 1.10), 0% 1.18 (0.92, 1.51), 32% 1.41 (0.99, 2.00), 58%
MEALS 0.89 (0.72, 1.10), 0% 1.16 (0.92, 1.46), 30% 1.37 (1.00, 1.87), 53%
MXY1 0.89 (0.72, 1.10), 0% 1.23 (0.97, 1.56), 38% 1.52 (1.09, 2.10), 70%
MXY2 0.89 (0.72, 1.10), 0% 1.37 (1.06, 1.77), 53% 1.80 (1.26, 2.58), 102%

 

57

CHAPTER V

DISCUSSION

Applied Epidemiology

Paneth (138) points out the importance of relating measures of response to
meaningful public health outcomes. Of similar value is the ability to link characterization
of exposure to investigations of dose-response relationships. This is not simply a matter
of converting units, or even necessarily quantifying available or ingested dose, since
many factors mediate the relationship between a meal of PCB-contaminated sport ﬁsh
and a hypothetical biologically effective dose. Study designs able to simultaneously
relate sport ﬁsh consumption to laboratory based measures of the dose and the response
are likely to contribute most meaningfully to risk assessment and hazard control. Of the
23 studies reviewed in this thesis (Table 2), only six report both measures of sport ﬁsh
consumption and biosample based measures. This limits the comparability of existing

studies and may partially explain inconsistent ﬁndings.

Findings
Discussion of this study’s ﬁndings will include consideration of theoretical validity,
criterion-related validity, sample selection, exposure measurement, questionable
criterion, theoretical model, strengths, and limitations.

Theoretical validity. The four measures of sport ﬁsh consumption compared in
this thesis provide a logical basis for relating reported exposure to actual absorbed dose
and a hypothetical biologically effective dose. The thesis ﬁndings show that the relative

increases in criterion-related validity, with R-Square delta as the criterion, correspond to

58

theory and suggest that this approach is theoretically valid. Overall, the ﬁndings support
the general assertion that indexes of cumulative sport ﬁsh consumption that account for
both rate and duration of exposure are more strongly associated with observed serum
PCB level than are measures that account for either rate or duration, but not both.

Thesis ﬁndings might be misinterpreted so as to suggest that increasing the
criterion-related validity of a measure is simply a matter of constructing more and more
complex measures of sport ﬁsh consumption. This is a false interpretation, since there is
certain to be a threshold at which the theoretical advantage of a more complex measure is
overcome by disadvantages of costs, lower participation rates, and the compounding of
measurement errors associated with the individual components of more complex
measures of exposure.

Criterion-related validity. Comparisons between predictors indicated statistically
signiﬁcant variation in criterion-related validity. Predicted mean serum PCB levels for
oversampling (Table 10) would increase the proportion of more highly exposed subjects
and result in a 30 to 100% increase in predicted mean serum PCB levels, and provide a
meaningful increase in statistical power for detecting exposure-response relationships.

The study ﬁndings indicate that the modiﬁed index for lifetime meals (MXY2)
offers a meaningful improvement over other predictors tested, based upon the proportion
of variance explained and predicted values. MXY2 explained a similar amount of
variance (R-Square, 9%) as the measure reported by Jacobson et al. (1983) (R-Square,
8%). There are only negligible differences between the relative costs of collecting and

calculating each of the four predictors.

59

Sample selection. The ﬁndings support use of questionnaire-based measures of
ﬁsh consumption as instruments in sampling strategies that screen target populations for
more highly exposed individuals. Sampling strategies that target special populations such
as charter boat captains, or minority groups, appear to result in more highly exposed
study samples. However, such sample strategies may not be appropriate for all study
questions. The high positive skew of the distribution of serum PCB level data in the
study sample mirrors the general population and illustrates the inefﬁciency of existing
sampling strategies. These predictors may be of use to screen target populations and over
sample for individuals more likely to exhibit elevated serum PCB body burdens, or may
be used when a researcher lacks sufﬁcient resources, or cannot justify analyzing a full set
of biosamples.

Exposure measurement. The relatively small proportion of variation explained
(R—Square, 4 to 9%) by the four questionnaire-based predictors of serum PCB level can
be attributed to at least four sources, including: 1) other unmeasured sources of PCB
exposure; 2) measurement error for serum PCB level; 3) important but unmeasured
pharmacodynamic factors; and 3) measurement error for the questionnaire-based
predictors.

In data from this thesis, twenty three (18%) subjects reported zero sport ﬁsh meals
in the previous 12 months. Of those, the average number of years eating sport ﬁsh was
16.8, and only one reported never eating sport ﬁsh (Table 1). The mean age of these
subjects was 34.2 years, so they were actually younger than the mean age for the
remainder of the cohort. This highlights a vulnerability of self-reported measures that do

not account for lifetime duration of ﬁsh cOnsumption. Questionnaire-based measures of

60

sport ﬁsh consumption limited to consumption during the previous 12 months misclassify
sport ﬁsh consumers as unexposed no matter how much sport ﬁsh they consmned in
previous years. This is a particular problem considering the participation bias for studies
that collect biosamples. Therefore evaluation of criterion-related validity in such samples
is questionable.

The ﬁndings of this study suggest that strength of association may be subject to
excessive exposure misclassiﬁcation. For example, Courval et a1. (93) report ﬁndings of
an association between male sport ﬁsh consumption and reduced time to pregnancy.
They based exposure upon the index of lifetime sport ﬁsh meals (MXY1). However, my
thesis shows this measure misclassiﬁes sport ﬁsh consumers who have not eaten sport
ﬁsh in the previous 12 months.

The ﬁndings also suggest that, questionnaire-based measures of ﬁsh consumption
are, by themselves, poor proxies measures of PCB body burden. With the R-Square
deltas explaining no better than 9% of the sample variation in serum total PCB level, this
suggests that excessively high measurement error that probably overrides any advantage
gained by case of acquisition.

