Alcohol Dosing and Total Mortality in Men and Women: An Updated Meta-analysis of 34 Prospective Studies

Background  Moderate consumption of alcohol is inversely related with coronary disease, but its association with mortality is controversial. We performed a meta-analysis of prospective studies on alcohol dosing and total mortality.

Methods  We searched PubMed for articles available until December 2005, supplemented by REFERENCES from the selected articles. Thirty-four studies on men and women, for a total of 1 015 835 subjects and 94 533 deaths, were selected. Data were pooled with a weighed regression analysis of fractional polynomials.

Results  A J-shaped relationship between alcohol and total mortality was confirmed in adjusted studies, in both men and women. Consumption of alcohol, up to 4 drinks per day in men and 2 drinks per day in women, was inversely associated with total mortality, maximum protection being 18% in women (99% confidence interval, 13%-22%) and 17% in men (99% confidence interval, 15%-19%). Higher doses of alcohol were associated with increased mortality. The inverse association in women disappeared at doses lower than in men. When adjusted and unadjusted data were compared, the maximum protection was only reduced from 19% to 16%. The degree of association in men was lower in the United States than in Europe.

Conclusions  Low levels of alcohol intake (1-2 drinks per day for women and 2-4 drinks per day for men) are inversely associated with total mortality in both men and women. Our findings, while confirming the hazards of excess drinking, indicate potential windows of alcohol intake that may confer a net beneficial effect of moderate drinking, at least in terms of survival.

An inverse association between moderate alcohol consumption and coronary heart disease has been shown in observational studies.^1-3 Mechanisms supporting this association include increased high-density lipoprotein cholesterol level and fibrinolysis, decreased platelet aggregation and coagulation factors,⁴ and beneficial effects on endothelial function and inflammation.⁵ Nonetheless, abuse of alcohol is unquestionably harmful.^6-8 As a consequence, strong interest exists about the possibility that at any dose, the benefit of alcohol can overcome its harmful effects.^7,8

The relationship between alcohol and mortality has been depicted as a J-shaped curve, attributed to a combination of beneficial and harmful effects.^9-11 Indeed, if low alcohol intake is inversely related to coronary heart disease, the other side of the coin shows an increased risk of certain cancers, cirrhosis, and death from accidents associated with increased alcohol consumption.⁶

Moreover, whether alcohol has a different role in men and women is still debated. In previous studies, a similar inverse association of low doses of alcohol with cardiovascular disease in men and women was noted,^7,12 but a meta-analysis of total mortality revealed a J-shaped relation only for men older than 34 years or women older than 54 years.⁹ Herein, we perform an updated meta-analysis of prospective studies to investigate the relationship between alcohol dosing and all-cause mortality, separately in men and women.

A PubMed search (www.pubmed.gov) identified studies published until December 2005, without a start date; studies were restricted to humans, and their titles and/or abstracts contained at least 1 of the following terms: alcohol, beer, wine, or spirits plus the term mortality or death. Assessment of the REFERENCES was also conducted. Seventy-three publications were identified. Studies were excluded if they considered only 1 category of risk (n = 4) or did not report mortality separately for the sexes (n = 5); if they considered mortality for specific causes (n = 3) or if they comprised multiple reports (n = 9) (the longer follow-up was considered); or if the reference category was not the one with the lowest alcohol intake (n = 4) or if relative risks or numbers of cases and person-years were not available (n = 14). A total of 34 reports were identified.^13-46 Fourteen studies^{13,15-18,21-23,26,30,33,34,36,38-40} reported results separately for the sexes; 1 study³⁷ reported data for 2 age groups, and 1 study²⁹ for wine and beer. These studies contributed 2 dose-response curves each. Two studies contributed 4 curves: 1 study³⁰ reported results separately for 2 ethnic groups and sexes, and another³⁶ for age groups and sexes. In the end, 56 independent curves were available for meta-analysis, 37 in men and 19 in women.

Whenever possible, adjusted relative risks were extracted; otherwise, crude relative risks and 95% confidence intervals (CIs) were calculated from the number of events.^{16,22,31,38,42}

The amount of a drink was taken as quantified by each author whenever possible; otherwise (7 studies) it was considered equivalent to 10 g of ethanol; considering a drink equivalent to either 12 or 14 g of ethanol did not change our results (data not shown).

