A review of the relationship between dimensions of alcohol consumption and the burden of disease: 2026 update including Mendelian randomisation studies

Abstract

Background and aims

Evidence on the causal impact and corresponding risk relationships between dimensions of alcohol consumption and health outcomes continues to evolve, with some contradictory findings across study designs. This review aimed to update current knowledge on causality and risk relationships to inform global and national comparative risk assessments for alcohol.

Methods

Fully alcohol‐attributable conditions were identified using International Classification of Diseases (ICD) 10th and 11th revision codes. We conducted a scoping review of meta‐analyses of cohort studies on average consumption and health outcomes (56 reviews), a systematic review of Mendelian randomisation (MR) studies on alcohol and ischaemic heart disease (IHD; 20 studies), and narrative syntheses on injuries, biological pathways, and reversibility of effects.

Results

ICD‐11 provides more detailed categories, defining 62 fully alcohol‐attributable conditions compared with 48 in ICD‐10. Meta‐analyses support monotonic increasing dose–response relationships between average consumption and most attributable health outcomes within infectious diseases, cardiovascular diseases, cancers, and digestive diseases. Relationships are J‐shaped for IHD, ischaemic stroke, and type 2 diabetes, with lower risk at low‐to‐moderate consumption (generally only without heavy episodic drinking). For dementia, heavy drinking is harmful and, among non‐heavy drinkers, relationships are age‐specific. MR evidence for IHD largely suggested null or harmful relationships, but only three studies tested non‐linear effects. In our view, the overall synthesis indicates that current MR evidence is insufficient to refute a J‐shaped relationship for IHD. Injury risk is driven primarily by acute intoxication and includes substantial harm to others. Acute risks are reversible with reductions in drinking or abstention, whereas many chronic disease processes are only partly reversible.

Conclusions

Epidemiological evidence to inform comparative risk assessments for alcohol is comprehensive, but prone to major limitations. Triangulation, alongside biological plausibility, can strengthen synthesis across cohort and Mendelian randomisation studies, and the target trial framework can help future studies avoid design‐induced biases. Comparative risk assessments with ischaemic heart disease should, at this point, prioritise evidence from cohort studies.

INTRODUCTION

Alcohol consumption is a major risk factor for global burden of disease, injury and premature death [1, 2] (for an overview, see Rehm and Imtiaz [3]) according to comparative risk assessments (CRAs) [4]. CRAs integrate population‐specific prevalence data on alcohol consumption and estimates of how specific dimensions of alcohol, such as average drinking level and frequency of heavy drinking episodes, relate to health risks. This integration yields an alcohol‐attributable fraction (AAF), defined as the proportion of a specific health condition within a population that is caused by alcohol. AAFs assume a counterfactual scenario, often defined by a scenario where alcohol is entirely absent [5], that is, if no one consumed alcohol.

CRAs are, therefore, based on the assumption that alcohol consumption causally impacts health risks. In the absence of long‐term randomised trials evaluating the causal effects of alcohol consumption on health outcomes, risk relationships used to calculate AAFs have primarily been estimated from observational data, particularly prospective cohort studies, which are combined via meta‐analyses [6]. Epidemiological research on alcohol's health impacts continues to evolve, increasingly benefiting from large‐scale cohort studies with extended follow‐up periods. However, alcohol exposure is still measured poorly in many cohorts, often only once at baseline and restricted to one dimension, mainly volume of drinking [7]. This is one of the reasons why findings from epidemiological research are often disputed.

Recently, a growing number of Mendelian randomisation (MR) studies, which use genetic variants as instrumental variables, with alcohol consumption as an exposure, have been published. These studies are theoretically less susceptible to the confounding biases inherent in other observational research [3] and have frequently reported risk estimates differing from those derived from cohort or case–control studies—often suggesting no associations, or even associations in opposite directions, with conditions associated with alcohol consumption in prior studies (e.g. Carr et al. [8] and van de Luitgaarden et al. [9]).

This review summarises the current evidence on the relationships between alcohol consumption and health outcomes in four sections, following a similar structure to the last three monographs published in Addiction [6, 10, 11]. First, we provide an overview of disease and injury categories that are fully attributable (100%) to alcohol consumption. Second, we summarise evidence for disease categories partially attributable to alcohol (i.e. conditions potentially caused by alcohol consumption, but also occurring independently), drawing primarily on cohort studies and, for ischaemic heart disease, also on MR studies. In this section, we also outline how evidence from cohort and MR studies can be integrated to inform CRAs, using ischaemic heart disease as an example. Third, we present an overview of injury categories partially attributable to alcohol consumption. Last, we summarise biological pathways and the potential reversibility of health effects of alcohol. Based on this overview, we will discuss implications for future research in alcohol epidemiology and for CRAs for alcohol.

METHODS

Terminology

Unless otherwise defined based on specific studies, we use the term ‘chronic heavy drinking’ to denote drinking at least 40 and 60 grams of pure alcohol per day for females and males, respectively. Heavy episodic drinking is used for the same sex‐specific amount per single occasion.

Identification of disease categories potentially causally related to alcohol

Disease and injury categories fully attributable to alcohol were identified using the World Health Organization (WHO) International Statistical Classification of Diseases and Related Health Problems, 10th (ICD‐10) and 11th (ICD‐11) revision databases, with the search term ‘alcohol*’. For disease categories causally related, but only partially attributable to alcohol, we started with the disease categories identified in the last Addiction monograph [6], which had been corroborated by WHO's Technical Advisory Group for Alcohol and Drug Epidemiology (TAG‐ADE) before the last Global Status Report on Alcohol and Health [1]. As part of this review, we re‐examine the past classification and report in more detail the newly selected disease outcomes and dose–response relationships discussed by the newly constituted WHO TAG‐ADE, which held its first meeting on 8 to 9 July 2025 [12].

Identification of the best meta‐analyses of cohort studies on alcohol consumption and health outcomes (potentially) partially attributable to alcohol

This study used the results of the Alcohol Intake and Health Study [13], which involved a systematic scoping review of recent systematic reviews and meta‐analyses of cohort studies estimating relationships between alcohol consumption and risks of conditions within selected disease categories (PROSPERO preregistration: CRD42024584948).

The systematic scoping review included mortality and morbidity from key disease conditions considered to be caused by alcohol consumption [6]. No systematic reviews to quantify the impact of alcohol consumption on human immunodeficiency virus (HIV) / acquired immunodeficiency syndrome (AIDS), other sexually transmitted diseases, cervical cancer or depression were found, although a causal relationship between alcohol consumption and the risk of these diseases and conditions has been established.

However, the available literature allowed to indirectly model the potential impact of alcohol on these disease categories via sexual behaviours, except for depression. For general considerations and the modelling approach based on HIV/AIDS refer to Rehm et al. [14]; for further considerations refer to Llamosas‐Falcón et al. and Morojele et al. [15, 16]. Cervical cancer falls under this category because it is primarily caused by persistent infection with human papilloma virus, which is transmitted through skin‐to‐skin sexual contact [17]. Depression, however, still lacks methodology to quantify the causal impact of alcohol while adequately accounting for potential bidirectionality and causation by other factors (see below).

Rather than conducting meta‐analyses ourselves, which we feared could be biased, especially in areas where there have been more than a dozen meta‐analyses, we decided to select independent experts for each area who could identify the best available meta‐analyses for the purpose of conducting CRAs and developing guidance documents. Expert panels were consulted to identify, for each disease category, the three most appropriate risk relationships from those reported in the identified systematic reviews and meta‐analyses [13]. Experts were identified by selecting authors with the highest number of first‐ or last‐author scientific publications in each disease category (cancer, cardiovascular diseases, digestive diseases, neurological disorders and infectious diseases) over the past 10 years, based on PubMed searches. Experts with a potential conflict of interest were excluded. Invited authors formed expert panels for their respective disease areas.

During expert panel consultations, each expert received systematic reviews and meta‐analyses identified from the systematic scoping review for their disease category and completed a questionnaire to select the most appropriate risk relationship for each condition within the category, justifying their choices. Based on these expert responses, the final risk relationships selected for inclusion in the Alcohol Intake and Health Study [13] were determined. The selected risk relationships were summarised separately by sex. If a chosen risk relationship was not sex‐specific, it was assumed to be the same for males and females.