Questionable criterion. It can be argued that poor criterion-related validity for
questionnaire-based measures of sport ﬁsh consumption in predicting serum PCB levels
is insufﬁcient evidence against their use as exposure measures. A meaningful proportion
of variance in serum PCB level is attributable to pharmacodynamic factors that mediate
the true cumulative dose. Even though regression models include control variables that
approximate pharmacodynamic factors (e. g., Gender and Body Mass Index), it is unlikely

that they are entirely effective and a meaningful portion of model error is almost certainly

61

attributable to this limitation. Given mediation by these factors, it is logical to infer that
no measure of cumulative exposure would be highly correlated with an unadjusted
endpoint such as point-in-time body burden, and other methods for evaluating the value
of questionnaire-based measure of sport ﬁsh consumption may be more appropriate.
Theoretical model. Figure 7 illustrates a theoretically plausible scenario
consistent with existing research in which neither self-report (E 1 ) nor biosample-based
(E2) measures of exposure are strongly associated with a hypothetical biologically
effective dose (D). In this scenario, the true strength of association (DR) is strongly
diluted by measurement error in both measures of exposure. Plausible examples of this
scenario include: 1) both E1 and E2 are confounded with an unidentiﬁed toxic agent, D,
that contaminates sport ﬁsh; 2) the true biologically effective dose is sensitive to timing,
resulting in serious measurement error in E1 and E2; or 3) one or both E1 and E2 are

spuriously associated with the response due to bias.

ElR

 

 

Figure 7. Diagram of a theoretically plausible relationship between self-reported

sport ﬁsh consumption (E1), biosample based measures (E2), relevant dose (D), and

response (R).

62

Strengths

This study has a number of strengths. This study contributes to the literature on
exposure assessment methods for the investigation of sport ﬁsh consumption. Four
representative questionnaire-based measures of sport caught ﬁsh consumption were
modeled providing: 1) evaluations of predictive power; 2) between-predictor
comparisons; and 3) predictions of performance. In the statistical analysis, special
consideration was given to compliance with model assumptions and where appropriate
utilized transformations; 4) standardizations to facilitate comparisons between predictors
with different units, and 5) an appropriate statistical test for comparing non-nested
models; 6) results show cohesive statistical analyses between both nonparametric
(Pearson correlation matrix) and parametric (multivariable regression models)
procedures; 7) ﬁndings are also consistent with theory, in that criterion-related validity
was positively associated with improved characterization of rate, duration, and

cumulative exposure.

Limitations

There are several limitations with this study: 1) we were unable to test predictive
values of indexes of PCB intake that adjust for variation in PCB exposure by species
consumed; 2) the use of serum PCB levels unadjusted for lipids probably contributes
meaningful measurement error; 3) among the available modeled predictors, no test for
model reliability was performed due to insufﬁcient sample size; 4) it is difﬁcult to
interpret how confounding of predictors with control variables may affect the validity of

the model. However, it is important to remember that the objectives of this thesis relate

63

to predictive models, and therefore they are not necessarily subject to the same strict
requirements as are models aimed at performing causal inference.

The application of these ﬁndings to other populations warrants special
consideration, since the predictive power of these measures very likely varies with the
characteristics of target populations. For example, among the very young, total years
probably carries less weight than meals consumed in the previous 12 months, and vice
versa, in an older population, total years consuming sport ﬁsh probably carries greater
weight than meals consumed in the previous 12 months. However, an interaction term
for age and meals in the previous 12 months was not statistically signiﬁcant (Model B
adding AGE*t_MEALS, p=0.78 1).

The study sample does not represent a random sample of the general population.
Study sample subjects are self selected from a deﬁned target population. They are know
to differ from the general population on demographic and exposure related characteristics
and may respond differently to self-reported information. Considering these limitations,
ﬁndings from this study can not be directly extrapolated to other populations but may be

considered suggestive.

Implications for Future Research

This study identiﬁes sources of substantial measurement error in many existing
studies that could account for the observed weak associations between exposure and
response. Similarly, if observed exposure-response relationships are indeed causal, the
true biologically effective dose-response relationship could be much stronger than

observed due to measurement error related dilution of the strength of association. If

64

future research is able to more effectively control these sources of error then it is
plausible that causal inferences may be made with much more conﬁdence.

Comparability. Between study differences in methodologies limit comparability
and interpretation of the consistency of ﬁndings. It would be advantageous if future
research reported comparable measures of exposure. Given the limitations and
misinterpretation of existing literature, it would be of beneﬁt for studies to adopt a
minimum standard for exposure measures based upon self-reported sport ﬁsh
consumption. Future research should consider adopting an approach that allows
evaluation of the predictive power and reliability of their sampling strategies.

Large sample size random digit dial telephone surveys are often used at the state
level to assess the exposure level of the general population. These surveys are also
performed to investigate the effectiveness of ﬁsh consumption advisories and to assess
the economic impact of sports ﬁshing.

Nested study design. Given all the special considerations (e. g., oversampling,
assessing differential participation bias, and desire for external validity), a possible study
design that takes best advantage of this approach is a design that nests screening of the
general population within a large scale ﬁsh consumption survey. The Charter Boat
Captains Cohort is the only existing cohort to utilize a nested study design (109). In a
nested design, interviewees who score highest on the measure of sport ﬁsh consumption
could then be recruited into the exposure-response study. This study design may act to
limit recall bias since subjects would response to the ﬁsh consumption survey would be
unaware that they may be contacted later for information about outcomes. Such a study

design offers several potential strengths: 1) Nesting sample selection with a survey

65

contacts a large population for oversampling; 2) allows assessment of non-participants
that permit evaluation of non-differential participation bias; 3) combining resources is

more cost effective.

Conclusions

The conﬁdence with which causal inferences may be made is limited in this body
of research. Limitations include problems of weak associations, vulnerability to bias,
loss of sensitivity, measurement error resulting in excessive exposure misclassiﬁcation,
and dilution of strength of association. One strategy to address these limitations, in future
research, is to identify measures of exposure that reduce measurement error. This study
investigated four common predictors. In a population of Great Lakes Fisheaters, the
index of modiﬁed lifetime sport ﬁsh meals (MXY2) is a simple and effective predictor of
human serum total PCB levels that offers a modest but meaningful improvement over
other predictors considered including lifetime years eating sport ﬁsh (YEARS), sport ﬁsh
meals in the prior 12 months (MEALS), or the index of lifetime sport ﬁsh meals
(MXY1). These ﬁndings are consistent with theory, but no test for reliability was
performed due to insufﬁcient data, and no validation has yet been performed on other

sources of data.

