Introduction
Wertheimer and Leeper (1) first reported an association between
wire configuration codes, a surrogate measure of residential magnetic field
exposure, and childhood cancer. A persistent concern has been the validity
of classifying exposure based on attributes of the power lines in the vicinity
of the home (2). Subsequent studies in both the USA (3,4)
and Europe (5,6) considered both characteristics of nearby distribution
and transmission lines and exposure indicators based on measured fields.
In general, although not as strong as Wertheimer and Leeper's (1)
original results, subsequent studies have reported relative risks of about
1.5-2.0 based on Wertheimer-Leeper wire codes (3,4). One of the most
surprising aspects of studies that included both measured magnetic fields
and wire codes was that associations with cancer were generally weaker for
measured fields than for wire codes (3,4).
Measured fields have consistently been correlated, albeit weakly, with
wire codes (1,3,7), providing empirical evidence that wire codes
are a proxy for contemporaneous magnetic field exposure in the home. If
the only goal was to determine present-day exposure to the home occupants,
measured fields would have to be considered superior to wire codes. However,
the goal of all epidemiologic studies conducted to date has been to characterize
long-term historical exposure. Given an interest in residential exposures
over a period as long as 15 or 20 years, the question of whether the present
measured field is a better surrogate than a wire code is not easily answered.
Distribution wiring is believed to change infrequently, so that the wire
code of the present is probably identical to the wire code that would have
been assigned in the past. In contrast, measured fields are influenced by
both stable attributes, such as the components produced by nearby power
lines, and transient sources, such as in-home wiring and electrical load
on the powerline grid at the time of measurement. Recent unpublished results
from Sweden addressing transmission line exposures (8) support the
contention that present-day measured fields may not be as useful a surrogate
for historical exposure as wiring characteristics, particularly when combined
with information on historical electric loads.
Wire codes not only may have the theoretical advantage of historical
stability, but they provide many logistical advantages in conducting epidemiologic
studies. Foremost is the opportunity to collect wire coding information
without disrupting the occupants of the home. Because entry into the home
or yard is seldom required to obtain information on wiring attributes, nonresponse
proportions may be markedly lower and data collection can be scheduled much
more efficiently than when respondent cooperation is needed, as is required
for obtaining in-home measurements. For example, in the study on which this
paper is based (3), measurements were obtained for 36% of case homes
and 75% of control homes, in contrast to wire codes, which were obtained
for 90% of cases and 93% of controls. Compared to measurements, which integrate
across all sources, wire codes have the advantage of more explicitly defining
the source of the magnetic field exposure. Finally, it is principally wire
codes that have been associated with cancer and, regardless of whether this
is reflective of an influence of magnetic fields or some other attribute
of the residence, wire codes are worthy of further study simply because
of the reported associations.
Because of the clear logistical advantages of wire codes, the possible
theoretical advantages, and their rather consistent association with risk
of childhood cancer, we have proposed refinements to the original Wertheimer-Leeper
code (Kaune and Savitz, submitted) using data from the previously reported
study of magnetic fields and childhood cancer (3). These refinements
make no sacrifice in the ability to predict measured fields, yet avoid the
most subjective elements of the coding method and create fewer categories.
By avoiding the major sources of inaccuracy in assessing wire thickness
and distinguishing among first-span, short first-span, and second-span secondary
lines, we may well reduce random error in classification. Wire thickness,
for example, is a subjective distinction based on a comparison among wires,
lacking any gold standard of validity. The resulting error, almost certain
to be nondifferential with respect to case or control status because the
data were collected blindly, would dilute measures of association. By creating
a smaller number of categories, the precision of the risk estimates is enhanced
relative to the previously reported five-level code. Therefore, we analyzed
the association between the modified Wertheimer-Leeper code and childhood
cancer in the Denver, Colorado, area and contrasted the results to those
from the original analysis.
