Follow-Up Data Improve the Estimation of the Prevalence of Heavy Alcohol Consumption

Follow-Up Data Improve the Estimation of the Prevalence of Heavy Alcohol Consumption Abstract Aims We aim to adjust for potential non-participation bias in the prevalence of heavy alcohol consumption. Methods Population survey data from Finnish health examination surveys conducted in 1987–2007 were linked to the administrative registers for mortality and morbidity follow-up until end of 2014. Utilising these data, available for both participants and non-participants, we model the association between heavy alcohol consumption and alcohol-related disease diagnoses. Results Our results show that the estimated prevalence of heavy alcohol consumption is on average of 1.5 times higher for men and 1.8 times higher for women than what was obtained from participants only (complete case analysis). The magnitude of the difference in the mean estimates by year varies from 0 to 9 percentage points for men and from 0 to 2 percentage points for women. Conclusion The proposed approach improves the prevalence estimation but requires follow-up data on non-participants and Bayesian modelling. INTRODUCTION Reliable information about the prevalence of heavy alcohol consumption is important because alcohol-related health problems and undesired social consequences (Klingemann and Gmel, 2001) cause significant costs in many countries (Rehm et al., 2009). Prevalence estimates can be obtained through health surveys but the low participation rates (Galea and Tracy, 2007) imperil the reliability of the results. If the participation is selective with respect to alcohol consumption, the estimates of alcohol use suffer from non-participation bias, which hinders their usability for decision-making. If non-participants have worse health than participants, the bias usually leads to an overly positive image of the health of the population. Empirical evidence suggests that participation is often selective concerning alcohol consumption. Studies from Canada (Zhao et al., 2009), England (Boniface et al., 2017), Finland (Karvanen et al., 2016; Kopra et al., 2017), Norway (Torvik et al., 2012), Scotland (Gorman et al., 2014) Sweden (Romelsjö, 1989) and the USA (Dawson et al., 2014) conclude that non-participants drink more alcohol than participants. Follow-up studies have shown that non-participants tend to have a higher risk of alcohol-related diseases (Romelsjö, 1989; Jousilahti et al., 2005; Gorman et al., 2014; Christensen et al., 2015; Karvanen et al., 2016), and increased risk of hospitalisations and death (Jousilahti et al., 2005; Christensen et al., 2015; Karvanen et al., 2016), which indicates that non-participants tend to use more alcohol than participants. A study from the Netherlands (Lahaut et al., 2002) found that non-participants are more often abstainers than participants, which is not directly interpretable as a conflicting result because many social factors may be associated with abstaining. An older study from Sweden did not find an indication of selective participation (Halldin, 1985). In addition to selective non-participation, bias may be introduced by imperfect coverage of the target population by the survey sampling frame and by questionnaire design. First, if some individuals of target population cannot be invited to a survey, the sample does not represent the population of interest and the estimates will be biased. Mäkelä and Huhtanen (2010) observed that in Finland, persons who cannot be invited to a survey due to missing home address have about four times higher risk for alcohol-related deaths. This caused a small bias in the population estimates. Second, (Livingston and Callinan, 2015) claim that quantity–frequency design of the alcohol use questions underestimates alcohol consumption by one-third compared with asking about drinking with a within-location beverage-specific design. Gmel (2000) reported that alcohol as a subject of survey study does not have an impact on participation in comparison to similar questionnaire without alcohol-related questions. Studies from the USA (Dawson et al., 2014) and Finland (Mäkelä, 2003) have shown that the non-participation bias cannot be adjusted using just weights depending on basic demographic variables. Some studies adjust for selective non-participation utilising continuum of resistance model (Zhao et al., 2009; Meiklejohn et al., 2012; Boniface et al., 2017) but there are also other methods (Karvanen et al., 2016; Kopra et al., 2017). In Finnish health surveys, participation has been decreasing, while reported alcohol consumption has been mainly increasing from the 1960s to 2007 (Mäkelä et al., 2012). Jousilahti et al. (2005) report that in Finland non-participants have a higher risk of alcohol-specific diseases and death (Jousilahti et al., 2005), which is why we expect the estimates of heavy alcohol consumption to be biased. The difference in the disease risk could be explained by heavier alcohol consumption among non-participants. From previous studies (Harald et al., 2007; Hirvonen, 2017), we know that participation in the FINRISK Study is affected by age, gender, area and education (Reinikainen et al., 2017). We aim to adjust for selective non-participation for heavy alcohol consumption and to estimate the prevalence of heavy alcohol consumption with reduced bias. We present a Bayesian solution that is based on mortality and morbidity follow-up data. METHODS Data We used data from the National FINRISK Study, which is a series of cross-sectional health examination surveys (Borodulin et al., 2017) conducted in Finland every fifth year since 1972. We analysed data for the years 1987–2007. Years 1972–82, as well as 2012, were excluded because the questions of alcohol consumption were too different from the questions in 1987–2007. In 1987 and 1992 studies, the questions were essentially the same. In 1997 study, questions regarding the usage of cider or mild wine (alc. vol under 5%) were added. Otherwise, the study remained the same as earlier. In 2002, the questions regarding the consumption of red wine and other wines were separated from each other. The total alcohol consumption was not measured several alcohol beverage-specific questions but the participants were advised to calculate the number of alcohol portions (standard drinks) consumed and to mark the corresponding category in the quantity–frequency table. In 2007, the alcohol questions were the same as in 2002, but the instructions for the calculation of daily alcohol consumption were improved. The surveys provide data on 25–74-year-old adults from six regions of Finland. We restrict the data to people aged 25–65 years since oldest age group 65–74 years old was not available in all areas until 2007. The survey consists of questionnaires and a health examination carried out at a local study clinic. The sample was drawn from the National Population Register and was stratified by region, gender and 10-year age group. In total, there are 44,317 invitees including 31,567 participants. The survey data contains self-reported alcohol consumption and background variables, age, gender, region and study year for the whole sample. The survey utilised beverage-specific quantity–frequency questions on alcohol use in the first three surveys and graduate frequency measure in the latter two surveys. The questions related to alcohol consumption are provided in Appendix B. The study questionnaires in 1987, 1992 and 1997 asked alcohol usage one type of alcohol beverages at a time: beer, spirits/vodka, long drink or cider (in 1997), wines and mild wines (in 1992 and 1997). In 2002 and 2007, the questionnaire was different and individuals reported their alcohol consumption for all beverage types in one question. From these questions, a number of standard drinks consumed per week during a previous 12 months was calculated. One standard drink equals 12 g of pure alcohol which is equivalent to, e.g. one bottle of beer (33cl, 4.7 volume percent of alcohol). Based on the number of standard drinks, the (self-reported) total amount of 100% alcohol consumed in the previous 12 months was calculated. Since our main interest was to estimate the prevalence of heavy alcohol users, we classified participants as heavy alcohol consumers and others (non-heavy alcohol consumers) as follows. The persons who reported consuming on average at least 24 standard drinks per week for men or at least 16 standard drinks per week for women during the one-year period before the examination were considered heavy alcohol consumers. The survey data were linked to three registers: The Register of Completed Education and Degrees (Statistics Finland, 2016), Care Register for Health Care (National Institute for Health and Welfare, 2017) and Cause of Death Register (Official Statistics of Finland, 2017) using personal identification code. The register-based data were available for both participants and non-participants. The level of education is categorised according to the International Standard Classification of Education (ISCED, 2011) levels: (1) high level (tertiary education, ISCED levels 5–8), (2) middle level (secondary education, ISCED levels 3–4) and (3) low level (primary education or less or unknown, ISCED levels 0–2). The Care Register gives data about the hospital visits with dates and ICD codes for both participants and non-participants. From the Causes of Death Register, we obtain data about dates and ICD codes of the cause of death. Follow-up data contains the time-to-event (age) and ICD code of the first alcohol-related disease diagnosis or death. The ICD codes we considered to be alcohol-related are listed in Table 1. The follow-up begins from the survey and ends at the end of 2014. Persons who have neither alcohol-related disease diagnosis nor the alcohol-related cause of death are censored at the end of the follow-up. Deaths not related to alcohol are treated as censorings. Table 1. The ICD-codes interpreted as alcohol-related events ICD-9  291 Alcohol-induced mental disorders  303 Alcohol dependence syndrome  357.5 Alcoholic polyneuropathy  425.5 Alcoholic cardiomyopathy  535.3 Alcoholic gastritis  571.0 Alcoholic fatty liver  571.1 Acute alcoholic liver disease  571.2 Alcoholic cirrhosis of liver  571.3 Alcoholic liver damage, unspecified  577.0D-F Alcoholic disease of the pancreas, acute  577.1C-D Alcoholic disease of the pancreas, chronic  980.0 Toxic effect ethyl alcohol  980.2 Toxic effect of isopropyl alcohol  980.8 Toxic effect of other specified alcohols  980.9 Toxic effect of other unspecified alcohol  E851 Accidental poisoning by alcohol ICD-10  F10 Mental and behavioural disorders due to use of alcohol  G31.2 Degeneration of nervous system due to alcohol  G62.1 Alcoholic polyneuropathy  G72.1 Alcoholic myopathy  I42.6 Alcoholic cardiomyopathy  K29.2 Alcoholic gastritis  K70 Alcoholic liver disease  K85.2 Alcohol-induced acute pancreatitis  K86.0 Alcohol-induced chronic pancreatitis  T51 Toxic effect of alcohol  X45 Accidental poisoning or other exposure to alcohol  Y15 Poisoning by and exposure to alcohol, undetermined intent ICD-9  291 Alcohol-induced mental disorders  303 Alcohol dependence syndrome  357.5 Alcoholic polyneuropathy  425.5 Alcoholic cardiomyopathy  535.3 Alcoholic gastritis  571.0 Alcoholic fatty liver  571.1 Acute alcoholic liver disease  571.2 Alcoholic cirrhosis of liver  571.3 Alcoholic liver damage, unspecified  577.0D-F Alcoholic disease of the pancreas, acute  577.1C-D Alcoholic disease of the pancreas, chronic  980.0 Toxic effect ethyl alcohol  980.2 Toxic effect of isopropyl alcohol  980.8 Toxic effect of other specified alcohols  980.9 Toxic effect of other unspecified alcohol  E851 Accidental poisoning by alcohol ICD-10  F10 Mental and behavioural disorders due to use of alcohol  G31.2 Degeneration of nervous system due to alcohol  G62.1 Alcoholic polyneuropathy  G72.1 Alcoholic myopathy  I42.6 Alcoholic cardiomyopathy  K29.2 Alcoholic gastritis  K70 Alcoholic liver disease  K85.2 Alcohol-induced acute pancreatitis  K86.0 Alcohol-induced chronic pancreatitis  T51 Toxic effect of alcohol  X45 Accidental poisoning or other exposure to alcohol  Y15 Poisoning by and exposure to alcohol, undetermined intent Table 1. The ICD-codes interpreted as alcohol-related events ICD-9  291 Alcohol-induced mental disorders  303 Alcohol dependence syndrome  357.5 Alcoholic polyneuropathy  425.5 Alcoholic cardiomyopathy  535.3 Alcoholic gastritis  571.0 Alcoholic fatty liver  571.1 Acute alcoholic liver disease  571.2 Alcoholic cirrhosis of liver  571.3 Alcoholic liver damage, unspecified  577.0D-F Alcoholic disease of the pancreas, acute  577.1C-D Alcoholic disease of the pancreas, chronic  980.0 Toxic effect ethyl alcohol  980.2 Toxic effect of isopropyl alcohol  980.8 Toxic effect of other specified alcohols  980.9 Toxic effect of other unspecified alcohol  E851 Accidental poisoning by alcohol ICD-10  F10 Mental and behavioural disorders due to use of alcohol  G31.2 Degeneration of nervous system due to alcohol  G62.1 Alcoholic polyneuropathy  G72.1 Alcoholic myopathy  I42.6 Alcoholic cardiomyopathy  