Systematic review of Mendelian randomisation studies on alcohol consumption and ischaemic heart disease

We conducted a systematic review to identify MR studies estimating and reporting associations between (genetically predicted) alcohol consumption and ischaemic heart disease. We selected ischaemic heart disease as a case example because findings from cohort and MR studies have been largely contradictory and have fuelled ongoing debates about the reliability and unbiasedness of each approach for alcohol epidemiology. Full details of the systematic review are provided in the Supplementary Appendix.

We systematically searched Embase from inception to 28 February 2025 using a search string (Table S1). We included MR studies in adult populations that explicitly stated using MR and either estimated and reported (1) a gene–outcome association using a genetic variant shown to be associated with alcohol consumption (to test a sharp causal null hypothesis); or (2) MR estimates for genetically predicted alcohol consumption and the outcome of interest. We excluded conference abstracts, reviews, editorials and preprints, as well as MR estimates for composite outcomes that included ischaemic heart disease alongside other outcomes. No language or geographical restrictions were applied.

Two authors independently screened titles/abstracts and full texts in Covidence, resolving disagreements through discussion. We extracted study characteristics, definitions of alcohol consumption and outcomes, data sources and ancestry, genetic instruments, MR methods and effect estimates using a standardised template. We report primary MR estimates as presented in the original studies. When multiple methods were used, or no primary method was specified, we report inverse‐variance weighted estimates. We also report multivariable and non‐linear MR results when available. Risk of bias was assessed using a tool adapted from Jabeen et al. [18] (Table S2), which ranks studies based on vulnerability to weak instrument bias, genetic and other confounding biases, horizontal pleiotropy and participant selection bias.

Both the systematic review of MR studies and the systematic scoping review of meta‐analyses of cohort studies followed Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) guidelines [19] (Tables S3 and S4) (please see the corresponding publication of the Alcohol Intake and Health Study [20] for the PRISMA checklist of the scoping review).

Integration of evidence from traditional epidemiological and Mendelian randomisation studies for comparative risk assessments

In a viewpoint, we discuss how evidence from cohort and MR studies can be integrated to inform CRAs, using ischaemic heart disease as a case study, and summarise key design‐related biases and potential strategies for interpreting and strengthening causal inference from current and future studies.

RESULTS

Disease and injury categories fully (100%) attributable to alcohol

Table S5 lists all identified conditions and injuries that are fully (100%) attributable to alcohol consumption, as defined in ICD‐10 (48 codes, counting only the most granular level listed) and ICD‐11 (62 codes). These fully alcohol‐attributable codes are mainly in the categories of non‐communicable diseases and injuries, with disease categories in the following major chapters (most prevalent disease categories in parentheses): endocrine, nutritional and metabolic disease, mental and behavioural disorders (with most ICD codes in this chapter centred around alcohol dependence and other use disorders), diseases of the nervous system, diseases of the circulatory system (alcoholic cardiomyopathy), diseases of the digestive system (alcoholic liver diseases such as cirrhosis), diseases related to pregnancy and the perinatal period (fetal alcohol syndrome and fetal alcohol spectrum disorders, denoting conditions of alcohol consumption that harm others rather than the drinker) and external causes and resulting injuries (alcohol poisoning).

Most fully alcohol‐attributable diseases require heavy drinking, either occasionally (e.g. alcohol poisoning) or chronically (most chronic disease categories linked to alcohol). Overall, ICD‐11 offers more detailed categories for the core categories of alcohol dependence and other use disorders (within the category 6C40) (see Table S5), but the concordance with ICD‐10 of broader categories of alcohol use disorders was still high in first tests [21].

The most important change was that the diagnosis of harmful use of alcohol (either episodic or chronic—or ‘sustained’ in ICD terminology) in ICD‐11 now includes harm to the health of others [22] acknowledging the fact that alcohol‐attributable health harm is not restricted to the drinker. It will remain to be seen whether this diagnosis will actually be used in clinical practice. Furthermore, its validity when assessed by self‐report is questionable [21].

Diseases and injuries that alcohol consumption may affect: evidence from cohort and other epidemiological studies, except Mendelian randomisation studies

A total of 7294 records were identified in the systematic scoping review (Figure S1) of systematic reviews and meta‐analyses evaluating the relationship between alcohol consumption and specific disease categories based on data from cohort studies, after duplicate removal. Of these, 7003 were excluded at the title and abstract screening stage, and an additional 235 were excluded following full‐text review. Fifty‐six unique systematic reviews and meta‐analyses were included: four focused on infectious diseases, 24 on cancer, 19 on cardiovascular diseases, five on digestive diseases, two on diabetes mellitus and two on epilepsy. As described in the Methods section above, the final meta‐analysis to be used in CRAs for each disease outcome was derived based on ratings from an expert panel.

The main results are summarised in Table 1 and in the following sections. The dose–response relationships estimated in the meta‐analyses selected by expert panels are shown in Figure 1.

TABLE 1.

Potentially alcohol‐attributable broad disease categories—Epidemiological indicators.

