a meta-analysis of alcohol consumption and the risk of 15 diseases
TRANSCRIPT
www.elsevier.com/locate/ypmed
Preventive Medicine 38 (2004) 613–619
A meta-analysis of alcohol consumption and the risk of 15 diseases
Giovanni Corrao, Ph.D.,a Vincenzo Bagnardi, Sc.D.,a
Antonella Zambon, Sc.D.,a and Carlo La Vecchia, M.D.b,c,*
aDipartimento di Statistica, Universita di Milano-Bicocca, Milan, Italyb Istituto di Statistica Medica e Biometria, Universita di Milano, Milan, Italy
c Istituto di Ricerche Farmacologiche ‘‘Mario Negri’’, Milan, Italy
Abstract
Background. To compare the strength of evidence provided by the epidemiological literature on the association between alcohol
consumption and the risk of 14 major alcohol-related neoplasms and non-neoplastic diseases, plus injuries.
Methods. A search of the epidemiological literature from 1966 to 1998 was performed by several bibliographic databases. Meta-
regression models were fitted considering fixed and random effect models and linear and nonlinear effects of alcohol intake. The effects of
some characteristics of the studies, including an index of their quality, were considered.
Results. Of the 561 initially reviewed studies, 156 were selected for meta-analysis because of their a priori defined higher quality,
including a total of 116,702 subjects. Strong trends in risk were observed for cancers of the oral cavity, esophagus and larynx, hypertension,
liver cirrhosis, chronic pancreatitis, and injuries and violence. Less strong direct relations were observed for cancers of the colon, rectum,
liver, and breast. For all these conditions, significant increased risks were also found for ethanol intake of 25 g per day. Threshold values were
observed for ischemic and hemorrhagic strokes. For coronary heart disease, a J-shaped relation was observed with a minimum relative risk of
0.80 at 20 g/day, a significant protective effect up to 72 g/day, and a significant increased risk at 89 g/day. No clear relation was observed for
gastroduodenal ulcer.
Conclusions. This meta-analysis shows no evidence of a threshold effect for both neoplasms and several non-neoplastic diseases. J-shaped
relations were observed only for coronary heart disease.
D 2004 The Institute For Cancer Prevention and Elsevier Inc. All rights reserved.
Keywords: Alcohol intake; Alcohol-related diseases; Meta-analysis
Introduction
Alcohol drinking is a major correlate of health and
disease, and it has been associated to cancer, cardiovascular
diseases, digestive tract conditions, accidents, and violence.
Although alcohol is not known to be carcinogenic in
animal experimentation, there is strong epidemiological
evidence that consumption of alcoholic beverages increases
the risk of cancers of the oral cavity and pharynx, esoph-
agus, and larynx [1,2]. The risks are essentially thought to
be related to ethanol content and appear to be linked to the
most commonly used alcoholic beverages in each popula-
tion. These risks tend to increase with the amount of ethanol
drunk, but it is still unclear whether there is any defined
0091-7435/$ - see front matter D 2004 The Institute For Cancer Prevention and
doi:10.1016/j.ypmed.2003.11.027
* Corresponding author. Istituto di Ricerche Farmacologiche ‘‘Mario
Negri’’, Via Eritrea 62, 20157 Milan, Italy. Fax: +39-02-3320-0231.
E-mail address: [email protected] (C. La Vecchia).
threshold below which no effect is evident [1,2]. Alcohol
drinking is also associated with primary liver cancer, al-
though this relation is difficult to investigate in epidemio-
logical studies because most alcohol-related liver cancers
follow a cirrhotic degeneration, which leads to a reduction
of alcohol drinking [3].
Alcohol drinking has also been linked to cancers of the
large bowel in both sexes [2] and breast cancer in women
[4]. Although these associations are still open to discussion,
these are the two major causes of cancer death in developed
countries after lung cancer [5,6], and therefore even a
moderate excess risk may have important public health
implications. Furthermore, alcohol drinking is strongly
related to mortality from liver cirrhosis, chronic pancreatitis,
hypertension, stroke, accidents, and violence [7,8].
