Cardiovascular disease is the primary cause of mortality in developed countries,
with the exception of Japan. In Comunidad Valenciana, an autonomous region
in Spain with approximately 4 million inhabitants, cardiovascular disease accounts
for 35% of total female mortality and 46% of total male mortality (Melchor
et al. 1998). Geographic variability within Comunidad Valenciana has been documented
(Ferrándiz et al. 2000, 2002; Nolasco et al. 1992).
Previously published scientific articles have reported a negative correlation
between drinking water hardness and cardiovascular mortality. Some results
obtained in ecologic studies (Cradford et al. 1971; Lacey and Shaper 1984;
Pocock et al. 1980) suggest that high levels of drinking water hardness (i.e.,
high concentrations of calcium and magnesium) are protective against cardiovascular
diseases, mainly against ischemic heart disease.
Several surveys based on individual cases (Hall and Jungner 1993; Van der
Vijver et al. 1992) have not confirmed the protective effect of calcium. Nevertheless,
Rylander et al. (1991) and Yang (1998) suggest the beneficial effect of magnesium
against coronary heart disease mortality as well as against cerebrovascular
mortality. The results of these studies, obtained at the aggregated level,
have been partially corroborated by case-control studies (Rubenowitz et
al. 1996, 2000). These case-control studies, conducted in 18 southern
municipalities in Sweden, found a protective effect of magnesium against acute
myocardial infarction mortality but failed to find this effect for the total
incidence in men.
A recent report by Marx and Neutra (1997) on the relationship between ischemic
disease and magnesium in drinking water presented an analysis of several ecologic
studies showing contradictory results, perhaps because the studies were not
sufficiently specific to find the associations the investigators were exploring.
The authors concluded their report by recommending that further studies be
conducted to evaluate the apparent benefit of drinking water with high magnesium
concentration.
A key issue to be addressed is the hypothetical temporal sequence between
exposure and adverse health effects. It is unclear when to measure these factors,
as there is no clear latency period. Some authors (Rubenowitz et al. 1996)
have indicated that 1 year is sufficient to produce observable magnesium effects.
However, other authors pointed out that longer periods of observations are
needed (Marx and Neutra 1997).
More recently, Ferrandiz et al. (2003) studied the relationship between cerebrovascular
mortality and calcium and magnesium concentrations in drinking water in 262
municipalities of the Valencia province in Spain from 1990 to 1995. They found
a decreasing temporal trend, suggesting the cumulative effect of this beneficial
factor, although this assertion needs further research.
Our research extends this last study, taking advantage of the Spanish Rapid
Inquiry Facility (RIF). The RIF is an analytical tool for quick assessment
that is applied to the data gathered in an information system developed within
the European Health and Environment Information System (EUROHEIS) project and
allows exposure analysis with covariates (Gómez et al. 2002).
First, we enlarged the period studied to 1991-1998. Second, we included
all municipalities of Comunidad Valenciana, not just those belonging to the
province of Valencia, thus enlarging our study from 263 to 538 municipalities.
This provides a wider range of values for the factors being studied. Third,
we considered ischemic [ICD-9 410-414 (World Health Organization 1978)] as
well as cerebrovascular [ICD-9 430-438; (World Health Organization 1978)] diseases.
Finally, data on drinking water hardness have also been completed and updated
as a result of efforts to build a comprehensive environmental database inside
the RIF.
Material and Methods
Data
One of the primary advantages of the RIF is the comprehensive database built
into it. Mortality/morbidity, demographic, socioeconomic, and environmental
data are assembled in a georeferenced system, allowing geographic representations
of these phenomena. All data used in our study came from this source.
When constructing and updating the RIF database, mortality counts were obtained
from the mortality registry of the Dirección General de Salud Pública.
These numbers correspond to residents of Comunidad Valenciana and include deaths
of those residents that occurred in or out of the region.
