Methodological biases in estimating the relationship between alcohol consumption and breast cancer: The role of drinker misclassification errors in meta-analytic results

Abstract

Background

While alcohol consumption has been linked to breast cancer in women, few studies have controlled for possible biases created by including former or occasional drinkers in the abstainer reference group. We explored the potential for such misclassification errors as sources of bias in estimates of the alcohol-breast cancer relationship.

Methods

Meta-analyses of population case-control, hospital case-control, and cohort studies to examine relationships between level of alcohol use and breast cancer morbidity and/or mortality in groups of studies with and without different misclassification errors.

Results

Of 60 studies identified, only 6 were free of all misclassification errors. The abstainer reference group was biased by the inclusion of former drinkers in 49 studies, occasional drinkers (<10g ethanol per week) in 22 and by both these groups in 18. Occasional drinkers were also mixed with light or hazardous level drinkers in 21 studies. Unbiased estimates of the OR for breast cancer were 1.011 (95% CI: 0.891-1.148) among former drinkers (n = 11) and 1.034 (95% CI: 1.003-1.064) among occasional drinkers (n=17). Hazardous level drinking (>20g<41g ethanol/day) was not significantly associated with breast cancer in studies with occasional drinker bias. However, in studies free from occasional drinker bias, the OR for breast cancer was 1.085 (95% CI: 1.015-1.160) for low level (<21g/day) drinkers (n=17), 1.374 (95% CI: 1.319-1.431) for hazardous level drinkers (n=26) and 1.336 (95% CI: 1.228-1.454) for harmful level (>40g/day) drinkers (n=9).

Conclusions

While the great majority of studies of the alcohol-breast cancer link include misclassification errors, only misclassification of occasional drinkers was found to bias risk estimates significantly. Estimates based on error free studies confirmed that low, hazardous and harmful levels of alcohol use each significantly increase the risk of breast cancer.

Introduction

A large body of literature exists on the association between alcohol consumption and risk of breast cancer (Swanson et al., 1997, Bowlin et al., 1997, Ferraroni et al., 1998, Maura et al., 1998, World Cancer Research Fund, 2007). Nevertheless, questions remain regarding this association and findings have often been conflicting, especially in relation to potential risks from low-level consumption.

The present paper uses meta-analysis to examine the alcohol-breast cancer association and addresses some key methodological challenges that are common in this literature. Specifically, we examine the potential impact of drinker misclassification errors. These errors occur when former and/or occasional drinkers are included in the ‘abstainer’ reference group, or when occasional drinkers are included in a low-level drinking group. Recent research has shown that including former and/or occasional drinkers with abstainers creates bias that can exaggerate protection against heart disease from low risk drinking (Chikritzhs et al., 2009, Fillmore et al., 2006, Fillmore et al., 2007, Roizen et al., 2012, Stockwell et al., 2012b). It is conceivable that such biases are operating in a similar manner in relation to other disease outcomes associated with alcohol consumption. In this paper we explore this possibility in relation to the alcohol-breast cancer association. This would have implications for the validity of findings, alcohol policy and advice to patients. We will review the current literature on alcohol consumption and breast cancer, and will discuss what types of drinker misclassification errors might be present and how these may affect results.

Review of the literature

Several studies have found increases in risk of breast cancer with heavier drinking, with risk estimates ranging from 1.69 (95% CI 1.19 to 2.40) to 1.83 (95% CI 1.18–2.85), (Gapstur et al., 1992; Graham et al., 1992). Other studies have suggested that there may also be increased risk at even one drink per day (containing 12-14g of ethanol). These studies have reported risk estimates ranging from 1.21 (0.95-1.55) to 1.52 (1.19-1.93) (Bowlin et al., 1997; Ferraroni et al., 1998; Maura et al., 1998; Swanson et al., 1997). Other research has found only weak associations between low level alcohol consumption and breast cancer, with similar odds ratios ranging from 1.06 (95% CI: 1.00, 1.11) to 1.10 (95% CI): 1.06, 1.14) (Ellison et al., 2001; Sneyd et al., 1991). As Shapiro (Shapiro, 1994) observed, even weak associations may be important and warrant investigation as they may translate into thousands of cases for a common illness such as breast cancer.