It has been suggested, in contrast, that moderate alcohol
consumption (i.e., <25 g/day) protects against coronary
heart disease [9]. A number of plausible mechanisms have
been formulated, including the relation between alcohol
Elsevier Inc. All rights reserved.
G. Corrao et al. / Preventive Medicine 38 (2004) 613–619614
drinking and cholesterol, high density lipoprotein, apolipo-
proteins, fibrinolysis, coagulation factors, and blood pres-
sure, but none has as yet been firmly established [10].
It has also been suggested that the protection is specific,
or greater, for selected types of alcoholic beverages, such as
red wine (due to phenolic compounds, flavonoids or tannin)
[11,12]. Independently of the population studied, and hence
of the most frequent type of alcohol beverages, however,
most studies found reduced risk of cardiovascular diseases in
moderate drinkers compared with nondrinkers [8,9,13–16].
Several comprehensive reviews are available on the effect
of alcohol on disease [1,17–23]. To evaluate the global
effect of alcohol on a population level, a more accurate
quantification of its effects and of its dose–risk relation on
various diseases is nonetheless required. This paper thus also
includes an original analysis of risks for moderate doses,
which are of specific interest for medical practice.
Methods
The following 15 conditions considered to be alcohol-
related were included in the meta-analysis: malignant neo-
plasms of the oral cavity and pharynx, of the esophagus, of
the colon, of the rectum, of the liver, of the larynx, and of the
breast, essential hypertension, coronary heart disease, hem-
orrhagic stroke, ischemic stroke, gastroduodenal ulcer, liver
cirrhosis, chronic pancreatitis, and injuries and violence. The
methods of this work are described in details elsewhere
[18,19].
Articles included in the meta-analysis were found through
a search of the literature published from 1966 to 1998. The
search was based on several bibliographic databases (MED-
DLINE, Current Contents, EMBASE CAB Abstracts, and
Core Biomedical Collection) supplemented by attention to
all references in the selected articles. Completeness was
verified by a hand search on the most relevant journals of
epidemiology and medicine and by comparing our search
with that of general reviews and meta-analyses published on
this issue [17,20–23].
For each of the above-reported conditions, the search
process involved combining the keywords ‘‘alcohol con-
sumption’’, ‘‘relative risk’’, and the specific ‘‘conditions’’.
These keywords were exploded in the search, thus to include
all the articles investigating the same issue, but reporting the
same term in different forms (i.e., ‘‘alcohol consumption’’ or
‘‘alcohol intake’’, ‘‘relative risk’’ or ‘‘risk ratio’’ or ‘‘odds
ratio’’, ‘‘cancer’’ or ‘‘malignant neoplasm’’ or ‘‘neoplasia’’,
‘‘cerebrovascular disease’’ or ‘‘stroke’’).
Each publication identified by this process was reviewed
and included in the analysis if the following criteria were
met: (i) case-control or cohort study published as an original
article; (ii) findings expressed as odds ratio or relative risk
(RR) considering at least three levels of alcohol consump-
tion; (iii) papers reporting the number of cases and noncases
and the estimates of the odds ratios or RR for each exposure
level. When the results of a study were published more than
once, only the most recent and complete article was included
in the analysis.
Two readers, blinded to the authors’ names and affilia-
tions and to the results pertaining to alcohol consumption,
independently determined the eligibility of each paper. The
same readers evaluated several characteristics of each study
and scored the quality of the studies according to criteria
elsewhere reported [18,19]. Questions related to the study
design (nine items), data collection methods for alcohol
consumption (four questions), and data analysis (two items).