Similarly, environmental data such as those on the quality of drinking water
were obtained from the Servicio de Calidad de Aguas at the Conselleria de Medio
Ambiente. This agency has been analyzing public water supplies on an annual
basis since 1989, although the frequency of the measurements varies between
municipalities and some data are still missing. The average number of measurements
in each municipality was 5.6 from 1991 to 1998. Fitting these data into the
RIF database required statistical imputation of their values. Bayesian analysis
of spatiotemporal models was used to perform this task as described in Abellán
et al. (2003). Nevertheless, calcium and magnesium concentrations in drinking
water were stable during the study, thereby minimizing the influence of the
imputation methodology on the results of the analysis. Any missing value at
a location was estimated by the average of the nearest 5 years at the same
location. This simple procedure was sufficient in previous exploratory analyses.
Finally, demographic and socioeconomic data were provided by the Valencian
Statistical Institute, where municipal statistics are updated regularly.
Exposure Analysis
We performed exposure analysis within the RIF by defining regions (bands)
composed of geographical units sharing similar levels of the risk factor under
study.
These degrees of risk can be based on distance to a putative origin of risk
(as in point source analysis) or on values of some environmental variables,
as in our study. For each of the calcium and magnesium concentrations, we defined
five bands, using as cut points the quintiles of their respective distributions
on the 538 municipalities studied. To achieve uniqueness of these bands during
the period studied, we used the values corresponding to 1991, the first year
of our study. This choice was based, among other reasons, on a special program
of water quality measurement used by the regional authorities that year. These
regional authorities used a methodology common to all of the municipalities
in Comunidad Valenciana.
Table 1
|
Table 1 shows the values of calcium and magnesium defining those bands, the
number of municipalities in each, and the percentage of total population of
the region. The unevenness shown by these values is due to multiple ties in
some of the cut points as well as to the variability of population sizes of
these 538 municipalities. Populated municipalities make appreciable contributions
to the band where they are allocated.
Assessing Effects of Risk Factors
We performed statistical analysis using the estimation of the relative risk
of each band i by the corresponding standardized mortality ratio SMRi = Oi/Ei of
observed (O) to expected (E) mortality counts. In the computation
of the expected counts Ei, standardization by age groups
was performed separately for each sex, as well as by levels of a deprivation
index based on three municipal indicators: the ratios of unemployed individuals,
the proportion of illiterates among individuals > 10 years of age, and number
of vehicles per individual inhabitants (Arias et al. 1993). This deprivation
index has been incorporated in the RIF as a new field attached to each municipality
register. As a comparison region for each municipality, deprivation index standardization
uses the band of the covariate that it belongs to and not the whole region
of study. Thus, the model incorporates and controls the fact that risk could
not be the same at different levels of the covariate.
In our study we performed indirect intrinsic standardization, using the population
of the whole region as the reference population for each of the periods studied.
The standardization procedure implicitly assumes that the expected rate in
each stratum of a region is equal to the product of the relative risk of the
region and a common mortality rate of this stratum. This is called the proportionality
assumption (Wakefield et al. 2000), which must be checked to obtain valid conclusions.
We performed a linear fit to the strata-specific rates of each municipality,
and we did not observe clear departures from the linear assumption (Ferrándiz
et al. 2003).
The output provided by the RIF includes the SMRs and their 95% confidence
intervals in each band for every sex group.
Significant relative risks, that is, those for which confidence intervals
do not include the value 1, are highlighted in the RIF output to facilitate
their detection by visual inspection. Thus, we can identify those levels of
the studied factors that correspond to unusually high or low relative risks.
Relying on this band-by-band inspection to identify the influence of a risk
factor as significant has a statistical drawback. Because we are performing
multiple tests to obtain a unified conclusion, we encounter the problem of
simultaneous inference; that is, we risk identifying the global effect of an
irrelevant factor as significant with a probability much higher than the nominal
5% level of each test. Thus, we have to protect against this global type I
error by increasing the confidence level of our intervals or by performing
a global test of homogeneity of bands before accepting any individual significant
result.