Methods
The data used here were collected between 1984 and 1986 as part of a
case-control study of childhood cancer and residential magnetic field exposure,
details of which are reported elsewhere (3). Briefly, all cases of
cancer in children under age 15 were identified among residents of the 1970
Denver Standard Metropolitan Statistical Area (Adams, Arapahoe, Boulder,
Denver, and Jefferson counties), from 1January 1976 through 31 December
1983. The Colorado Central Cancer Registry provided most of the cases, with
complete registration for the period 1979- 1982, supplemented by data from
area hospitals before and after that time. Diagnostic accuracy was assured
through microscopic confirmation (over 95%) and review by a pediatric oncologist.
Incidence rates in the study area were similar to those reported in the
Surveillance, Epidemiology and End Results (SEER) registries (9),
suggesting essentially complete ascertainment.
We chose controls through random-digit dialing by starting with each
case's telephone number at the time of diagnosis and randomly replacing
the last two digits. A child of similar age (±3 years) and sex was
sought (3). An undesirable aspect of identifying controls some years
after the cases had been diagnosed is that only eligible controls who remained
in the home occupied at the time of the case's diagnosis were identifiable.
Therefore, controls were more residentially stable than cases. In the analysis,
we examined a marker of residential stability as a potential confounder
or effect modifier.
We conducted a structured interview addressing a wide range of potential
childhood cancer risk factors with the mother or alternate respondent for
those cases whose physicians permitted us to contact them and who agreed
to participate in the study, along with identified, cooperative controls.
We assigned wire configuration codes based on residence at the time of diagnosis
for cases or the assigned "diagnosis" date for controls, regardless
of duration of occupancy. Lists of the addresses of interest were provided
to a coder who was unaware of the case or control status of the occupant
of each residence. The coder sketched a map of the distribution and transmission
lines in the vicinity of the homes and used a coding sheet to record detailed
information on wiring.
As reported in the original analysis (3), we analyzed the Wertheimer-Leeper
code using the four levels plus buried wire category (10), as well
as by dichotomizing the wire codes as low (ordinary low current configuration,
very low current configuration, buried wire) or high (very high current
configuration, ordinary high current configuration). The modified wire codes
analyzed in this paper, whose levels are low wire code (LWC), medium wire
code (MWC), and high wire code (HWC), were derived based on their relations
with measured fields. This new coding system assigns HWC to homes within
20 m of a transmission or three-phase primary line and MWC to homes that
are not HWC but have a transmission line or three-phase primary line within
46 m or an open secondary line within 26 m (Kaune and Savitz, submitted).
Both the original Wertheimer-Leeper code and modified code were able to
account for approximately 20% of the variance in measured fields, but there
is a clear separation in the means across the wire code categories in both
instances (Kaune and Savitz, submitted).
We calculated crude odds ratios and 95% confidence intervals for total
cases and subgroups of acute lymphocytic leukemia (ALL), brain tumors, lymphomas,
soft tissue tumors, and other cancers. With the five-level code, we contrasted
each of the four upper levels with buried wire as the referent. For the
dichotomous codes, we contrasted high with low, and for the modified Wertheimer-Leeper
code, both HWC and MWC were contrasted with LWC.
Potential confounding was examined for age at diagnosis (0-4, 5-9, 10-14
years), gender (male, female), race (white, black/other), mother's age (<20
versus >=21 years), mother's smoking (yes, no), father's education (<16
versus >=16), per capita income (<$7,000 versus >=$7,000 per year),
year of diagnosis (before 1979, 1980 or later), residential stability (stable
residence from birth to diagnosis, moved between birth and diagnosis), and
residence type (single family dwelling, other). We computed adjusted odds
ratios and confidence intervals using the Mantel-Haenszel technique (11).
Logistic regression models were constructed that simultaneously adjusted
for age at diagnosis, race, mother's smoking, per capita income, year of
diagnosis, and residential stability. Age at diagnosis, gender, father's
education, per capita income, year of diagnosis, residential stability,
and residence type were also examined as potential effect modifiers.