The quality score for a study was obtained by adding up the
points given for individual questions. The readers also
evaluated whether each study reported RR adjusted for major
known risk factors (e.g., gender and age for all the con-
ditions, smoking for upper aerodigestive tract cancers, breast
cancer and stroke, hepatitis virus infections for liver cancer
and cirrhosis, hormonal and reproductive factors for breast
cancer, etc.. . .). Discrepancies between readers in the choice
to include an article and in quality score assignment were
resolved in conference.
Pooled estimates of the effect of alcohol consumption on
the risk of each investigated condition were based on a four-
step procedure. In the first step, several weighed least
squares regression models were fitted by prepooling the
results of all studies included, taking into account the
correlation between estimates within each study [24]. A
family of second-degree models was generated by power
transformation of the exposure variable, and the best-fitting
model was chosen to summarize the relation of interest [25].
Depending on both the magnitude and the sign of the
estimated parameters, the resulting function was able to
identify J-, U-, or L-shaped curves, and to accommodate
relations asymptotically tending towards a RR value for high
exposure levels.
In the second step, with the aim of testing the hypothesis
that the qualitative characteristics of the studies may modify
the effect of alcohol, several meta-regression models were
fitted [26]. Models included as covariate: (i) the term(s) of
alcohol consumption identified in the first step of the
analysis; (ii) the interaction term between alcohol intake
and a covariate describing a qualitative characteristic of the
study. The qualitative characteristics were the quality score,
the availability of adjusted RR, and the study design.
In the third step, with the aim of yielding more reliable
functions, studies with higher quality score (the cut-off was
placed differently for each condition, considering studies
with a score equal or above the condition-specific median
value), those conducted with a cohort design or those
reporting estimates adjusted for relevant covariates, were
selected only if one or more of these qualitative character-
istics resulted as significant effect modifiers.
Pooled RR and the corresponding 95% confidence inter-
vals (CI) were derived from the parameters of the meta-
regression models described above and by the corresponding
standard errors. For each condition, RR was modelled to
G. Corrao et al. / Preventive Medicine 38 (2004) 613–619 615
obtain an estimate of the risk associated to specific doses of
alcohol with respect to nondrinkers (model-based RR).
Finally, in the fourth step, the consistency of the model-
based RR was evaluated with reference studies reporting
relative risks for light consumption (V25 g/day). Pooled RR
and the corresponding 95% CI were estimated as simple
average of the reported RR weighed for the inverse of the
corresponding variance [27]. A test of homogeneity of the
RR across studies was provided by the Q statistics [27], and
random effects models were used when there was evidence
of significant heterogeneity [28].
Results
A total of 561 studies were retrieved. Of these, 240 were
included in the analysis (because they gave information on
number of subjects, on RRs for more than two levels of
alcohol consumption, and were not included in subsequent
reports), and 156 were selected for final analysis because of
their higher quality, and the results refer therefore to 156
studies. Of the selected studies, 148 reported adjusted
estimates for the main risk covariates of interest, 99 were
case-control, and 57 cohort studies, including a total of
116,702 cases of the 15 conditions considered.