A systematic review of alcohol drinking and breast cancer risk in the Japanese population (Nagata et al., 2007) reported results from three cohort studies and eight case-control studies, but with mixed results. The Collaborative Group on Hormonal Factors in Breast Cancer (Ferraroni and La Vecchia, 2002) reported on 53 case-control and cohort studies; they found that compared to non-drinking women, the risk of breast cancer was 1.32 (1.19-1.45, p=0.000) for those who drank 35-44g of alcohol per day, and 1.46 (1.33-1.61, p=0.000) for women who consumed >45g of alcohol per day. Longnecker (1994) conducted a meta-analysis of 38 epidemiologic studies to clarify the association between alcohol intake and breast cancer risk. Relative to non-drinkers, he reported a risk of 1.1 (95% CI, 1.1 to 1.2) for one drink per day (13g of ethanol), a risk of 1.2 (95% CI, 1.1 to 1.3) for two drinks per day and a risk of 1.4 (95% CI, 1.2 to 1.6) for three or more drinks per day. The author concluded that the modest size of the association and variation in results across studies leave the causal role of alcohol in question.

While the above meta-analyses and systematic reviews indicate the presence of an increased risk of breast cancer with increased alcohol consumption none of them addressed the issue of drinker misclassification errors and possible resulting biases in estimates.

Drinker misclassification error

In observational studies, such as studies on alcohol and breast cancer, it is impossible to completely eliminate all sources of bias. There is always the possibility of unidentified and uncontrolled confounding; if, for instance, many studies based on different methods are convergent in producing significantly elevated relative risk estimates, we are inclined to think of those results as valid and stable. However, modest changes in risk may also be due to the same underlying bias consistently present in most of these studies resulting in the appearance of spurious statistical stability. This may also discourage further inquiry as it may appear as if the results are robust by virtue of their consistency and that the research question has therefore been sufficiently addressed (Shapiro, 1994).

In the present paper we focus on just one potential source of bias that may stem from misclassification of occasional or former drinkers into either the reference group of ‘abstainers’ or into other drinking groups. Other potential sources of confounding include type of outcome assessed (morbidity or mortality), other study design features (e.g. cohort or case-control study) and population characteristics such as age and ethnicity. These misclassification problems and related issues, among others, have been mostly discussed in relation to purported health benefits of low-level alcohol use but may have broader relevance (Chikritzhs et al., 2009, Fekjær, 2013, Fillmore et al., 2007). Significant bias may be caused by three types of drinker misclassification: (i) including former drinkers with abstainers; (ii) including occasional drinkers with abstainers, and (iii) including occasional drinkers with low or hazardous level drinkers. These misclassifications may create bias via several avenues. Former drinkers may be those who stopped drinking due to health problems (Fillmore et al., 2006). In some populations it has been shown that as people age and experience declining health, they are inclined not only to give up drinking but also to reduce their drinking (Shaper et al., 1988). Thus, including such occasional drinkers with abstainers could make hazardous risk drinkers look healthier by comparison, i.e. cause bias (Fillmore et al., 2006). This practice also prevents estimates being made of relative risk for adverse outcomes among occasional drinkers (Shaper et al., 1988).

Occasional drinkers in some populations may also have a better health status than abstainers for reasons unrelated to their drinking which could cause bias from another direction (Stockwell et al., 2012). That is, a number of studies combine occasional with moderate risk drinkers, thus giving the latter the appearance of lower risk when compared with abstainers; including occasional drinkers in the low risk drinker group will make the latter appear healthier by comparison. One way to minimise these potential biases is to avoid drinker misclassification by reporting results separately not only for former but also for occasional drinkers. It is also been suggested that in order to avoid bias, former drinkers should in fact be placed into drinking groups which reflect their past exposure levels prior to analysis (i.e. an intention-to-treat approach) (Liang and Chikritzhs, 2013). Unfortunately, however, there is usually insufficient information on participant's past levels of drinking to enable investigation of dose-response risk relationships.

Many epidemiological studies appear to indicate that low to moderate risk alcohol use protects against CHD. Consequently, some physicians actively advise their patients that “moderate” use of alcohol is beneficial for heart health (Mukamal et al., 2010). Recent evidence suggests that apparent protective effects of low to moderate level alcohol use on CHD, dementia and all-cause mortality may be over-stated due to the presence of methodological biases (including drinker misclassification bias) in the majority of applicable studies

(Chikritzhs et al., 2009, Fillmore et al., 2006, Fillmore et al., 2007, Roizen et al., 2012, Stockwell et al., 2012b). Similarly, a study on cardiovascular disease (CVD) suggests that some or all of the apparent protective effect of moderate alcohol consumption on CVD may be due to residual or unmeasured confounding by other lifestyle factors (Naimi et al., 2005). In particular, this study suggests that even using lifetime abstainers as a reference group may create bias due to this group's tendency toward less than optimal health (e.g., many abstain due to health problems). Furthermore, recent research from the UK indicates that young adults who elect to be complete abstainers tend to have poorer health and lower economic status than drinkers (Ng Fat and Shelton, 2012). It follows, then, that other alcohol-disease relationships may be subject to the same biases, likely resulting in the underestimation of disease risk. Sensitivity to artificially suppressed risk may be especially important with respect to low to moderate level drinking categories, where past studies show weaker dose-response impacts or risk associations (Chen et al., 2011).

