An exploration of the dynamic longitudinal relationship between mental health and alcohol consumption: a prospective cohort study

The Whitehall II prospective cohort study started with a sample of 10,308 British civil servants (6,895 men and 3,413 women), who were aged 34 to 56 years at entry into the study (1985 to 88) [38]. The current investigation uses of data from three clinical phases: 5 (1997 to 1999; referred to hereafter as 'baseline'), 7 (2002 to 2004) and 9 (2007 to 2009). At baseline, the total number of eligible participants was 7,870. Those who had not consumed alcohol in the year before baseline and additionally those with missing values for either alcohol consumption or mental health variables at baseline were excluded from the analytic sample (n = 548 and n = 1,036 respectively; categories were not mutually exclusive). This provided approximately 10 years of follow-up information for 6,330 participants who had consumed alcohol in the year before baseline.

The University College London Medical School Committee on the ethics of human research approved the Whitehall II study.

Participants were asked to report the number of drinks they had consumed in the previous week, quoting separately for beer/cider (pints), wine (glasses), and spirits (measures). Drinks were converted into UK units of alcohol (one unit is equivalent to 8 g of ethanol) using a conservative estimate of one UK unit for each measure of spirits and glass of wine, and two UK units for each pint of beer. These converted measurements were then summed to define the total weekly number of UK units consumed.

Mental health (combining symptoms of depression and anxiety) was assessed using the mental health component score [39, 40] (MCS) of the Short Form (SF)-36 questionnaire [41]. The SF-36 refers to symptoms experienced in the previous 4 weeks. The MCS has been validated using UK data sources [42], and reliability estimates typically exceed values of 0.90 for Cronbach’s α [39]. The MCS uses a scale of 0 to 100, with higher scores indicating better functioning.

Adjustment was made for several baseline covariates, including age, sex, ethnicity, socioeconomic status, marital status, highest educational qualification, economic activity, social network [43], smoking status, level of physical activity [44]. and use of anti-depressant medication. Problematic alcohol consumption (defined by the CAGE questionnaire [45]) was used to adjust for the possibility that problem drinking may be driving any observed relationship [46].

Poor physical health could influence both alcohol consumption [47, 48] and mental health [49] trajectories. Therefore, general physical health was accounted for by adjusting for several chronic conditions. A combination of self-report and validated clinical health events [50] were included, such as self-reported long-standing physical illness and belonging to the lowest sex-specific SF-36 physical health component quartile [51], as well as known diabetes mellitus, coronary heart disease (CHD), stroke, transient ischemic attack (TIA), total serum cholesterol, systolic and diastolic blood pressure, resting heart rate over 80 beats/min [52], and body mass index (BMI).

Bivariate latent change score (LCS) models [35, 53‐57] were used to explore the dominant temporal sequence in the longitudinal relationship between alcohol consumption and mental health symptoms. LCS models are an extension of standard growth curve models [58] (also referred to as random effects models) and acknowledge that repeated measures on the same individual are correlated. A general overview of the underlying assumptions and specification of LCS models are presented (see Additional file 1), but a comprehensive outline of the mathematical and statistical properties [55, 57], as well as a comparison of LCS models with other multivariate longitudinal models [35] can be found elsewhere.

Briefly, there are three primary parameters of interest: 1) the slope parameter (α), which refers to the additive sum of changes during follow-up; 2) the autoproportional parameter (β), which refers to the lagged effect of a variable on an upcoming change in itself (self-feedback); and 3) the coupling parameter (γ), which describes the lagged effect of one variable on the upcoming change in the alternate variable.

Both the intercept and the slope were fitted as random effects, allowing for them to vary between individuals. Intercepts and slopes (as well as their random effects) were correlated (ρ) both within a single process (for example, the alcohol consumption intercept with the alcohol slope) and between processes (for example, the mental health intercept with the alcohol slope). Intercepts and slopes were estimated conditional on the baseline covariates described above.