Table 1
Selection process and main characteristics of the studies selected for the meta-an
Condition History of studies selection Study design
Retrievedb Includedc Selectedd Case-control Coh
Neoplastic conditions (cancer site)
Oral cavity and pharynx 58 24 15 14 1
Esophagus 51 28 14 13 1
Larynx 38 20 20 20 0
Colon 16 16 12 4
Rectum }49 14 6 4 2
Liver 43 19 10 8 2
Breast 72 48 29 24 5
Non neoplastic conditions
Essential hypertension 11 3 2 0 2
Coronary heart disease 196 51 28 0 28
Ischemic stroke 7 6 3 3
Hemorrhagic stroke }56 9 9 6 3
Gastroduodenal ulcer 9 3 2 1 1
Liver cirrhosis 27 15 9 6 3
Chronic pancreatitis 4 2 2 2 0
Injures and violence 34 18 12 1 11
Total 561 240 156 99 57
Model-based pooled relative risks and corresponding 95% confidence interval foa Pooled relative risk and corresponding 95% confidence interval (CI) directly ob
alcohol drinkers.b All the 561 case-control or cohort studies published as original articles were firsc Causes of exclusion were: only two alcohol categories were reported; number o
characterize exposure in terms of grams of alcohol per day; the article considere
included in meta-analysis.d Criteria for selection were: high quality score (malignant neoplasm of oral cavi
disease, and liver cirrhosis); reporting estimates adjusted for the main risk indicato
ulcer, and injures and adverse effects); or performed with a prospective cohort de
The main characteristics of the studies selected for meta-
analysis are given in Table 1, together with the model-based
pooled RR and 95% CI for 25, 50, and 100 g/day of alcohol
intake. Direct trends in risk were observed for cancers of the
oral cavity and pharynx, esophagus, and larynx. Direct
relations were also observed for cancers of the colon, rectum,
and liver, as well as for breast cancer. Among non-neoplastic
conditions, strong direct trends in risk were derived for
hypertension, liver cirrhosis, chronic pancreatitis, and inju-
ries and violence. For all these conditions, significant in-
creased risks were found starting from the lowest dose of
alcohol considered (25 g/day, corresponding to about two
drinks per day). Significant increased risks were found only
at 100 g/day for coronary heart disease and ischemic stroke,
and at 50 g/day for hemorrhagic stroke. A significant
protective action was observed at 25–50 g/day for coronary
heart disease. No clear relation was observed for gastrodu-
odenal ulcer.
The corresponding RR functions, with 95% CI, are
plotted in Fig. 1. Most functions were linear on a logarithmic
scale, except for diseases more strongly related to alcohol
(oral and pharyngeal cancer and liver cirrhosis), with flatten
curves at higher doses, and for cardiovascular diseases with
evidence of J-shaped relations. For ischemic and hemorrhag-
ic strokes, the nadirs were reached at 15 and 3 g/day,
alysis
No. of RR (and 95% CI) for selected doses of alcohol intakea
ortcases
25 g/day 50 g/day 100 g/day
4507 1.86 (1.76–1.96) 3.11 (2.85–3.39) 6.45 (5.76–7.24)
3233 1.39 (1.36–1.42) 1.93 (1.85–2.00) 3.59 (3.34–3.87)
3789 1.43 (1.38–1.48) 2.02 (1.89–2.16) 3.86 (3.42–4.35)
5360 1.05 (1.01–1.09) 1.10 (1.03–1.18) 1.21 (1.05–1.39)
1420 1.09 (1.08–1.12) 1.19 (1.14–1.24) 1.42 (1.30–1.55)
1321 1.19 (1.12–1.27) 1.40 (1.25–1.56) 1.81 (1.50–2.19)
32,175 1.25 (1.20–1.29) 1.55 (1.44–1.67) 2.41 (2.07–2.80)
5801 1.43 (1.33–1.53) 2.04 (1.77–2.35) 4.15 (3.13–5.52)
49,640 0.81 (0.79–0.83) 0.87 (0.84–0.90) 1.13 (1.06–1.21)
893 0.90 (0.75–1.07) 1.17 (0.97–1.44) 4.37 (2.28–8.37)
1192 1.19 (0.97–1.49) 1.82 (1.46–2.28) 4.70 (3.35–6.59)
425 0.98 (0.77–1.25) 0.97 (0.59–1.57) 0.93 (0.35–2.45)
2202 2.90 (2.71–3.09) 7.13 (6.35–8.00) 26.52 (22.26–31.59)
247 1.34 (1.16–1.54) 1.78 (1.34–2.36) 3.19 (1.82–5.59)
4501 1.12 (1.06–1.18) 1.26 (1.13–1.40) 1.58 (1.27–1.95)
116,706 – – –
r selected doses of alcohol consumption are also reported.
tained from the coefficients of the meta-regression model; reference: non-
tly retrieved. A study may include information on more than one condition.
f cases or noncases were not reported; article reported insufficient data to
d only partial results subsequently reported in complete form in an article
ty, of rectum, of liver, and of breast, essential hypertension, coronary heart
rs (malignant neoplasm of esophagus, ischemic stroke, gastric and duodenal
sign (coronary heart disease).