Prospective observational studies indicating protective effects from low to moderate level alcohol use encompass an assortment of conditions including many for which a causal mechanism involving alcohol is implausible. A recent review identified studies reporting apparent health benefits from light to moderate level drinking for Alzheimer's, asthma, the common cold, depression, gallstones, deafness, urinary tract symptoms, osteoporosis, obesity, arthritis, various cancers and hip fracture and even liver cirrhosis (Fekjær, 2013). The great majority of studies contain drinker misclassification errors and contain other design errors that have been identified in relation to studies of CHD and stroke (Stockwell et al., 2012b).

In the present study we examine the possible impact of different types of drinker misclassification error on the association between alcohol consumption and breast cancer at different drinking levels. Our principal strategy was to contrast breast cancer risks separately for both former and occasional drinkers with long-term abstainers in studies relatively free of misclassification error. Based on the literature reviewed above, it was predicted that both former and occasional drinkers would exhibit increased risk of breast cancer compared with lifetime abstainers. Finding significantly increased or reduced risk for these drinking categories would be indicative of their potentially harbouring biased results from studies which have misclassified drinkers into the reference group of “abstainers”. Finally, we report estimates of breast cancer risk at three levels of alcohol consumption (low, hazardous and harmful) in studies free of misclassification error.

Methods

We performed a meta-analysis on 60 epidemiological studies (34 hospital case-control or population case-control and 26 cohort studies) that had quantitatively evaluated associations between alcohol use and multiple selected disease outcomes for breast cancer. We then performed separate meta-analysis on subgroups of these studies divided according to the presence or absence of different drinker misclassification errors. Our team searched MEDLINE using key the words “breast cancer” and “alcohol” and variations (e.g., “carcinoma of breast”, “ethyl alcohol”) to identify studies published in English up to and including 2013. We supplemented these with additional studies cited as references in previous studies, systematic reviews and meta-analyses. The following initial inclusion/exclusion criteria were applied:

1. There was a clearly defined alcohol variable that quantified the association between alcohol intake and breast cancer at two or more levels of consumption in terms of odds ratios, risk ratios or relative risk.

2. Meta-analyses were excluded, as were studies in which the cohort was men, and studies in which the outcome was benign breast disease.

3. We also excluded studies that used the same cohort as another study already in our set of included studies.

A total of 60 published research articles were identified from our systematic literature search after excluding 49 papers from the 116 located on the basis of reviews of 1515 abstracts. Reasons for exclusion included being duplicate studies, reviews or meta-analyses and type of disease outcome, as summarised in Figure 1.

Figure 1. Flow of study selection from literature search to inclusion in meta-analysis.

The included studies were then coded for presence or absence of each form of drinker misclassification error:

1. Former drinker misclassification: the “abstainer” reference group was defined in terms of the absence of current drinkers without reporting results separately for former drinkers or otherwise specifically excluding them (n=49);

2. Occasional drinker misclassification: the “abstainer” reference group specifically included some category of occasional drinkers up to the equivalent of 1.43 g ethanol per day or had loose definitions of abstention such as drinking “never or almost never” (n=22);

3. Reverse occasional drinker misclassification: all occasional drinkers by the above definition were excluded from the abstainer reference group but then included in a low or hazardous level drinking category i.e. there was no separate risk estimate for a category of occasional drinking (n=21).

Two trained coders each independently coded studies using a standardized codebook. Coders recorded sample type, study characteristics, demographics, alcohol consumption quantification, breast cancer outcomes, measures of association, bias types and control variables. A senior investigator then checked the coding sheet for accuracy, completeness and consistency. Coding discrepancies were discussed among the team until all issues were resolved and inclusion criteria were finalized.

Analyses

Drinking categories were defined as follows, using grams of ethanol consumed per day: (i) former drinkers now completely abstaining; (ii) occasional drinkers, consuming less than one drink per week (0.01-1.43 g/day), (iv) low level drinkers, 1 drink/month to 2 drinks/day (1.5-24 g/day); (v) hazardous level drinkers, 2-4 drinks/day (25-44 g/day); and, (vi) harmful level drinkers, more than 4 drinks/day (>44 g/day).