66

CHAPTER VI

APPENDICES

67

APPENDIX A: SOURCE QUESTION FOR YEARS

L.3. I will read a list of age groups. Please tell me at which ages you ate sport-caught
ﬁsh from any of Michigan’s Great Lakes and tributaries. Answer yes only if you ate at
least one Great Lakes sport-caught ﬁsh meal each year while you were in that age
group. Did you eat any sport-caught ﬁsh when you were...

(CIRCLE ANY THA T APPLY. DO NOT ASK ABOUT PERIODS BEYOND

RESPONDENT ’S AGE)

No Yes
a. 5 — 9 yrs old 0 1
b. 10- 14 yrs old 0 1
c. 15 - 19 yrs old 0 1
d. 20 - 24 yrs old 0 1
e. 25 - 29 yrs old 0 1
f 30 - 34 yrs old 0 1
g. 35 + yrs old 0 1

68

APPENDIX B: SOURCE QUESTION FOR MEALS

L.4. Please describe your usual sport-caught ﬁsh consumption during each of the
past 12 months. Considering only ﬁsh that you, a family member, or an

acquaintance caught in Michigan waters; how many meals did you eat in the three

 

 

months of...
None 1to3 ' 4to6 7to9 Mia 12 13 or
meals meals meals meals more
April, May, and
June

 

July, August, &

September

 

October, November

& December

 

January, February.
& March

 

 

 

 

 

 

 

 

 

69

APPENDIX C: INFLUENCE PLOT FOR MODEL A

Inﬂuence plot of SERUMPCB modeled on GENDER, AGE, BMI and t_YEARS

(n-IZBJ

 

 

 

 

 

 

 

 

..
a:
.f,
E O O 0 °.'
7: o
.2 .- 00 6030 0 000
1. o 3,
S f 3" £950 09 0 (60128-2
LE 0.. -. . _
"E o 5‘ o o O O (0) O
.2 00% 0m 0 0(2) 00 OO O o (34 o1®2
-l O
@ 00
-2 O ....... 6GP" ._ ....................
-3-1 ' I ' I ' I ' I ' I ' I ' l ' I ' l ' I ' I ' I ' 1
.01 .oz .03 .04 .05 .os .07 .011 .09 .1o .11 .12 .13 .14

Leverage (Hat Value)

70

APPENDIX D: INFLUENCE PLOT FOR MODEL B

lnfluence plot of SERUMPCB modded on GENDER, AGE, BMI and t_MEALS

(n-l20)

 

as e -
-l @“méD. ............ .m’ ..... ......... .......... ............ , ...........

E 0692
E 680 o 00 % O o O
a: . 0 0 O
'13 0 .: ......... '. .......... .. ......................................... 0145.; .............................................................
.2 to” o o o. 0 0 3-2 “30129-2
: o 0 ° 0 o Cb O 01
.. -, (55’ O O
0 C8)

 

 

 

Leverloo ("It Value)

71

APPENDIX E: INFLUENCE PLOT FOR MODEL C

Inﬂuence plot of SERUMPCB modeled on GENDER, AGE, BMI and t_MXYl

(n-IZO)

 

a -1
105-

«02 «e
2‘ ......... @..0,Qg§z..i”®...o ..... ............. .................................................................
i g .
— o 0 0 O ;
g 11 Q g 5
'2 1 if 00 oo 0000 O O o (g) f 0.2
E 0.. ....... ..... if: ..... ..... ..................... 7.....;._.§..,,,e_, ......................................... .............
E: 0 2° 0:100; g 0 ° 0 OOQO 0 Q 0 0 019-2 0'8-3'8-2 \
g -.« (it: 00 8 9 0 <9

00000 Ce 00 O O
............................. ...................................................... ........................ q ........................... .................

 

 

 

T I
.01 .02 .03 .04 .05 .06 .0? .08 .09 .10 .11 .I2 .13

Leverage (Net Value)

72

APPENDIX F: INFLUENCE PLOT FOR MODEL D

Inﬂuence plot of SERUMPCB modded on GENDER, AGE, BMI and t_MXY2

(n-120)

 

 

 

 

a-
2. ..... 2 ........................ 1' u.uu.uu.n.un. ”.Hé.nu.uu.uu.uu.h..HH.HH.H.HH.WHE .......................
e O
_ C9 <99 0 O
g '1 O 0 «£04
:3 O o O E i 63%
3 °°°oo 0 o o o O 0 g s 0' B
«x ..‘ . q o a. o . ' i
:3 o .............. .....o ..... .. ...o .......... 0° .................... m ..... o ....... 01311412. ............... o'a'waz ............................
Z 1 °8 °o 0° 0° 0113-2 ' 3
..., 8 Q. 8 o g 8
5'“ 00 on) 00 (R °®2 ’
Cb O O i i
09 Q? ..
_2‘ ....................... .0" ........................ . ............. . ...... .............................. ........... , ...........
'3' I I I I I I I I I I I I I
.01 .oz .03 .04 .05 .06 .07 .00 .09 .10 .11 .12 .13 .14

Leveraoo (Hat Value)

73

10.

ll.

12.

REFERENCES

Lione A. Polychlorinated biphenyls and reproduction. Reprod Toxicol
l988;2(2):83-9.

ATSDR. Toxicological proﬁle for polychlorinated biphenyls (PCBs). Atlanta,
Georgia: Agency for Toxic Substances and Disease Registry, Public Health
Service, US. Department of Health and Human Services; 2000.

Erickson MD. Analytical chemistry of PCBs. Boca Raton: Lewis Publishers;
1 997.