K29.2 Alcoholic gastritis  K70 Alcoholic liver disease  K85.2 Alcohol-induced acute pancreatitis  K86.0 Alcohol-induced chronic pancreatitis  T51 Toxic effect of alcohol  X45 Accidental poisoning or other exposure to alcohol  Y15 Poisoning by and exposure to alcohol, undetermined intent ICD-9  291 Alcohol-induced mental disorders  303 Alcohol dependence syndrome  357.5 Alcoholic polyneuropathy  425.5 Alcoholic cardiomyopathy  535.3 Alcoholic gastritis  571.0 Alcoholic fatty liver  571.1 Acute alcoholic liver disease  571.2 Alcoholic cirrhosis of liver  571.3 Alcoholic liver damage, unspecified  577.0D-F Alcoholic disease of the pancreas, acute  577.1C-D Alcoholic disease of the pancreas, chronic  980.0 Toxic effect ethyl alcohol  980.2 Toxic effect of isopropyl alcohol  980.8 Toxic effect of other specified alcohols  980.9 Toxic effect of other unspecified alcohol  E851 Accidental poisoning by alcohol ICD-10  F10 Mental and behavioural disorders due to use of alcohol  G31.2 Degeneration of nervous system due to alcohol  G62.1 Alcoholic polyneuropathy  G72.1 Alcoholic myopathy  I42.6 Alcoholic cardiomyopathy  K29.2 Alcoholic gastritis  K70 Alcoholic liver disease  K85.2 Alcohol-induced acute pancreatitis  K86.0 Alcohol-induced chronic pancreatitis  T51 Toxic effect of alcohol  X45 Accidental poisoning or other exposure to alcohol  Y15 Poisoning by and exposure to alcohol, undetermined intent Complete case analysis The complete case analysis (e.g. mean estimate from the participants) assumes that participation is not selective concerning alcohol consumption. Violations of this assumption lead to bias. We compared the results of complete case analysis with a Bayesian approach, which relies on more realistic assumptions and allows for selective non-participation concerning heavy alcohol use. Modelling approach We applied a Bayesian approach introduced in Kopra et al. (2018) to estimate the prevalence of heavy alcohol consumption. The Bayesian model consists of three sub-models which are fitted simultaneously. The sub-models are: participation model, risk factor model and survival model. The mathematical formulas for the models are given in Appendix A. The participation model describes which variables affect participation. Participation is defined as a binary indicator (0 or 1) for the availability information on alcohol consumption. This model is a logistic regression model with linear covariates for study year and age, and categorical variables for the region (4 levels), education (3 levels) and the alcohol consumption (binary). The model also takes into account the possible interactions of gender and study year, gender and alcohol consumption, and study year and alcohol consumption. The risk factor model describes how alcohol consumption (heavy or non-heavy) varies by background variables. By background variables, we mean age, gender, region, study year and education. The model is a logistic regression model with interactions for the year of birth with gender, region, study year and education. The survival model describes the relationship between alcohol consumption and alcohol-related diseases. All disease events are combined and modelled as one survival outcome. The survival model is a piecewise constant hazard model with one-year baseline hazard period terms. The model assumes monotonically increasing baseline hazard, which is accomplished using prior specification. In addition to baseline hazard, alcohol consumption is used as a regressor. Both baseline hazard terms and the regression coefficient are gender-specific. Our estimation method requires that the average disease risk of non-participants must be between the average risks of heavy alcohol consumer participants and other participants. This technical requirement can be evaluated from the data. Prior distributions We used weakly informative prior distributions that reflect the existing knowledge but have variances large enough to allow for surprises. This approach is recommended in textbooks on Bayesian statistics (Gelman et al., 2014) and there exists guidelines for elicitation of prior distributions (O’Hagan et al., 2006). The participation model needed an informative prior for the effect of heavy alcohol consumption on the participation, i.e. for the strength of selectivity mechanism. Some degree of subjectivity cannot be avoided in the prior specification. To define a weakly informative prior, we took a 45-year-old non-heavy alcohol consumer who participates with probability 0.7 as a reference and considered the prior probability for a heavy alcohol consumer who is otherwise similar. We elicit that there is 25% chance that person participates with probability p lower than 0.5 (P(p ≤ 0.5) = 0.25), 35% chance for p ≤ 0.6, and 50% chance for p ≤ 0.7. The functional form of the prior distribution was chosen to be logistic distribution. These elicitations lead to logistics prior distribution with expected value zero and variance 1/2.05. In the survival model, we applied monotonically increasing baseline hazards separately for men and women. The prior distribution for the first hazard term (25–26-year-olds) was a uniform distribution with a range from 0 to 20. From second hazard term (26–27-year-olds) to the last hazard term (99–100-year-olds), each term had a uniform distribution with the lower limit being the value of previous baseline hazard term and upper limit 20. All the remaining model parameters had normally distributed priors with zero mean and variance 1000. The prior distributions are presented using mathematical notation in Table A1. Imputations and model fitting Alcohol consumption was missing for the non-participants. These missing values (heavy or non-heavy) were imputed simultaneously with Bayesian model fitting using data augmentation (Tanner and Wong, 1987). The model was fitted using Markov chain Monte Carlo (MCMC) (Robert and Casella, 2004) and implemented with Just Another Gibbs Sampler (JAGS)-software (Plummer, 2003) and R software (R Foundation for Statistical Computing, 2017) with rjags-package (Plummer, 2015). The convergence of MCMC chains was investigated using Brooks–Gelman Rˆ diagnostics (Brooks and Gelman, 1998) and all the Rˆs were ≤1.01 which indicates convergence. The model fitting utilised computational resources of IT Center for Science Ltd (CSC). RESULTS Descriptive statistics In Table 2, we present the descriptive statistics on age, education and gender. These variables are examined by study year, and comparisons can be made between participants and non-participant as well as between heavy alcohol consumers and other alcohol consumers. Table 2. Description of background information by study year for non-participants, participants, and heavy and non-heavy alcohol consumers among participants Year Non-participants Participants All Heavy alcohol consumers Moderate alcohol consumers Average age  1987 42.5 44.4 41.7 44.4  1992 41.8 44.7 44.5 44.7  1997 42.8 45.0 45.0 45.0  2002 42.3 44.9 45.7 44.8  2007 41.9 45.6 47.3 45.5 High education (%)  1987 13.6 18.5 23.0 18.4  1992 18.4 26.6 30.2 26.5  1997 24.7 29.9 31.4 29.8  2002 25.6 35.7 32.7 35.8  2007 27.7 38.6 34.3 38.8 Middle education (%)  1987 30.6 31.8 31.0 31.8  1992 35.6 34.9 34.4 34.9  1997 37.7 38.1 37.6 38.1  2002 41.8 40.6 43.5 40.4  2007 45.3 44.2 45.5 44.1 Low education (%)  1987 55.8 49.7 46.0 49.8  1992 46.1 38.5 35.4 38.6  1997 37.6 32.1 31.0 32.1  2002 32.6 23.8 23.9 23.8  2007 27.0 17.2 20.2 17.1 Women (%)  1987 42.3 52.0 15.9 52.7  1992 40.5 53.0 21.2 54.1  1997 43.2 52.9 22.8 54.3  2002 41.6 53.6 29.0 54.9  2007 42.7 53.9 27.8 55.3 Year Non-participants Participants All Heavy alcohol consumers Moderate alcohol consumers Average age  1987 42.5 44.4 41.7 44.4  1992 41.8 44.7 44.5 44.7  1997 42.8 45.0 45.0 45.0  2002 42.3 44.9 45.7 44.8  2007 41.9 45.6 47.3 45.5 High education (%)  1987 13.6 18.5 23.0 18.4  1992 18.4 26.6 30.2 26.5  1997 24.7 29.9 31.4 29.8  2002 25.6 35.7 32.7 35.8  2007 27.7 38.6 34.3 38.8 Middle education (%)  1987 30.6 31.8 31.0 31.8  1992 35.6 34.9 34.4 34.9  1997 37.7 38.1 37.6 38.1  2002 41.8 40.6 43.5 40.4  2007 45.3 44.2 45.5 44.1 Low education (%)  1987 55.8 49.7 46.0 49.8  1992 46.1 38.5 35.4 38.6  1997 37.6 32.1 31.0 32.1  2002 32.6 23.8 23.9 23.8  2007 27.0 17.2 20.2 17.1 Women (%)  1987 42.3 52.0 15.9 52.7  1992 40.5 53.0 21.2 54.1  1997 43.2 52.9 22.8 54.3  2002 41.6 53.6 29.0 54.9  2007 42.7 53.9 27.8 55.3 Table 2. Description of background information by study year for non-participants, participants, and heavy and non-heavy alcohol consumers among participants Year Non-participants Participants All Heavy alcohol consumers Moderate alcohol consumers Average age  1987 42.5 44.4 41.7 44.4  1992 41.8 44.7 44.5 44.7  1997 42.8 45.0 45.0 45.0  2002 42.3 44.9 45.7 44.8  2007 41.9 45.6 47.3 45.5 High education (%)  1987 13.6 18.5 23.0 18.4  1992 18.4 26.6 30.2 26.5  1997 24.7 29.9 31.4 29.8  2002 25.6 35.7 32.7 35.8  2007 27.7 38.6 34.3 38.8 Middle education (%)  1987 30.6 31.8 31.0 31.8  1992 35.6 34.9 34.4 34.9  1997 37.7 38.1 37.6 38.1  2002 41.8 40.6 43.5 40.4  2007 45.3 44.2 45.5 44.1 Low education (%)  1987 55.8 49.7 46.0 49.8  1992 46.1 38.5 35.4 38.6  1997 37.6 32.1 31.0 32.1  2002 32.6 23.8 23.9 23.8  2007 27.0 17.2 20.2 17.1 Women (%)  1987 42.3 52.0 15.9 52.7  1992 40.5 53.0 21.2 54.1  1997 43.2 52.9 22.8 54.3  2002 41.6 53.6 29.0 54.9  2007 42.7 53.9 27.8 55.3 Year Non-participants Participants All Heavy alcohol consumers Moderate alcohol consumers Average age  1987 42.5 44.4 41.7 44.4  1992 41.8 44.7 44.5 44.7  1997 42.8 45.0 45.0 45.0  2002 42.3 44.9 45.7 44.8  2007 41.9 45.6 47.3 45.5 High education (%)  1987 13.6 18.5 23.0 18.4  1992 18.4 26.6 30.2 26.5  1997 24.7 29.9 31.4 29.8  2002 25.6 35.7 32.7 35.8  2007 27.7 38.6 34.3 38.8 Middle education (%)  1987 30.6 31.8 31.0 31.8  1992 35.6 34.9 34.4 34.9  1997 37.7 38.1 37.6 38.1  2002 41.8 40.6 43.5 40.4  2007 45.3 44.2 45.5 44.1 Low education (%)  1987 55.8 49.7 46.0 49.8  1992 46.1 38.5 35.4 38.6  1997 37.6 32.1 31.0 32.1  2002 32.6 23.8 23.9 23.8  2007 27.0 17.2 20.2 17.1 Women (%)  1987 42.3 52.0 15.9 52.7  1992 40.5 53.0 21.2 54.1  1997 43.2 52.9 22.8 54.3  2002 41.6 53.6 29.0 54.9  2007 42.7 53.9 27.8 55.3 The average age of the non-participants was lower than the average age of the participants. Over the years, the average age appears not to have changed much for the non-participants but it has slightly increased for the participants. Among participants, the average age of heavy alcohol consumers has increased more rapidly than for non-heavy alcohol consumers. The average age of heavy alcohol consumers was 41.7 in 1987 (44.4 for non-heavy) and it has increased between each study being 47.3 for heavy alcohol consumers and 45.5 for non-heavy alcohol consumers in 2007. The average age of non-heavy alcohol consumers has also increased between the studies, except between the 1997 and 2002 when it decreased by 0.2 years. The level of education has increased for both participants and non-participants during the study period. The non-participants tend to have low education more often than participants, and participants tend to have high education more often than non-participants. In 1987, there were a higher proportion of highly educated participants among heavy alcohol consumers than among non-heavy alcohol consumers. In 2007, the situation was opposite; the proportion of highly educated persons is higher for non-heavy alcohol consumers than for the heavy alcohol consumers. The proportion of women among participants has slightly increased from 52.0% to 53.4% during 1987–2007. Among non-heavy alcohol consumers, the proportion is higher: 52.7–55.3%. Women are a minority among heavy alcohol consumers. There were 15.9% women among heavy alcohol consumers in 1987, and the proportion has notably increased being 27.8% in 2007. The proportion of women was higher among the participants than among non-participants. The proportion of women among participating heavy alcohol users has been rapidly increasing over the years, while the corresponding proportion had not increased by much among non-heavy alcohol consumers. The number of invitees, the participation rate and the number of events for both participant and non-participant men and women are presented in Table 3. During the study period, the proportion of heavy alcohol consumers has increased for both men and women among participants, and simultaneously the participation rate has decreased. Table 3. Number of invitees, the participation rate, the prevalence of heavy alcohol consumption based on participants and Bayesian modelling (posterior mean), and the number of alcohol-related incident events (per 1000 follow-up years) for the non-participant and the participant men and women Year Invited N Participation rate (%) Prevalence for participants (%) Posterior mean (%) Alcohol-related incident events (per 1000 follow-up years) Participants Non-participants Men  1987 3910 79.5 5.0 9.6 202 (2.8) 91 (5.2)  1992 3888 73.3 7.7 15.0 168 (2.9) 128 (6.5)  1997 4034 70.0 9.2 9.7 150 (3.2) 103 (5.4)  2002 3955 66.5 14.4 22.9 118 (3.6) 62 (3.8)  2007 3202 61.8 11.0 12.8 47 (3.1) 35 (3.7) Women  