One standard drink was defined as one Australian/New Zealand standard drink of 10 grams of ethanol. The midpoint was used to calculate the upper limit and was estimated for open ended upper drinking categories by adding three-quarters of the range of the preceding drinking category to the minimum value. All studies had an open-ended hazardous level drinking group and all upper limits of quantity consumed per day were accepted as valid.

Models were estimated with the drinking variable in categorical form. We tested hypotheses regarding the impact of different drinker misclassification errors with the potential to exaggerate protective effects, mask “true” associations and influence the size and significance of the results. Prior to analyses we log-transformed our data to address the distributional differences between studies and to render a more accurate averaging across studies. Study was treated as a random effect. Studies were weighted by the inverse of the estimated variance of the log odds, as derived from reported standard errors or confidence intervals. Results are expressed as odds ratios. For all analyses we used SAS 9.3 (SAS Institute Inc., 2011).

For all analyses, we chose mixed effects models rather than fixed effects (FE) models, since application of a FE model is based on the assumption that all studies being analysed are homogenous at the level of the study population effect sizes. This assumption could not be justified with the studies included in our analyses due to heterogeneity. We carried out mixed effects regressions applying the proc mixed procedure, with the ln(OR) for breast cancer risk as the dependent variable.

Our first set of analyses included all 60 available studies (pooled abstainer model) and used a loose “abstainer” group definition. Specifically, this pooled abstainer group included all studies whether or not they were free of any of the three types of drinker misclassification error. We then investigated the potential impact of the different drinker misclassification errors by conducting additional meta-analyses using stricter definitions of the ‘abstainer’ group for the reference category (i.e., only former drinkers but not occasional drinkers included in the reference, occasional but not former drinkers included). In the ‘long-term abstainer’ model, the reference group was defined as women who abstained for a year or longer and did not include former or occasional drinkers. We used this model as the comparison model in significance testing.

We identified 7 studies that controlled for drinker misclassification bias by having a ‘clean’ reference group of long-term abstainers, though one of these studies included occasional drinkers within a low level drinking category i.e. contained reverse occasional drinker misclassification error. There were also 4 studies which not only included occasional drinkers in the reference group but also individuals who drank up to 10g of ethanol per day. Due to the small number of studies identified without misclassification error it was not possible to explore the effects of other potential confounding variables such as type of outcome, study design and ethnicity.

We assessed publication bias using funnel and precision plots and regression analysis (Sterne et al., 2000). In addition, since our meta-analysis included a range of study sizes including medium-size studies, we used Begg and Mazumdar's rank correlation test (Begg and Mazumdar, 1994) to report the rank correlation (Kendall's tau) between the standardized effect size and standard errors of these effects. We examined our set of studies for heterogeneity using the I² statistic.

Results

Table 1 shows included studies and their characteristics, as well as breast cancer overall odds ratios with 95% confidence intervals for all drinkers, compared to non-drinkers. We identified only 7 studies that had used long-term abstainers (no drinking in past 12 months or longer) as the reference category (i.e., no reference group misclassification), one of which included occasional drinkers in the low-risk drinking group (i.e., had reverse occasional drinker misclassification). Figure 2 shows the systematic elimination of biases from different drinker misclassification errors in the 60 included studies.

Table 1. Summary of included studies (odds ratios for any drinking versus none).