Disease category	ICD‐10 codes for cause of death a	ICD‐11 codes for cause of death	Causality and reference to meta‐analyses/selected systematic reviews	Effect
Infectious diseases
Tuberculosis	A10‐A14, A15‐A19.9, B90‐B90.9, K67.3, K93.0, M49.0, P37.0	1B10.0, 1B10.1, 1B10.Z, 1B11.0, 1B11.2, 1B11.Z, 1B12.Y, 1B12.40, 1B12.5–1B12.8, 1B12.1–1B12.3, XA0NE9/5A74.0, 1B12.Y, 1B13.Z, 1B13.0, 1B13.1, 1B13.Z, 1G80, FB0Z, KA61.0	Causality: Rehm et al. [23] Meta‐analysis: Simou et al. [24]	Detrimental
HIV/AIDS	B20‐B24.9	1C62.Z, 1C62.1, 1C62.3Y, 1C62.2, 1C62.0, 1C62	Causality: Rehm et al. [14]; Williams et al. [25]; Morojele et al. [16] CRA calculations will be based on: Rehm et al. [14]	Detrimental
Other sexually transmitted diseases	A50–A58, A60–A60.9, A63–A63.8, B63, I98.0, K67.0–K67.2, M03.1, M73.0–M73.1, N70–N71.9, N73–N74.8	1A60.Z, 1A60.0, 1A60.1, 1A60.2, 1A60.3, 1A60.5, 1A61.Z, 1A61.0, 1A61.1‐1A61.5, 1A62.Z, 1A62.1, 1A62.01, 1A62.00, 1A62.0Z, 1A62.2Z, 1A62.Y, 1A6Z, 1A63, 1A7Z, 1A70.0Z, 1A70.1, 1A71, 1A72.4, 1A72.0, 1A72.3, 1A73, 1A80, 1A81.Z, 1A81.0, 1A81.1, 1A81.Y, 1A90, 1A91, 1A94.Z, 1A94.0, 1A94.1, 1A9Z, 1A95.Z, BE2Y, DC5Z, 1C21, 1A62.21/FB50.0, GA07.Z, GA07.0, GA07.1, GA01.Z, GA01.0Z, GA01.1Z, GA05.Z, GA05.0, GA05.1, GA05.2, GA06, 1B12.5/GA04, 1B12.5, 1A71/GA05.Z, 1A81.1/GA05.Z	Causality: Cook and Clark [26]; Llamosas‐Falcón et al. [15] CRA calculations: the behavioural causal pathway via alcohol's impact on decision‐making is the same as for HIV/AIDS [14, 25], so we suggest using the same alcohol‐attributable fractions as for HIV/AIDS, but without the effect of alcohol consumption on mortality via medication non‐adherence	Detrimental
Lower respiratory infections: pneumonia	A48.1, A70, J09–J15.8, J16‐J16.9, J20–J21.9, P23.0–P23.4	1C19.Z, 1C22, 1E31, 1E30, 1E32, CA40.1Z, CA40.00‐CA40.08, CA40.10–CA40.13, CA40.0Z, CA40.0Y, CA40.Z, CA42.Z, CA42.Y&XN4NV, CA42.0–CA42.5, CA42.Y&XN2TU, CA41.Z, CA41.0, CA41.Y&XN513, KB24	Causality: Samokhvalov et al. [27]; Traphagen et al. [28], for heavy drinking und alcohol use disorders: Simet and Sisson [29] Meta‐analysis: Simou et al. [30]	Detrimental
Cancers
Lip and oral cavity cancer	C0‐C08.9, D00.00–D00.07, D10.0–D10.5, D11–D11.9, D37.01–D37.04, D37.09 c	2B60.Z, 2B6Y, 2B61.Z, 2B61.Z&XA0HQ3, 2B62.Z, 2B62.1Z, 2B63.Z, 2B64.Z, 2B65.Z, 2B66.Z, 2B67.Z, 2B68.Z, 2E60.0, 2E90.0, 2E90.1, 2E90.2, 2E90.3, 2E90.4, 2E90.5, 2E91.Z, 2E91.0, 2E91.1, D37.0	Causality: IARC [31, 32]: sufficient evidence for carcinogenicity in humans b Meta‐analysis: World Cancer Research Fund International [33]	Detrimental
Other pharynx cancer	C09–C10.9, C12–C13.9, D10.7 c	2B69.Z, 2B6A.Z, 2B6C.Z, 2B6D.Z, 2E90.7	Causality: IARC [31, 32]: sufficient evidence for carcinogenicity in humans b Meta‐analysis: World Cancer Research Fund International [33]	Detrimental
Oesophagus cancer	C15–C15.9, D00.1, D13.0 c	2B70.Z, 2E60.1, 2E92.0	Causality: IARC [31, 32]: sufficient evidence for carcinogenicity in humans b Meta‐analysis: World Cancer Research Fund International [34]	Detrimental
Colon and rectum cancer	C18–C21.9, D01.0–D01.3, D12‐D12.9, D37.3–D37.5 c	2B90.Z, 2B90.0Z&XA6J68, 2B81.Z, 2B90.0Z, 2B90.2Z, 2B90.1Z, 2B90.3Z, 2B91.Z, 2B92.Z, 2C00.Z, 2E61.0, 2E61.0&XA33J5, 2E61.1, 2E61.2, 2E92.Z, 2E92.4Y, 2E92.4Z, 2E92.5	Causality: IARC [31, 32]: sufficient evidence for carcinogenicity in humans b Meta‐analysis: Jun et al. [35]	Detrimental
Liver cancer	C22–C22.9, D13.4 c	2C12.1, 2C12.02, 2C12.10, 2C12.01, 2B56.3, 2B5F.2, 2C12.0Z, 2E92.7	Causality: IARC [31, 32]: sufficient evidence for carcinogenicity in humans b Meta‐analysis: World Cancer Research Fund International [36]	Detrimental
Larynx cancer	C32–C32.9, D02.0, D14.1, D38.0 c	2C23.Z, 2C23.1Y, 2C23.2Y, 2C23.3Y, 2C23.4, 2C23.5, 2E62.0, 2F00.Z	Causality: IARC [31, 32]: sufficient evidence for carcinogenicity in humans b Meta‐analysis: World Cancer Research Fund International [33]	Detrimental
Female breast cancer	C50–C50.929, D05–D05.92, D24–D24.9, D48.6‐D48.62, D49.3, N60–N60.99 c	2C6Z, 2E65.Z, 2E65.0, 2E65.2, 2F30.Z	Causality: IARC [31, 32]: sufficient evidence for carcinogenicity in humans b Meta‐analysis: Sohi et al. [37]	Detrimental
Cervix uteri cancer	C53	2C77.Z	Causality: Rehm et al. [14]; Williams et al. [25]; Morojele et al. [16] CRA calculations will be based on: Rehm et al. [14]	Detrimental
Diabetes mellitus
Diabetes mellitus	E10–E10.11, E10.3–E11.1, E11.3–E12.1, E12.3–E13.11, E13.3–E14.1, E14.3–E14.9, P70.0–P70.2, R73–R73.9	5A10, 5A11, 5A12, 5A13.Y, 5A14, KB60.0, KB60.1, KB60.2Z, MA18.0Z, MA18.00	Causality: Howard et al. [38]; Bonnet et al. [39] Meta‐analysis: Llamosas‐Falcón et al. [40]	Beneficial or detrimental, depending on level and patterns of drinking as well as the population
Neuropsychiatric disorders
Alzheimer's disease and other dementias	F00–F03.91, G30–G31.1, G31.8–G31.9	6D80.Z, 6D80.0, 6D80.1, 6D81, 6D8Z, 6D83, 6D85.5, 6D85.1, 6D85.0, 6D85.3, 8A20, 8A2Z, 8E7Y, MB21.0	Causality: Pervin and Stephen [41]; Rehm et al. [42] CRA calculations will be based on Lancet commission on dementia [43, 44]	Detrimental for heavy drinking
Epilepsy	G40–G41.9	8A6Z, 8A61.Z, 8A61.4Z, 8A60.Z, 8A68.4, 8A66.Z, 8A66.0, 8A66.1Z, 8A67	Causality: Bartolomei [45]; Barclay et al. [46]; Leach et al. [47] Meta‐analysis: Woo et al. [48]	Detrimental
Cardiovascular diseases
Hypertensive heart disease	I11–I11.9	BA01, BD1Z	Causality: Puddey and Beilin [49]; O'Keefe et al. [50]; in addition, we have good evidence that interventions leading to reductions of alcohol consumption subsequently lead to reductions in blood pressure and hypertension; Xin et al. [51]; Roerecke et al. [52] Meta‐analysis: Cecchini et al. [53]	Detrimental
Ischaemic heart disease	I20–I25.9	BA40.Z, BA40.0, BA85.Z, BA41.Z, BA42.Z, BA60.Z, BB24, BC40.0, BA60.3, BA60.4, BA60.6, BA60.7, BA4Z, BA43, BA60.0, BA5Z, BA52.Z, BA50, BA5Y, BA51.Z	Causality: Mukamal and Rimm [54]; Collins et al. [55]; Roerecke and Rehm [56]; Lee et al. [57] Meta‐analysis: Zhao et al. [58]	Beneficial or detrimental, depending on level and patterns of drinking
Cardiomyopathy	A39.52, B33.2–B33.24, D86.85, I40–I43.9, I51.4–I51.5	1D85.Z, BC42.Z, BC42.1, BC43.Z, BC43.0Z, BC43.12, BC43.1Z, BC43.20, BC43.3, BC43.01, BC43.4, BC43.Y, BC4Z	Causality: Iacovoni et al. [59]; George and Figueredo [60]; Rehm et al. [61]; Lee et al. [57] No meta‐analyses were identified. There is a separate category for alcoholic cardiomyopathy, which is responsible for 3%–40% of all cardiomyopathies [59]. CRA calculations will be based on: Manthey et al. [62]	Detrimental
Atrial fibrillation and flutter	I48–I48.92	BC81.Z, BC81.30, BC81.31, BC81.32, BC81.20	Causality: Rosenqvist [63]; Rosenberg and Mukamal [64]; Lee et al. [57] Meta‐analysis: Jiang et al. [65]	Detrimental
Ischaemic stroke	G45–G46.8, I63–I63.9, I65–I66.9, I67.2–I67.3, I67.5–I67.6, I69.3–I69.398	8B10.Z, 8B10.Y, 8B22.Y, 8B10.0, MB21.12, 8B26.Z, 8B26.2, 8B26.3, 8B26.4, 8B26.0, 8B26.1, 8B26.50, 8B26.51, 8B26.5Z, 8B11.3, 8B11.0, 8B11.2Z, 8B11.50, 8B11.1, 8B11.51, BD55, BD55&XA2K99, 8B2Z, BD55&XA2JH8, BD55&XA1VB0, BD55&XA7C50, 8B22.B, 8B22.1, 8B25.0	Causality: Puddey et al. [66]; Mazzaglia et al. [67]; Collins et al. [55]; Lee et al. [57] Meta‐analysis: Larsson et al. [68]	Beneficial or detrimental, depending on level and patterns of drinking (similar to ischaemic heart disease)
Haemorrhagic and other non‐ischaemic stroke	I60‐I61.9, I62.0–I62.03, I67.0–I67.1, I68.1‐I68.2, I69.0–I69.298	8B01.2, 8B01.0, 8B00.Z, 8B00.0, 8B00.1, 8B00.2, 8B00.3, 8B00.4, 8B00.5, 8B02, 8B22.0, 8B22.5, 8B22.7Y, 8B23, 8B25.2, 8B25.1, 8B25.3	Causality: Puddey et al. [66]; Mazzaglia et al. [67]; Lee et al. [57] Meta‐analysis: Zhang et al. [69]	Detrimental
Gastrointestinal diseases
Cirrhosis of the liver	B18–B18.9, I85–I85.9, I98.2, K70–K70.9, K71.3–K71.51, K71.7, K72.1–K74.69, K74.9, K75.8–K76.0, K76.6–K76.7, K76.9	1E51.Z, 1E51.2, 1E51.0Z, 1E51.1, 1E51.Y, DA26.01, DA26.0Z, DA26.00, DB94.Z, DB94.0, DB94.1Z, DB94.2, DB94.3, DB95.1Z, DB95.5, DB99.8, DB99.7, DB97.2, DB93.Y, DB93.0, DB96.1Z, DB93.1, DB97.Z, DB9Z, DB92.Z, DB98.7Z, DB99.2	Causality: a causal impact of alcohol is by definition, as there are alcoholic sub‐categories for many liver diseases in the ICD (see Table S5); pathogenesis: Gao and Bataller [70] Meta‐analyses: Llamosas‐Falcon et al. [71] (sex‐specific); Llamosas‐Falcon et al. [72]	Detrimental
Pancreatitis	K85–K86.9	DC31.Z, DC31.0, DC31.2, DC31.1, DC31.3, DC3Z, DC32.3, DC32.Z, DC30.0, DC30.1	Causality: specific impact assessment not necessary, as there are two conditions of pancreatitis which are 100% alcohol‐attributable (see Table S5); for pathogenesis: Braganza et al. [73]; Yadav et al. [74]; Lankisch et al. [75]; Majumder and Chari [76] Meta‐analysis selected: Samokhvalov et al. [77]	Detrimental