EPA. Lake Michigan Mass Balance. Chicago IL: USEPA, Great Lakes National
Program Ofﬁce; 1997.

Eisenreich SJ, Looney BB, Thmton JD. Airborne organic contaminants in the
Great Lakes ecosystem. Environ Sci Technol 1981;15:30—38.

Offenberg JH, Baker J E. Aerosol size distributions of polycyclic aromatic
hydrocarbons in urban and over-water atmospheres. Environmental Science and
Technology 1999;33:3324-3331.

Zhang H, Eisenreich SJ, Franz TR, Baker J E, Offenberg JH. Evidence for
increased gaseous PCB ﬂuxes to Lake Michigan from Chicago. Environmental
Science and Technology 1999;33:2129-2137.

Norstrom RJ. Understanding bioaccumulation of POPs in food webs. Chemical,
biological, ecological and environmental considerations. Environ Sci Pollut Res
Int 2002;9(5):300-3.

Newman MC, Unger MA. Fundamentals of ecotoxicology. 2 ed. Boca Raton, FL:
Lewis Publishers; 2003.

Jensen S. Report of a New Chemical Hazard. New Scientist 1966;32:612.

Larsson P, Collvian, L., Okla, L., Meyer, G. Lake productivity and water
chemistry as govemers of the uptake of persistent pollutants in ﬁsh. Envionmental
Science and Technology 1992;26:346-352.

Berglund O, Larsson P, Ewald G, Okla L. Inﬂuence of trophic status on PCB
distribution in lake sediments and biota. Environ Pollut 2001;113(2):l99-210.

74

 

 

13.

14.

15.

16.

17.

18.

19.

20.

21.

22.

23.

Johnson BL, Hicks HE, De Rosa CT. Key environmental human health issues in
the Great Lakes and St. Lawrence River basins. Environ Res 1999;80(2 Pt 2):SZ-
S12.

Taylor WW. Great Lakes Fisheries Policy and Management. Lansing, Michigan:
Michigan State Univerisity Press; 1999.

Hoekstra PF, Letcher RJ, O'Hara TM, Backus SM, Solomon KR, Muir DC.
Hydroxylated and methylsulfone-containing metabolites of polychlorinated
biphenyls in the plasma and blubber of bowhead whales (Balaena mysticetus).

Environ Toxicol Chem 2003;22(11):2650-8.

Humphrey HE. Evaluation of changes of the level of polychlorinated biphenyls
(PCBs) in human tissue. Lansing, MI: Michigan Department of Public Health;
1976 June.

Fujiwara K. Environmental and food contamination with PCB's in Japan. Sci
Total Environ 1975;4(3):219-47.

Newland M, Paletz, EM. Animal Studies of Methymercury and PCBs: What Do
They Tell Us About Expected Effects in Humans? Neurotoxicology
2000;21(6):1003-1028.

EPA. Health effects of PCBs. In: United States: Environmental Protection
Agency: http://www.epggov/opptintr/ncb/effects.html; 2002.

Swain WR. Effects of organochlorine chemicals on the reproductive outcome of
humans who consumed contaminated Great Lakes ﬁsh: an epidemiologic
consideration. J Toxicol Environ Health 1991;33(4):587-639.

Guo YL, Lambert GH, Hsu CC, Hsu MM. Yucheng: health effects of prenatal
exposure to polychlorinated biphenyls and dibenzofurans. Int Arch Occup
Environ Health 2004;77(3):153-8.

EPA. Federal Registry: Polychlorinated Biphenyls (PCB's) In: United States:
Environmental Protection Agency; 1979. Report No.: 44:3 1 5 1 5-3 15 19.

Connelly NA, Knuth BA. Great Lakes Fish Consumption Health Advisories:
Angler Response to Advisories and Evaluation of Communication Techniques.
Ithaca, NY: Human Dimensions Research Unit, Department of Natural
Resources, New York State College of Agriculture and Life Sciences; 1993
February 1993. Report No.: HDRU Series No. 93-3.

75

24.

25.

26.

27.

28.

29.

30.

31.

32.

33.

34.

MESB. Critical Review of a Proposed Uniform Great Lakes Fish Advisory
Protocol. Lansing, Michigan: Michigan Environmental Science Board (MESB);
1995 September.

Kamrin MA, Fischer LJ. Current status of sport ﬁsh consumption advisories for
PCBs in the Great Lakes. Regul Toxicol Pharmacol 1999;29(2 Pt l):175-81.

EPA. Fish Advisories. In: United States: Environmental Protection Agency:
http://wwwepagov/ost/ﬁshh 2004.

EPA. Update: National Listing of Fish and Wildlife Advisories. Washington, DC:
US. Environmental Protection Agency (EPA); 2003 May. Report No.: EPA-823-
F-03-003.

Stow CA, Larnon EC, Qian SS, Schrank CS. Will Lake Michigan lake trout meet
the Great Lakes Strategy 2002 PCB reduction goal? Environ Sci Technol
2004;38(2):359-63.

Laden F, Neas LM, Spiegelman D, Hankinson SE, Willett WC, Ireland K, et al.
Predictors of plasma concentrations of DDE and PCBs in a group of US. women.
Environ Health Perspect 1999;107(1):75-81.

Rylander L, Dyremark E, Stromberg U, Ostrnan C, Hagmar L. The impact of age,
lactation and dietary habits on PCB in plasma in Swedish women. Sci Total
Environ 1997;207(1):55-6l.

Loomis D, Browning SR, Schenck AP, Gregory E, Savitz DA. Cancer mortality
among electric utility workers exposed to polychlorinated biphenyls. Occup
Environ Med l997;54(10):720—8.

Luotarno M, J arvisalo J, Aitio A. Analysis of polychlorinated biphenyls (PCBs) in
human serum. Environ Health Perspect 1985;60:327-32.

Sweeney AM, Syrnanski E, Burau KD, Kim Y], Humphrey HE, Smithci MA.
Changes in serum PBB and PCB levels over time among women of varying ages
at exposure. Environ Res 2001;86(2):128-39.