Four separate models were estimated: 1) no coupling (baseline) model; 2) alcohol consumption producing change in mental health model; 3) mental health producing change in alcohol consumption model; and 4) dynamic/reciprocal change model. Nested models were compared to determine the best-fitting model (that is, to justify the inclusion of either or both coupling parameters over a baseline model that did not contain them) using a χ² difference test.

The relative fit of the hypothesized models compared with the observed data was assessed using the Tucker–Lewis index (TLI), the comparative fit index (CFI), and the root mean squared error of approximation (RMSEA). Cut-off values close to 0.95 were used to determine a good fit for TLI and CFI, while a cut-off value close to 0.06 was used for RMSEA [59].

Models were estimated in Mplus v6.12 [60] using the full information maximum likelihood (FIML) estimator [61]. An α level of 0.05 was considered statistically significant for all analyses.

Table 1 displays descriptive statistics for the complete analytic sample (based on observed information only). The mean age of participants was 56 years, over 70% were male, and approximately 6% were non-white. Most were of high to intermediate socioeconomic status, almost four-fifths were married or cohabiting, over 60% had post-secondary or university level qualifications, and around 65 were economically active. Almost 11% were current smokers, 11% were identified as problem drinkers, and 70% were physically active. In terms of physical health, 6% of the sample had known CHD, 4% had known diabetes mellitus, 0.5% had experienced a stroke, 0.7% had experienced a TIA, almost 3% were currently being prescribed anti-depressant medication, and approximately half reported a long-standing illness.At baseline, participants consumed on average 14.5 UK units of alcohol per week, and this figure had reduced to 11 UK units by the end of follow-up. Mental health scores started at an average of 51 and increased to 54 (a random selection of observed (A) mental health and (B) alcohol consumption trajectories are displayed in Figure 1).

Indices related to model fit and statistics concerning model comparison are shown in Table 2. All models specified were well fitting according to commonly accepted thresholds of model fit as outlined above [59]. Detailed estimates for the best-fitting model are presented in Table 3,while only the fixed effect parameters are presented for other models specified in Table 4 (random effects for these models can be found in Additional file 2: Table S2A).

As the association was robust to adjustment for confounding factors, only the fully adjusted estimates will be discussed here (but age and sex, as well as fully adjusted estimates are presented). Furthermore, only parameters of primary interest will be highlighted.

A series of LCS models were estimated to test lag-leading and reciprocal relationships between weekly number of UK units consumed and mental health. In both minimally adjusted and fully adjusted models, it was found that a model in which mental health was specified as the leading indicator of change gave the best fit (Table 2, Table 3).Plotting the parameters of this model in a vector field (Figure 2) demonstrates the complex relationship between weekly alcohol consumption and mental health, and also helps to visualize the correlation between the mental health intercept and alcohol slope, which is difficult to interpret in isolation. It shows that participants who initially had poor mental health and low alcohol consumption increased their consumption between phases, whereas those with good to adequate mental health who drank at higher levels tended to decrease their consumption between phases while their mental health scores remained relatively stable. Furthermore, it shows that participants with poor mental health and high alcohol consumption on the previous occasion had a shallower decline in their consumption than those with good mental health drinking the same amount. This indicates that in addition to fuelling increases in alcohol consumption among low-level consumers, poor mental health may also be a maintaining factor for heavy alcohol consumption.