Abbreviations: AIDS, acquired immune deficiency syndrome; CRA, comparative risk assessment; HIV = human immunodeficiency virus; IARC, International Agency for Research on Cancer; ICD, International Statistical Classification of Diseases and Related Health Problems.

^aICD codes for non‐fatal disease outcomes are slightly different than those used in the Global Burden of Diseases, Risk Factors and Injuries study, but for this overview table, we did not want to introduce this distinction (for the respective ICD codes used in the Global Burden of Diseases, Injuries, and Risk Factors study, see [78).

^bFor definitions, see International Agency for Research on Cancer IARC monographs on the evaluation of carcinogenic risks to humans Lyon, France International Agency for Research on Cancer 2016 [79].

^cThe risk relationships between alcohol consumption and the respective cancer sites are based on studies with ICD‐10 C codes; the D codes were only listed as we wanted them to be compatible with the Global Burden of Diseases, Injuries, and Risk Factors study.

FIGURE 1.

Relationships between average volume of alcohol consumption (grams per day) and risk of health outcomes, based on the best available meta‐analyses of cohort studies selected by expert panels as well as other relevant studies (see Table 1 for meta‐analyses and text for other relevant studies). AIDS, acquired immune deficiency syndrome; HED, heavy episodic drinking; HIV = human immunodeficiency virus; unint. = unintentional.

Infectious disease

The causal impact of alcohol on the four broad categories of infectious disease from the last Global Status Report on Alcohol and Health [1] has been confirmed in recent reviews [15, 16]. The main biological mechanism involves alcohol‐induced liver dysfunction, which disrupts both non‐specific innate and adaptive immune responses through acute and chronic alcohol consumption [80, 81, 82, 83, 84]. Lowered immune responses increase susceptibility to communicable diseases. More specific biological pathways are described in the references listed in Table 1.

Behavioural pathways for HIV and other sexually transmitted diseases involve impaired decision‐making regarding sexual behaviours, which has been demonstrated experimentally [14, 25]. It is worth noting that the estimated AAF for HIV may be overestimated in current CRAs because they do not account for the impact of newer medications reducing the viral load and, consequently, the risk of infectious disease transmission.

Non‐communicable disease

There are five broad categories of non‐communicable diseases related to alcohol consumption, each with different biological pathways: cancer, cardiovascular disease, diabetes mellitus, neuropsychiatric disease and gastrointestinal disease.

Cancer

Alcohol has long been established as a major carcinogen and ranks as the second or third leading risk factor for cancer in many countries and regions (e.g. Soerjomataram et al. [85], Islami et al. [86] and Fink et al. [87]). The International Agency for Research on Cancer (IARC) determined in 2010 and 2012 that alcohol consumption causally impacts seven cancer sites [31, 32], which have been included in all CRAs.

The biological pathways of the impact of alcohol consumption on cancer vary by cancer site. The most important pathways include (see Rumgay et al. [89] and United States Department of Health and Human Services: Office of the US Surgeon General Alcohol and Cancer risk [88]):

1. When alcohol is metabolised in the human body, it breaks down into acetaldehyde, a compound that damages DNA in multiple ways, therefore, increasing the risk of cancer.

2. Alcohol induces oxidative stress, causing DNA, protein and cell damage as well as increasing inflammation, which are all mechanisms that elevate cancer risk.

3. Alcohol alters levels of multiple hormones, including oestrogen, which contributes to increased breast cancer risk [37].

4. Alcohol causes greater absorption of other carcinogens, such as those from cigarette smoke.

Although the causal impact of alcohol on all seven cancer sites determined by IARC has been confirmed in recent reviews and meta‐analyses (Table 1), there are clear indications that additional cancer sites may also be causally impacted by alcohol consumption [89]: first, several meta‐analyses find stable dose–response relationships between alcohol and cancers of the stomach [90], biliary tract [91], pancreas [92] and lung [92], although the risk for lung cancer almost disappeared when estimated in never smokers [93]. Moreover, large hospital‐based cohort studies found stable and harmful associations between alcohol dependence and these additional cancer sites, as well as a reduction of risk for these cancer sites after alcohol treatment interventions [94]. At this point, only the systematic evaluation of causality by IARC is needed to decide which cancer sites should be included in CRAs for alcohol, and it should be undertaken urgently for all sites with stable epidemiological associations that have not yet been included.

Diabetes mellitus

The relationship between alcohol and diabetes mellitus (Type 2) has been estimated in various meta‐analyses, yielding mixed results (e.g. Llamosas‐Falcón et al. [40], Baliunas et al. [95] and Knott et al. 96]). Although all found some evidence for protective effects, more recent meta‐analyses suggest that this benefit may be restricted to females, particularly those who are overweight or obese (body mass index of 25 kg/m² or greater) [40]. The impact of non‐heavy alcohol consumption with enhanced insulin sensitivity, reduced basal insulin secretion rate and lower fasting plasma glucagon concentration in females seems to indicate a plausible biological pathway for this sex‐specific effect [39, 97].

Neuropsychiatric disease

Even though alcohol consumption and alcohol use disorders are associated with most neuropsychiatric disease categories (e.g. Puddephatt et al. [98]), it is rarely established that these associations reflect causal effects and quantifying such effects is fraught with challenges. As a result, the most recent CRA of the WHO [1] only includes two conditions: alcohol use disorders (i.e. alcohol dependence and harmful use of alcohol), where causality is inherent by definition (see above), and epilepsy (see Table 1). Recent systematic reviews and meta‐analyses support the inclusion of epilepsy (e.g. Woo et al. [48]) into WHO's CRA.