Schantz SL, Jacobson J L, Humphrey HE, Jacobson SW, Welch R, Gasior D.
Determinants of polychlorinated biphenyls (PCBs) in the sera of mothers and
children from Michigan farms with PCB-contaminated silos. Arch Environ Health
1994;49(6):452-8.

76

35.

36.

37.

38.

39.

40.

41.

42.

43.

44.

Birmingham B, Gilman A, Grant D. PCDD/PCDF multimedia exposure analysis
for the Canadian population: detailed exposure estimation. Chemosphere
1989;19:637-42.

Newhook RC. Polybrominated biphenyls: multimedia exposure analysis. Ottawa,
Canada: Department of National Health and Welfare; 1988.

Fitzgerald EF, Brix KA, Deres DA. Polychlorinated biphenyl (PCB) and
dichlorodiphenyl dichloroethylene (DDE) exposure among Native American men
from contaminated Great Lakes ﬁsh and wildlife. Toxicol Ind Health
1996;12:361-8.

Hites RA, Foran J A, Carpenter DO, Hamilton MC, Knuth BA, Schwager SJ.
Global assessment of organic contaminants in farmed salmon. Science
2004;303(5655):226-9.

Anderson HA, F alk C, Hanrahan L. Proﬁles of Great Lakes critical pollutants: a
sentinel analysis of human blood and urine--The Great Lakes Consortium.
Environ Health Perspectives 1998;106(5):279-289.

Dar E, Kanarek MS, Anderson HA, Sonzogni WC. Fish consumption and
reproductive outcomes in Green Bay, Wisconsin. Environ Res 1992;59(l):189-
201.

Dellinger J A, Meyers RM, Gebhardt KJ, K. HL. The Ojibwa Health Study: ﬁsh
residue comparisons for Lakes Superior, Michigan, and Huron. Toxicol Ind
Health 1996;12(3-4):393-402.

Schantz SL, Sweeney AM, Gardiner J C, Humphrey HE, McCaffrey RJ, Gasior
DM, et al. Neuropsychological assessment of an aging population of Great Lakes
ﬁsheaters. Toxicol Ind Health 1996;12(3-4):403-17.

Gerstenberger SL, Tavris DR, Hansen LK, Pratt-Shelley J, Dellinger J A.
Concentrations of blood and hair mercury and serum PCBs in an Ojibwa
population that consumes Great Lakes region ﬁsh. J Toxicol Clin Toxicol
1997;35(4):377-86.

Hanrahan LP, F alk C, Anderson HA, Draheim L, Kanarek MS, Olson J. Serum
PCB and DDE levels of frequent Great Lakes sport ﬁsh consumers-a ﬁrst look.
The Great Lakes Consortium. Environ Res 1999;80(2 Pt 2):SZ6-S37.

77

45.

46.

47.

48.

49.

50.

51.

52.

53.

54.

Stewart P, Darvill T, Lonky E, Reihman J, Pagano J, Bush B. Assessment of
prenatal exposure to PCBs from maternal consumption of Great Lakes ﬁsh: an
analysis of PCB pattern and concentration. Environ Res 1999;80(2 Pt 2):S87-S96.

He JP, Stein AD, Humphrey HE, Paneth N, Courval JM. Time trends in sport-
caught Great Lakes ﬁsh consumption and serum polychlorinated biphenyl levels
among Michigan Anglers, 1973-1993. Environ Sci Technol 2001;35(3):435-40.

He JP. Sport-caught Great Lakes ﬁsh consumption and human serum
polychlorinated biphenyl levels: cross-sectional and longitudinal relations in
Michigan anglers [Masters]. East Lansing, Michigan: Michigan State University;
1998.

Tee PG, Sweeney AM, Symanski E, Gardiner J C, Gasior DM, Schantz SL. A
longitudinal examination of factors related to changes in serum polychlorinated
biphenyl levels. Environ Health Perspect 2003;111(5):702-7.

Courval JM, DeHoog JV, Holzman CB, Tay EM, Fischer L, Humphrey HE, et al.
Fish consumption and other characteristics of reproductive-aged Michigan

anglers--a potential population for studying the effects of consumption of Great
Lakes ﬁsh on reproductive health. Toxicol Ind Health l996;12(3-4):347-59.

Lonky E, Reihman J, Darvill T, Mather J, Daly H. Neonatal behavioral
assessment scale performance in humans inﬂuenced by maternal consumption of
environmentally contaminated Lake Ontario ﬁsh. J Great Lakes Res 1996;22:198-
212.

Waller DP, Presperin C, Drum ML. Great Lakes ﬁsh as a source of maternal and
fetal exposure to chlorinated hydrocarbons. Toxicol Ind Health 1996;12:335-45.

Tilden J, Hanrahan LP, Anderson H, Palit C, Olson J, Kenzie WM. Health
advisories for consumers of Great Lakes sport ﬁsh: is the message being
received? Environ Health Perspect 1997;105(12): 1360-5.

Fitzgerald EF, Hwang SA, Langguth K, Cayo M, Yang BZ, Bush B, et a1. Fish
consumption and other environmental exposures and their associations with
serum PCB concentrations among Mohawk women at Akwesasne. Environ Res
2004;94(2): 160-70.

Beehler GP, McGuinness BM, Vena J E. Characterizing Latino anglers'
environmental risk perceptions, sport ﬁsh consumption, and advisory awareness.
Med Anthropol Q 2003;17(1):99-l 16.

78

55.

56.

57.

58.

59.

60.

61.

62.

63.

64.

65.

Kuwabara K, Yakushiji T, Watanabe I, Yoshida S, Yoyarna K, Kunita N. Increase
in the human blood PCB levels promptly following ingestion of ﬁsh containing
PCBs. Bull Environ Contam Toxicol l979;21(l-2):273-8.

Matthews HB, Dedrick RL. Pharmacokinetics of PCBs. Annu Rev Pharmacol
Toxicol 1984;24:85-103.

Hesse J. State of the Great Lakes, 1996 Annual Report: PCB Levels in the Blood
of an Avid Great Lakes Angler. Lansing, Michigan: Ofﬁce of the Great Lakes -
Michigan Department of Environmental Quality; 1997 April.