1987 3961 85.1 0.7 2.4 52 (0.6) 29 (2.0)  1992 3951 81.0 2.5 4.4 46 (0.7) 23 (1.5)  1997 4031 75.8 3.7 5.3 25 (0.5) 37 (2.3)  2002 4019 75.4 5.7 5.9 34 (0.9) 13 (1.0)  2007 3278 71.3 4.0 5.8 12 (0.7) 10 (1.3) Year Invited N Participation rate (%) Prevalence for participants (%) Posterior mean (%) Alcohol-related incident events (per 1000 follow-up years) Participants Non-participants Men  1987 3910 79.5 5.0 9.6 202 (2.8) 91 (5.2)  1992 3888 73.3 7.7 15.0 168 (2.9) 128 (6.5)  1997 4034 70.0 9.2 9.7 150 (3.2) 103 (5.4)  2002 3955 66.5 14.4 22.9 118 (3.6) 62 (3.8)  2007 3202 61.8 11.0 12.8 47 (3.1) 35 (3.7) Women  1987 3961 85.1 0.7 2.4 52 (0.6) 29 (2.0)  1992 3951 81.0 2.5 4.4 46 (0.7) 23 (1.5)  1997 4031 75.8 3.7 5.3 25 (0.5) 37 (2.3)  2002 4019 75.4 5.7 5.9 34 (0.9) 13 (1.0)  2007 3278 71.3 4.0 5.8 12 (0.7) 10 (1.3) Table 3. Number of invitees, the participation rate, the prevalence of heavy alcohol consumption based on participants and Bayesian modelling (posterior mean), and the number of alcohol-related incident events (per 1000 follow-up years) for the non-participant and the participant men and women Year Invited N Participation rate (%) Prevalence for participants (%) Posterior mean (%) Alcohol-related incident events (per 1000 follow-up years) Participants Non-participants Men  1987 3910 79.5 5.0 9.6 202 (2.8) 91 (5.2)  1992 3888 73.3 7.7 15.0 168 (2.9) 128 (6.5)  1997 4034 70.0 9.2 9.7 150 (3.2) 103 (5.4)  2002 3955 66.5 14.4 22.9 118 (3.6) 62 (3.8)  2007 3202 61.8 11.0 12.8 47 (3.1) 35 (3.7) Women  1987 3961 85.1 0.7 2.4 52 (0.6) 29 (2.0)  1992 3951 81.0 2.5 4.4 46 (0.7) 23 (1.5)  1997 4031 75.8 3.7 5.3 25 (0.5) 37 (2.3)  2002 4019 75.4 5.7 5.9 34 (0.9) 13 (1.0)  2007 3278 71.3 4.0 5.8 12 (0.7) 10 (1.3) Year Invited N Participation rate (%) Prevalence for participants (%) Posterior mean (%) Alcohol-related incident events (per 1000 follow-up years) Participants Non-participants Men  1987 3910 79.5 5.0 9.6 202 (2.8) 91 (5.2)  1992 3888 73.3 7.7 15.0 168 (2.9) 128 (6.5)  1997 4034 70.0 9.2 9.7 150 (3.2) 103 (5.4)  2002 3955 66.5 14.4 22.9 118 (3.6) 62 (3.8)  2007 3202 61.8 11.0 12.8 47 (3.1) 35 (3.7) Women  1987 3961 85.1 0.7 2.4 52 (0.6) 29 (2.0)  1992 3951 81.0 2.5 4.4 46 (0.7) 23 (1.5)  1997 4031 75.8 3.7 5.3 25 (0.5) 37 (2.3)  2002 4019 75.4 5.7 5.9 34 (0.9) 13 (1.0)  2007 3278 71.3 4.0 5.8 12 (0.7) 10 (1.3) The probabilities for not having alcohol-related disease diagnosis up to the given age for men and women are presented by Kaplan–Meier survival plots (Fig. 1). The top row shows that the non-participants were more likely to have alcohol-related diagnoses than participants. The lower row shows that the risk for non-participants lies between the risks of heavy and non-heavy alcohol consumers, which is a requirement for the utilised Bayesian model. The number of persons with a disease diagnosed in each group is reported next to the survival curve in the Fig. 1. Fig. 1. View largeDownload slide Kaplan–Meier survival plots for men and women comparing the probabilities of not having alcohol-related diagnoses among participants and non-participants (upper panels) and for heavy, non-heavy alcohol consumers and non-participants (lower panels). The number of persons with a disease diagnosed in each group is reported within parenthesis. Fig. 1. View largeDownload slide Kaplan–Meier survival plots for men and women comparing the probabilities of not having alcohol-related diagnoses among participants and non-participants (upper panels) and for heavy, non-heavy alcohol consumers and non-participants (lower panels). The number of persons with a disease diagnosed in each group is reported within parenthesis. Adjusted prevalences of heavy alcohol consumption Figure 2 presents the trends of the prevalence of heavy alcohol consumption, based on complete case analysis and the Bayesian modelling. It can be seen that the mean estimates of the Bayesian approach lie above the estimates of the complete case analysis. The numeric values are presented in Table 3. Fig. 2. View largeDownload slide Comparison of prevalence estimates of complete case analysis and Bayesian multiple imputations adjusted for education. Note that the scales of the vertical axis for men and women are different from each other. Fig. 2. View largeDownload slide Comparison of prevalence estimates of complete case analysis and Bayesian multiple imputations adjusted for education. Note that the scales of the vertical axis for men and women are different from each other. To compare the prevalence estimates based on participants only, and the posterior estimate for the prevalence of entire survey, absolute and relative differences can be calculated. For men, the absolute difference of the yearly prevalence estimates for 1987–2007 are 4.6, 7.3, 0.5, 8.6 and 1.8 percentage points calculated from Table 3, respectively. Those lead to an average difference of 4.6 percentage points. The corresponding relative differences for men are 1.93 (i.e. almost a two-fold difference), 1.95, 1.06, 1.6 and 1.17, respectively, and the average relative difference is 1.5. For women, the corresponding values are yearly absolute differences; 1.7, 1.9, 1.6, 0.3 and 1.9, respectively, leading to average absolute difference of 1.5 percentage points. The yearly relative differences are 3.39, 1.77, 1.42, 1.04 and 1.47, respectively, leading to average relative difference of 1.8, see Table 3. For men, the mean estimates based on Bayesian model vary year by year, but the credible intervals do not exclude the possibility of a monotonically increasing trend from 1987 to 2002. The complete case estimates are outside of the 90% credible interval of Bayesian trends in 1987, 1992 and 2002. The credible intervals are narrower for women than for men. For women, the complete case prevalence estimates are outside of the 90% credible intervals of Bayesian trends in 1987 and 1992, and are within the credible interval in 1997, 2002 and 2007. DISCUSSION There is evidence that non-participation in a survey asking about alcohol consumption is selective with respect to heavy alcohol consumption in Finland and in many other countries. We studied the prevalence of heavy alcohol consumption based on data from the National FINRISK Study, which suffer from selective non-participation. In FINRISK data, the average self-reported alcohol consumption for men was equal to 5.9 l and for women 1.9 l of pure 100% alcohol per year. For comparison, the national consumption statistics by National Institute for Health and Welfare (2016) show that the average yearly consumption of 100% alcohol for persons at least 15 years old was in the range of 10–13 l per person during 1987–2007. Thus, in FINRISK data the self-reported consumption is about 60–70% lower what has been reported in the national consumption statistics (which were not used in our modelling in any way). Although many reasons can partly explain the differences between the consumption statistics and self-reported data, e.g. questionnaire design and imperfect matching of survey frame with the target population, the differences between non-participants and participants in the follow-up data summarised in Fig. 1 suggest that selection bias is present. We observed differences in alcohol-related events for participants and non-participants. Non-participants had significantly increased risk for alcohol-related disease or death compared with participants, and men had a higher risk than women. This phenomenon has also been observed for other data, see (Romelsjö, 1989; Gorman et al., 2014; Christensen et al., 2015). When participation is selective with respect to variables to be studied, which is the case for alcohol use, the estimates from complete case analysis are affected by non-participation bias and the real level of uncertainty is hidden, e.g. confidence intervals are not wide enough when complete case analysis is used. Mäkelä (2003) and Dawson et al. (2014) demonstrated that this kind of bias cannot be reduced for alcohol data with demographic information. Gorman et al. (2017) utilised morbidity and mortality data from Scotland to assess the magnitude of bias in the estimates of alcohol consumption. We compared the estimates obtained by a complete case analysis with estimates obtained by adjusting for non-participation with a full Bayesian modelling approach. The Bayesian approach gave a higher estimate of heavy alcohol consumption than the complete case analysis. Our approach reduced the bias and made the uncertainty visible. We estimated that the magnitude of bias is 0–9 percentage points for men and 0–2 percentage points for women in the FINRISK data. The Bayesian mean estimate was on average 1.5 times higher for men and 1.8 times higher for women compared with participants. The modelling approach can also be extended to estimate associations between heavy alcohol consumption and other variables. The use of our approach requires follow-up data and background variables for the entire invited sample (including non-participants), follow-up time long enough to observe alcohol-related disease events and Bayesian modelling. The first requirement cannot be fulfilled in many countries because of lack of register data or legal restrictions for data linkage. The second requirement means that the prevalence estimates will be available only several years after the survey. This requirement may be relaxed if there exist earlier surveys that can be assumed to share the same model parameters with the current survey. The third requirement is the easiest to fulfil because it only calls for statistical expertise that is widely available. To conclude, the prevalence of heavy alcohol consumption based on survey participants only appears to be biased downward for both men and women. The magnitude of observed absolute bias was larger for men than women. The proposed non-participation adjustment approach is useful in context of alcohol research when follow-up data on non-participants are available and the modelling requirements are met. The follow-up data can be used to improve the estimation of the prevalence of heavy alcohol consumption. CONFLICT OF INTEREST STATEMENT None declared. FUNDING This work was supported by the Finnish Foundation for Alcohol Studies and Academy of Finland [grant numbers 266251 and 311877]. REFERENCES Boniface S , Scholes S , Shelton N , et al. . ( 2017 ) Assessment of non-response bias in estimates of alcohol consumption: applying the continuum of resistance model in a general population survey in England . PLoS One 12 : e0170892 . Google Scholar CrossRef Search ADS PubMed Borodulin K , Tolonen H , Jousilahti P , et al. . ( 2017 ) Cohort Profile: the National FINRISK Study . Int J Epidemiol http://dx.doi.org/10.1093/ije/dyx239 (11 March 2018, date last accessed) http://dx.doi.org/10.1093/ije/dyx239 . Brooks SP , Gelman A . ( 1998 ) General methods for monitoring convergence of iterative simulations . J Comput Graph Stat 7 : 434 – 55 . Christensen AI , Ekholm O , Gray L , et al. . ( 2015 ) What is wrong with non-participants? Alcohol-, drug- and smoking related mortality and morbidity in a 12-year follow up study of participants and non-participants in the Danish Health and Morbidity Survey . Addiction 110 : 1505 – 12 . Google Scholar CrossRef Search ADS PubMed Dawson DA , Goldstein RB , Pickering RP , et al. . ( 2014 ) Nonresponse bias in survey estimates of alcohol consumption and its association with harm . J Stud Alcohol Drugs 75 : 695 – 703 . Google Scholar CrossRef Search ADS PubMed Galea S , Tracy M . ( 2007 ) Participation rates in epidemiologic studies . Ann Epidemiol 17 : 643 – 53 . Google Scholar CrossRef Search ADS PubMed Gelman A , Carlin JB , Stern HS , et al. . ( 2014 ) Bayesian Data Analysis , 3rd edition . Boca Raton, FL, USA : Chapman & Hall/CRC . Gmel G . ( 2000 ) The effect of mode of data collection and of nonresponse on reported alcohol consumption: a split sample study in Switzerland . Addiction 95 : 123 – 34 . Google Scholar CrossRef Search ADS PubMed Gorman E , Leyland AH , McCartney G , et al. . ( 2014 ) Assessing the representativeness of population-sampled health surveys through linkage to administrative data on alcohol-related outcomes . Am J Epidemiol 180 : 941 – 8 . Google Scholar CrossRef Search ADS PubMed Gorman E , Leyland AH , McCartney G , et al. . ( 2017 ) Adjustment for survey non-representativeness using record-linkage: refined estimates of alcohol consumption by deprivation in Scotland . Addiction 112 : 1270 – 80 . Google Scholar CrossRef Search ADS PubMed Halldin J . ( 1985 ) Alcohol consumption and alcoholism in an urban population in central Sweden . Acta Psychiat Scand 71 : 128 – 40 . Google Scholar CrossRef Search ADS PubMed Harald K , Salomaa V , Jousilahti P , et al. . ( 2007 ) Non-participation and mortality in different socioeconomic groups: the FINRISK population surveys in 1972-92 . J Epidemiol Commun H 61 : 449 – 54 . Google Scholar CrossRef Search ADS Hirvonen E . ( 2017 ) Puuttuvuuden mallintaminen FINRISKI -tutkimuksessa (in Finnish). Masters thesis, University of Jyväskylä, http://urn.fi/URN:NBN:fi:jyu-201706192945 (11 March 2018, date last