There is now enough evidence to also include dementia in the next CRA of the WHO [12] or elsewhere. Almost all systematic reviews and meta‐analyses concur that heavy drinking is associated with increased risks of cognitive decline and various forms of dementia [42, 43, 44, 99]. This is further corroborated by a large population‐based French study of all hospital patients followed for several years [100]. Because hospital data typically lack self‐reports on drinking, this study defined diagnosis of alcohol dependence as the main exposure, which may serve as a proxy for heavy drinking [101]. Overall, alcohol dependence was strongly associated with all forms of dementia, including, but not limited to, alcohol‐related brain disease and the majority of early‐onset dementia (defined as onset before 65 years). In this cohort, the relative risk of 3 associated with alcohol dependence/heavy drinking exceeds all relative risks of all other risk factors for dementia identified by the Lancet commission [43]. There are well‐established pathways from heavy drinking to dementia, such as neuroinflammation, oxidative stress, impaired cholinergic function and damage to the blood–brain barrier [41]. The detrimental effects of heavy drinking on cardiovascular diseases (discussed below) also apply to vascular dementia.

For non‐heavy drinking, evidence on dementia risk is less consistent. Observational studies suggest harmful effects in younger adults, whereas in older adults, low to moderate consumption may confer potential benefits—as reflected by a J‐shaped relationship [42, 99]. Studies on potential mechanisms also show divergent results (e.g. Collins et al. [55] and Wiegmann et al. [102]). Given these inconsistencies, along with the fact that brain damage occurs even below heavy drinking levels (e.g. Topiwala et al. [103, 104]), we propose a conservative approach in which only the detrimental effects of heavy drinking are modelled and restricted to early‐onset dementia. This would align with the approach taken in a recent modelling study by Kilian et al. [105].

The overall dilemma of quantifying the impact of alcohol consumption on mental disorders can be exemplified by its relationship with depression. Clearly, causality is bidirectional: alcohol consumption can cause depression, and vice versa, with genetic vulnerability contributing to both [106, 107]. Because of this bidirectional impact, major depressive episodes are now classified as either alcohol‐induced or independent [108]. Early attempts to quantify the impact of alcohol consumption on depression for a CRA relied on temporality only [106], but this was insufficient to establish causality. Meanwhile, several improvements have been made [107, 109, 110], and we expect inclusion of depression in WHO CRAs in the future.

Cardiovascular disease

Alcohol consumption has causal effects on many cardiovascular diseases through multiple biological pathways. Let us start with the detrimental risk relationships with chronic heavy drinking occasions [57, 111]. First, in individuals with chronic heavy drinking, alcohol acts as a toxin that weakens the heart muscle [112, 113, 114], directly contributing to alcohol‐related cardiomyopathy. Second, alcohol consumption raises blood pressure in a well‐established dose‐dependent manner (see Table 1) [53, 115], and reductions in consumption lower blood pressure, indicating that initial harms from drinking are reversible [51, 52]. Actually, alcohol interventions have been used as a first‐line treatment for incident hypertension in heavy drinkers [116, 117]. These hypertensive effects contribute not only to hypertensive heart disease, but also to other cardiovascular outcomes, especially haemorrhagic stroke. Third, alcohol consumption impairs vascular function through endothelial dysfunction and oxidative stress [118]. In contrast, the cardioprotective effects of low to moderate consumption can be explained by favourable changes in several surrogate biomarkers for cardiovascular risk, such as higher levels of adiponectin, reduced fibrinogen levels, inhibition of platelet aggregation and reduced inflammation [66, 111, 119, 120]. These mechanisms mainly impact ischaemic outcomes, such as ischaemic heart disease and ischaemic stroke. However, heavy episodic drinking, irrespective of the average drinking volume, confers no cardiovascular benefits [57, 111]—even for otherwise non‐heavy drinkers [56, 121]. The biological pathways are similar to those seen in chronic heavy drinking: an increase in blood pressure, adverse effects on blood lipids and disruption of the cardiac conducting system that increases the risk of atrial fibrillation, arrhythmias and sudden cardiac death [122, 123]. Heavy episodic drinking also reverses any beneficial effects of moderate consumption on platelets, and instead promotes blood clotting.

In sum, there are plausible biological explanations for the J‐shaped relationships consistently observed for ischaemic outcomes (see Table 1 and Carr et al. [8]). However, the population‐level relevance of these estimates is less clear, as most cohort studies are based on middle‐class populations with a lower prevalence of heavy episodic drinking than in the general population. Therefore, when estimating risk relationships based on the average level of drinking, patterns of heavy episodic drinking should always be considered (e.g. by separately estimating relationships for those who report and those who do not report heavy episodic drinking). Furthermore, CRAs should take into account that relative risks for cardiovascular disease outcomes attenuate to 1 with older age [124].

Gastrointestinal disease

Establishing the causal impact of alcohol consumption on both liver cirrhosis and pancreatitis is not necessary, because both conditions include subtypes that are fully attributable to alcohol by definition. However, it is important to note that alcohol consumption also affects the progression of both disease categories, independent of the original aetiology [125, 126]. For instance, a recent meta‐analysis found that approximately 40% of all complications related to liver disease caused by hepatitis C (i.e. liver cirrhosis, decompensated liver cirrhosis and death) could be attributable to heavy drinking [127]. For any global CRA, this implies that relative risk functions should be applied to all liver cirrhosis outcomes, irrespective of the original aetiology [72, 127].

Evidence from Mendelian randomisation studies on alcohol consumption and ischaemic heart disease

Before we present findings from our systematic review of MR studies on alcohol consumption and ischaemic heart disease, we briefly explain MR (with alcohol consumption as the exposure) and the key underlying assumptions it relies on for readers who may be unfamiliar with the approach. MR studies use an instrumental variable approach, using one or more genetic variants that predict alcohol consumption as instruments to estimate the effect of alcohol consumption on health outcomes. This approach is designed to reduce risk of reverse causation and bypass the assumption of no unmeasured confounding between exposure and outcome, which can never be completely ruled out when analysing observational data using conventional methods [128, 129]. To test the null hypothesis of no causal effect of alcohol consumption on an outcome for anyone in the population (the ‘sharp causal null’), without yet interpreting the direction or magnitude of such an effect, the genetic variants must be valid instruments [130]. That is, each variant must meet the following three assumptions: (1) the variant is associated with alcohol consumption (‘relevance’); (2) the variant causes the outcome only via alcohol consumption, not through other pathways (‘exclusion restriction’ or ‘no horizontal pleiotropy’); and (3) the variant does not share unmeasured common causes with the outcome (‘exchangeability’, ‘exogeneity’ or ‘independence’). A valid genetic instrument may either directly cause variation in alcohol consumption (causal instrument) or serve as a proxy for an unmeasured causal genetic variant (surrogate instrument) [130]. A fourth assumption, often under discussed in the MR literature, but critical for interpreting the direction and magnitude of the point estimates obtained from MR, is effect homogeneity or monotonicity [130, 131]. Assuming homogeneity (of which several versions exist) allows estimation of the average causal effect in the entire population. Alternatively, assuming monotonicity allows estimation of the average causal effect among ‘compliers’—an unknown subgroup of the population who would increase alcohol consumption if genetically predisposed to higher use and decrease consumption if genetically predisposed to lower use.

Characteristics of included Mendelian randomisation studies

We identified 20 MR studies that estimated an association between (genetically predicted) alcohol consumption and ischaemic heart disease [132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151]. Study selection is summarised in the PRISMA flow diagram (Figure S2) and detailed study characteristics (exposure/outcome definitions, data sources, ancestry, instruments and methods) and findings are provided in Table S6.

Across studies, nine used a one‐sample MR design [132, 133, 136, 138, 139, 141, 143, 147, 148], eight used two‐sample MR [135, 137, 140, 145, 146, 149, 150, 151] and three used both [134, 142, 144]. Most studies were conducted in predominantly European‐ancestry samples (8 studies; [134, 137, 138, 140, 142, 146, 150, 151]), followed by Chinese‐ancestry samples (5 studies; [132, 133, 141, 147, 148]). One study each used Japanese [139] and Korean populations [136], and five included mixed‐ancestry samples [135, 143, 144, 145, 149]. Nine studies evaluated ischaemic heart disease only [132, 133, 136, 138, 140, 141, 142, 143, 145], three evaluated myocardial infarction only [137, 144, 150] and eight evaluated both [134, 135, 139, 146, 147, 148, 149, 151].