Fig. 1. Relative risk functions and corresponding 95% confidence intervals describing the dose-response relationship between alcohol consumption and the risk of 15 alcohol-related conditions obtained by fitting
meta-regression models.
G.Corra
oet
al./Preven
tiveMedicin
e38(2004)613–619
616
Fig. 3. Comparison between model-based and point-based relative risks and corres
25 g/day.
Fig. 2. Relative risk function and corresponding 95% confidence intervals
describing the dose-response relationship between alcohol consumption and
the risk of coronary heart disease.
G. Corrao et al. / Preventive Medicine 38 (2004) 613–619 617
respectively, but no evidence of significant protective effect
was observed. Significant increased risks were obtained
starting from 53 to 28 g/day, respectively. The RR function
of coronary heart disease is also plotted in Fig. 2. The
minimum (RR = 0.80) was reached at 20 g/day, a significant
protective effect was observed up to 72 g/day, while a
significant increased risk was obtained starting from 89
g/day (RR = 1.05).
Fig. 3 compares model-based and point-based (i.e., based
on studies providing direct information on intake <25 g/day)
pooled RR for light alcohol intake. Chronic pancreatitis was
not considered because no studies reported RR for light
intake. Owing to the smaller number of studies included and
to the weaker statistical assumptions, point-based RR
showed wider CI than model-based estimates. Consistent
and significant increased RRs were observed for cancers of
the oral cavity and pharynx, esophagus, larynx, colon, liver,
and breast, hypertension, liver cirrhosis, and injuries and
ponding 95% confidence intervals associated with alcohol intake lower than
G. Corrao et al. / Preventive Medicine 38 (2004) 613–619618
violence. The RR was above unit—though nonsignificant-
ly—for rectal cancer, too. The protective action of low dose
of alcohol on the risk of coronary heart disease was also
confirmed. Ischemic and hemorrhagic stroke, and gastrodu-
odenal ulcer were not associated to moderate alcohol intake
according to both analyses. The point-based RRs were
homogeneous across studies for all conditions, except inju-
ries and violence.
Discussion
This work has some of the limitations, but also most
strengths, of meta-analyses of published studies. Thus, bias
towards selective publications of studies showing inverse
relations with moderate levels of alcohol drinking was
apparent for coronary heart disease and is likely to have
occurred—in various directions—for other diseases, includ-
ing esophageal and laryngeal cancer, too. The overall pooled
estimates may be systematically influenced by the study
characteristics. We used three approaches to control for this.
Firstly, because an important part of the heterogeneity was
explained by the quality of the studies, we only used data
from a selection of 156 studies meeting a priori-defined
quality criteria. Secondly, we investigated reasons for het-
erogeneity. At least two of these, that is, geographic area and
gender, may reflect real effect modification due to different
patterns of drinking in southern Europe as compared to
northern Europe and the United States, as well as potential
differences in alcohol metabolism in women and men
[18,19]. Thirdly, because for some diseases a large amount
of unexplained heterogeneity remained even after these
controls, random effect models were used to derive pooled
RR estimates at moderate intake. With such an approach,
heterogeneous estimates still remained only for injuries and
violence. This finding is likely attributable to the broad
spectrum of conditions included in this category (e.g., road
injuries, suicides, homicides, accidental falls, etc.. . .).It is possible that alcohol drinking has been systemati-
cally underreported in several studies, mostly for specific
conditions, such as liver cirrhosis or accidents and violence.