accessed). International Standard Classification of Education: ISCED 2011 . UIS, Montreal, Quebec. Jousilahti P , Salomaa V , Kuulasmaa K , et al. . ( 2005 ) Total and cause specific mortality among participants and non-participants of population based health surveys: a comprehensive follow up of 54 372 Finnish men and women . J Epidemiol Commun H 59 : 310 – 5 . Google Scholar CrossRef Search ADS Karvanen J , Tolonen H , Härkänen T , et al. . ( 2016 ) Selection bias was reduced by re-contacting nonparticipants . J Clin Epidemiol 76 : 209 – 17 . Google Scholar CrossRef Search ADS PubMed Klingemann H , Gmel G (eds) . ( 2001 ) Mapping the Social Consequences of Alcohol Consumption . Dordrecht, The Netherlands : Kluwer Academic Publishers . Kopra J , Härkänen T , Tolonen H , et al. . ( 2017 ) Adjusting for selective non-participation with re-contact data in the FINRISK 2012 survey . Scand J Public Health doi:10.1177/1403494817734774 . Advance online publication. Kopra J , Karvanen J , Härkänen T . ( 2018 ) Bayesian models for data missing not at random in health examination surveys . Stat Model 18 : 113 – 28 . doi:10.1177/1471082×17722605. Google Scholar CrossRef Search ADS Lahaut VM , Jansen HA , Van de Mheen D , et al. . ( 2002 ) Adjusting for selective nonparticipation with recontact and hospitalisation history data . Alcohol Alcoholism 37 : 256 – 60 . Google Scholar CrossRef Search ADS PubMed Livingston M , Callinan S . ( 2015 ) Underreporting in alcohol surveys: whose drinking is underestimated? J Stud Alcohol Drugs 76 : 158 – 64 . Google Scholar CrossRef Search ADS PubMed Meiklejohn J , Connor J , Kypri K . ( 2012 ) The effect of low survey response rates on estimates of alcohol consumption in a general population survey . PLoS One 7 : e35527 . Google Scholar CrossRef Search ADS PubMed National Institute for Health and Welfare . ( 2016 ) Alcoholic Beverage Consumption 2016 [e-publication]. Helsinki, Finland. https://www.thl.fi/fi/tilastot/tilastot-aiheittain/paihteet-ja-riippuvuudet/alkoholi/alkoholijuomien-kulutus (11 March 2018, date last accessed). National Institute for Health and Welfare . ( 2017 ) Care Register for Health Care. Helsinki, Finland. http://www.thl.fi/en/web/thlfi-en/statistics/information-on-statistics/register-descriptions/care-register-for-health-care (11 March 2018, date last accessed). Mäkelä P . ( 2003 ) Impact of correcting for nonresponse by weighting on estimates of alcohol consumption . J Stud Alcohol 64 : 589 – 96 . Google Scholar CrossRef Search ADS PubMed Mäkelä P , Huhtanen P . ( 2010 ) The effect of survey sampling frame on coverage: the level of and changes in alcohol related mortality in Finland as a test case . Addiction 105 : 1935 – 41 . Google Scholar CrossRef Search ADS PubMed Mäkelä P , Tigerstedt C , Mustonen H . ( 2012 ) The Finnish drinking culture: change and continuity in the past 40 years . Drug Alcohol Rev 31 : 831 – 40 . Google Scholar CrossRef Search ADS PubMed O’Hagan A , Buck CE , Daneshkhah A , et al. . ( 2006 ) Uncertain Judgements: Eliciting Experts’ Probabilities . Chichester, England : John Wiley & Sons . Official Statistics of Finland (OSF) : Causes of death [e-publication]. Helsinki, Finland. http://tilastokeskus.fi/til/ksyyt/index_en.html (11 March 2018, date last accessed). Plummer M . ( 2003 ) JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling. In Proceedings of the 3rd International Workshop on Distributed Statistical Computing. 124, p. 125 . Wien, Austria: Technische Universit at Wien. Plummer M . ( 2015 ) rjags: Bayesian Graphical Models using MCMC. R package version 3-15. https://CRAN.R-project.org/package=rjags (11 March 2018, date last accessed). R Foundation for Statistical Computing . ( 2018 ) R: A Language and Environment for Statistical Computing. Vienna, Austria. https://www.R-project.org/ (11 March 2018, date last accessed). Robert CP , Casella G . ( 2004 ) Monte Carlo Statistical Methods . Vienna, Austria : Springer Texts in Statistics . Rehm J , Mathers C , Popova S , et al. . ( 2009 ) Global burden of disease and injury and economic cost attributable to alcohol use and alcohol-use disorders . Lancet 373 : 2223 – 33 . Google Scholar CrossRef Search ADS PubMed Reinikainen J , Tolonen H , Borodulin K , et al. . ( 2017 ) Participation rates by educational levels have diverged during 25 years in Finnish health examination surveys . Eur J Public Health , Advance online publication doi:10.1093/eurpub/ckx151 . Romelsjö A . ( 1989 ) The relationship between alcohol consumption and social status in Stockholm. Has the social pattern of alcohol consumption changed? Int J Epidemiol 18 : 842 – 51 . Google Scholar CrossRef Search ADS PubMed Statistics Finland . The Register of Completed Education and Degrees. Helsinki, Finland. https://www.stat.fi/til/kou_en.html (11 March 2018, date last accessed). Tanner M , Wong W . ( 1987 ) The calculation of posterior distributions by data augmentation . J Am Stat Assoc 82 : 528 – 40 . Google Scholar CrossRef Search ADS Torvik FA , Rognmo K , Tambs K . ( 2012 ) Alcohol use and mental distress as predictors of non-response in a general population health survey: the HUNT study . Soc Psych Psych Epid 47 : 805 – 16 . Google Scholar CrossRef Search ADS Zhao J , Stockwell TIM , MacDonald S . ( 2009 ) Nonresponse bias in alcohol and drug population surveys . Drug Alcohol Rev 28 : 648 – 57 . Google Scholar CrossRef Search ADS PubMed APPENDIX A The statistical model is based on work presented in Kopra et al. (2017a), which utilises similar model to estimate the prevalence of smoking. Notation for the data For each individual i = 1,…,N invited to a survey, we denote Mi being indicator of participation (Mi = 1 for participants and Mi = 0 for non-participants). Background information Xi consists of both survey frame and education information and is available for both participants and non-participants. The survey frame has variables gender gi, region ri, age ai and study year si. The education is denoted by ei, and values 1, 2 and 3 corresponds to high, middle and low education, respectively. Thus, Xi = (gi, ri, ai, si, ei). The heavy alcohol consumption Yi is a binary variable such that heavy alcohol consumers have Yi = 1 and non-heavy alcohol consumers have Yi = 0. The variable Ti is the age at the first diagnosis of any of the alcohol-related diseases. The Ti is right censored and left-truncated at the age when the person entered the study. Participation model The participation model logit(P(Mi=1|Xi,Yi))=α0[gi,si]+α1[gi,si,ei]+η[gi,si]Yi+α2[gi,Yi](ai−45)+α3[ri], (1) is a logistic regression model with following parameters. First, parameter α0[gi,si] is a constant where notation [gi,si] indicates that there are independent α0 parameters for all levels of gender gi and study year si. Second, parameter α1[gi,si,ei] is the regression coefficient for education levels. For the lowest education level, this parameter is forced to be 0. The parameter η[gi,si] describes how heavy alcohol consumption affects participation. For this parameter, we need an informative prior. The parameter α2[gi,Yi] describes how age at study affect participation. Finally, α3[ri] is a term for the region. For one of the regions, this parameter is forced to be 0. We selected a model that included important factors affecting participation while ensuring the convergence of the MCMC chains in Bayesian inference. Table A1. Prior distributions Notation Distribution Interpretation Participation model  η Logistic(0, τ = 2.05) How heavy alcohol consumption affects the participation.  α0, α1, α2, α3 N(0, 1000−1) Other parameters. Risk factor model  β0, β1 N(0,1000−1) Other parameters. Survival model  h0,0(25) Unif(0, 20) Hazard for men at age 25–26.  h0,1(25) Unif(0,20) Hazard for women at age 25–26.  h0,0(t), t = 26, 27, … Unif(h0,0(t–1), 20) Hazard for men at age t.  h0,1(t), t = 26, 27, … Unif(h0,1(t–1), 20) Hazard for women at age t.  γ1 N(0, 1000−1) How heavy alcohol consumption affects hazard for men.  γ2 N(0, 1000−1) How heavy alcohol consumption affects hazard for women. Notation Distribution Interpretation Participation model  η Logistic(0, τ = 2.05) How heavy alcohol consumption affects the participation.  α0, α1, α2, α3 N(0, 1000−1) Other parameters. Risk factor model  β0, β1 N(0,1000−1) Other parameters. Survival model  h0,0(25) Unif(0, 20) Hazard for men at age 25–26.  h0,1(25) Unif(0,20) Hazard for women at age 25–26.  h0,0(t), t = 26, 27, … Unif(h0,0(t–1), 20) Hazard for men at age t.  h0,1(t), t = 26, 27, … Unif(h0,1(t–1), 20) Hazard for women at age t.  γ1 N(0, 1000−1) How heavy alcohol consumption affects hazard for men.  γ2 N(0, 1000−1) How heavy alcohol consumption affects hazard for women. View Large Table A1. Prior distributions Notation Distribution Interpretation Participation model  η Logistic(0, τ = 2.05) How heavy alcohol consumption affects the participation.  α0, α1, α2, α3 N(0, 1000−1) Other parameters. Risk factor model  β0, β1 N(0,1000−1) Other parameters. Survival model  h0,0(25) Unif(0, 20) Hazard for men at age 25–26.  h0,1(25) Unif(0,20) Hazard for women at age 25–26.  h0,0(t), t = 26, 27, … Unif(h0,0(t–1), 20) Hazard for men at age t.  h0,1(t), t = 26, 27, … Unif(h0,1(t–1), 20) Hazard for women at age t.  γ1 N(0, 1000−1) How heavy alcohol consumption affects hazard for men.  γ2 N(0, 1000−1) How heavy alcohol consumption affects hazard for women. Notation Distribution Interpretation Participation model  η Logistic(0, τ = 2.05) How heavy alcohol consumption affects the participation.  α0, α1, α2, α3 N(0, 1000−1) Other parameters. Risk factor model  β0, β1 N(0,1000−1) Other parameters. Survival model  h0,0(25) Unif(0, 20) Hazard for men at age 25–26.  h0,1(25) Unif(0,20) Hazard for women at age 25–26.  h0,0(t), t = 26, 27, … Unif(h0,0(t–1), 20) Hazard for men at age t.  h0,1(t), t = 26, 27, … Unif(h0,1(t–1), 20) Hazard for women at age t.  γ1 N(0, 1000−1) How heavy alcohol consumption affects hazard for men.  γ2 N(0, 1000−1) How heavy alcohol consumption affects hazard for women. View Large Risk factor model The model for risk factor (heavy alcohol consumption) is logit(P(Yi=1|Xi))=β0[gi,ri,si,ei]+(si−ai−1938)β1[gi,ri,si,ei]. (2) The risk factor model is stratified by gender gi, region ri, study year si and education ei using similar notation as in (1). The parameter β0[gi,ri,si,ei] is constant for persons born in 1938. The parameter β1[gi,ri,si,ei] determines how the heavy alcohol consumption prevalence changes with the year of birth. Survival model Let dNi(t) be the number of new events (increment) for the individual i at the time t. The increment follows a Poisson distribution with intensity parameters λi(t). The intensity λi(t) is modelled independently for both genders consisting of one-year period piecewise-constant baseline hazard terms h0,0(t) for men and h0,1(t) for women, and heavy alcohol consumption term exp(γ1Yi) and exp(γ2Yi) indicating the effect of heavy alcohol consumption for men and women, respectively dNi(t)∼Poisson(λi(t)) λi(t)={exp(γ1Yi)h0,0(t),giventhatTi≥tandg=0exp(γ2Yi)h0,1(t),giventhatTi≥tandg=10,Ti<t. Prior distributions The prior distributions are specified in Table A1. The prior distributions for piecewise constant hazard terms h0,0(t) and h0,1(t) are specified such that the hazard becomes increasing function with respect to t. APPENDIX B The study questions in 1987 Consumption of Alcohol 1. Do you use any alcoholic drinks, even occasionally (f. ex. beer, wine or spirits)? 1 yes 2 no, but I have not quitted completely 3 no, because I quit using alcohol…… years ago 4 I have never used alcohol If you have quitted alcohol use, please specify, why did you quit? no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 2. Have you during the past year (last 12 months) had any alcohol (beer, wine or spirits)? 1 yes 2 no (for your part, the questions are completed) 3. How often do you usually drink beer (III or IV A)? 1 daily 2 a few times a week 3 about once a week 4 few times a month 5 about once a month 6 about once in a few months 7 3–4 times a year 8 twice a year 9 once a year or more seldom 0 never 4. How much do you usually drink beer at a time? 1 less than one bottle 2 1 bottle 3 2 bottles 4 3 bottles 5 4–5 bottles 6 6–9 bottles 7 10–14 bottles 8 15 bottles or more 9 I do not drink beer 5. How often do you usually drink wine (light or strong, also home made)? 1 daily 2 a few times a week 3 about once a week 4 a few times a month 5 about once a month 6 about once in a few months 7 3–4 times a year 8 twice a year 9 once a year or more seldom 0 never 6. How much do you usually drink wine at a time? 1 half a glass 2 one glass 3 two glasses 4 about half a big bottle 5 a little less than one big bottle 6 about one big bottle 7 from one to two big bottles 8 more than two big bottles 9 I do not drink wine 7. How often do you usually drink spirits? 