The instrument choice largely reflected the underlying population ancestry. Studies in populations of (predominantly) European ancestry commonly used variants in alcohol‐metabolising genes, including ADH1B rs1229984 alone [140] or with other variants [143, 144] and ADH1C rs698 [138]. Studies in Asian populations frequently used ALDH2 rs671 either alone [132, 133, 136, 139, 141] or combined with ADH1B rs1229984 [147, 148]. Several studies used multi‐variant instruments (between 3 and 110; [134, 137, 142, 146, 149]), and two studies did not report the number of variants used [135, 151].

Alcohol exposure definitions were heterogeneous. Ten studies modelled alcohol volume continuously (e.g. drinks/week or grams/day; [133, 135, 136, 141, 144, 145, 146, 147, 148, 149]). Two studies evaluated(log‐transformed) intake frequency [150, 151], and another also problematic alcohol consumption [146]. One study assessed any alcohol consumption [137], three studies used the Alcohol Use Disorders Identification Test‐Consumption (AUDIT‐C) [134, 142, 143] and four studies relied on genotype‐based comparisons only [132, 138, 139, 140].

The primary estimators were most often inverse‐variance weighting [135, 137, 145, 146, 149, 150, 151] or two‐stage least squares [133, 136, 141, 143]. One study used both estimators with different instruments [144]. Five studies reported sharp causal null (gene–outcome) comparisons [132, 138, 139, 140], with one study additionally using inverse‐variance weighting with multiple instruments [142]. Use of non‐linear MR methods was rare. In addition to inverse‐variance weighting, one study applied the residual stratification method [134] and two studies (using the same underlying data but different outcome definitions) explored non‐linearity by stratifying on genotype‐area combinations using a Cox proportional hazards model [147, 148].

Summary of findings

Of the studies that estimated an association between genetically predicted alcohol volume or intake frequency and ischaemic heart disease, one found a significant positive association (i.e. higher alcohol consumption was associated with higher risk [145]), two reported mixed results [141, 144] and seven reported null associations [133, 136, 146, 147, 148, 149, 151]. One study reported a negative association between alcohol intake frequency and myocardial infarction (i.e. higher consumption was associated with lower risk [150]).

Of studies using genetically predicted AUDIT‐C, two reported significant positive associations with ischaemic heart disease outcomes [134, 142], whereas one reported no association [143].

Five studies tested a sharp causal null hypothesis. Two studies found lower risk among individuals with genetic variants associated with faster alcohol metabolism (ADH1B rs1229984 and rs1229984‐T [140, 142]). In contrast, one study comparing GG versus AA at ALDH2 rs671 reported a lower risk in the GG genotype, which is associated with faster acetaldehyde metabolism and higher alcohol consumption [139]. Two studies reported no association for alleles at ADH1C rs698 [138] and ALDH2 rs671 [132], respectively.

Multivariable Mendelian randomisation

Five studies reported multivariable MR analyses adjusting for other risk factors, and one of these did not treat alcohol as the primary exposure, but only adjusted for it in multi‐variable MR [135]. Adjusting for smoking initiation did not materially change the null associations between log‐transformed drinks/week and ischaemic heart disease outcomes obtained without adjustment [145, 149]. After joint adjustment for systolic blood pressure and smoking, a previously positive association between drinks/day and myocardial infarction in one study was rendered non‐significant [144], and the null association between AUDIT‐C and ischaemic heart disease remained unchanged [143]. After adjustment for cannabis use disorder, no association was observed between drinks/week and myocardial infarction or ischaemic heart disease [135].

Non‐linear Mendelian randomisation

Three studies assessed potential non‐linearity. Biddinger et al. [134] estimated a positive, approximately log‐linear association with ischaemic heart disease using the residual stratification method, consistent with their inverse‐variance weighted estimate. Millwood et al. [148] stratified by genotype‐area combinations and largely confirmed the overall null association per 280 g/week of genetically predicted consumption, while reporting significant positive associations for a few consumption categories with fatal myocardial infarction and ischaemic heart disease [147].

Risk of bias assessment

Four studies were judged at low risk of bias, one at moderate risk and 15 at high risk, driven by high risk in at least one domain (weak instruments, genetic confounding, non‐genetic confounding, pleiotropy or selection bias). Risk of bias assessments are summarised visually in Figure S3. Studies at low to moderate risk of bias reported positive [134, 140] or null associations [137, 147, 148] between alcohol consumption and ischaemic heart disease.

Collectively, the available MR evidence does not support a protective effect of alcohol consumption on ischaemic heart disease risk at low levels of consumption, in contrast with major meta‐analyses of cohort studies that estimate J‐shaped relationships. However, the interpretation is constrained because most MR studies tested log‐linear effects, only three evaluated non‐linearity and none separated dimensions of drinking (e.g. average consumption from heavy episodic drinking)—limiting the ability of MR to directly assess whether risk is lower at low consumption levels compared to non‐drinking.

Integration of evidence from cohort and Mendelian randomisation studies to inform comparative risk assessments for ischaemic heart disease (and other conditions) and a potential path forward: A viewpoint

We will use ischaemic heart disease as a case study for thinking about evidence integration across cohort and MR studies in alcohol research, with the expectation that some of these themes will carry over across outcome categories. Overall, we are faced with a situation where alcohol consumption is clearly associated with ischaemic heart disease risk in cohort studies—irrespective of the exact shape of the curve—but MR studies predominantly suggest no association (see Carr et al. [8] and the results of our systematic review above).

Reconciliation of the evidence provided by cohort and MR studies requires reconciliation of the limitations of each approach. Given that these limitations are well‐documented elsewhere, we reiterate just a few points related to alcohol studies specifically (also summarised in Table S7): ill‐defined drinking strategies and related questions of measurement, time‐varying alcohol consumption and time‐varying confounding and misaligned time zero. Ill‐defined drinking strategies arise in part because alcohol consumption is a multi‐dimensional phenotype. At a minimum, a precise exposure definition must consider both average volume of drinking (composed of usual frequency and quantity per occasion) and pattern of consumption (for ischaemic heart disease, mainly heavy drinking occasions) [56]. This issue is further compounded by the reliance on self‐reported alcohol consumption in much of the epidemiologic literature. Alcohol consumption also varies over time—often quite dramatically—requiring analytic techniques that handle time‐varying confounding without introducing bias [152]. Finally, defining ‘time zero’ of an alcohol study is non‐trivial across both designs, and it is possible for misalignment of study eligibility, exposure initiation or measurement and beginning of follow‐up, which can introduce selection and/or measurement bias [153]. These sources of bias can be related, but they can also manifest in different ways across study designs and analytic strategies.

In the face of these pervasive limitations across both designs, plausibility and replicability may serve as useful metrics for evidence integration and evaluation. At this point, we see good reason to assume a causal effect of average level of alcohol consumption on ischaemic heart disease for the following reasons: (1) there are established biological pathways for such an effect [66, 120, 154]; (2) hundreds of potential confounders have been tested over the years, as alcohol and ischaemic heart disease is one of the most explored dose–response relationships [8, 155] and the J‐shape persisted; and (3) alcohol policy measures such as restrictions of alcohol availability were associated with reductions in cardiovascular and ischaemic heart disease mortality (e.g. Stumbrys et al. [156]).

This leaves us with the task of determining the exact shape of the dose–response relationship. The overwhelming majority of cohort studies find a lower risk of ischaemic heart disease at low levels of average alcohol consumption, if no heavy drinking occasions are present [56] and a higher risk at higher levels, characterising the J‐shaped relationship. This J‐shape is observed in all major meta‐analyses of cohort studies, but the lower risk at low consumption levels becomes less pronounced as more potential biases are controlled [58].

In our systematic review, most MR studies did not assess non‐linearity (17/20), limiting their ability to directly evaluate a J‐shaped relationship. The three studies that did explore non‐linearity reported contradictory patterns, but importantly, none provided evidence consistent with a beneficial effect of low average consumption on ischaemic heart disease risk. Biddinger et al. [134] estimated a positive, approximately log‐linear association, whereas Millwood et al. reported largely null associations with (fatal) myocardial infarction and ischaemic heart disease [147, 148]. Overall, these three studies—in which several biases cannot be ruled out (e.g. because of the use of non‐linear MR methods that have been shown to yield spurious, implausible associations [157] as well as biases described above and in Table S7)—do not seem sufficient to overturn the use of evidence from cohort studies as the main source for CRAs. We would need to find genetic markers that can distinguish heavy drinking occasions from average level of drinking, and then, test for any protective effect in the latter. Moreover, methods for modelling non‐linear associations in MR are still relatively nascent and undergoing active methodological development, and it is possible that future contributions from MR will aid greatly in clarifying the shape of the curve.