First Author	Year	Country	Study Design	OR	95% CI Lower	95% CI Upper	Outcome
Adami	1988	Sweden, Norway	Population cc	0.686	0.506	0.866	morbidity
Allen	2009	UK	Cohort	1.096	1.020	1.172	morbidity
Berstad	2007	USA	Population cc	1.159	0.929	1.390	morbidity
Bessaoud	2008	France	Population cc	0.415	0.153	0.676	morbidity
Bowlin	1997	USA	Population cc	1.364	1.146	1.582	morbidity
Breslow	2011	USA	Cohort	0.961	0.783	1.139	mortality
Britton	2002	USA	Population cc	1.052	0.940	1.164	both
Brown	2010	USA	Population cc	0.954	0.749	1.160	morbidity
Chen	2002	USA	Cohort	1.194	1.149	1.239	both
Cooper	1989	Australia	Population cc	0.940	0.699	1.181	morbidity
Croghan	2009	USA	Hospital cc	1.090	0.822	1.359	morbidity
Dumeaux	2004	France, Norway	Cohort	1.379	1.158	1.600	morbidity
Enger	1999	USA	Population cc	0.930	0.766	1.094	morbidity
Feigelson	2003	USA	Cohort	1.075	0.938	1.211	both
Ferraroni	1998	Italy	Hospital cc	1.227	1.136	1.317	morbidity
Friedenreich	1993	Canada	Cohort	1.077	0.917	1.238	morbidity
Gapstur	1992	USA	Cohort	1.245	1.085	1.406	morbidity
Garland	1999	USA	Cohort	0.945	0.804	1.086	both
Hiatt	1988	USA	Cohort	1.423	1.066	1.779	morbidity
Holmberg	1995	Sweden	Population cc	1.488	1.079	1.897	morbidity
Horn-Ross	2012	USA	Cohort	1.121	1.028	1.222	morbidity
Islam	2013	Japan	Hospital cc	1.198	1.060	1.355	morbidity
Kabat	2010	USA	Cohort	0.864	0.759	0.969	mortality
Katsouyanni	1994	Greece	Case control	1.206	0.934	1.477	morbidity
Keogh	2012	UK	Population cc	1.360	1.171	1.579	morbidity
Kinney	2000	USA	Case control	0.937	0.818	1.055	morbidity
Kropp	2001	Germany	Population cc	0.846	0.632	1.060	morbidity
Lash	2000	USA	Population cc	1.016	0.765	1.268	both
Le	1984	Case control	Population cc	1.231	0.993	1.468	morbidity
Levi	1996	Switzerland	Population cc	1.595	1.177	2.013	morbidity
Lew	2009	USA	Cohort	1.133	1.074	1.191	morbidity
Li	2003	USA	Population cc	1.194	1.126	1.261	morbidity
Li	2010	USA	Cohort	1.079	1.021	1.136	morbidity
Lin	2005	Japan	Cohort	1.136	0.597	1.676	both
Longnecker	1995	USA	Population cc	1.113	0.862	1.364	morbidity
Martin-Moreno	1993	Spain	Population cc	1.487	1.230	1.744	morbidity
Mattisson	2004	Sweden	Cohort	0.982	0.765	1.261	morbidity
McCarty	2012	USA	Population cc	1.477	1.219	1.789	morbidity
McDonald	2004	USA	Population cc	1.101	1.008	1.195	morbidity
McTiernan	1986	USA	Population cc	1.439	1.141	1.737	morbidity
Meara	1989	UK	Hospital cc	0.978	0.852	1.104	morbidity
Morch	2007	Denmark	Cohort	1.320	1.144	1.524	morbidity
Nasca	1990	USA	Population cc	1.110	0.985	1.235	morbidity
Rohan	2000	Canada	Cohort	1.064	0.888	1.239	morbidity
Rosenberg	1990	Canada	Population cc	0.784	0.645	0.923	morbidity
Schatzkin 87	1987	USA	Cohort	1.590	1.183	1.996	morbidity
Simon	1991	USA	Cohort	1.063	0.772	1.353	both
Smith	1994	UK	Population cc	0.951	0.629	1.274	morbidity
Sprague	2008	USA	Population cc	1.184	1.039	1.330	morbidity
Stolzenberg-Solomon	2006	Not clear	Cohort	1.142	0.981	1.303	both
Suzuki	2005	Sweden	Cohort	1.142	1.039	1.246	both
Swanson	1997	USA	Population cc	1.122	0.973	1.272	morbidity
Terry	2006	USA	Population cc	0.948	0.763	1.134	morbidity
Thygesen	2008	Denmark	Cohort	1.061	0.917	1.206	both
Tjonneland	2003	Denmark	Cohort	1.196	1.033	1.386	both
Toniolo	1989	Italy	Population cc	1.169	0.925	1.413	morbidity
Van den Brandt	1995	Netherlands	Cohort	1.323	1.104	1.541	morbidity
Van'T Veer	1989	Netherlands	Population cc	0.697	0.436	0.959	morbidity
Zhang	1999	USA	Cohort	0.733	0.618	0.849	both
Zhang	2006	USA	Cohort	1.079	1.030	1.129	both

Population cc=population case-control, Hospital cc=hospital case-control,

Figure 2. Numbers of studies eliminated due to former or occasional drinker misclassification errors out of 60 studies.

Publication bias

Inspection of funnel and precision plots indicated that the majority of studies clustered around the mean effect size. The bottom of the funnel plots showed slight asymmetry, indicating that publication bias might be present to some degree. If asymmetry is caused by publication bias we would expect that high standard errors (small studies) would be associated with larger effect sizes (Rothstein et al., 2006). Kendall's Tau was 0.07 (p=0.287), indicating that publication bias would be unlikely.