1 daily 2 a few times a week 3 about once a week 4 a few times a month 5 about once a month 6 about once in a few months 7 3–4 times a year 8 twice a year 9 once a year or more seldom 0 never 8. How much do you usually drink spirits at a time? 1 less than one restaurant measure (less than 4 cl) 2 one restaurant measure (about 4 cl) 3 two restaurant measures (about 8 cl) 4 3–4 restaurant measures 5 5–6 restaurant measures (about quarter liter) 6 7–10 restaurant measures 7 about a half liter bottle 6 more than a half liter bottle 7 I do not drink spirits 9. How often have you during the last 12 months had so much beer, wine or spirits that you have felt intoxicated? 1 a few times a week or more often 2 about once a week 3 a few times a month 4 about once a month 5 about once in two months 6 5 times a year 7 2–3 times a year 8 once a year 9 not even once The changes in questions from 1987 to 1992 The questions 1, 6 and 7 have with changes in text. We have highlighted the removed text with strikeout font (e.g. removed) and added text with italic font (e.g. added). The changes are in comparison with the previous survey. 1. Do you use any alcoholic drinks, even occasionally (f. ex. beer, wine or spirits)? 1 yes yes, at least once a month 2 no, but I have not quitted completelyyes, less than once a month 3 no, because I quit using alcohol …… years ago 4 I have never used alcohol If you have quitted alcohol use, please specify, why did you quit? no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 6. How much do you usually drink wine at a time? 1 half a glass 2 one glass 3 two glasses 4 about one small bottleabout half a big bottle 5 a little less than one big bottle 6 about one big bottle 7 from one to two big bottles 8 more than two big bottles 9 I do not drink wine 7. How much do you usually drink spirits at a time? 1 less than one restaurant measure (less than 4 cl) 2 one restaurant measure (about 4 cl) 3 two restaurant measures (about 8 cl) 4 3–4 restaurant measures 5 5–6 restaurant measures (about quarter liter) 6 7–10 restaurant measures 7 about a half liter bottle 6 more than a half liter bottle 7 I do not drink spirits The changes in questions from 1992 to 1997 The questions 4 and 6 have with changes in text. We have highlighted the removed text with strikeout font (e.g. removed) and added text with italic font (e.g. added). The changes are in comparison with the previous survey. 4. How much do you usually drink beer at a time?(1 bottle = 1/3 liters.) 1 less than one bottle 2 1 bottle 3 2 bottle 4 3 bottles 5 4–5 bottles 6 6–9 bottles 7 10–14 bottles 8 15 bottles or more 9 I do not drink beer 6. How much do you usually drink wine at a time? 1 half a glass 2 one glass (1 glass = c. 12 cl) 3 two glasses 4 about half a big bottle (1 bottle = 0,75 l) 5 a little less than one big bottle 6 about one big bottle 7 from one to two big bottles 8 more than two big bottles 9 I do not drink wine The changes in questions from 1997 to 2002 In 2002 the questions 3–8 have been replaced with a new question number 3. 3. How often did you drink the following amounts in one day during the last 12 months? Instruction: Start answering from the first row. Mark (x) the most suitable ‘How often?’ alternative. Then continue row at a time down in the same manner. Please mark only one alternative per row. 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses Doses per day Never Once a month or more seldom 2–3 times a month About once a week 2–3 times a week 4–5 times a week 6–7 times a week 15 or more □ □ □ □ □ □ □ 13–14 □ □ □ □ □ □ □ 11–12 □ □ □ □ □ □ □ 9–10 □ □ □ □ □ □ □ 7–8 □ □ □ □ □ □ □ 5–6 □ □ □ □ □ □ □ 3–4 □ □ □ □ □ □ □ 1–2 □ □ □ □ □ □ □ Doses per day Never Once a month or more seldom 2–3 times a month About once a week 2–3 times a week 4–5 times a week 6–7 times a week 15 or more □ □ □ □ □ □ □ 13–14 □ □ □ □ □ □ □ 11–12 □ □ □ □ □ □ □ 9–10 □ □ □ □ □ □ □ 7–8 □ □ □ □ □ □ □ 5–6 □ □ □ □ □ □ □ 3–4 □ □ □ □ □ □ □ 1–2 □ □ □ □ □ □ □ Doses per day Never Once a month or more seldom 2–3 times a month About once a week 2–3 times a week 4–5 times a week 6–7 times a week 15 or more □ □ □ □ □ □ □ 13–14 □ □ □ □ □ □ □ 11–12 □ □ □ □ □ □ □ 9–10 □ □ □ □ □ □ □ 7–8 □ □ □ □ □ □ □ 5–6 □ □ □ □ □ □ □ 3–4 □ □ □ □ □ □ □ 1–2 □ □ □ □ □ □ □ Doses per day Never Once a month or more seldom 2–3 times a month About once a week 2–3 times a week 4–5 times a week 6–7 times a week 15 or more □ □ □ □ □ □ □ 13–14 □ □ □ □ □ □ □ 11–12 □ □ □ □ □ □ □ 9–10 □ □ □ □ □ □ □ 7–8 □ □ □ □ □ □ □ 5–6 □ □ □ □ □ □ □ 3–4 □ □ □ □ □ □ □ 1–2 □ □ □ □ □ □ □ The changes in questions from 2002 to 2007 In 2007 the new question number 4 has been updated with a small change in the instructions and a change in the categories of consumed doses per day. 4. How often did you drink the following amounts in one day during the last 12 months? Instruction: Start answering from the first row. Mark (x) the most suitable ‘How often?’ alternative. Then continue row at a time down in the same manner. Please mark only one alternative per row. 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses Doses per day At least 4 times a week 2–3 times a week About once a week 1–2 times a month 3–10 times a year 1–2 times a year Never 18 or more 1 2 3 4 5 6 7 13–17 1 2 3 4 5 6 7 8–12 1 2 3 4 5 6 7 5–7 1 2 3 4 5 6 7 3–4 1 2 3 4 5 6 7 1–2 1 2 3 4 5 6 7 Doses per day At least 4 times a week 2–3 times a week About once a week 1–2 times a month 3–10 times a year 1–2 times a year Never 18 or more 1 2 3 4 5 6 7 13–17 1 2 3 4 5 6 7 8–12 1 2 3 4 5 6 7 5–7 1 2 3 4 5 6 7 3–4 1 2 3 4 5 6 7 1–2 1 2 3 4 5 6 7 Doses per day At least 4 times a week 2–3 times a week About once a week 1–2 times a month 3–10 times a year 1–2 times a year Never 18 or more 1 2 3 4 5 6 7 13–17 1 2 3 4 5 6 7 8–12 1 2 3 4 5 6 7 5–7 1 2 3 4 5 6 7 3–4 1 2 3 4 5 6 7 1–2 1 2 3 4 5 6 7 Doses per day At least 4 times a week 2–3 times a week About once a week 1–2 times a month 3–10 times a year 1–2 times a year Never 18 or more 1 2 3 4 5 6 7 13–17 1 2 3 4 5 6 7 8–12 1 2 3 4 5 6 7 5–7 1 2 3 4 5 6 7 3–4 1 2 3 4 5 6 7 1–2 1 2 3 4 5 6 7 © The Author(s) 2018. Medical Council on Alcohol and Oxford University Press. All rights reserved. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Alcohol and Alcoholism Oxford University Press

Follow-Up Data Improve the Estimation of the Prevalence of Heavy Alcohol Consumption

Alcohol and Alcoholism , Volume 53 (5) – Sep 1, 2018

Loading next page...
 
/lp/ou_press/follow-up-data-improve-the-estimation-of-the-prevalence-of-heavy-YIe10u6zGb
Publisher
Oxford University Press
Copyright
© The Author(s) 2018. Medical Council on Alcohol and Oxford University Press. All rights reserved.
ISSN
0735-0414
eISSN
1464-3502
D.O.I.
10.1093/alcalc/agy019
Publisher site
See Article on Publisher Site

Abstract

Abstract Aims We aim to adjust for potential non-participation bias in the prevalence of heavy alcohol consumption. Methods Population survey data from Finnish health examination surveys conducted in 1987–2007 were linked to the administrative registers for mortality and morbidity follow-up until end of 2014. Utilising these data, available for both participants and non-participants, we model the association between heavy alcohol consumption and alcohol-related disease diagnoses. Results Our results show that the estimated prevalence of heavy alcohol consumption is on average of 1.5 times higher for men and 1.8 times higher for women than what was obtained from participants only (complete case analysis). The magnitude of the difference in the mean estimates by year varies from 0 to 9 percentage points for men and from 0 to 2 percentage points for women. Conclusion The proposed approach improves the prevalence estimation but requires follow-up data on non-participants and Bayesian modelling. INTRODUCTION Reliable information about the prevalence of heavy alcohol consumption is important because alcohol-related health problems and undesired social consequences (Klingemann and Gmel, 2001) cause significant costs in many countries (Rehm et al., 2009). Prevalence estimates can be obtained through health surveys but the low participation rates (Galea and Tracy, 2007) imperil the reliability of the results. If the participation is selective with respect to alcohol consumption, the estimates of alcohol use suffer from non-participation bias, which hinders their usability for decision-making. If non-participants have worse health than participants, the bias usually leads to an overly positive image of the health of the population. Empirical evidence suggests that participation is often selective concerning alcohol consumption. Studies from Canada (Zhao et al., 2009), England (Boniface et al., 2017), Finland (Karvanen et al., 2016; Kopra et al., 2017), Norway (Torvik et al., 2012), Scotland (Gorman et al., 2014) Sweden (Romelsjö, 1989) and the USA (Dawson et al., 2014) conclude that non-participants drink more alcohol than participants. Follow-up studies have shown that non-participants tend to have a higher risk of alcohol-related diseases (Romelsjö, 1989; Jousilahti et al., 2005; Gorman et al., 2014; Christensen et al., 2015; Karvanen et al., 2016), and increased risk of hospitalisations and death (Jousilahti et al., 2005; Christensen et al., 2015; Karvanen et al., 2016), which indicates that non-participants tend to use more alcohol than participants. A study from the Netherlands (Lahaut et al., 2002) found that non-participants are more often abstainers than participants, which is not directly interpretable as a conflicting result because many social factors may be associated with abstaining. An older study from Sweden did not find an indication of selective participation (Halldin, 1985). In addition to selective non-participation, bias may be introduced by imperfect coverage of the target population by the survey sampling frame and by questionnaire design. First, if some individuals of target population cannot be invited to a survey, the sample does not represent the population of interest and the estimates will be biased. Mäkelä and Huhtanen (2010) observed that in Finland, persons who cannot be invited to a survey due to missing home address have about four times higher risk for alcohol-related deaths. This caused a small bias in the population estimates. Second, (Livingston and Callinan, 2015) claim that quantity–frequency design of the alcohol use questions underestimates alcohol consumption by one-third compared with asking about drinking with a within-location beverage-specific design. Gmel (2000) reported that alcohol as a subject of survey study does not have an impact on participation in comparison to similar questionnaire without alcohol-related questions. Studies from the USA (Dawson et al., 2014) and Finland (Mäkelä, 2003) have shown that the non-participation bias cannot be adjusted using just weights depending on basic demographic variables. Some studies adjust for selective non-participation utilising continuum of resistance model (Zhao et al., 2009; Meiklejohn et al., 2012; Boniface et al., 2017) but there are also other methods (Karvanen et al., 2016; Kopra et al., 2017). In Finnish health surveys, participation has been decreasing, while reported alcohol consumption has been mainly increasing from the 1960s to 2007 (Mäkelä et al., 2012). Jousilahti et al. (2005) report that in Finland non-participants have a higher risk of alcohol-specific diseases and death (Jousilahti et al., 2005), which is why we expect the estimates of heavy alcohol consumption to be biased. The difference in the disease risk could be explained by heavier alcohol consumption among non-participants. From previous studies (Harald et al., 2007; Hirvonen, 2017), we know that participation in the FINRISK Study is affected by age, gender, area and education (Reinikainen et al., 2017). We aim to adjust for selective non-participation for heavy alcohol consumption and to estimate the prevalence of heavy alcohol consumption with reduced bias. We present a Bayesian solution that is based on mortality and morbidity follow-up data. METHODS Data We used data from the National FINRISK Study, which is a series of cross-sectional health examination surveys (Borodulin et al., 2017) conducted in Finland every fifth year since 1972. We analysed data for the years 1987–2007. Years 1972–82, as well as 2012, were excluded because the questions of alcohol consumption were too different from the questions in 1987–2007. In 1987 and 1992 studies, the questions were essentially the same. In 1997 study, questions regarding the usage of cider or mild wine (alc. vol under 5%) were added. Otherwise, the study remained the same as earlier. In 2002, the questions regarding the consumption of red wine and other wines were separated from each other. The total alcohol consumption was not measured several alcohol beverage-specific questions but the participants were advised to calculate the number of alcohol portions (standard drinks) consumed and to mark the corresponding category in the quantity–frequency table. In 2007, the alcohol questions were the same as in 2002, but the instructions for the calculation of daily alcohol consumption were improved. The surveys provide data on 25–74-year-old adults from six regions of Finland. We restrict the data to people aged 25–65 years since oldest age group 65–74 years old was not available in all areas until 2007. The survey consists of questionnaires and a health examination carried out at a local