This second alternative seems easier from the output of the RIF. It provides
the observed Oi and expected Ei counts
so that we can perform a 
2 test of homogeneity of the number of
(n) bands by computing the statistic
[1]
to be compared with the quantiles of the 2 distribution
with n - 1
degrees of freedom. In Equation 1, r is the ratio of total observed
to total expected cases in the entire region, the maximum likelihood estimator
of the common relative risk under the assumption of homogeneity of bands and
Poisson-distributed counts.
Handling Multiple Covariates
Standardizing mortality/morbidity rates by levels of a covariate as we have
with age groups and deprivation index is a way of filtering its influence to
allow the resulting statistics to be free from its effects. The remaining variability,
if any, will be due to sources other than this covariate.
Covariate analysis, an option available within the RIF environment, performs
this task. Once we have stipulated the desired bands of the covariate under
study, the RIF computes the relevant statistics of each band, as described
in the preceding section. Then we can ask the program to build a new index
with these levels to standardize rates in future studies. [See Gómez
et al. (2002) for computational details.]
In each analysis we performed within the RIF, we can compare results obtained
before and after standardization by levels of a covariate. For example, we
want to know if calcium concentration in drinking water is a relevant covariate
once we have considered the magnesium concentration. Thus, we have compared
bands defined from calcium levels after standardization by levels of magnesium.
Heterogeneity of these bands will indicate that calcium provides relevant information
beyond that supplied by magnesium. Furthermore, comparison of calcium bands
before and after standardizing by levels of magnesium will illustrate the interaction
of both factors.
Disease Mapping
One main objective of epidemiologic surveillance tasks is the detection of
regions that have unusually high risk. Disease mapping is a powerful tool designed
to this end, especially when we are dealing with environmental risk factors.
Because environmental phenomena are linked to geography, the influence of these
risk factors can be detected by geographic representations of relative risks.
[See Lawson and Williams (2001) for an introductory text and Lawson et al.
(1999) for a deeper insight.]
Disease mapping deals typically with small geographic units. If the influence
of hidden environmental factors extends over several units, mortality/morbidity
counts will be correlated. Therefore, to analyze these units we need statistical
models allowing for spatial correlation.
Furthermore, the small populations attached to these geographic units produce
unstable estimates of relative risks, thus requiring more robust statistical
methods.
The RIF addresses both problems by resorting to the empirical Bayes analysis
of a hierarchical Poisson-gamma model similar to that of Clayton and Kaldor
(1987). Computational details are described in the statistical appendix of
Aylin et al. (1999).
From a surveillance perspective, we want to determine if removing the effects
of a covariate changes the geographical pattern of relative risks. To this
end we can perform disease mapping before and after standardization by levels
of a covariate. By comparing the resulting maps, we can verify whether high-risk
regions move to lower levels of risk or if they remain high, indicating that
factors other than this covariate are still affecting population health status.
There could be hidden factors not included in the study. The geographic pattern
can help us determine the nature of these hidden factors.
Results
To delimit the size of the studied phenomena, we first considered the annual
rates per 100,000 inhabitants for the whole region. The annual rates for cerebrovascular
disease from 1991 to 994 are 153.37 for women and 114.26 for men. From 1995
to 1998 these rates are 129.03 and 97.17, respectively.
The annual rates for ischemic heart disease per 100,000 inhabitants are 80.13
for women and 121.46 for men from 1991 to 1994, whereas they are 86.41 and
126.10, respectively, from 1995 to 1998.
Cerebrovascular disease rates are higher in women than in men; for ischemic
heart disease the rates are higher in men. Comparing both periods, we observe
a decrease in the rate of cerebrovascular disease and an increase in the rate
of ischemic heart disease.
The subsequent analysis focuses on relative risks rather than on the rates
and is based on the routine output of the RIF.