Data collected were as follows: (1) the value x of alcohol intake (measured in grams per day) assigned as the midpoint of the reported ranges; x was defined as 1.2 times the lower boundary for the open-ended upper categories⁴⁷; (2) frequency counts, adjusted relative risks, and 95% CIs for each x level; and (3) covariates describing the characteristics of the study. Inverse variance–weighted methods, taking into account the correlation between estimates within each study, were used.⁴⁷ The models to be fitted were selected among fractional polynomial curves of the second order.⁴⁸ Fractional polynomials are a family of models considering power transformations of a continuous exposure variable, restricted to a predefined set of integer and noninteger exponents.⁴⁹ The regression models were log(relative risk [RR]) = β₁x^p + β₂x^q and the exponents p and q were selected among the following set: {−2.0, −1.0, −0.5, 0.0, 0.5, 1.0, 2.0}. When p = 0, x^p is replaced by log(x), and when p = q, the model becomes log(RR|x) = β₁x^p+β₂x^q log(x).¹¹ These choices account for a rich set of possible functions, including J-shaped relations and conventional polynomials. The best fit was defined as that with the highest likelihood. This method assumes that the only source of variability is within study (fixed-effects model). To consider differences among studies as a further source of random variability, an additional component of the variance was added in weighing each observation (random-effects model).⁵⁰ The effects of covariates were evaluated, including appropriate interaction terms between covariate and exposure.⁴⁷ To make some allowance for multiple comparisons, 99% CIs were used in subgroup analyses, and pairwise contrasts were adjusted following the Sidak method, as outlined by Ludbrook.⁵¹ All analyses were carried out using an SAS macro,¹¹ version 8.12 for Windows (SAS Institute Inc, Cary, NC).

Thirty-four studies provided 56 independent dose-response curves for alcohol intake and mortality, for a total of 1 015 835 subjects and 94 533 deaths from any cause. Thirty-seven curves (705 596 subjects and 78 592 deaths) concerned men, and 19 (310 239 subjects and 15 941 deaths), women. Characteristics of the studies are listed in Table 1. Briefly, 28 curves were from European, 17 from American, and 11 from other populations. Almost half of the studies had a median follow-up longer than 10 years. A category of no alcohol intake that also excluded former drinkers was considered as the reference category for most curves (n = 30); the others curves used as reference those subjects who declared either no alcohol use (n = 12) or occasional use (n = 14); former drinkers were not excluded from either group.

The best fitting model was obtained at p = q = 0.5, corresponding to the model log(RR) = β₁√x + β₂√x × log(x). The fitted parameters (SE) were β₁ = −0.1592 (0.0056) (P<.001) and β₂ = 0.0421 (0.0014) (P<.001) (Table 2 and Table 3). The relationship observed has to be interpreted as a J-shaped curve (Figure 1) since, after an initial decrease in mortality by increasing alcohol intake (a shape depending on the negative value of β₁), the curve reaches a plateau and reverts at higher amounts (β₂>0). The association with a lower mortality was apparent up to 42 g/d (about 4 drinks per day), and the lowest mortality was seen at 6 g/d, or about half a drink daily (RR, 0.81 [95% CI, 0.80-0.83]). The meta-regression analysis was repeated using random-effects models. The model with p = q = 0.5 was again selected; the curve was very similar (Figure 1), but as expected, CIs were larger. The deviances of fixed and random effects models fell from 879.4 to 154.0 (P<.001 for the difference), suggesting evidence of heterogeneity among studies. In the following analyses, the role of study characteristics in explaining the interstudy heterogeneity was explored. The model with p = q = 0.5 was consistently fitted.