study clinic. The sample was drawn from the National Population Register and was stratified by region, gender and 10-year age group. In total, there are 44,317 invitees including 31,567 participants. The survey data contains self-reported alcohol consumption and background variables, age, gender, region and study year for the whole sample. The survey utilised beverage-specific quantity–frequency questions on alcohol use in the first three surveys and graduate frequency measure in the latter two surveys. The questions related to alcohol consumption are provided in Appendix B. The study questionnaires in 1987, 1992 and 1997 asked alcohol usage one type of alcohol beverages at a time: beer, spirits/vodka, long drink or cider (in 1997), wines and mild wines (in 1992 and 1997). In 2002 and 2007, the questionnaire was different and individuals reported their alcohol consumption for all beverage types in one question. From these questions, a number of standard drinks consumed per week during a previous 12 months was calculated. One standard drink equals 12 g of pure alcohol which is equivalent to, e.g. one bottle of beer (33cl, 4.7 volume percent of alcohol). Based on the number of standard drinks, the (self-reported) total amount of 100% alcohol consumed in the previous 12 months was calculated. Since our main interest was to estimate the prevalence of heavy alcohol users, we classified participants as heavy alcohol consumers and others (non-heavy alcohol consumers) as follows. The persons who reported consuming on average at least 24 standard drinks per week for men or at least 16 standard drinks per week for women during the one-year period before the examination were considered heavy alcohol consumers. The survey data were linked to three registers: The Register of Completed Education and Degrees (Statistics Finland, 2016), Care Register for Health Care (National Institute for Health and Welfare, 2017) and Cause of Death Register (Official Statistics of Finland, 2017) using personal identification code. The register-based data were available for both participants and non-participants. The level of education is categorised according to the International Standard Classification of Education (ISCED, 2011) levels: (1) high level (tertiary education, ISCED levels 5–8), (2) middle level (secondary education, ISCED levels 3–4) and (3) low level (primary education or less or unknown, ISCED levels 0–2). The Care Register gives data about the hospital visits with dates and ICD codes for both participants and non-participants. From the Causes of Death Register, we obtain data about dates and ICD codes of the cause of death. Follow-up data contains the time-to-event (age) and ICD code of the first alcohol-related disease diagnosis or death. The ICD codes we considered to be alcohol-related are listed in Table 1. The follow-up begins from the survey and ends at the end of 2014. Persons who have neither alcohol-related disease diagnosis nor the alcohol-related cause of death are censored at the end of the follow-up. Deaths not related to alcohol are treated as censorings. Table 1. The ICD-codes interpreted as alcohol-related events ICD-9  291 Alcohol-induced mental disorders  303 Alcohol dependence syndrome  357.5 Alcoholic polyneuropathy  425.5 Alcoholic cardiomyopathy  535.3 Alcoholic gastritis  571.0 Alcoholic fatty liver  571.1 Acute alcoholic liver disease  571.2 Alcoholic cirrhosis of liver  571.3 Alcoholic liver damage, unspecified  577.0D-F Alcoholic disease of the pancreas, acute  577.1C-D Alcoholic disease of the pancreas, chronic  980.0 Toxic effect ethyl alcohol  980.2 Toxic effect of isopropyl alcohol  980.8 Toxic effect of other specified alcohols  980.9 Toxic effect of other unspecified alcohol  E851 Accidental poisoning by alcohol ICD-10  F10 Mental and behavioural disorders due to use of alcohol  G31.2 Degeneration of nervous system due to alcohol  G62.1 Alcoholic polyneuropathy  G72.1 Alcoholic myopathy  I42.6 Alcoholic cardiomyopathy  K29.2 Alcoholic gastritis  K70 Alcoholic liver disease  K85.2 Alcohol-induced acute pancreatitis  K86.0 Alcohol-induced chronic pancreatitis  T51 Toxic effect of alcohol  X45 Accidental poisoning or other exposure to alcohol  Y15 Poisoning by and exposure to alcohol, undetermined intent ICD-9  291 Alcohol-induced mental disorders  303 Alcohol dependence syndrome  357.5 Alcoholic polyneuropathy  425.5 Alcoholic cardiomyopathy  535.3 Alcoholic gastritis  571.0 Alcoholic fatty liver  571.1 Acute alcoholic liver disease  571.2 Alcoholic cirrhosis of liver  571.3 Alcoholic liver damage, unspecified  577.0D-F Alcoholic disease of the pancreas, acute  577.1C-D Alcoholic disease of the pancreas, chronic  980.0 Toxic effect ethyl alcohol  980.2 Toxic effect of isopropyl alcohol  980.8 Toxic effect of other specified alcohols  980.9 Toxic effect of other unspecified alcohol  E851 Accidental poisoning by alcohol ICD-10  F10 Mental and behavioural disorders due to use of alcohol  G31.2 Degeneration of nervous system due to alcohol  G62.1 Alcoholic polyneuropathy  G72.1 Alcoholic myopathy  I42.6 Alcoholic cardiomyopathy  K29.2 Alcoholic gastritis  K70 Alcoholic liver disease  K85.2 Alcohol-induced acute pancreatitis  K86.0 Alcohol-induced chronic pancreatitis  T51 Toxic effect of alcohol  X45 Accidental poisoning or other exposure to alcohol  Y15 Poisoning by and exposure to alcohol, undetermined intent Table 1. The ICD-codes interpreted as alcohol-related events ICD-9  291 Alcohol-induced mental disorders  303 Alcohol dependence syndrome  357.5 Alcoholic polyneuropathy  425.5 Alcoholic cardiomyopathy  535.3 Alcoholic gastritis  571.0 Alcoholic fatty liver  571.1 Acute alcoholic liver disease  571.2 Alcoholic cirrhosis of liver  571.3 Alcoholic liver damage, unspecified  577.0D-F Alcoholic disease of the pancreas, acute  577.1C-D Alcoholic disease of the pancreas, chronic  980.0 Toxic effect ethyl alcohol  980.2 Toxic effect of isopropyl alcohol  980.8 Toxic effect of other specified alcohols  980.9 Toxic effect of other unspecified alcohol  E851 Accidental poisoning by alcohol ICD-10  F10 Mental and behavioural disorders due to use of alcohol  G31.2 Degeneration of nervous system due to alcohol  G62.1 Alcoholic polyneuropathy  G72.1 Alcoholic myopathy  I42.6 Alcoholic cardiomyopathy  K29.2 Alcoholic gastritis  K70 Alcoholic liver disease  K85.2 Alcohol-induced acute pancreatitis  K86.0 Alcohol-induced chronic pancreatitis  T51 Toxic effect of alcohol  X45 Accidental poisoning or other exposure to alcohol  Y15 Poisoning by and exposure to alcohol, undetermined intent ICD-9  291 Alcohol-induced mental disorders  303 Alcohol dependence syndrome  357.5 Alcoholic polyneuropathy  425.5 Alcoholic cardiomyopathy  535.3 Alcoholic gastritis  571.0 Alcoholic fatty liver  571.1 Acute alcoholic liver disease  571.2 Alcoholic cirrhosis of liver  571.3 Alcoholic liver damage, unspecified  577.0D-F Alcoholic disease of the pancreas, acute  577.1C-D Alcoholic disease of the pancreas, chronic  980.0 Toxic effect ethyl alcohol  980.2 Toxic effect of isopropyl alcohol  980.8 Toxic effect of other specified alcohols  980.9 Toxic effect of other unspecified alcohol  E851 Accidental poisoning by alcohol ICD-10  F10 Mental and behavioural disorders due to use of alcohol  G31.2 Degeneration of nervous system due to alcohol  G62.1 Alcoholic polyneuropathy  G72.1 Alcoholic myopathy  I42.6 Alcoholic cardiomyopathy  K29.2 Alcoholic gastritis  K70 Alcoholic liver disease  K85.2 Alcohol-induced acute pancreatitis  K86.0 Alcohol-induced chronic pancreatitis  T51 Toxic effect of alcohol  X45 Accidental poisoning or other exposure to alcohol  Y15 Poisoning by and exposure to alcohol, undetermined intent Complete case analysis The complete case analysis (e.g. mean estimate from the participants) assumes that participation is not selective concerning alcohol consumption. Violations of this assumption lead to bias. We compared the results of complete case analysis with a Bayesian approach, which relies on more realistic assumptions and allows for selective non-participation concerning heavy alcohol use. Modelling approach We applied a Bayesian approach introduced in Kopra et al. (2018) to estimate the prevalence of heavy alcohol consumption. The Bayesian model consists of three sub-models which are fitted simultaneously. The sub-models are: participation model, risk factor model and survival model. The mathematical formulas for the models are given in Appendix A. The participation model describes which variables affect participation. Participation is defined as a binary indicator (0 or 1) for the availability information on alcohol consumption. This model is a logistic regression model with linear covariates for study year and age, and categorical variables for the region (4 levels), education (3 levels) and the alcohol consumption (binary). The model also takes into account the possible interactions of gender and study year, gender and alcohol consumption, and study year and alcohol consumption. The risk factor model describes how alcohol consumption (heavy or non-heavy) varies by background variables. By background variables, we mean age, gender, region, study year and education. The model is a logistic regression model with interactions for the year of birth with gender, region, study year and education. The survival model describes the relationship between alcohol consumption and alcohol-related diseases. All disease events are combined and modelled as one survival outcome. The survival model is a piecewise constant hazard model with one-year baseline hazard period terms. The model assumes monotonically increasing baseline hazard, which is accomplished using prior specification. In addition to baseline hazard, alcohol consumption is used as a regressor. Both baseline hazard terms and the regression coefficient are gender-specific. Our estimation method requires that the average disease risk of non-participants must be between the average risks of heavy alcohol consumer participants and other participants. This technical requirement can be evaluated from the data. Prior distributions We used weakly informative prior distributions that reflect the existing knowledge but have variances large enough to allow for surprises. This approach is recommended in textbooks on Bayesian statistics (Gelman et al., 2014) and there exists guidelines for elicitation of prior distributions (O’Hagan et al., 2006). The participation model needed an informative prior for the effect of heavy alcohol consumption on the participation, i.e. for the strength of selectivity mechanism. Some degree of subjectivity cannot be avoided in the prior specification. To define a weakly informative prior, we took a 45-year-old non-heavy alcohol consumer who participates with probability 0.7 as a reference and considered the prior probability for a heavy alcohol consumer who is otherwise similar. We elicit that there is 25% chance that person participates with probability p lower than 0.5 (P(p ≤ 0.5) = 0.25), 35% chance for p ≤ 0.6, and 50% chance for p ≤ 0.7. The functional form of the prior distribution was chosen to be logistic distribution. These elicitations lead to logistics prior distribution with expected value zero and variance 1/2.05. In the survival model, we applied monotonically increasing baseline hazards separately for men and women. The prior distribution for the first hazard term (25–26-year-olds) was a uniform distribution with a range from 0 to 20. From second hazard term (26–27-year-olds) to the last hazard term (99–100-year-olds), each term had a uniform distribution with the lower limit being the value of previous baseline hazard term and upper limit 20. All the remaining model parameters had normally distributed priors with zero mean and variance 1000. The prior distributions are presented using mathematical notation in Table A1. Imputations and model fitting Alcohol consumption was missing for the non-participants. These missing values (heavy or non-heavy) were imputed simultaneously with Bayesian model fitting using data augmentation (Tanner and Wong, 1987). The model was fitted using Markov chain Monte Carlo (MCMC) (Robert and Casella, 2004) and implemented with Just Another Gibbs Sampler (JAGS)-software (Plummer, 2003) and R software (R Foundation for Statistical Computing, 2017) with rjags-package (Plummer, 2015). The convergence of MCMC chains was investigated using Brooks–Gelman Rˆ diagnostics (Brooks and Gelman, 1998) and all the Rˆs were ≤1.01 which indicates convergence. The model fitting utilised computational resources of IT Center for Science Ltd (CSC). RESULTS Descriptive statistics In Table 2, we present the descriptive statistics on age, education and gender. These variables are examined by study year, and comparisons can be made between participants and non-participant as well as between heavy alcohol consumers and other alcohol consumers. Table 2. Description of background information by study year for non-participants, participants, and heavy and non-heavy alcohol consumers among participants Year Non-participants Participants All Heavy alcohol consumers Moderate alcohol