Phillips DL, Smith AB, Burse VW, Steele GK, Needham LL, Harmon WH. Half-
life of polychlorinated biphenyls in occupationally exposed workers. Arch
Environ Health 1989;44(6):35 1-4.

Shirai JH, Kissel J C. Uncertainty in estimated half-lives of PCBS in humans:
impact on exposure assessment. Sci Total Environ 1996;187(3):199-210.

Yakushiji T, Watanabe I, Kuwabara K, Tanaka R, Kashimoto T, Kunita N, et a1.
Rate of decrease and half-life of polychlorinated biphenyls (PCBs) in the blood of
mothers and their children occupationally exposed to PCBs. Arch Environ
Contam Toxicol 1984;13(3):341-5.

Taylor PR, Reilly AA, Stelma JM, Lawrence CE. Estimating serum
polychlorinated biphenyl levels in highly exposed workers: an empirical model. J
Toxicol Environ Health 1991 ;34(4):413-22.

Wolff MS, Fischbein A, Selikoff U. Changes in PCB serum concentrations among
capacitor manufacturing workers. Environ Res 1992;59(1 ):202-16.

Mes J, Arnold DL, Bryce F. The elimination and estimated half-lives of speciﬁc
polychlorinated biphenyl congeners from the blood of female monkeys after
discontinuation of daily dosing with Aroclor 1254. Chemosphere l995;30(4):789-
800.

Ryan JJ, Levesque D, Panopio LG, Sun WF, Masuda Y, Kuroki H. Elimination of
polychlorinated dibenzofurans (PCDFs) and polychlorinated biphenyls (PCBs)
from human blood in the Yusho and Yu-Cheng rice oil poisonings. Arch Environ
Contam Toxicol 1993;24(4):504-12.

Masuda Y, Kuroki H, Haraguchi K, Ryan JJ, Shu ST. [Elimination of PCDF and
PCB congeners in the blood of patients with PCB poisoning in Taiwan]. Fukuoka
Igaku Zasshi 1991;82(5):262-8.

79

66.

67.

68.

69.

70.

71.

72.

73.

74.

75.

76.

Schildkraut JM, Demark-Wahnefried W, DeVoto E, Hughes C, Laseter J L,
Newman B. Environmental contaminants and body fat distribution. Cancer
Epidemiol Biomarkers Prev 1999;8(2):179-83.

Schmid K, Lederer P, Goen T, Schaller KH, Strebl H, Weber A, et al. Internal
exposure to hazardous substances of persons from various continents:
investigations on exposure to different organochlorine compounds. Int Arch
Occup Environ Health 1997;69(6):399-406.

Dobson S, van Esch GJ. Polychlorinated biphenyls and terphenyls. Geneva:
World Health Organization; 1993.

Falk C, Hanrahan L, Anderson HA, Kanarek MS, Draheim L, Needham L, et al.
Body burden levels of dioxin, furans, and PCBs among frequent consumers of
Great Lakes sport ﬁsh. The Great Lakes Consortium. Environ Res 1999;80(2 Pt
2):Sl9-825.

Feeley MM. Biomarkers for Great Lakes priority contaminants: halogenated
aromatic hydrocarbons. Environ Health Perspect 1995;103 Suppl 9:7-16.

Hanrahan L, Anderson HA, Falk C, Olson J. Reproductive predictors of serum
PCB and DDE levels among frequent Great Lakes sport ﬁsh consumers: the role
of gender, births and breastfeeding. Eur J Oncol 1999;4(5):585-593.

Hovinga ME, Sowers M, Humphrey HE. Environmental exposure and lifestyle
predictors of lead, cadmium, PCB, and DDT levels in Great Lakes ﬁsh eaters.
Arch Environ Health 1993;48(2):98-104.

Kearney JP, Cole DC, F erron LA, Weber JP. Blood PCB, p,p'-DDE, and mirex
levels in Great Lakes ﬁsh and waterfowl consumers in two Ontario communities.
Environ Res 1999;80(2 Pt 2):S138-S149.

Kreiss K. Studies on populations exposed to polychlorinated biphenyls. Environ
Health Perspect 1985;60: 193-9.

Pelletier C, Irnbeault P, Tremblay A. Energy balance and pollution by
organochlorines and polychlorinated biphenyls. Obes Rev 2003;4(1):17-24.

Pelletier C, Doucet E, Irnbeault P, Tremblay A. Associations between weight
loss-induced changes in plasma organochlorine concentrations, serum T(3)
concentration, and resting metabolic rate. Toxicol Sci 2002;67(1):46-51.

80

77.

78.

79.

80.

81.

82.

83.

84.

85.

86.

87.

Patandin S, Dagnelie PC, Mulder PG, Op de Coul E, van der Veen J E, Weisglas-

Kuperus N, et al. Dietary exposure to polychlorinated biphenyls and dioxins from
infancy until adulthood: A comparison between breast-feeding, toddler, and long-
term exposure. Environ Health Perspect l999;107(1):45-51.

Schecter A, Ryan JJ, Papke O. Decrease in levels and body burden of dioxins,
dibenzofurans, PCBS, DDE, and HCB in blood and milk in a mother nursing
twins over a thirty-eight month period. Chemosphere 1998;37(9-12):1807-16.

Rothman KJ, Sander G. Modern epidemiology. 2nd ed. Philadelphia: Lippincott-
Raven; 1998.

Armstrong B, K, White, E, Saracci, R. Principles of Exposure Measurement in
Epidemiology. New York: Oxford University Press; 1992.

Rothman K, Poole, C. A strengthening program for weak associations. Int J
Epidemiol 1988;17(Suppl):955-959.

Rice DC. PCBs and behavioral impairment: are there lessons we can learn from
lead? Neurotoxicol Teratol 1996;18(3):229-32; discussion 271-6.

Pryor J, Hughes, C, Foster, W, Hales, BF, Robaire, B. Critical windows of
exposure for children's health: the reproductive system in animals and humans.
Environ Health Perspect 2000;108(Suppl 3):491-503.