Forty-eight curves (908 182 subjects and 86 941 deaths) were adjusted at least for age; among them, 28 were adjusted for social status too, and 10 for social status and dietary markers (Table 1). Figure 2A shows the pooled curves for different levels of adjustment. The difference was highly significant (P<.001), showing that part of the heterogeneity is amenable to adjustment. The association with lower mortality decreased in adjusted studies, maximum protection falling from 36% to 17% (Table 2 and Table 3), but it remained substantial and statistically significant. Adjustment for social status and dietary markers did not affect the results (Table 2 and Table 3 and Figure 2A). In addition, we compared adjusted with nonadjusted data derived from the same studies. Nonadjusted RRs were available for 34 curves. Using nonadjusted data, the pooled curve predicted a maximum protection of 19% (99% CI, 17%-21%) in comparison with 16% (99% CI, 14%-18%) obtained from adjusted data (Figure 2B). Further analyses were conducted on the 48 adjusted curves.

Among adjusted studies, 32 curves were on men and 16 on women. Overall, the curves for men and women were different (P<.001); in particular, β₂ was greater in women, whereas β₁ was not (Table 2 and Table 3). As a consequence, the protection was apparent up to 3 drinks per day in men but only up to 2 drinks per day in women (Table 2 and Table 3 and Figure 3A); on the contrary, the maximum risk reduction was similar in men (17%; 99% CI, 15%-19%) and women (18%; 99% CI, 13%-22%). Thirteen studies provided separate curves for men and women recruited from the same population; the pooled curves for men and women were different for the range at which alcohol remained protective but comparable regarding the maximum protection (data not shown).

In women, pooled curves obtained using data from the United States or Europe or other countries (Australia, Japan, and/or China) were comparable (Figure 3B and Table 2 and Table 3) (P>.54 for differences between countries). In contrast, strong differences were observed in men (Figure 3C and Table 2 and Table 3) (P<.003 for each pairwise comparison, showing that part of heterogeneity in men is attributable to the set of the study). In particular, maximum risk reduction was in the range 20% to 28% in European but 14% to 19% in US studies, and the protection extended up to 6 drinks per day in European but only up to 3 drinks per day in US studies.

Further subgrouping of US data according to ethnicity provided no evidence of heterogeneity (data not shown).

In studies that used as reference the category of no alcohol intake and excluded former drinkers, the protection was significantly lower (P<.001) (Figure 4A and Table 2 and Table 3).

The pooled curves were similar in larger and smaller studies (Figure 4B and Table 2 and Table 3) (P = .61).

In studies published before 1998 or with follow-up longer than 10 years, the protection was slightly but significantly lower (Figure 4C and D and Table 2 and Table 3) (P<.001).

In this updated meta-analysis of 34 prospective studies, findings were pooled from more than 1 million subjects and almost 100 000 deaths from any cause. We observed a J-shaped relationship between total mortality and alcohol intake, showing that a low level of alcohol consumption is significantly associated with reduced total mortality, while high-level consumption is associated with increased mortality. Our meta-analysis, including 10 articles published after 2000 that could not be considered in former meta-analyses,^9-11 also took advantage of a novel approach recently developed by a team of researchers, including one of us (V.B.).¹¹ Special attention was paid to differences between sexes and to the role of confounding.

The dose-response curves are similar for both sexes when alcohol intake is light, but they differ with heavier intake; in fact, the inverse association in women apparently disappears at doses lower than in men, in agreement with previous findings.¹⁰ Women are more exposed than men to death for any cause at moderate to high levels of alcohol consumption, probably owing to increasing risk of cancer.⁶ Experimental evidence shows that when men and women consume the same amount of alcohol, women experience higher blood alcohol concentrations. Women metabolize ethanol differently and have a lower gastric alcohol dehydrogenase activity, resulting in higher blood ethanol levels and higher risk of liver disease.⁵² Finally, because premenopausal women have a low incidence of cardiovascular disease, the benefits of alcohol on total mortality may appear to be reduced.

The degree of association was lower in adjusted studies, as might be expected in view of several confounding factors characterizing observational studies on drinking habits^3,8; however, the benefit of light to moderate drinking remained in a range of undoubted public health value (15%-18%). Although residual confounding cannot be excluded,⁸ it would be very unlikely to modify the scenario in a substantial manner. We found indeed that when adjusted and unadjusted data derived from the same studies were compared, the maximum protection conferred by light to moderate drinking only decreased from 19% to 16%; we can thus presume that, even in the pessimistic hypothesis that residual confounding would have the same strength in lowering the protection as that of known confounding, the “real” (maximum) protection against total mortality associated with low levels of alcohol consumption would still be higher than 10%. A similar reasoning would also apply to the harm associated with heavier drinking.