consumers Average age  1987 42.5 44.4 41.7 44.4  1992 41.8 44.7 44.5 44.7  1997 42.8 45.0 45.0 45.0  2002 42.3 44.9 45.7 44.8  2007 41.9 45.6 47.3 45.5 High education (%)  1987 13.6 18.5 23.0 18.4  1992 18.4 26.6 30.2 26.5  1997 24.7 29.9 31.4 29.8  2002 25.6 35.7 32.7 35.8  2007 27.7 38.6 34.3 38.8 Middle education (%)  1987 30.6 31.8 31.0 31.8  1992 35.6 34.9 34.4 34.9  1997 37.7 38.1 37.6 38.1  2002 41.8 40.6 43.5 40.4  2007 45.3 44.2 45.5 44.1 Low education (%)  1987 55.8 49.7 46.0 49.8  1992 46.1 38.5 35.4 38.6  1997 37.6 32.1 31.0 32.1  2002 32.6 23.8 23.9 23.8  2007 27.0 17.2 20.2 17.1 Women (%)  1987 42.3 52.0 15.9 52.7  1992 40.5 53.0 21.2 54.1  1997 43.2 52.9 22.8 54.3  2002 41.6 53.6 29.0 54.9  2007 42.7 53.9 27.8 55.3 Year Non-participants Participants All Heavy alcohol consumers Moderate alcohol consumers Average age  1987 42.5 44.4 41.7 44.4  1992 41.8 44.7 44.5 44.7  1997 42.8 45.0 45.0 45.0  2002 42.3 44.9 45.7 44.8  2007 41.9 45.6 47.3 45.5 High education (%)  1987 13.6 18.5 23.0 18.4  1992 18.4 26.6 30.2 26.5  1997 24.7 29.9 31.4 29.8  2002 25.6 35.7 32.7 35.8  2007 27.7 38.6 34.3 38.8 Middle education (%)  1987 30.6 31.8 31.0 31.8  1992 35.6 34.9 34.4 34.9  1997 37.7 38.1 37.6 38.1  2002 41.8 40.6 43.5 40.4  2007 45.3 44.2 45.5 44.1 Low education (%)  1987 55.8 49.7 46.0 49.8  1992 46.1 38.5 35.4 38.6  1997 37.6 32.1 31.0 32.1  2002 32.6 23.8 23.9 23.8  2007 27.0 17.2 20.2 17.1 Women (%)  1987 42.3 52.0 15.9 52.7  1992 40.5 53.0 21.2 54.1  1997 43.2 52.9 22.8 54.3  2002 41.6 53.6 29.0 54.9  2007 42.7 53.9 27.8 55.3 Table 2. Description of background information by study year for non-participants, participants, and heavy and non-heavy alcohol consumers among participants Year Non-participants Participants All Heavy alcohol consumers Moderate alcohol consumers Average age  1987 42.5 44.4 41.7 44.4  1992 41.8 44.7 44.5 44.7  1997 42.8 45.0 45.0 45.0  2002 42.3 44.9 45.7 44.8  2007 41.9 45.6 47.3 45.5 High education (%)  1987 13.6 18.5 23.0 18.4  1992 18.4 26.6 30.2 26.5  1997 24.7 29.9 31.4 29.8  2002 25.6 35.7 32.7 35.8  2007 27.7 38.6 34.3 38.8 Middle education (%)  1987 30.6 31.8 31.0 31.8  1992 35.6 34.9 34.4 34.9  1997 37.7 38.1 37.6 38.1  2002 41.8 40.6 43.5 40.4  2007 45.3 44.2 45.5 44.1 Low education (%)  1987 55.8 49.7 46.0 49.8  1992 46.1 38.5 35.4 38.6  1997 37.6 32.1 31.0 32.1  2002 32.6 23.8 23.9 23.8  2007 27.0 17.2 20.2 17.1 Women (%)  1987 42.3 52.0 15.9 52.7  1992 40.5 53.0 21.2 54.1  1997 43.2 52.9 22.8 54.3  2002 41.6 53.6 29.0 54.9  2007 42.7 53.9 27.8 55.3 Year Non-participants Participants All Heavy alcohol consumers Moderate alcohol consumers Average age  1987 42.5 44.4 41.7 44.4  1992 41.8 44.7 44.5 44.7  1997 42.8 45.0 45.0 45.0  2002 42.3 44.9 45.7 44.8  2007 41.9 45.6 47.3 45.5 High education (%)  1987 13.6 18.5 23.0 18.4  1992 18.4 26.6 30.2 26.5  1997 24.7 29.9 31.4 29.8  2002 25.6 35.7 32.7 35.8  2007 27.7 38.6 34.3 38.8 Middle education (%)  1987 30.6 31.8 31.0 31.8  1992 35.6 34.9 34.4 34.9  1997 37.7 38.1 37.6 38.1  2002 41.8 40.6 43.5 40.4  2007 45.3 44.2 45.5 44.1 Low education (%)  1987 55.8 49.7 46.0 49.8  1992 46.1 38.5 35.4 38.6  1997 37.6 32.1 31.0 32.1  2002 32.6 23.8 23.9 23.8  2007 27.0 17.2 20.2 17.1 Women (%)  1987 42.3 52.0 15.9 52.7  1992 40.5 53.0 21.2 54.1  1997 43.2 52.9 22.8 54.3  2002 41.6 53.6 29.0 54.9  2007 42.7 53.9 27.8 55.3 The average age of the non-participants was lower than the average age of the participants. Over the years, the average age appears not to have changed much for the non-participants but it has slightly increased for the participants. Among participants, the average age of heavy alcohol consumers has increased more rapidly than for non-heavy alcohol consumers. The average age of heavy alcohol consumers was 41.7 in 1987 (44.4 for non-heavy) and it has increased between each study being 47.3 for heavy alcohol consumers and 45.5 for non-heavy alcohol consumers in 2007. The average age of non-heavy alcohol consumers has also increased between the studies, except between the 1997 and 2002 when it decreased by 0.2 years. The level of education has increased for both participants and non-participants during the study period. The non-participants tend to have low education more often than participants, and participants tend to have high education more often than non-participants. In 1987, there were a higher proportion of highly educated participants among heavy alcohol consumers than among non-heavy alcohol consumers. In 2007, the situation was opposite; the proportion of highly educated persons is higher for non-heavy alcohol consumers than for the heavy alcohol consumers. The proportion of women among participants has slightly increased from 52.0% to 53.4% during 1987–2007. Among non-heavy alcohol consumers, the proportion is higher: 52.7–55.3%. Women are a minority among heavy alcohol consumers. There were 15.9% women among heavy alcohol consumers in 1987, and the proportion has notably increased being 27.8% in 2007. The proportion of women was higher among the participants than among non-participants. The proportion of women among participating heavy alcohol users has been rapidly increasing over the years, while the corresponding proportion had not increased by much among non-heavy alcohol consumers. The number of invitees, the participation rate and the number of events for both participant and non-participant men and women are presented in Table 3. During the study period, the proportion of heavy alcohol consumers has increased for both men and women among participants, and simultaneously the participation rate has decreased. Table 3. Number of invitees, the participation rate, the prevalence of heavy alcohol consumption based on participants and Bayesian modelling (posterior mean), and the number of alcohol-related incident events (per 1000 follow-up years) for the non-participant and the participant men and women Year Invited N Participation rate (%) Prevalence for participants (%) Posterior mean (%) Alcohol-related incident events (per 1000 follow-up years) Participants Non-participants Men  1987 3910 79.5 5.0 9.6 202 (2.8) 91 (5.2)  1992 3888 73.3 7.7 15.0 168 (2.9) 128 (6.5)  1997 4034 70.0 9.2 9.7 150 (3.2) 103 (5.4)  2002 3955 66.5 14.4 22.9 118 (3.6) 62 (3.8)  2007 3202 61.8 11.0 12.8 47 (3.1) 35 (3.7) Women  1987 3961 85.1 0.7 2.4 52 (0.6) 29 (2.0)  1992 3951 81.0 2.5 4.4 46 (0.7) 23 (1.5)  1997 4031 75.8 3.7 5.3 25 (0.5) 37 (2.3)  2002 4019 75.4 5.7 5.9 34 (0.9) 13 (1.0)  2007 3278 71.3 4.0 5.8 12 (0.7) 10 (1.3) Year Invited N Participation rate (%) Prevalence for participants (%) Posterior mean (%) Alcohol-related incident events (per 1000 follow-up years) Participants Non-participants Men  1987 3910 79.5 5.0 9.6 202 (2.8) 91 (5.2)  1992 3888 73.3 7.7 15.0 168 (2.9) 128 (6.5)  1997 4034 70.0 9.2 9.7 150 (3.2) 103 (5.4)  2002 3955 66.5 14.4 22.9 118 (3.6) 62 (3.8)  2007 3202 61.8 11.0 12.8 47 (3.1) 35 (3.7) Women  1987 3961 85.1 0.7 2.4 52 (0.6) 29 (2.0)  1992 3951 81.0 2.5 4.4 46 (0.7) 23 (1.5)  1997 4031 75.8 3.7 5.3 25 (0.5) 37 (2.3)  2002 4019 75.4 5.7 5.9 34 (0.9) 13 (1.0)  2007 3278 71.3 4.0 5.8 12 (0.7) 10 (1.3) Table 3. Number of invitees, the participation rate, the prevalence of heavy alcohol consumption based on participants and Bayesian modelling (posterior mean), and the number of alcohol-related incident events (per 1000 follow-up years) for the non-participant and the participant men and women Year Invited N Participation rate (%) Prevalence for participants (%) Posterior mean (%) Alcohol-related incident events (per 1000 follow-up years) Participants Non-participants Men  1987 3910 79.5 5.0 9.6 202 (2.8) 91 (5.2)  1992 3888 73.3 7.7 15.0 168 (2.9) 128 (6.5)  1997 4034 70.0 9.2 9.7 150 (3.2) 103 (5.4)  2002 3955 66.5 14.4 22.9 118 (3.6) 62 (3.8)  2007 3202 61.8 11.0 12.8 47 (3.1) 35 (3.7) Women  1987 3961 85.1 0.7 2.4 52 (0.6) 29 (2.0)  1992 3951 81.0 2.5 4.4 46 (0.7) 23 (1.5)  1997 4031 75.8 3.7 5.3 25 (0.5) 37 (2.3)  2002 4019 75.4 5.7 5.9 34 (0.9) 13 (1.0)  2007 3278 71.3 4.0 5.8 12 (0.7) 10 (1.3) Year Invited N Participation rate (%) Prevalence for participants (%) Posterior mean (%) Alcohol-related incident events (per 1000 follow-up years) Participants Non-participants Men  1987 3910 79.5 5.0 9.6 202 (2.8) 91 (5.2)  1992 3888 73.3 7.7 15.0 168 (2.9) 128 (6.5)  1997 4034 70.0 9.2 9.7 150 (3.2) 103 (5.4)  2002 3955 66.5 14.4 22.9 118 (3.6) 62 (3.8)  2007 3202 61.8 11.0 12.8 47 (3.1) 35 (3.7) Women  1987 3961 85.1 0.7 2.4 52 (0.6) 29 (2.0)  1992 3951 81.0 2.5 4.4 46 (0.7) 23 (1.5)  1997 4031 75.8 3.7 5.3 25 (0.5) 37 (2.3)  2002 4019 75.4 5.7 5.9 34 (0.9) 13 (1.0)  2007 3278 71.3 4.0 5.8 12 (0.7) 10 (1.3) The probabilities for not having alcohol-related disease diagnosis up to the given age for men and women are presented by Kaplan–Meier survival plots (Fig. 1). The top row shows that the non-participants were more likely to have alcohol-related diagnoses than participants. The lower row shows that the risk for non-participants lies between the risks of heavy and non-heavy alcohol consumers, which is a requirement for the utilised Bayesian model. The number of persons with a disease diagnosed in each group is reported next to the survival curve in the Fig. 1. Fig. 1. View largeDownload slide Kaplan–Meier survival plots for men and women comparing the probabilities of not having alcohol-related diagnoses among participants and non-participants (upper panels) and for heavy, non-heavy alcohol consumers and non-participants (lower panels). The number of persons with a disease diagnosed in each group is reported within parenthesis. Fig. 1. View largeDownload slide Kaplan–Meier survival plots for men and women comparing the probabilities of not having alcohol-related diagnoses among participants and non-participants (upper panels) and for heavy, non-heavy alcohol consumers and non-participants (lower panels). The number of persons with a disease diagnosed in each group is reported within parenthesis. Adjusted prevalences of heavy alcohol consumption Figure 2 presents the trends of the prevalence of heavy alcohol consumption, based on complete case analysis and the Bayesian modelling. It can be seen that the mean estimates of the Bayesian approach lie above the estimates of the complete case analysis. The numeric values are presented in Table 3. Fig. 2. View largeDownload slide Comparison of prevalence estimates of complete case analysis and Bayesian multiple imputations adjusted for education. Note that the scales of the vertical axis for men and women are different from each other. Fig. 2. View largeDownload slide Comparison of prevalence estimates of complete case analysis and Bayesian multiple imputations adjusted for education. Note that the scales of the vertical axis for men and women are different from each other. To compare the prevalence estimates based on participants only, and the posterior estimate for the prevalence of entire survey, absolute and relative differences can be calculated. For men, the absolute difference of the yearly prevalence estimates for 1987–2007 are 4.6, 7.3, 0.5, 8.6 and 1.8 percentage points calculated from Table 3, respectively. Those lead to an average difference of 4.6 percentage points. The corresponding relative differences for men are 1.93 (i.e. almost a two-fold difference), 1.95, 1.06, 1.6 and 1.17, respectively, and the average relative difference is 1.5. For women, the corresponding values are yearly absolute differences; 1.7, 1.9, 1.6, 0.3 and 1.9, respectively, leading to average absolute difference of 1.5 percentage points. The yearly relative differences are 3.39, 1.77, 1.42, 1.04 and 1.47, respectively, leading to average relative difference of 1.8, see Table 3. For men, the mean estimates based on Bayesian model vary year by year, but the credible intervals do not exclude the possibility of a monotonically increasing trend from 1987 to 2002. The complete case estimates are outside of the 90% credible interval of Bayesian trends in 1987, 1992 and 2002. The credible intervals are narrower for women than for men. For women, the complete case prevalence estimates are outside of the 90% credible intervals of Bayesian trends in 1987 and 1992, and are within the credible interval in 1997, 2002 and 2007. DISCUSSION There is evidence that non-participation in a survey asking about alcohol consumption is selective with respect to heavy alcohol consumption in Finland and in many other countries. We studied the prevalence of heavy alcohol consumption based on data from the National FINRISK Study, which suffer from selective non-participation. In FINRISK data, the average self-reported alcohol consumption for men was equal to 5.9 l and for women 1.9 l of pure 100% alcohol per year. For comparison, the national consumption statistics by National Institute for Health and Welfare (2016) show that the average yearly consumption of 100% alcohol for persons at least 15 years old was in the range of 10–13 l per person during 1987–2007. Thus, in FINRISK data the self-reported consumption is about 60–70% lower what has been reported in the national consumption statistics (which were not used in our modelling in any