Cok I, Hakan Satiroglu M. Polychlorinated biphenyl levels in adipose tissue of
primiparous women in Turkey. Environ Int 2004;30(1):7-10.

Muscat JE, Britton JA, Djordjevic MV, Citron ML, Kemeny M, Busch-
Devereaux E, et al. Adipose concentrations of organochlorine compounds and
breast cancer recurrence in Long Island, New York. Cancer Epidemiol

Biomarkers Prev 2003;12(12):1474-8.

Enrique MO, Morales V, Ngoumgna E, Prescilla R, Tan E, Hernandez E, et al.
Prevalence of fetal exposure to environmental toxins as determined by meconium
analysis. Neurotoxicology 2002;23(3):329-39.

Hong Z, Gunter M, Randow FF. Meconium: a matrix reﬂecting potential fetal
exposure to organochlorine pesticides and its metabolites. Ecotoxicol Environ Saf
2002;51(1):60-4.

81

88.

89.

90.

91.

92.

93.

94.

95.

96.

97.

98.

Whyatt RM, Barr DB. Measurement of organophosphate metabolites in
postpartum meconium as a potential biomarker of prenatal exposure: a validation
study. Environ Health Perspect 2001;109(4):417-20.

Seegal RF. Can epidemiological studies discern subtle neurological effects due to
perinatal exposure to PCBs? Neurotoxicol Teratol l996;18(3):251-4; discussion
271 -6.

Glynn AW, Granath F, Anne M, Atuma S, Damerud PO, Bjerselius R, et a1.
Organochlorines in Swedish women: determinants of serum concentrations.
Environ Health Perspect 2003;111(3):349-55.

DeVoto E, Fiore BJ, Millikan R, Anderson HA, Sheldon L, Sonzogni WC, et a1.
Correlations among human blood levels of speciﬁc PCB congeners and
implications for epidemiologic studies. Am J Ind Med 1997;32(6):606-13.

Gladen BC, Longnecker MP, Schecter AJ. Correlations among polychlorinated
biphenyls, dioxins, and furans in humans. Am J Ind Med 1999;35(l):15-20.

Courval JM, DeHoog JV, Stein AD, Tay EM, He J, Humphrey HE, et al. Sport-
caught ﬁsh consumption and conception delay in licensed Michigan anglers.
Environ Res 1999;80(2 Pt 2):S l 83-S l 88.

Hagrnar L, Linden K, Nilsson A, Norrving B, Akesson B, Schutz A, et al. Cancer
incidence and mortality among Swedish Baltic Sea ﬁshermen. Scand J Work
Environ Health 1992;] 8(4):217-24.

Jacobson J L, Jacobson SW. Evidence for PCBs as neurodevelopmental toxicants
in humans. Neurotoxicology 1997;18(2):415-24.

Rylander L, Hagrnar L. Mortality and cancer incidence among women with a high
consumption of fatty ﬁsh contaminated with persistent organochlorine
compounds. Scand J Work Environ Health l995;21(6):419-26.

Schwartz PM, Jacobson SW, Fein G, Jacobson JL, Price HA. Lake Michigan ﬁsh
consumption as a source of polychlorinated biphenyls in human cord serum,
maternal serum, and milk. Am J Public Health 1983;73(3):293-6.

Fein GG, Jacobson J L, Jacobson SW, Schwartz PM, Dowler JK. Prenatal
exposure to polychlorinated biphenyls: effects on birth size and gestational age. J
Pediatr 1984;105(2):315-20.

82

99.

100.

101.

102.

103.

104.

105.

106.

107.

108.

109.

110.

Mendola P, Buck GM, Vena JE, Zielezny M, Sever LE. Consumption of PCB-
contarninated sport ﬁsh and risk of spontaneous fetal death. Environ Health
Perspect 1995;103(5):498-502.

Wicklund Glynn A, Willett, LB. Food Frequency Questionnaires as a Measure of
Exposure to Organochlorines in Epidemiological Studies. In. Special Circular
163-99 ed; 1999.

Paneth N. Human reproduction after ingestion of PCB-contaminated ﬁsh. Health
Environ Digest 1991;5:4-6.

Sexton K, Adgate J L, Eberly LE, Clayton CA, Whitmore RW, Pellizzari ED, et
al. Predicting children's short-terrn exposure to pesticides: results of a
questionnaire screening approach. Environ Health Perspect 2003;111(1):123-8.

Cole DC, Sheeshka J, Murkin EJ, Kearney J, Scott F, Ferron LA, et a1. Dietary
intakes and plasma organochlorine contaminant levels among Great Lakes ﬁsh
eaters. Arch Environ Health 2002;57(5):496-509.

Woodruff T, Wolff MS, Davis DL, Hayward D. Organochlorine exposure
estimation in the study of cancer etiology. Environ Res 1994;65(1):132-44.

Hutchinson R, Kraﬁ, C.E. Hmong ﬁshing activity and ﬁsh consumption. Journal
of Great Lakes Research l994;20(2):471-478.

Jacobson SW, F ein GG, Jacobson J L, Schwartz PM, Dowler JK. The effect of
intrauterine PCB exposure on visual recognition memory. Child Dev
1985;56(4):853-60.

Vena JE, Buck GM, Kostyniak P, Mendola P, Fitzgerald E, Sever L, et al. The
New York Angler Cohort Study: exposure characterization and reproductive and
developmental health. Toxicol Ind Health l996;12(3-4):327-34.

Daly HB, Stewart PW, Lunkenheimer L, Sargent D. Maternal consumption of
Lake Ontario salmon in rats produces behavioral changes in the offspring. Toxicol
Ind Health 1998;14(1-2):25-39.

Anderson H, Falk C, F iore B, Hanrahan L, Humphrey HE, Kanarek M, et a1.
Consortium for the Health Assessment of Great Lakes Sport Fish Consumption.
Toxicol Ind Health 1996; 12(3-4):369-73.