Consequently, all the RRs would be underestimated in case
of selective underreporting by cases. Further, although this
meta-analysis included a total of more than 116,000 cases,
absolute numbers were relatively limited for some diseases
(i.e., gastroduodenal ulcer) or levels of alcohol drinking
(i.e., high levels for breast cancer in women). Exclusion of
cross-sectional studies has also substantially limited the
information available on hypertension [18]. Finally, the
estimates may be grossly biased for some of the diseases
considered, including liver cancer, due to preexisting cir-
rhosis and consequent reduced alcohol consumption [3,29].
These limitations notwithstanding, this meta-analysis still
include most published information on alcohol and disease,
and consequently provide the most accurate estimates of risk
for most common conditions considered to be alcohol-
related. Some of the findings from this meta-analysis are
innovative and of specific relevance, including the absence
of a threshold effect for any of the seven cancer sites
considered from both model-based and point-based RR,
and the apparently stronger association for oral and pharyn-
geal as compared to esophageal or laryngeal cancer across
subsequent levels of alcohol drinking. The absence of
evidence of association with gastroduodenal ulcer is of
interest, although it may be due to decreased alcohol drink-
ing following symptoms of the disease. The strong associ-
ations for hypertension and, mostly, for liver cirrhosis are
also of interest for a more accurate quantification of the dose-
risk relation and of the consequent public health relevance.
Of even greater importance is the definition of a risk
function between alcohol drinking and coronary heart dis-
ease, based on data from almost 50,000 cases, and providing
an estimate of a 20% risk reduction for 20 g/day of alcohol,
and an RR below unity until 70 to 80 g/day, but significantly
above unity more than 90 g/day. Finally, there was evidence
of a threshold effect for both ischemic and hemorrhagic
stroke.
These figures will therefore assist in any risk or benefit
assessment for alcohol drinking on an individual basis and a
population level. To put these findings in a context of
absolute risk, the baseline incidence of various diseases in
the population and the individual’s age should be taken into
account.
Acknowledgments
Supports for this study came from the Italian Ministry of
Health, the Italian Ministry of the University and Scientific
and Technologic Research, COFIN 2003, and the Italian
Association for Research on Cancer.
References
[1] International Agency for Research on Cancer. Alcohol drinking. IARC
Monogr Eval Carcinogen Risks Hum. Lyon: IARC; 1988. p. 416.
[2] Doll R, Forman D, La Vecchia C, Woutersen R. Alcoholic beverages
and cancers of the digestive tract and larynx. In: Macdonald I, editor.
Health issue related to alcohol consumption. Second ed. Oxford:
Blackwell; 1995. p. 351–93.
[3] La Vecchia C, Negri E, Cavalieri L, Franceschi S. Liver cirrhosis and
the risk of primary liver cancer. Eur J Cancer Prev 1998;7: 315–20.
[4] Schatzkin A, Longnecker MP. Alcohol and breast cancer. Where are
we now and here do we go from here? Cancer 1994;74:1101–10.
[5] Levi F, Lucchini F, La Vecchia C. Worldwide patterns of cancer
mortality, 1990–1994. Eur J Cancer Prev 1999;8:381–400.
[6] Levi F, Lucchini F, Boyle P, Negri E, La Vecchia C. Cancer incidence
and mortality in Europe, 1988–1992. J Epidemiol Biostat 1998;3:
295–373.
[7] Brewer RD, Morris PD, Cole TB, Watkins S, Patetta MJ, Popkin C.
The risk of dying in alcohol-related automobile crashes among habit-
ual drink drivers. N Engl J Med 1994;331:513–7.
[8] Thun MJ, Peto R, Lopez AD, Monaco JH, Henley SJ, Heath Jr CW,
et al. Alcohol consumption and mortality among middle-aged and
elderly U.S. adults. N Engl J Med 1997;337:1705–14.