What can MR—and other approaches—add? Given limitations in both designs when studying causal effects of alcohol consumption, this creates an opportunity to triangulate findings from cohort and MR studies to qualitatively assess the strength of evidence for causal effects [158]. Triangulation involves comparing results from approaches with different (ideally unrelated) key sources of bias and considering whether those biases would be expected to shift estimates in the same or opposite directions. For triangulation to be informative, the approaches should, as far as possible, address the same underlying causal question and the main potential biases in each approach should be stated explicitly, including their expected direction (where feasible). When conclusions align across approaches despite differing bias structures, confidence in a causal interpretation is strengthened.

Other complementary paths forward also exist. Whereas triangulation can aid in the synthesis of existing evidence, the target trial framework and target trial emulation [159] can aid future studies in eliminating design‐induced and other self‐inflicted biases. Emulating a (hypothetical) pragmatic trial, or target trial, provides a useful framework for identifying and mitigating some biases [160] by making explicit what the pragmatic trial would be to answer the causal question of interest that could be reasonably emulated with observational data. Such a framework is compatible with the ‘fixes’ suggested in Table S7 regarding some of the key sources of bias in alcohol studies.

In sum, although we clearly should work on improving the methodology, at this point, the MR evidence base is too weak to rule out the cardioprotective effect at low levels of consumption and the harmful effect at high levels established in cohort studies and supported by biological pathways. For the specific purpose of calculating AAFs needed in CRAs to estimate population‐level burden attributable to alcohol, we recommend continuing to rely primarily on cohort‐based risk relationships (e.g. those described in Table 1), because they provide more directly interpretable evidence on effect magnitude required for CRA modelling. This can be complemented by triangulation with MR to identify discrepancies between approaches and, where observed, to interpret AAFs and alcohol‐attributable burden estimates with appropriate caution.

Overview of different dimensions of alcohol consumption and injury outcomes

Alcohol is a neurotoxin and central nervous system depressant that impairs balance, visual focus, reaction time and executive functioning, even at low to moderate levels of consumption [161]. In the context of injuries, blood alcohol concentrations (BACs) as low as 0.03 g/dl have been shown to impair reaction time and other cognitive functions, with effects intensifying at higher BAC levels [162]. The risk of injury is influenced not only by the amount of alcohol consumed, but also by the person involved, the vehicle or agent causing the injury, the environment in which the injury occurs and co‐use of other psychoactive substances [163, 164, 165, 166].

Evidence supporting a causal role of alcohol in injuries comes from laboratory experiments [167], driving simulation studies [168] and emergency department case‐crossover studies [169]. Furthermore, time‐series analyses suggest that changes in alcohol policies and interventions can significantly impact population‐level injury rates [170]. In the context of self‐harm, both acute and chronic alcohol consumption contribute causally by triggering the onset and exacerbating major depressive episodes [107].

The intoxicating effects of alcohol increase the risk of a wide range of injuries, including road traffic injuries, falls, drowning, burns, poisonings, self‐harm and violence [6]. The role of alcohol consumption in injuries resulting from exposure to forces of nature, or conflict and terrorism, is less clear. However, given alcohol's effects on balance, visual focus, reaction time and executive functioning, it may also contribute to these types of deaths from injuries [161].

Alcohol consumption also contributes to injuries inflicted on others. In cases of violence, including intimate partner violence and sexual assault, alcohol increases the risk of both perpetrating and experiencing harm. This is largely because of its effects on executive functioning, such as impaired judgment, reduced inhibitions and alcohol‐induced myopia [171]. In both road traffic and violent injuries, the risk is not limited to the drinker, but also extends to others who may or may not have been consuming alcohol.

Average volume of alcohol consumption, patterns of drinking, biological pathways and reversibility of effects

All CRAs in the WHO tradition to date have modelled the impact of alcohol consumption based on two dimensions: average volume of consumption and patterns of drinking, most often conceptualised as heavy episodic drinking occasions (see Rehm et al. [5], and Llamosas‐Falcón et al. [172] for early discussions of these dimensions). Figure 1 of the previous Addiction monograph in this series [6] attempted to summarise the status quo of knowledge in this area.

On the one extreme are most alcohol‐attributable cancer sites (except for breast and cervical cancer) (see Table 1 and below), where risk is mostly driven by overall alcohol exposure. For these outcomes, risk can be reasonably approximated by average consumption levels, regardless of drinking pattern. On the other extreme are injuries (e.g. traffic‐related injuries), where risk largely depends on BAC, often approximated by the frequency of heavy drinking occasions [173], although average drinking also contributes to this risk [174, 175]. The differences between the impact of average drinking and heavy drinking occasions largely reflect distinct biological pathways, which also determine the potential reversibility of alcohol‐attributable outcomes.

Consider again the cancer example: one of the major pathways involves the metabolism of alcohol, which creates acetaldehyde, leading to multiple mutagenic effects in deoxyribonucleic acid (DNA) [176, 177]. Although this mechanism may contribute to increased breast cancer risk, there is also an important impact on hormones, which are influenced by patterns of drinking as well [37, 89]. In other words, the relative impact of average consumption versus patterns of drinking will depend on which biological pathways are most relevant for a given outcome.

The health impacts of alcohol consumption are, in part, reversible. Some health risks, particularly those of heavy episodic drinking and alcohol intoxication (e.g. injuries or contracting sexually transmitted diseases), are transient and only elevated during periods of intoxication.

For infectious diseases other than sexually transmitted diseases, the reversibility of alcohol‐related health risks is more complex and depends on the specific condition. Alcohol affects nearly all components of the innate immune system, with both immediate and long‐term effects [178]. Acute alcohol intoxication impairs immune function by blocking the differentiation and maturation of granulocytes (i.e. granulopoiesis) during infection [179] and by suppressing the myeloid proliferative response during bacterial infections [180]. These effects are generally transient and tend to reverse shortly after abstaining from alcohol. In contrast, chronic alcohol consumption has more lasting consequences. It reduces populations of natural killer T cells, which play critical immunoregulatory roles [179]. Although long‐term abstinence can improve immune function, these impairments may only be partially reversible, depending on the volume and duration of alcohol consumption [181].

Many chronic conditions caused by alcohol are not fully reversible, even after reducing or stopping consumption. The liver, for instance, experiences both immediate and long‐term impacts. A reduction in alcohol consumption leads to an immediate decrease in the risk of liver cirrhosis. This has been demonstrated on a population level, for instance, in the sharp declines in liver cirrhosis mortality following large reductions in alcohol consumption during the Gorbachev reforms or the German occupation of Paris, when alcohol was confiscated (see Zatoński et al. [182] for historical examples). Similarly, even modest reductions in alcohol consumption through policies like increased taxation have been associated with reduced liver cirrhosis mortality [183, 184]. Although alcohol‐related cirrhosis is often irreversible [185], reducing consumption may slow disease progression and decrease the risk of death.

Some cardiovascular effects of alcohol may also be reversible. Reducing consumption lowers cardiovascular disease risk by preventing acute effects like alcohol‐induced increases in heart rate [186]. Alcohol control policies that limit access to alcohol—especially during times with frequent heavy drinking like late‐night hours—have been shown to decrease the risk of fatal cardiovascular events, including the ‘holiday heart syndrome’ [156, 187]. Other cardiovascular risks, such as hypertension, may take longer to reverse, but improvements can be seen within days to weeks of abstinence [188].

In the case of cancer, alcohol‐related damage may already be too advanced for reductions in consumption to yield meaningful benefits. Nonetheless, reductions in alcohol consumption at the population level and increases in abstention rates are associated with improved outcomes, even for cancers [94, 189] and other chronic diseases (e.g. Roerecke et al. [190]). In addition to causing cancer, alcohol may contribute to its progression by promoting tumour growth via acetate production, facilitating angiogenesis, invasion, metastasis and impairing immune response [191].