Exposure Analysis
Figure 1. 95% Confidence intervals and means of
SMRs for cerebrovascular
mortality of males (A,C) and females (B,D) in bands defined for
calcium (A,B) and magnesium (C,D). B, band.
|
Figure 2. 95% Confidence intervals and means of SMRs
for ischemic heart
mortality of males (A,C) and females (B,D) in bands defined for
calcium (A,B) and magnesium (C,D). B, band. |
Table 2
|
Figure 3. 95% Confidence intervals and means of SMRs
for cerebrovascular
mortality of males (A,C) and females (B,D) in bands defined from
calcium (A,B) and magnesium (C,D). B, band. |
Figure 4. 95% Confidence intervals and means of SMRs
for ischemic heart
mortality of males (A,C) and females (B,D) in bands defined for calcium
(A,B) and magnesium (C,D), after standardization by calcium C,D)
and magnesium (A,B). B, band. |
Figure 5. Disease mapping of total cerebrovascular
mortality for the whole
period: smoothed SMRs (A,C) and significance of 95% confidence intervals
(B,D) after standardization by age, sex, and deprivation index (A,B)
and further standardization by Mg (C,D). Inf, infinity. Municipalities
illustrating the change of risk level when adjusting for the covariate are circled
in blue. |
Table 3
|
Figure 6. Comparing settings of magnesium bands for
cerebrovascular mortality
in women. Output obtained with (A) three bands, (B) five bands,
and (C) seven bands. B, band. |
Figures 1 and 2 are a comparison of the relative risk of bands defined from
calcium and magnesium concentration levels. They display the SMRs and the 95%
confidence intervals from the output obtained with the RIF. Figure 1 illustrates
cerebrovascular mortality and Figure 2 illustrates ischemic heart mortality.
All SMRs for these figures have been computed after standardization by age
and deprivation index.
For each disease we have constructed four plots according to sex and covariate.
In each band, both 1991-1994 and 1995-1998 are represented side by
side for better comparison of temporal variation.
The horizontal line at SMR = 1 allows quick recognition of those intervals
not containing this particular value, that is, those intervals that were not
significant because the corresponding band shows a significantly high or low
relative risk. Because we have performed intrinsic indirect standardization,
distance from the SMR = 1 indicates a difference with respect to the average
behavior of the whole region. The presence of significant confidence intervals
is a clear sign of the heterogeneity of the bands.
As we discussed in "Material and Methods," this information has to be complemented
with testing the homogeneity of the bands. The resulting chi-square statistic
G and corresponding p-values are displayed in Table 2 under the headings
G and p-value.
Magnesium after Calcium and Calcium after Magnesium
To see the additional effect of each covariate once the other has been taken
into account, we have repeated the analysis of the preceding section. This
time, however, SMRs have been computed after standardization by the covariate
not explicitly present in the exposure analysis. Consequently, Figure 3 has
to be compared with Figure 1 and Figure 4 with Figure 2. The corresponding
homogeneity tests appear in Table 2 under the headings G and p-value.
From these comparisons we can see that trends are similar in general but
that confidence intervals become less significant. Many more intervals intersect
the horizontal line SMR = 1 when we standardize by the covariate not present
in the exposure analysis. This loss of significance is apparent as well from
columns p-value and p-value of Table 2. We verify there that p-values
increase notably from from first column to second column, indicating that the
hidden covariate contributes to the heterogeneity between bands. Conversely,
small p-values suggest the covariate that defines the bands still provides
useful information beyond that of the covariate used in standardization.
Disease Mapping
Rapid Inquiry Facility output gives tabulated SMRs for all municipalities.
Shown for each sex group (males, females, and males + females) for each municipality
are observed and expected number of cases, the corresponding SMR and the 95%
confidence interval, and the smoothed estimation of this SMR based on the empirical
Bayes procedure mentioned in preceding sections. These rows are duplicated
to show standardized and nonstandardized results.
Because we are working with 538 municipalities and two diseases, the textual
output is more than 500 pages for each of the studied covariates. Although
interesting for detailed consultation purposes, it does not fit in the reduced
space of a scientific paper. Maps better summarize these results. They facilitate
the capture of essential aspects of health status. However, the entire set
of maps for the present study is excessive, and we will restrict ourselves
to some of the most illustrative results.