G. Corrao et al. / Preventive Medicine 38 (2004) 613–619 619
[9] Miller GJ, Beckles GLA, Maude GH, Carson DC. Alcohol consump-
tion: protection against coronary heart disease and risks to health. Int J
Epidemiol 1990;19:923–30.
[10] Moore RD, Pearson TA. Moderate alcohol consumption and coronary
artery disease. A review. Medicine 1986;65:242–67.
[11] Frankel EN, Kanner J, German JB, Parks E, Kinsella JE. Inhibition of
oxidation of human low-density lipoprotein by phenolic substances in
red wine. Lancet 1993;341:454–7.
[12] La Vecchia C. Alcohol in the Mediterranean diet: assessing risks and
benefits. Eur J Cancer Prev 1995;4:3–5.
[13] Boffetta P, Garfinkel L. Alcohol drinking and mortality among men
enrolled in an American cancer society prospective study. Epidemi-
ology 1990;1:342–8.
[14] Bianchi C, Negri E, La Vecchia C, Franceschi S. Alcohol consump-
tion and the risk of acute myocardial infarction in women. J Epide-
miol Community Health 1993;47:308–11.
[15] Doll R, Peto R, Hall E, Wheatley K, Gray R. Mortality in relation to
consumption of alcohol: 13 years’ observations on male British doc-
tors. Br Med J 1994;309:911–8.
[16] Tavani A, La Vecchia C, Negri E, D’Avanzo B, Franzosi MG, Tog-
noni G. Alcohol intake and risk of myocardial infarction in Italian
men. J Epidemiol Biostat 1996;1:31–9.
[17] Macdonald I, editor. Health issue related to alcohol consumption.
Second ed. Oxford: Blackwell; 1995. p. 459.
[18] Corrao G, Bagnardi V, Zambon A, Arico S. Exploring the dose-
response relationship between alcohol consumption and the risk of
several alcohol-related conditions: a meta-analysis. Addiction
1999;94:1551–73.
[19] Corrao G, Rubbiati L, Bagnardi V, Zambon A, Poikolainen K. Alco-
hol and coronary heart disease: a meta-analysis. Addiction 2000;
95:1505–23.
[20] Longnecker MP, Orza MJ, Adams ME, Vioque J, Chalmers TC. A
meta-analysis of alcoholic beverage consumption in relation to risk of
colorectal cancer. Cancer Cause Control 1990;1:59–68.
[21] Maclure M. Demonstration of deductive meta-analysis: ethanol intake
and risk of myocardial infarction. Epidemiol Rev 1993;15:328–51.
[22] Longnecker MP. Alcoholic beverage consumption in relation to risk
of breast cancer: meta-analysis and review. Cancer Cause Control
1994;5:73–82.
[23] Corrao G, Bagnardi V, Zambon A, Torchio P. A meta-analysis of
alcohol intake in relation to risk of liver cirrhosis. Alcohol Alcohol
1998;33:381–92.
[24] Greenland S, Longnecker MP. Methods for trend estimation from
summarized dose-response data, with applications to meta-analysis.
Am J Epidemiol 1992;135:1301–9.
[25] Royston P, Ambler G, Sauerbrei W. The use of fractional polynomials
to model continuous risk variables in epidemiology. Int J Epidemiol
1999;28:964–74.
[26] Berlin JA, Longnecker MP, Greenland S. Meta-analysis of epidemi-
ologic dose-response data. Epidemiology 1993;4:218–28.
[27] Greenland S. Quantitative methods in the review of epidemiologic
literature. Epidemiol Rev 1987;9:1–30.
[28] DerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin
Trials 1986;7:177–88.
[29] Arico S, Corrao G, Torchio P, Galatola G, Tabone M, Valenti M, et al.
A strong negative association between alcohol consumption and the
risk of hepatocellular carcinoma in cirrhotic patients: a case-control
study. Eur J Epidemiol 1994;10:251–7.