Heterogeneity

The full set of 60 included studies displayed high heterogeneity, resulting in an overall I² of 76.256 (p=0.000) denoting 76% between-study heterogeneity. We addressed this by using mixed effects models and performing subgroup analyses. These modifications reduced between-study heterogeneity for the majority of the drinking groups within each set of subgroup analyses.

Pooled meta-analysis of all 60 studies

The pooled analysis shown in Table 2 indicated significantly higher risk of breast cancer for low, hazardous and harmful level drinkers when compared to the loosely defined “abstainer” reference group. In this analysis, where separate results were available for former or occasional drinkers, the breast cancer risks for these drinking categories were not significantly different from the reference group.

Table 2. Pooled analysis.

Group	N Stud.	Pooled OR	95% CI Lower	95% CI Upper	p	Tau 2	I2
Model for all studies, both with and without bias, pooled (n=60)
Pooled abstainers		1.000
Former drinkers	11	1.011	0.891	1.148	0.708	0.018	50.622
Occasional drinkers	21	1.024	0.955	1.099	0.501	0.014	70.705
Low level drinkers	60	1.095	1.057	1.133	0.000	0.010	74.162
Hazardous level drinkers	39	1.328	1.260	1.400	0.000	0.011	47.436
Harmful level drinkers	12	1.342	1.179	1.528	0.000	0.003	4.729

Former drinker misclassification error

Table 3 presents results from two meta-analyses: one of the subgroup of studies which contained only former drinker misclassification error in the abstainer reference group (n=32); and one of the subgroup of studies which reported separate results for former drinkers and abstainers (n=11). In the latter group, there was no significantly elevated risk of breast cancer for former drinkers. Furthermore, the OR for low-level drinkers in this error free model (1.162) was slightly higher than in the first model (1.061). This is the opposite of what would have been expected had former drinker misclassification error imparted a measurable impact on abstainer breast cancer risk; as their poorer health would have elevated the risk among the abstainer reference group and thus reduced any apparent difference in risk relative to low level drinkers. As it is, however, studies which distinguished former drinkers did not indicate that they were at increased risk of breast cancer compared to abstainers and therefore it is logically consistent that mixing the two groups should yield similar outcomes.

Table 3. Models comparing studies with and without former drinker reference group bias.

Group	N Stud.	Pooled OR	95% CI Lower	95% CI Upper	P	Tau 2	I2
Model for all studies with only former drinker reference group bias (n=31)
Abstainers		1.000
Former drinkers	n/a
Occasional drinkers	13	1.004	0.922	1.093	0.926	0.012	64.253
Low level drinkers	31	1.061	1.015	1.109	0.009	0.008	76.884
Hazardous level drinkers	20	1.302	1.194	1.419	0.000	0.013	48.917
Harmful level drinkers	8	1.362	1.250	1.484	0.000	0.011	54.547
Model for all studies without former drinker reference group bias (n=11)
Abstainers		1
Former drinkers	11	1.011	0.891	1.148	0.867	0.018	50.622
Occasional drinkers	6	0.990	0.842	1.164	0.861	0.031	80.985
Low level drinkers	10	1.162	1.116	1.210	0.000	0.020	58.987
Hazardous level drinkers	8	1.319	1.110	1.568	0.002	0.034	68.629
Harmful level drinkers	2	-	-	-	-	-	-

Occasional drinker misclassification error

The results in Table 4 provide some evidence that misclassifying occasional drinkers as abstainers could be the source of downward bias in estimates of breast cancer risk from alcohol consumption. The second model presented shows a modest but significantly elevated risk of breast cancer for occasional drinkers in studies that separate these from other types of abstainer. The first model investigated four studies with occasional drinker misclassification error. As would be expected, the overall OR for hazardous level drinkers was higher among studies free of misclassification error; however, the reverse occurred for low-level drinkers. There was also substantial between study heterogeneity.

Table 4. Models comparing studies with and without occasional drinker reference group bias.

Group	N Studies	Pooled OR	95% CI Lower	95% CI Upper	p	Tau 2	I2
Model for all studies with only occasional drinker reference group bias (n=4)
Pooled abstainers		1.000
Former drinkers	4	0.945	0.601	1.486	0.806	0.147	77.528
Occasional drinkers	2	-	-	-	-	-	-
Low level drinkers	3	1.217	1.001	1.345	0.000	0.009	75.425
Hazardous level drinkers	2	-	-	-	-	-	-
Harmful level drinkers	1	-	-	-	-	-	-
Model for all studies without occasional drinker reference group bias (n=38)
Abstainers		1.000
Former drinkers	7	1.036	0.961	1.118	0.354	0.087	76.577
Occasional drinkers	17	1.034	1.003	1.064	0.000	0.012	68.154
Low level drinkers	17	1.085	1.015	1.160	0.017	0.009	68.638
Hazardous level drinkers	26	1.374	1.319	1.431	0.000	0.016	54.598
Harmful level drinkers	9	1.336	1.228	1.454	0.000	0.101	12.604