The relationship with mortality appears to be lower in US-based than in European studies, but only in men. Such a difference has no obvious explanation. Would women living in either continent follow more comparable drinking habits than men? Would it be linked to the lower amounts of alcohol consumed more regularly by women all over the world? Both the amount and the pattern of drinking is important for the effect of alcohol⁵³; in particular, drinking wine at meals (a typical habit in European Mediterranean areas) enhances the ability of wine to prevent the development of atheromatous lesions.⁵⁴ In contrast, binge or irregular drinking, most likely exerted by men, is an unhealthy habit.⁵⁵ Moreover, influence of genetics on the effect of ethanol has also been described.⁵⁶ Possible differences in pattern of drinking and/or in the distribution of genetic factors might explain different results observed in studies conducted in American or European men.

The results of any meta-analysis may be plagued by publication bias; nevertheless, we considered only follow-up studies on total mortality, and it is hard to hypothesize that high-quality studies would not have been published because they reported negative results. We believe therefore that publication bias—if any—might have only weakly altered our findings.

Underreporting of alcohol consumption would result in a tendency for RRs to be biased toward the null hypothesis, and this may have distorted the shape of the J-curve and the apparent threshold for harm.

The selection of nondrinkers as a reference group has been questioned because this group may include ex-drinkers who stopped drinking because of health problems.^3,57 A subgroup analysis restricted to studies that excluded either ex-drinkers or very light drinkers from the reference group generated a pooled curve that indeed predicted a lower (though statistically significant) protection, confirming the importance of properly selecting the reference group in studies on alcohol and health.^3,10,57

Duration of the follow-up and year of publication have been identified as other sources of heterogeneity. However, stratification analyses for these characteristics resulted in pooled curves that consistently predicted a substantial reduction in total mortality, within comparable alcohol dose ranges.

Randomized controlled trials offer a more solid answer than observational studies to many questions in medicine, mainly restricted, however, to the efficacy of drugs; controlled intervention trials on diet in general and on alcohol in particular are difficult and ethically questionable to perform.^7,57 One has therefore to rely on observational studies such as those analyzed here or prospective studies where participants spontaneously decrease alcohol consumption or stop drinking altogether. Interestingly, the first study of the latter type⁵⁸ supports the inverse relation of moderate alcohol intake with coronary heart disease.

In conclusion, this meta-analysis confirms the hazards of excess drinking but also indicates the existence of potential windows of alcohol intake that may confer a net beneficial effect of drinking, at least in terms of survival, both in men and in women. Heavy drinkers should be urged to cut their consumption, but people who already regularly consume low to moderate amounts of alcohol should be encouraged to continue.

Correspondence: Licia Iacoviello, MD, PhD, Laboratory of Genetic and Environmental Epidemiology, “John Paul II” Center for High Technology Research and Education in Biomedical Sciences, Catholic University, Largo Gemelli 1, 86100 Campobasso, Italy (licia.iacoviello@rm.unicatt.it).

Accepted for Publication: August 30, 2006.

Author Contributions: All authors have read the final version and approved submission of the manuscript. Study concept and design: Di Castelnuovo, Donati, Iacoviello, and de Gaetano. Acquisition of data: Di Castelnuovo and Costanzo. Analysis and interpretation of data: Di Castelnuovo, Costanzo, Bagnardi, and Iacoviello. Drafting of the manuscript: Di Castelnuovo, Costanzo, and de Gaetano. Critical revision of the manuscript for important intellectual content: Bagnardi, Donati, Iacoviello, and de Gaetano. Statistical analysis: Di Castelnuovo and Bagnardi. Obtained funding: Donati. Administrative, technical, and material support: Costanzo. Study supervision: Donati, Iacoviello, and de Gaetano.

Financial Disclosure: None reported.

Funding/Support: Research for this article was funded by the Italian Ministry of University, Research, Education (MIUR), Decreto No. 1588-19/11/2004.

Role of the Sponsor: The sponsor of the study had no involvement in study design; data collection, analysis, or interpretation; writing of the report; or in the decision to submit the paper for publication.