Our findings contradict the most recently published review on the relationship between alcohol use and depression [10], which concluded that increasing involvement with alcohol raises the risk of depression by two-fold. This review was, however, met with criticism [64, 65] for primarily being based on previous work by the authors themselves [9]. Furthermore, the review focused on alcohol-use disorders and major depression. Our work is therefore not directly comparable. As outlined earlier, we chose to focus on sub-syndrome symptoms of mental health and alcohol consumption (not problem consumption) as there has been a distinct lack of work exploring actual alcohol consumption (that is, what people drink) in this relationship; previous interest has largely been on the relationship between alcohol-use disorders and major depression. This makes drawing comparisons between our work and others complicated. It may be that there is something about the symptoms of problematic alcohol consumption that increases the risk of having [6, 7, 66‐71] or developing [9, 10, 72, 73] depression, independent of the amount of alcohol consumed [46, 74]. Recent work has shown that individuals who self-medicate symptoms of anxiety [75] or depression [12] with alcohol have an increased risk of developing (persistent) alcohol dependence. Therefore, it could be that the relationship we observed is part of a larger complex system involving a transition from sub-syndromal symptoms of mental health influencing changes in alcohol consumption (as in our analyses) until a certain threshold is reached, at which symptoms of alcohol dependence take over and increase the risk of developing clinical disorders [9, 10]; that is, there are two separate dynamic systems at play that influence alcohol consumption and mental health pre and post clinical disorder.

The approach that we took to modeling the relationship between alcohol use and mental health longitudinally utilized multiple measurement occasions to model change in both variables over time, which is known to improve the accuracy of estimated change [58, 76]. Previous work has also shown that it is important to consider variability in alcohol consumption [77], and the LCS model methodology directly incorporated individual change both in the total weekly alcohol consumption and in mental health. Furthermore, the method we used allowed for the effect of alcohol consumption on mental health and vice versa to be estimated simultaneously in the same model.

There are, however, several limitations of our study. First, data from phase 5 were used as the starting point, and it is possible that selective attrition may have occurred between the 'true' baseline (phase 1) and the baseline used in these analyses. This would result in a healthier cohort of participants being used to estimate the final model parameters, reducing the generalizability of the findings [78, 79]. Similarly, we used data from the Whitehall II cohort of British civil servants, which is not a representative sample of the general population. Work published using Whitehall II data has been highly influential in epidemiology and public health, shaping research agendas on social inequalities in health [80] and improving the understanding of the etiology of disease [81] but this limitation should be noted when considering the generalizability of our findings.

Second, one of the major concerns in alcohol epidemiology is measurement error in self-reported alcohol consumption [82]. It is acknowledged that self-reported measures of consumption are likely to be biased [82‐88], and therefore effect estimates obtained may actually be underestimates of the true association of interest. The use of latent variables (upon which LCS models are based; see Additional file 1) has been advocated in the field of alcohol epidemiology [89] to account for this known measurement error. Additionally, the MCS scale of the SF-36 is not solely concerned with psychiatric symptoms but also with mental health-related quality of life (although evidence exists to suggest that high MCS scores are associated with clinical depression [40, 90, 91]). It is possible that the relationship between alcohol intake and mental health might differ if other psychiatric questionnaires were used to define symptoms of mental health, or if the distinction was made between symptoms of depression and anxiety. However, it is argued that in practice it is difficult to effectively determine specific characteristics of depression from symptoms of, for example, anxiety using self-report measures of symptoms because of the considerable heterogeneity of symptoms between disorders (that is, self-reported symptoms often reflect a comorbidity between depression and other mood/stress-related disorders [92‐95]). This has led some investigators to conclude that self-report measures of mental health symptoms at a population level merely reflect a single underlying latent construct of psychological distress [96‐99].

Another issue concerning the main measures used in this study is that they refer to different time periods; information on alcohol consumption pertained to the previous week whereas information on mental health symptoms referred to the previous 4 weeks. It is possible that this discrepancy in the period of reference may have biased our findings. For example, smaller studies looking at the relationship between mood and alcohol on a daily basis have shown that increased alcohol consumption is associated with decreased happiness on the following day, and that symptoms of sadness are associated with decreased consumption on the next day [100]. These findings contrast with our own, and highlight the importance of the timeframe used in determining the best-fitting temporal sequence between alcohol consumption and mental health.