Meta-analysis of studies with no reference group misclassification errors

Table 5 shows two meta-analyses: one of studies with both reference group biases (former and occasional drinkers in reference group) and the second with neither. Although study numbers were relatively small, both meta-analyses indicated significantly elevated risks both for low and hazardous level drinkers with only small differences in ORs. Of the seven studies used in the second model, one included reverse occasional drinker misclassification. When this study was removed in a sensitivity test, the ORs and p values remained almost unchanged. As with the other models, there was substantial cross study heterogeneity in the estimates suggesting the influence of unmeasured variables.

Table 5. Studies containing both misclassification errors and studies free of reference group misclassification.

Group	N Stud.	Pooled OR	95% CI Lower	95% CI Upper	p	Tau 2	I2
Model for all studies with both former and occasional drinker reference group bias (n=18)
Pooled abstainers		1.0
Low level drinkers	18	1.134	1.046	1.229	0.017	0.059	57.010
Hazardous level drinkers	11	1.315	1.254	1.380	0.000	0.007	25.132
Harmful level drinkers	2	-	-	-	-	-	-
Model for all studies with neither former or occasional drinker reference group bias (n=7)
Long-term abstainers		1.0
Former drinkers	7	1.036	0.961	1.118	0.354	0.066	69.541
Occasional drinkers	5	1.036	0.875	1.226	0.681	0.048	32.877
Low level drinkers	7	1.151	1.101	1.204	0.000	0.054	72.087
Hazardous level drinkers	6	1.379	1.079	1.762	0.010	0.054	71.350
Harmful level drinkers	1	-	-	-	-	-	-

Discussion

Meta-analyses of the relationship between alcohol consumption and breast cancer risk were conducted in order to investigate the possible presence of bias in risk estimates due to drinker misclassification errors. Our hypothesis that drinker misclassification errors would lead to downward bias in risk estimates was only partly supported.

A systematic literature search identified 60 unique studies quantifying the alcohol breast cancer relationship. In keeping with many previous studies, the main pooled meta-analysis of all 60 studies found significantly elevated risk of breast cancer for low, hazardous and harmful level drinkers. The significant 9.5% elevated risk estimated for drinkers we categorised as “low-level” (ie 1.5 to 24g per day) was, not surprisingly, similar to those in previous meta-analysis of this literature OR=1.06, CI: 1.00-1.11 (Ellison et al., 2001) (OR=1.06, CI: 1.00-1.11) and OR=1.09, CI: 1.04-1.13 (Boyle and Boffetta, 2009).

Only 7 of the 60 studies identified were completely free of both former and occasional drinker misclassification errors but one of these still contained reverse occasional drinker bias. However, there was no evidence of significant bias caused by the most common type of misclassification, namely the inclusion of former drinkers in the abstainer reference group (49 studies). None of the meta-analyses reported significantly elevated risk of breast cancer among former drinkers, including that of the 11 studies with estimates for former drinkers. By contrast, there was modest evidence of the potential for downward bias in risk estimates due to misclassification of occasional drinkers into the abstainer reference group. In studies that distinguished occasional drinkers, there was a small but significant 3.4% increase in risk estimated for occasional drinkers versus abstainers. It follows therefore, that results from studies which combined occasional drinkers with the abstainer reference group would tend towards the null as the reference category risk is biased upwards making the drinker group risks appear less marked by comparison. The potential for such downward bias was also suggested by non-significant ORs estimated (Table 4) for hazardous level alcohol use in studies with occasional drinker misclassification (26.3%) versus the significant elevated risk (37.4%) in studies free of this error.

While confirming the potential for drinker misclassification error to cause slight downward bias in estimates of the risk of breast cancer from consuming alcohol, all models consistently supported the conclusion that low-level alcohol intake is associated with an increased risk of breast cancer. If one discounts the role of former drinker misclassification error and relies on estimates from meta-analysis of all the 38 studies free from occasional drinker reference group bias, best estimates of the elevated risk of breast cancer at each drinking level of those are contained in the second model reported in Table 4. This indicates significantly elevated risks in the order of 8.5%, 37.4% and 33.6% respectively for low, hazardous and harmful-level drinkers. As discussed below, upper-levels of alcohol consumption in studies of women's health are poorly defined and this last estimate has wide confidence intervals.