Figure 5 presents disease mapping of total (males + females) cerebrovascular
mortality for 1991-1998. Figure 5A and C represent smoothed municipal
relative risks. Figure 5B and D distinguish between significantly high, significantly
low, and nonsignificant 95% confidence intervals of SMRs, as the value SMR
= 1 is below, above, or inside the interval. Thus, we have an estimate of the
relative risk (Figure 5A,C) jointly with a measure of our confidence that the
value represents a real risk and is not being produced by mere chance (Figure
5B,D).
In Figure 5A and B, SMRs have been standardized by age, sex, and deprivation
index. In Figure 5C and D, standardization has included magnesium levels as
well.
Discussion
Effects of Calcium and Magnesium
The results displayed in Table 2 suggest a relationship between calcium and
magnesium and the data on mortality from cerebrovascular and ischemic diseases.
According to the p-value, testing homogeneity of bands shows clear evidence
of this association for cerebrovascular disease in women. All p-values
are below 0.0001 for both periods studied and both covariates. Evidence of
this association is not as strong for this same disease in men, because there
is no clear significant result from 1991 to 1994 with magnesium and from 1995
to 1998 with calcium.
For ischemic heart disease significant heterogeneous results are not achieved
if the threshold is set to 0.001 to declare a p-value significant. Nevertheless,
the p-values are quite small, with most between 0.001 and 0.05. A cautious
conclusion could be not to discard the possibility of this association without
further consideration.
We can examine the nature of those relationships. Focusing on the plot for
women and magnesium in Figure 1D (p-value = 1.12 
10-7),
the one most significant is in Table 2, where we can see a descending trend
with increasing levels of magnesium from bands 1 to 4. Band 5 breaks this trend,
giving a U-shaped aspect to this plot.
A similar pattern can be seen for ischemic heart mortality in women and magnesium
levels, although in this case heterogeneity has not been so significant (p-value
= 0.031).
Regarding the interaction between calcium and magnesium, Table 2 reveals
clearly how the effects of both covariates are partially confounded because
they produce in each other a loss of significance when used in the previous
standardization. The correlation coefficient for both covariates is 0.59. In
this situation it is difficult to assess the independent effect of each one,
and it is best to refer to the effect of hardness of drinking water.
In summary, we can say that this study provides statistical evidence of a
relationship between mortality from cardiovascular diseases and hardness of
drinking water. This relationship is stronger in cerebrovascular disease than
in ischemic heart disease, is more pronounced in women than in men, and is
more apparent with magnesium than with calcium. Nevertheless, the protective
nature of these two factors is not clearly established. Although the results
obtained suggest this possibility, they are not conclusive because of the irregular
trend in the series of confidence intervals and because many of the results
are not significant. Hidden socioeconomic and environmental factors not controlled
with the deprivation index or the studied covariates may remain. As suggested
by a reviewer, these could be caused by an ecologic bias associated with this
region.
Temporal Trend
We have paired the confidence intervals corresponding to 1991-1994 and
1995-1998 for each band in every plot in Figures 1-4. Direct inspection
of these charts reveals the stability of the SMRs during the entire period
studied. The majority of these pairs have a large intersection, with both intervals
sharing a large portion of their range of values.
Although one may have the impression that small decreases predominate in
these sets of paired intervals, the overlapping areas are so important that
the evidence of temporal variation is negligible.
Spatial Distribution of Risk
No spatial trend is apparent from maps presented in Figure 5, but several
clusters of different sizes are scattered over the entire region. For illustrative
purposes we focused on two particular regions, which are circled in the figure.
The upper circle shows a cluster of municipalities with high SMR and significant
confidence intervals. This is obvious in the map of smoothed SMRs standardized
by age, sex, and deprivation index (Figure 5A) and in the map with significance
of confidence intervals (Figure 5B). The lower circle shows another less extreme
cluster.