Results
As noted in the earlier report (3), 356 cases were identified,
among whom wire codes at diagnosis were obtained for 320 (90%), and interviews
were obtained for 252 (71%). For controls, 378 were identified, with an
estimated 79% response in the random-digit dialing phase, among whom wire
codes were obtained for 259 (93%) and interviews for 222 (80%). In conducting
the subject-by-subject review for this reanalysis, we identified two additional
wire codes, one for a case and one for a control, increasing the totals
to 321 cases and 260 controls. Among the cases, there were 103 leukemias
(98 with wire codes), 83 acute lymphocytic leukemias (78 with wire codes),
67 brain tumors (59 with wire codes), 35 lymphomas (30 with wire codes),
32 soft tissue sarcomas (all with wire codes), and 119 other cancer types
(102 with wire codes). The results are presented initially for all eligible
subjects with wire codes, but for analyses examining confounding or effect
modification, only the subset of subjects with both wire codes and completed
interviews could be included.
Table 1 summarizes the results of the original analysis by five-level
wire code. An increase in risk with higher wire codes was found, restricted
largely to the very high-current configuration residences [odds ratios (ORs)
of 1.6-3.3]. These measures are rather imprecise, particularly for the specific
cancer types. Suggestions of increased risk among ordinary high-current
configuration homes were found for total cancers, leukemias, brain tumors,
and other cancer, with ORs of 1.4-2.0. The dichotomous wire code (Table
2) yielded much more precise evidence of elevated risks for all cancers
except lymphomas (OR = 0.8, 95% CI: 0.3-2.2). Total cancers, leukemias,
and other cancers yielded ORs around 1.5, with brain tumors showing a stronger
association (OR = 2.1, 95% CI: 1.1-3.8). The impact of restricting the data
to interviewed subjects is presented in Table 2 as well, indicating that
in the absence of any adjustments for confounding, interviewed subjects
had slightly higher ORs for leukemia and lower ORs for brain tumors.
The mapping of the original five-level Wertheimer-Leeper code to the
modified three-level Wertheimer-Leeper code is described in Table 3. Because
the modified wire codes are hypothesized to be the more accurate, we examined
the potential misclassification introduced in reassigning homes from the
modified to the Wertheimer-Leeper codes. Overall, the pattern of movement
was as expected, with LWC homes predominantly migrating to the lowest three
levels of Wertheimer-Leeper wire code, MWC homes going to the middle groups,
and HWC homes moving to the two higher current-configuration categories.
Examination of the pattern of movement of cases versus controls yields rather
consistent results for the two groups. Focusing on the dichotomy of the
lowest three to the upper two categories, 35 of 266 LWC and MWC cases combined
(13%) were in the upper group versus 26 of 234 LWC and MWC controls (11%).
The likelihood of an HWC subject being assigned to the very high current
configuration category was also similar, with 35% of cases and only 31%
of controls moving in that direction. The misclassification thus appears
to be basically nondifferential.
Analysis of cancer risk in relation to the modified Wertheimer-Leeper
code yields ORs that are more precise than the results based on the five-level
wire code analysis and more markedly elevated than the results based on
the dichotomous codes (Table 4). In the HWC versus LWC contrast, total cancers
showed nearly a 2-fold increased risk, with leukemias (OR = 2.9, 95% CI:
1.5-5.5) and brain tumors (OR = 2.5, 95% CI: 1.1-5.5) particularly strongly
associated. In contrast to the 7 leukemias and 3 brain tumor cases in the
very high current configuration group, there were 22 leukemias and 12 brain
tumor cases in the HWC category, markedly improving the precision of the
estimates. Lymphoma and soft tissue sarcoma showed more modest and imprecise
ORs of 1.6 and 1.7, respectively, and the aggregation of other cancers showed
no association (OR = 1.0, 95% CI: 0.5-2.2). The contrast of MWC with LWC
yielded no association for total cancers, but leukemias and brain tumor
cases showed a modest association (ORs of 1.3 and 1.2, respectively). The
risk elevations were basically restricted to the HWC group.