Furthermore, the competing models we specified allowed only for the previous occasion's alcohol and/or mental health score to influence change in the alternative variable by the next occasion. It is plausible that the relationship might have differed if we had allowed for longer lag specifications, as it may be that the relationship between alcohol intake and mental health takes longer to manifest (that is, the current specification of a single cross-lagged effect may fit the relationship between mental health influencing alcohol intake better than the relationship between alcohol consumption influencing mental health symptoms).

Additionally, there was greater variation in the measure of weekly alcohol intake than in that of mental health. It could be argued that this could also be a possible explanation as to why alcohol consumption was not found to be significantly related to changes in mental health. It may be that within a dynamic system that it is more difficult to effectively predict changes in one variable using a highly erratic alternative exposure. It is conceivable that the reciprocal model might have been of best fit had both measures been relatively stable over time.

A further methodological limitation is that we controlled only for baseline covariate values, and it is possible that their values changed over time. For example, comorbidities may have developed after the first measurement occasion. Health behaviors such as physical activity and smoking could also vary over time, and the changing status of these variables could all be confounders of the subsequent effects of alcohol on mental health and vice versa. However, factoring in changes in the covariate structure over time could be problematic within the current framework, because changes in some values, for example, systolic blood pressure, could be a direct consequence of previous alcohol consumption or mental health status, and thus be considered as intermediate confounders [101].

We identified that the dominant process underlying the dynamic relationship between alcohol consumption and mental health at a population level is mental health. Consequently, it could be inferred that targeting interventions to those with poor mental health (as well as introducing measures to ensure that those with normal/good mental health do not deteriorate) would have a beneficial effect in terms of reducing heavy drinking. This may also elicit favorable knock-on effects in terms of improving general physical health and reducing the risk of chronic diseases, as heavy drinking itself is associated with an increased risk of a range of health problems [102, 103] including cardiovascular disease [104‐109], cancer, [110, 111] and mortality [47, 112‐114]. Furthermore, the finding that mental health affects alcohol consumption may shed some light on the growing literature examining common mental disorders as risk factors for cardiovascular disease [115‐119] and all-cause mortality [120‐122], because alcohol consumption may be one of many mediators in this relationship.

This work provides further support that on-going efforts to improve mental health at a population level are vital to public health [123, 124]. The proposed implementation strategy [124] seeks both to tackle he social determinants of mental health [125] and to target individuals who are at high risk. To do so, a number of avenues will be pursued, including tackling inequalities in access to services (and ensuring equality in the level of service provided). In addition, conscious efforts are being made to tackle the stigma surrounding mental health issues; perhaps if individuals feel more comfortable talking about their mental health problems or seeking treatment for them, then they will not turn to alcohol as a form of self-medication.

Others may, however, be more cynical of our findings and take them to indicate that 1) consuming large amounts of alcohol is acceptable as it does not increase the risk of developing mental health problems, and 2) that it is reasonable to self-medicate with alcohol in response to psychological distress, as it will not worsen symptoms. However, it would be unwise to use our findings as a justification for drinking in a hazardous manner. Although a person’s mental health may not worsen, as highlighted above, increased alcohol consumption would heighten their risk of developing other disorders.

Regarding future work, it is important to examine the role of drinking pattern as well as to provide closer scrutiny of age (e.g. adolescent and elderly populations), sex, socioeconomic, and cultural differences in the dynamic relationship between alcohol consumption and mental health. It is also important that subsequent studies should examine the extent to which time-varying/modified confounding may explain the association observed using appropriate analytic methods [101]. Furthermore, it is also imperative that potential physiological and psychosocial mechanisms, both occurring alongside and precipitating (immediately or earlier in life) the parallel development of both trajectories are studied. This has been acknowledged by others in the field [10, 64, 65]. Understanding the factors that trigger increased alcohol consumption in the presence of poor mental health will allow for more effective interventions to be developed, both in terms of treatment and primary prevention.