When we include magnesium levels in the standardization calculus, we remove
the effect of this covariate in some sense. When we compare Figure 5A and B
with Figure 5C and D, we can see the effect of this removal. For example, the
upper circle shows that this change produces even more municipalities with
significant confidence intervals than before (so that the situation is worse
than previously thought). In the lower circles, the opposite is true. Some
municipalities decrease the significance of their SMRs (so that their previous
high relative risk has been partially explained by their level of magnesium).
Some Methodological Issues
The strategy of analysis followed in this study conforms to the exploratory
nature of the RIF as a tool to get quick and flexible insight into epidemiologic
surveillance problems.
One primary concern with the type of exposure analysis described here is
the sensitivity of results to the number and cut points of exposure bands.
An advisable practice is to try various configurations of these bands. We still
lack a clear recommendation about this topic.
In our study we used three, five, and seven bands in each of the 12 studies
(disease-covariate-sex combinations). Table 3 shows results obtained
in one of these comparisons--cerebrovascular mortality in women with bands
defined from magnesium levels. Figure 6 displays the SMRs and their 95% confidence
intervals from the output obtained with three, five, and seven bands. We can
see that although concrete numerical results vary, the general conclusions
remain.
Considering these 12 comparisons of different band settings, we found that
three bands tend to produce less significant heterogeneous results than with
five and seven bands, whereas there is little difference between these last
settings. Therefore, we have presented results with five bands.
Conclusions
Water in the Comunidad Valenciana is very hard. Because of the infrequent
occurrence, water with > 200 mg/L calcium and water containing > 10 mg/L
magnesium are considered rare by the World Health Organization (1996). Nearly
20% of Comunidad Valenciana municipalities and population have water supplies
with calcium concentrations > 200 mg/L. The case of magnesium is more striking--90%
of the water supply contains magnesium concentrations > 10 mg/L. This could
make identification of dose-response patterns and comparison between our
results and those from other studies difficult. Nevertheless, we observed considerable
consistency in a detailed analysis of the results obtained in the most recent
studies of this issue.
In other studies of calcium, the rank of calcium concentration distribution
in water is closer than in our study. For example, in the case-control
study performed in Sweden (Rubenowitz et al. 1996), the interval is 22-225
mg/L, and in a recent ecologic study conducted in France (Marque et al. 2003),
the calcium concentration interval is 9-146 mg/L. As in our study, mortality
results for cardiovascular noncerebrovascular diseases are less clear than
those for cerebrovascular diseases. Conversely, the case-control study
performed in Sweden shows a nonmonotonic U-shaped relationship between calcium
and heart attack risk mortality, as the authors obtained odds ratio < 1
in intermediate calcium concentrations between 34 and 81 mg/L (Rubenowitz et
al. 1996). In other words, they found a relationship between heart attack risk
mortality calcium levels that correspond to the second level of our distribution,
in which we found a significant relationship, with SMR < 1 at concentrations
between 65 and 89 mg/L. In a later study in Sweden, the results for calcium
were not conclusive (Rubenowitz et al. 2000). Nevertheless, in the French study,
the shape of the relationship between calcium and cardiovascular mortality
presents a clear biologic gradient, with less mortality risk at higher calcium
concentrations in water (Marque et al. 2003).
On the other hand, our results support the protective effect assumption of
magnesium and the mortality risk due to cardiovascular diseases (Rubenowitz
et al. 1996, 2000). In addition, the results of the study performed in France
(Marque et al. 2003) show a U-shaped relation between magnesium concentration
in drinking water and cerebrovascular mortality, with lower risk in intermediate
values of magnesium, in the same way described in our study.
Briefly, the results of our study in Valencia support the assumption of association
between magnesium and mortality risk due to cardiovascular diseases. However,
results for calcium are less clear. The current lack of studies and the ecologic
nature and limitations of the exposure valuation used suggest that these study
results should be explored further with more suitable designs. This could be
achieved in the moving cohorts studies framework that addresses the role of
different nutrients and other factors in cardiovascular health.