Potential confounding was examined (Table 5) and generally found to be
absent. Only results for the HWC versus LWC are presented, with the contrast
of MWC versus LWC remaining near the null value with or without adjustment.
As in the original analysis, the baseline risk estimates were different
after restriction to interviewed subjects, with the overall ORs rising from
1.9 to 2.1 for all cases, 2.9 to 3.5 for leukemias, and falling from 2.5
to 1.9 for brain tumor cases. Thus, confounding was evaluated relative to
that benchmark.
Adjusted ORs for total cancers, leukemias, and acute lymphocytic leukemia
showed no evidence of substantial confounding, defined as adjusted estimates
more than 10% larger or smaller than the unadjusted estimates. For brain
tumors, a modest decrease in the OR was found after adjustment by race and
a modest increase with adjustment for year of diagnosis. Logistic regression
models were constructed for total cancers, leukemias, and brain cancer,
with age at diagnosis, race, maternal smoking, per capita income, year of
diagnosis, and residential stability included as potential confounders.
In those models, the adjusted ORs contrasting HWC and LWC were 2.0 for total
cancers (95% CI: 1.0-4.0) , 3.8 for leukemias (95% CI: 1.6-9.0), and 2.4
for brain tumors (95% CI: 0.8-7.6). With adjustment, the modest elevations
for medium versus low modified wire code were eliminated, with ORs of 0.8
for total cancers (95% CI: 0.5-1.2) , 0.9 for leukemias (95% CI: 0.4-1.8),
and 0.8 for brain tumors (95% CI: 0.4-1.9).
Indications of effect modification were examined for the cancers with
the most adequate numbers, namely, total cancers, leukemias, and brain tumors
(Table 6). For total cancers, and especially for brain tumor cases, the
increased risk for HWC was concentrated among the older cases. Females had
higher ORs than males, especially for leukemias. Social class indicators
have received considerable attention because of possible selection bias
in the manner in which controls were chosen and recruited (12). Risk
estimates were somewhat greater in the lower paternal education and lower
income strata but were still present in the higher education and higher
income strata. Residential stability is also of particular concern given
the restrictions that were imposed on the controls but not on the cases.
For total cancers and leukemia cases, ORs were less markedly elevated for
stable cases (same residence from birth to diagnosis), with a tendency in
the opposite direction for brain tumor cases. Finally, the strongest indication
of effect modification was found for year of diagnosis, in which much of
the elevation in risk was found for cases and controls with diagnosis dates
of 1980 or later.
Discussion
The principal result of this reanalysis of data on wire configuration
codes and childhood cancer is the stronger relations found for the modified
HWC as compared to the analysis based on the original dichotomized Wertheimer
and Leeper wire configuration coding system. In fact, the ORs based on the
modified HWC were similar in magnitude but markedly more precise than the
ORs previously noted for very high current configuration alone. The original
dichotomous Wertheimer-Leeper code yielded ORs of 1.5 and 2.1 for leukemias
and brain cancers, respectively, and the HWC produced ORs of 2.9 and 2.5,
respectively. A modestly elevated risk was associated with MWC compared
to LWC, but that elevation was eliminated by adjusting for potential confounding
factors.
Additional contrasts from the earlier analysis concern the magnitude
of risk associated with different cancer types. The analysis of the dichotomous
Wertheimer-Leeper codes produced rather consistent 1.5-fold increased risks
for all cancer types, with a slightly larger risk for brain cancer, and
no increased risk for lymphomas. The modified high wire code was notably
more strongly linked to leukemia (OR = 2.9) and brain cancer (OR = 2.5)
than to lymphoma (OR = 1.6), soft tissue sarcoma (OR = 1.7), or to the heterogeneous
other cancer group (OR = 1.0). Small numbers of lymphoma and soft tissue
sarcoma cases limit the certainty of this pattern, however.