This finding underlines concern that that low level drinking guidelines and physicians advice to patients both need to highlight risks from even low-level consumption which may not always be compensated for by apparent health benefits that may apply for other conditions such as coronary heart disease and diabetes (Latino-Martel et al., 2011, Stockwell et al., 2012a).

Strengths of the present study

The present study is the first to investigate the potential impact of drinker misclassification errors on meta-analytic outcomes specifically for the alcohol-breast cancer relationship. We provide results of pooled and subgroup analyses by misclassification type in comparison to a group of rigorous studies without such misclassifications. The confirmation of elevated breast cancer risk from even low-level drinking is important for alcohol and public health policy, especially given the prevalence of both low risk drinking among women and breast cancer. Our meta-analysis emphasizes the important methodological issue that failure to control for occasional drinker bias leads to underestimating risks from hazardous drinking.

Limitations

Only 6 of the 60 included studies rigorously controlled for all forms of drinker misclassification errors by separating out former and occasional drinkers not only from the abstainer reference group, but also from low risk drinkers. This limited number of misclassification error-free studies also proved to be a challenge for some subgroup analyses because of reduced power. Additional subgroup analyses of the drinker misclassification error free studies by other variables of interest, such as study design, disease outcome and ethnicity might have added useful information to the overall understanding of the alcohol-breast cancer link. The presence of substantial between study heterogeneity in the risk estimates of most models highlights the need for more in-depth investigation of the possible role of other factors such as ethnicity, socio-economic status, and other lifestyle risk factors as well as study design.

A particular methodological problem to highlight in the alcohol and breast cancer literature is the limited number of drinking categories most studies use - very often only two categories: low and hazardous drinking level. Arguably, one can only use the low level drinking category as a yardstick against which to compare estimates across the different study groups, since this category is reliably represented in each study.

Conclusions

While demonstrated to be a concern in relation to other categories of disease outcomes such as all-cause mortality and coronary heart disease (Fillmore et al., 2006), former drinker misclassification does not appear to be a significant concern in the breast cancer literature. There is, however, evidence of possible modest downward bias in risk estimates of the alcohol breast cancer relationship when there is occasional drinker misclassification present. This pattern of results may apply more generally to females across a range of health conditions or may be restricted to breast cancer. We are not aware of studies which have investigated gender differences in the impact of drinker misclassification errors and further research will need to determine this.

Importantly, and contrary to some other recent studies (Allen et al., 2009, Brown et al., 2010, Kabat et al., 2010), our meta-analysis shows consistently increased breast cancer risk for low-level women drinkers (between 1 drink/month to 2 drinks/day or 1.5 to 24 g/day). Our findings therefore strengthen the case for the conclusion that even low-level alcohol consumption poses a significantly increased risk for breast cancer.

Future directions and recommendations

Larger numbers of misclassification error-free studies (Breslow et al., 2011, Kinney et al., 2000, Li et al., 2003, Li et al., 2010, Swanson et al., 1997) are needed to derive more stable and nuanced conclusions regarding the alcohol-breast cancer relationship. The impact of drinker misclassifications on the accuracy of risk estimates for light drinking needs to be examined. Further, the theoretical risk remains that including occasional drinkers with low- risk drinkers may blur findings. We recommend, therefore, that in future researchers collect detailed information on participants' drinking which would enable the separation of former and occasional drinkers from abstainers and low-level drinkers. More preferable still, would be the detailed reporting of former drinker past drinking patterns and levels that would enable an intention-to-treat approach.

A final consideration is that even lifetime abstainers may not be the optimal comparison reference group in prospective studies on alcohol and health. It has been shown that lifetime abstainers tend to include those with ill health or low socioeconomic status (Fillmore et al., 1998). Recently, it has been confirmed that such a bias applies even to young adults who are lifetime abstainers (Ng Fat and Shelton, 2012). The preferred reference group may be an occasional drinker group consuming between 0.34 to 1.43 grams of alcohol per day (Rehm et al., 2007, Shaper et al., 1988) but this raises additional complexities and requires further investigation. The main implication is that risk estimates from all levels of alcohol consumption are likely underestimated due to unmeasured biases affecting the overall health outcomes and prospects of lifetime abstainers. We suggest that until there is further resolution on this point and more studies are conducted free of misclassification error that researchers, practitioners and policy makers maintain a precautionary approach in their recommendations and adopt a critical stance regarding the accuracy of the risk relationships so far estimated for different various disease outcomes, including breast cancer.