Alcohol's impact on the brain is also complex and only partially reversible. Acute conditions such as Wernicke–Korsakoff syndrome can be reversed if diagnosed and treated promptly [192]. Chronic alcohol consumption and heavy episodic drinking can cause brain atrophy, which increases the risk of dementia. However, some alcohol‐induced brain changes, such as cerebral atrophy, may be partially reversible. Studies using computed tomography have shown partial recovery of brain volume after sustained periods of abstinence from alcohol [193, 194].

Figure 2 provides an overview of the relationships between different dimensions of alcohol consumption, exemplary biological pathways and potential reversibility of effects.

FIGURE 2.

Overview of the relationships between different dimensions of alcohol consumption, exemplary biological pathways and reversibility of effects. BAC, blood alcohol concentration; DNA, deoxyribonucleic acid.

DISCUSSION

Overall, alcohol consumption causes many health outcomes, as supported by evidence from multiple study designs. Where findings disagree between approaches, particularly for J‐shaped relationships such as with ischaemic diseases, we need to weigh the evidence against the limitations of each study design. In these settings, triangulating epidemiological evidence across approaches, alongside considering biological plausibility, can help clarify whether discrepancies may be due to bias.

Before further discussing the implications of the current research, we would like to point out limitations of our work. Most importantly, every systematic review is only as valid as the validity of the underlying research. As outlined above, alcohol epidemiological research, whether based on cohort or MR studies, is susceptible to risk of severe biases. Our review also has its own limitations. For cohort studies, the systematic scoping review focused only on average levels of consumption, while drinking patterns and biological pathways were summarised through a narrative review. Although fetal alcohol spectrum disorder is fully attributable to alcohol consumption (Table S5), accounted for in the burden estimates for alcohol use disorders in the Global Health Estimates [195], and harm to others was discussed in the section on injuries, our treatment of these outcomes was not systematic. Future iterations will need to include a more systematic approach, as the health burden on other individuals is of similar magnitude to the burden on the drinker in some studies [196]. Our systematic review of MR studies also has several limitations. First, we focused on primary analysis results (e.g. obtained using the inverse‐variance weighted method) and did not systematically assess sensitivity analyses or within‐study multiple testing when evaluating the evidence. Therefore, our summary does not reflect whether alternative MR methods used within each study estimated consistent or contradicting results. Second, we identified potentially relevant studies based on whether their titles and abstracts contained terms from our search strings. Because MR studies often assess multiple outcomes without explicitly reporting them in the title or abstract, some relevant studies may have been missed. Where possible, this was addressed through manual addition of studies. Last, we only included studies that explicitly stated to have used MR, and therefore, may have missed studies, especially those that tested only gene‐outcome associations.

Several paths forward could help address current limitations in the evidence on alcohol's health impacts. First, we need new or expanded cohort studies that are more diverse in terms of the distributions of other risk factors that may interact with alcohol consumption, disease occurrence, causes of death, patterns of drinking and genetic profiles. Participants should be repeatedly invited to report on their alcohol consumption, including patterns of drinking, using more reliable measures [197]. Ideally, cohorts would begin in early‐ to mid‐adulthood and be followed for long periods, which would help reduce potential survival bias.

Second, we can make better use of existing observational data. Many large prospective cohorts are available, reflect a wide range of drinking patterns and often span many years or even decades. As discussed above, the target trial framework offers a practical way forward to reduce design‐induced biases and make the causal question explicit; yet, very few studies have emulated a target trial to evaluate the health effects of drinking strategies (for a recent review, see Hansford et al. [198]). Although largely underused in alcohol epidemiology, we believe this approach holds considerable potential to strengthen causal inference on the health effects of alcohol consumption. It is important to acknowledge that, with this framework, inherent biases such as residual and unmeasured confounding cannot be directly addressed. As large cohorts increasingly include genome sequencing and repeated measures of alcohol consumption, similar causal questions should also be addressed using MR and findings triangulated with evidence from other approaches.

Finally, we should randomise where possible. Currently, the University of Navarra Alumni Trialists Initiative (UNATI) is the only ongoing major randomised trial, examining the effect that an intervention recommending a reduction in alcohol consumption has on major disease and mortality [199]. This trial includes women 55 to 75 years old and men 50 to 70 years old who currently consume alcohol, with follow‐up planned for 4 years. Findings from trials like UNATI will be essential for the ‘alcohol‐in‐moderation’ debate. However, such trials are inherently limited by their specific study populations, relatively short follow‐up durations (if no funds are available for extension) and alcohol interventions designed to minimise harm and maximise adherence. Furthermore, for ethical and practical reasons, long‐term trials cannot randomise individuals to initiate or increase alcohol consumption or to engage in heavy episodic drinking.

The field of alcohol epidemiology can make substantial progress by refining its approaches and methods to provide clearer answers on the causal effects of alcohol consumption on health outcomes, the reversibility of those effects and the resulting alcohol‐attributable burden, all of which are central for designing effective interventions. Many of our studies on alcohol consumption and health are still plagued by the same problems that were pointed out 20 years ago—even though they are, in principle, avoidable.

AUTHOR CONTRIBUTIONS

Sinclair Carr: Conceptualization (equal); data curation (equal); formal analysis (equal); investigation (equal); methodology (equal); project administration (equal); validation (equal); visualization (equal); writing—original draft (equal); writing—review and editing (equal). Ana Lucia Espinosa Dice: Conceptualization (equal); data curation (equal); formal analysis (equal); investigation (equal); methodology (equal); validation (equal); writing—original draft (equal); writing—review and editing (equal). Gerhard E. Gmel Sr.: Conceptualization (equal); data curation (equal); formal analysis (equal); investigation (equal); methodology (equal); validation (equal); writing—original draft (equal); writing—review and editing (equal). Ahmed S. Hassan: Conceptualization (equal); data curation (equal); formal analysis (equal); investigation (equal); methodology (equal); validation (equal); writing—original draft (equal); writing—review and editing (equal). Kevin D. Shield: Conceptualization (equal); data curation (equal); formal analysis (equal); investigation (equal); methodology (equal); validation (equal); visualization (equal); writing—original draft (equal); writing—review and editing (equal). Jürgen Rehm: Conceptualization (equal); data curation (equal); formal analysis (equal); funding acquisition (lead); investigation (equal); methodology (equal); project administration (lead); validation (equal); visualization (equal); writing—original draft (equal); writing—review and editing (equal).

DECLARATION OF INTERESTS

None.

Supporting information

Figure S1. PRISMA flow diagram for the systematic scoping review to identify systematic reviews and meta‐analyses that evaluated the risk relationship between alcohol consumption and conditions within selected disease categories, based on the Alcohol Intake and Health Study [2].

Figure S2. PRISMA flow diagram for the systematic review to identify Mendelian randomisation studies of alcohol consumption and ischaemic heart disease.

Figure S3. Traffic light plot of risk of bias assessments for Mendelian randomisation studies on alcohol consumption and ischaemic heart disease.

Table S1. Search strategy used to identify Mendelian randomisation studies on alcohol consumption and ischaemic heart disease in Embase Classic+Embase up to 28 February 2025.

Table S2. Risk of bias assessment tool adapted from Jabeen and colleagues [1] with explanation of each domain.

Table S3. PRISMA 2020 for abstracts checklist for the systematic review of Mendelian randomisation studies.

Table S4. PRISMA 2020 checklist for the systematic review of Mendelian randomisation studies.

Table S5. Disease and injury categories fully (100%) attributable to alcohol consumption, defined according to International Statistical Classification of Diseases and Related Health Problems, 10th and 11th revision codes.

Table S6. Characteristics and main analysis results of Mendelian randomisation studies on alcohol consumption and ischaemic heart disease.

Table S7. Key (design‐based) biases in cohort and Mendelian randomisation studies of alcohol consumption, and ways to address them where feasible.

Data S1. Supplementary Information.

Carr S, Espinosa Dice AL, Gmel GE Sr., Hassan AS, Shield KD, Rehm J. A review of the relationship between dimensions of alcohol consumption and the burden of disease: 2026 update including Mendelian randomisation studies. Addiction. 2026;121(8):1998–2018. 10.1111/add.70435

DATA AVAILABILITY STATEMENT

N.A.