Longnecker MP, Rogan WJ, Lucier G. The human health effects of DDT
(dichlorodiphenyltrichloroethane) and PCBS (polychlorinated biphenyls) and an

83

111.

112.

113.

114.

115.

116.'

117.

118.

119.

120.

121.

overview of organochlorines in public health. Annu Rev Public Health
1997;18:211-44.

Swanson GM, Ratcliffe HE, Fischer 1.]. Human exposure to polychlorinated
biphenyls (PCBs): a critical assessment of the evidence for adverse health effects.
Regul Toxicol Pharmacol 1995;21(1):136-50.

Schell JD, Jr., Budinsky RA, Wemke MJ. PCBs and neurodevelopmental effects
in Michigan children: an evaluation of exposure and dose characterization. Regul
Toxicol Pharmacol 2001;33(3):300-12.

Hruska KS, Furth PA, Seifer DB, Sharara Fl, Flaws J A. Environmental factors in
infertility. Clin Obstet Gynecol 2000;43 (4) : 82 1-9.

Golden R, Doull J, Waddell W, Mandel J. Potential human cancer risks from
exposure to PCBs: a tale of two evaluations. Crit Rev Toxicol 2003;33(5):543-80.

Negri E, Bosetti C, Fattore E, La Vecchia C. Environmental exposure to
polychlorinated biphenyls (PCBs) and breast cancer: a systematic review of the
epidemiological evidence. Eur J Cancer Prev 2003;12(6):509-16.

Hagrnar L. Polychlorinated biphenyls and thyroid status in humans: a review.
Thyroid 2003;l3(11):1021-8.

Stein AD, Tay E, Courval J M. Absence of nonresponse bias in a study of sport-
caught Great Lakes ﬁsh consumption and conception failure. Environ Res
1999;80(3):287-93.

Larsen SB, Abell A, Bonde JP. Selection bias in occupational sperm studies. Am
J Epidemiol 1998;147(7):68l-5.

Last JM. A dictionary of epidemiology. 4th ed. New York: Oxford University
Press; 2001.

Chu S, Covaci A, Jacobs W, Haraguchi K, Schepens P. Distribution of methyl
sulfone metabolites of polychlorinated biphenyls and p,p'-DDE in human tissues.
Environ Health Perspect 2003;] 1 1(9): 1222-7.

Chevrier J, Dewailly E, Ayotte P, Mauriege P, Despres JP, Tremblay A. Body
weight loss increases plasma and adipose tissue concentrations of potentially toxic
pollutants in obese individuals. Int J Obes Relat Metab Disord 2000;24(10):1272-
8.

84

122.

123.

124.

125.

126.

127.

128.

129.

130.

131.

132.

133.

Imbeault P, Chevrier J, Dewailly E, Ayotte P, Despres JP, Tremblay A, et al.
Increase in plasma pollutant levels in response to weight loss in humans is related
to in vitro subcutaneous adipocyte basal lipolysis. Int J Obes Relat Metab Disord
2001;25(11):1585-91.

Walford RL, Mock D, MacCallum T, Laseter J L. Physiologic changes in humans
subjected to severe, selective calorie restriction for two years in biosphere 2:
health, aging, and toxicological perspectives. Toxicol Sci 1999;52(2 Suppl):61-5.

Fitzgerald EF, Weinstein AL, Youngblood LG, Standfast SJ, Melius JM. Health
effects three years after potential exposure to the toxic contaminants of an
electrical transformer ﬁre. Arch Environ Health 1989;44(4):214-21.

Noren K. Levels of organochlorine contaminants in human milk in relation to the
dietary habits of the mothers. Acta Paediatr Scand 1983;72(6):811-6.

Deutch B, Pedersen HS, Jorgensen EC, Hansen J C. Smoking as a determinant of
high organochlorine levels in Greenland. Arch Environ Health 2003;58(1):30-6.

Deutch B, Hansen J C. High human plasma levels of organochlorine compounds
in Greenland. Regional differences and lifestyle effects. Dan Med Bull
2000;47(2):132-7.

Deutch B, Hansen J C. High blood levels of persistent organic pollutants are
statistically correlated with smoking. Int J Circumpolar Health 1999;58(3):214-9.

Lackmann GM, Angerer J, Tollner U. Parental smoking and neonatal serum levels
of polychlorinated biphenyls and hexachlorobenzene. Pediatr Res
2000;47(5):598-601.

Lindstrom G, Haug LS, Nicolaysen T, Dybing E. Comparability of world-wide
analytical data of PCDDs, PCDFs and non-ortho PCBs in samples of chicken,
butter and salmon. Chemosphere 2002;47(2):139-46.

Rosner B. Fundamentals of Biostatistics. 4 ed. Belmont, California: Duxbury
Press; 1995.

Kleinbaum DG, Kupper LL, Muller KE, Nizam A. Applied regression analysis
and other multivariable methods. Paciﬁc Grove: Doxbury; 1998.

Friendly M. SAS macro programs: inﬂplot.
http://wwwmathyorkuca/SCS/sasmac/inﬂplot.html; 2001.

85

134.

135.

136.

137.

138.

Carroll RJ, Ruppert D. Transformation and weighting in regression. New York:
Chapman and Hall; 1988.

Cook RD, Weisberg S. An Introduction to Regression Graphics. New York: John
Wiley & Sons; 1994.

SAS. Computing a Bootstrap Conﬁdence Interval. In: SAS Institue Inc., Cary,
NC, USA; 1999.

Graybill F, Hariharan, KI. Regression Analysis: Concepts and Applications.
Belmont, California: Duxbury Press; 1994.

Paneth N. Adopting a public health approach to developmental neurotoxicity.
Neurotoxicol Teratol 1996;1 8(3):233-4.

86

 

 

I I‘ll: 1.0.1-

 

;l

[ ‘J :1 J‘J LIBHA‘RIES
ii iii; ‘11
01

E umvERsrrr
1 1,
i
i
l

iii

(I!

“w
.‘l ll

4
4
O
5
1111.2
O
3
9
2
1

ll“

ii”
3

l

l