In addition to the hypothesized sources of spurious elevations in risk
associated with wire configuration code, a principal source of dilution
in risk measures has been assumed to be nondifferential misclassification
of exposure (12,13). Under the assumption that an association is
truly present, reducing the extent of that misclassification would yield
stronger indices of association. The modified wire code is believed to be
more valid as an exposure marker than the original code because the most
subjective elements in the data collection process are eliminated, and the
relation between the wire code and field measurements is similar. In fact,
the pattern of results relative to the Wertheimer-Leeper wire code is consistent
with the hypothesis that nondifferential misclassification has been reduced.
Inherent uncertainties resulting from lack of information on exposures outside
the home and lack of detailed information about the use of appliances remain.
However, regardless of the absolute level of accuracy in wire codes as a
proxy of exposure, which is currently unknown, the improvement in wire codes
yielded a corresponding increase in the measures of association for leukemia
and brain cancer. Such a pattern is consistent with an underlying causal
association not only between wire codes and cancer but, because the modified
code was derived from spot measurements of magnetic fields, it is consistent
with the possibility of an effect of average magnetic field on childhood
cancer.
A number of aspects of the data call into question the presence of an
etiologic relation between wire code and childhood cancer. The design of
the study yielded controls who were more residentially stable than cases
(3), with residential stability potentially related to wire configuration
code. Instead of stable subjects (those who lived in the same home from
birth to diagnosis) yielding stronger associations, as might be expected
if longer duration of occupancy improved discrimination of relevant exposure
levels, the risk estimates for residentially stable subjects were actually
somewhat lower than for unstable subjects in relation to leukemia (ORs of
2.7 and 4.7) and modestly higher for brain cancer (ORs of 2.1 and 1.7, respectively).
However, residential stability is closely associated with age at diagnosis,
so that the lower risk for stable subjects reflects in large part a lower
risk for younger subjects.
Another aspect of control selection that raises questions about the validity
of our results is the loss of potential controls in the random-digit dialing
process, either because potential candidates lacked telephones, refused
at the screening stage, or refused to be interviewed. Wire code information
was available for most subjects who made it past the screening stage, but
the others are not included in the control group. It might be postulated
that such persons would be most prevalent in the lower socioeconomic status
stratum. Measured by either father's education or per capita income, ORs
for leukemia were greater among lower socioeconomic status subjects, but
not markedly so. In the higher education and income strata, ORs were still
2.5 or greater, and for brain cancer, there was no difference in the risk
estimates across education and income strata.
One unexplained source of effect modification was related to year of
diagnosis. Virtually all of the increase in risk for HWC homes was among
subjects diagnosed in 1980 or later, for whom the ORs were strikingly large
and imprecise. The ORs were 3.7 (95% CI: 1.4-9.7) for total cancers, 7.1
(95% CI: 2.3-22.1) for leukemias, and 3.8 (95% CI: 1.5-9.9) for brain cancer.
It might be speculated that wire codes are somehow more valid exposure markers
for more recent cases and controls, e.g., if modifications to the distribution
system were unexpectedly frequent over time, but the reason for enhanced
risks in more recently diagnosed subjects remains unknown.
At a minimum, this reanalysis should serve to encourage other investigators
to consider the modified wire code as an addition to, and perhaps even an
alternative to, the original Wertheimer-Leeper coding system. The pursuit
of this research avenue is largely driven by epidemiologic findings, and
the empirical basis for further study of the modified code is more compelling
than for the original wire configuration code based on the Denver study
analyzed here. The logistical advantages of the method (Kaune and Savitz,
submitted) add to the incentive to restrict studies to the modified code
alone.
The generalizability of these results to the other studies that have
used the Wertheimer-Leeper method (1,4,10) would be of great interest.
In a single study, the potential for inexplicable and perhaps random processes
to yield a pattern must be acknowledged. If the modified coding method enhanced
associations with cancer risk across study settings, its validity would
markedly increased.