way). Although many reasons can partly explain the differences between the consumption statistics and self-reported data, e.g. questionnaire design and imperfect matching of survey frame with the target population, the differences between non-participants and participants in the follow-up data summarised in Fig. 1 suggest that selection bias is present. We observed differences in alcohol-related events for participants and non-participants. Non-participants had significantly increased risk for alcohol-related disease or death compared with participants, and men had a higher risk than women. This phenomenon has also been observed for other data, see (Romelsjö, 1989; Gorman et al., 2014; Christensen et al., 2015). When participation is selective with respect to variables to be studied, which is the case for alcohol use, the estimates from complete case analysis are affected by non-participation bias and the real level of uncertainty is hidden, e.g. confidence intervals are not wide enough when complete case analysis is used. Mäkelä (2003) and Dawson et al. (2014) demonstrated that this kind of bias cannot be reduced for alcohol data with demographic information. Gorman et al. (2017) utilised morbidity and mortality data from Scotland to assess the magnitude of bias in the estimates of alcohol consumption. We compared the estimates obtained by a complete case analysis with estimates obtained by adjusting for non-participation with a full Bayesian modelling approach. The Bayesian approach gave a higher estimate of heavy alcohol consumption than the complete case analysis. Our approach reduced the bias and made the uncertainty visible. We estimated that the magnitude of bias is 0–9 percentage points for men and 0–2 percentage points for women in the FINRISK data. The Bayesian mean estimate was on average 1.5 times higher for men and 1.8 times higher for women compared with participants. The modelling approach can also be extended to estimate associations between heavy alcohol consumption and other variables. The use of our approach requires follow-up data and background variables for the entire invited sample (including non-participants), follow-up time long enough to observe alcohol-related disease events and Bayesian modelling. The first requirement cannot be fulfilled in many countries because of lack of register data or legal restrictions for data linkage. The second requirement means that the prevalence estimates will be available only several years after the survey. This requirement may be relaxed if there exist earlier surveys that can be assumed to share the same model parameters with the current survey. The third requirement is the easiest to fulfil because it only calls for statistical expertise that is widely available. To conclude, the prevalence of heavy alcohol consumption based on survey participants only appears to be biased downward for both men and women. The magnitude of observed absolute bias was larger for men than women. The proposed non-participation adjustment approach is useful in context of alcohol research when follow-up data on non-participants are available and the modelling requirements are met. The follow-up data can be used to improve the estimation of the prevalence of heavy alcohol consumption. CONFLICT OF INTEREST STATEMENT None declared. FUNDING This work was supported by the Finnish Foundation for Alcohol Studies and Academy of Finland [grant numbers 266251 and 311877]. REFERENCES Boniface S , Scholes S , Shelton N , et al. . ( 2017 ) Assessment of non-response bias in estimates of alcohol consumption: applying the continuum of resistance model in a general population survey in England . PLoS One 12 : e0170892 . Google Scholar CrossRef Search ADS PubMed Borodulin K , Tolonen H , Jousilahti P , et al. . ( 2017 ) Cohort Profile: the National FINRISK Study . Int J Epidemiol http://dx.doi.org/10.1093/ije/dyx239 (11 March 2018, date last accessed) http://dx.doi.org/10.1093/ije/dyx239 . Brooks SP , Gelman A . ( 1998 ) General methods for monitoring convergence of iterative simulations . J Comput Graph Stat 7 : 434 – 55 . Christensen AI , Ekholm O , Gray L , et al. . ( 2015 ) What is wrong with non-participants? Alcohol-, drug- and smoking related mortality and morbidity in a 12-year follow up study of participants and non-participants in the Danish Health and Morbidity Survey . Addiction 110 : 1505 – 12 . Google Scholar CrossRef Search ADS PubMed Dawson DA , Goldstein RB , Pickering RP , et al. . ( 2014 ) Nonresponse bias in survey estimates of alcohol consumption and its association with harm . J Stud Alcohol Drugs 75 : 695 – 703 . Google Scholar CrossRef Search ADS PubMed Galea S , Tracy M . ( 2007 ) Participation rates in epidemiologic studies . Ann Epidemiol 17 : 643 – 53 . Google Scholar CrossRef Search ADS PubMed Gelman A , Carlin JB , Stern HS , et al. . ( 2014 ) Bayesian Data Analysis , 3rd edition . Boca Raton, FL, USA : Chapman & Hall/CRC . Gmel G . ( 2000 ) The effect of mode of data collection and of nonresponse on reported alcohol consumption: a split sample study in Switzerland . Addiction 95 : 123 – 34 . Google Scholar CrossRef Search ADS PubMed Gorman E , Leyland AH , McCartney G , et al. . ( 2014 ) Assessing the representativeness of population-sampled health surveys through linkage to administrative data on alcohol-related outcomes . Am J Epidemiol 180 : 941 – 8 . Google Scholar CrossRef Search ADS PubMed Gorman E , Leyland AH , McCartney G , et al. . ( 2017 ) Adjustment for survey non-representativeness using record-linkage: refined estimates of alcohol consumption by deprivation in Scotland . Addiction 112 : 1270 – 80 . Google Scholar CrossRef Search ADS PubMed Halldin J . ( 1985 ) Alcohol consumption and alcoholism in an urban population in central Sweden . Acta Psychiat Scand 71 : 128 – 40 . Google Scholar CrossRef Search ADS PubMed Harald K , Salomaa V , Jousilahti P , et al. . ( 2007 ) Non-participation and mortality in different socioeconomic groups: the FINRISK population surveys in 1972-92 . J Epidemiol Commun H 61 : 449 – 54 . Google Scholar CrossRef Search ADS Hirvonen E . ( 2017 ) Puuttuvuuden mallintaminen FINRISKI -tutkimuksessa (in Finnish). Masters thesis, University of Jyväskylä, http://urn.fi/URN:NBN:fi:jyu-201706192945 (11 March 2018, date last accessed). International Standard Classification of Education: ISCED 2011 . UIS, Montreal, Quebec. Jousilahti P , Salomaa V , Kuulasmaa K , et al. . ( 2005 ) Total and cause specific mortality among participants and non-participants of population based health surveys: a comprehensive follow up of 54 372 Finnish men and women . J Epidemiol Commun H 59 : 310 – 5 . Google Scholar CrossRef Search ADS Karvanen J , Tolonen H , Härkänen T , et al. . ( 2016 ) Selection bias was reduced by re-contacting nonparticipants . J Clin Epidemiol 76 : 209 – 17 . Google Scholar CrossRef Search ADS PubMed Klingemann H , Gmel G (eds) . ( 2001 ) Mapping the Social Consequences of Alcohol Consumption . Dordrecht, The Netherlands : Kluwer Academic Publishers . Kopra J , Härkänen T , Tolonen H , et al. . ( 2017 ) Adjusting for selective non-participation with re-contact data in the FINRISK 2012 survey . Scand J Public Health doi:10.1177/1403494817734774 . Advance online publication. Kopra J , Karvanen J , Härkänen T . ( 2018 ) Bayesian models for data missing not at random in health examination surveys . Stat Model 18 : 113 – 28 . doi:10.1177/1471082×17722605. Google Scholar CrossRef Search ADS Lahaut VM , Jansen HA , Van de Mheen D , et al. . ( 2002 ) Adjusting for selective nonparticipation with recontact and hospitalisation history data . Alcohol Alcoholism 37 : 256 – 60 . Google Scholar CrossRef Search ADS PubMed Livingston M , Callinan S . ( 2015 ) Underreporting in alcohol surveys: whose drinking is underestimated? J Stud Alcohol Drugs 76 : 158 – 64 . Google Scholar CrossRef Search ADS PubMed Meiklejohn J , Connor J , Kypri K . ( 2012 ) The effect of low survey response rates on estimates of alcohol consumption in a general population survey . PLoS One 7 : e35527 . Google Scholar CrossRef Search ADS PubMed National Institute for Health and Welfare . ( 2016 ) Alcoholic Beverage Consumption 2016 [e-publication]. Helsinki, Finland. https://www.thl.fi/fi/tilastot/tilastot-aiheittain/paihteet-ja-riippuvuudet/alkoholi/alkoholijuomien-kulutus (11 March 2018, date last accessed). National Institute for Health and Welfare . ( 2017 ) Care Register for Health Care. Helsinki, Finland. http://www.thl.fi/en/web/thlfi-en/statistics/information-on-statistics/register-descriptions/care-register-for-health-care (11 March 2018, date last accessed). Mäkelä P . ( 2003 ) Impact of correcting for nonresponse by weighting on estimates of alcohol consumption . J Stud Alcohol 64 : 589 – 96 . Google Scholar CrossRef Search ADS PubMed Mäkelä P , Huhtanen P . ( 2010 ) The effect of survey sampling frame on coverage: the level of and changes in alcohol related mortality in Finland as a test case . Addiction 105 : 1935 – 41 . Google Scholar CrossRef Search ADS PubMed Mäkelä P , Tigerstedt C , Mustonen H . ( 2012 ) The Finnish drinking culture: change and continuity in the past 40 years . Drug Alcohol Rev 31 : 831 – 40 . Google Scholar CrossRef Search ADS PubMed O’Hagan A , Buck CE , Daneshkhah A , et al. . ( 2006 ) Uncertain Judgements: Eliciting Experts’ Probabilities . Chichester, England : John Wiley & Sons . Official Statistics of Finland (OSF) : Causes of death [e-publication]. Helsinki, Finland. http://tilastokeskus.fi/til/ksyyt/index_en.html (11 March 2018, date last accessed). Plummer M . ( 2003 ) JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling. In Proceedings of the 3rd International Workshop on Distributed Statistical Computing. 124, p. 125 . Wien, Austria: Technische Universit at Wien. Plummer M . ( 2015 ) rjags: Bayesian Graphical Models using MCMC. R package version 3-15. https://CRAN.R-project.org/package=rjags (11 March 2018, date last accessed). R Foundation for Statistical Computing . ( 2018 ) R: A Language and Environment for Statistical Computing. Vienna, Austria. https://www.R-project.org/ (11 March 2018, date last accessed). Robert CP , Casella G . ( 2004 ) Monte Carlo Statistical Methods . Vienna, Austria : Springer Texts in Statistics . Rehm J , Mathers C , Popova S , et al. . ( 2009 ) Global burden of disease and injury and economic cost attributable to alcohol use and alcohol-use disorders . Lancet 373 : 2223 – 33 . Google Scholar CrossRef Search ADS PubMed Reinikainen J , Tolonen H , Borodulin K , et al. . ( 2017 ) Participation rates by educational levels have diverged during 25 years in Finnish health examination surveys . Eur J Public Health , Advance online publication doi:10.1093/eurpub/ckx151 . Romelsjö A . ( 1989 ) The relationship between alcohol consumption and social status in Stockholm. Has the social pattern of alcohol consumption changed? Int J Epidemiol 18 : 842 – 51 . Google Scholar CrossRef Search ADS PubMed Statistics Finland . The Register of Completed Education and Degrees. Helsinki, Finland. https://www.stat.fi/til/kou_en.html (11 March 2018, date last accessed). Tanner M , Wong W . ( 1987 ) The calculation of posterior distributions by data augmentation . J Am Stat Assoc 82 : 528 – 40 . Google Scholar CrossRef Search ADS Torvik FA , Rognmo K , Tambs K . ( 2012 ) Alcohol use and mental distress as predictors of non-response in a general population health survey: the HUNT study . Soc Psych Psych Epid 47 : 805 – 16 . Google Scholar CrossRef Search ADS Zhao J , Stockwell TIM , MacDonald S . ( 2009 ) Nonresponse bias in alcohol and drug population surveys . Drug Alcohol Rev 28 : 648 – 57 . Google Scholar CrossRef Search ADS PubMed APPENDIX A The statistical model is based on work presented in Kopra et al. (2017a), which utilises similar model to estimate the prevalence of smoking. Notation for the data For each individual i = 1,…,N invited to a survey, we denote Mi being indicator of participation (Mi = 1 for participants and Mi = 0 for non-participants). Background information Xi consists of both survey frame and education information and is available for both participants and non-participants. The survey frame has variables gender gi, region ri, age ai and study year si. The education is denoted by ei, and values 1, 2 and 3 corresponds to high, middle and low education, respectively. Thus, Xi = (gi, ri, ai, si, ei). The heavy alcohol consumption Yi is a binary variable such that heavy alcohol consumers have Yi = 1 and non-heavy alcohol consumers have Yi = 0. The variable Ti is the age at the first diagnosis of any of the alcohol-related diseases. The Ti is right censored and left-truncated at the age when the person entered the study. Participation model The participation model logit(P(Mi=1|Xi,Yi))=α0[gi,si]+α1[gi,si,ei]+η[gi,si]Yi+α2[gi,Yi](ai−45)+α3[ri], (1) is a logistic regression model with following parameters. First, parameter α0[gi,si] is a constant where notation [gi,si] indicates that there are independent α0 parameters for all levels of gender gi and study year si. Second, parameter α1[gi,si,ei] is the regression coefficient for education levels. For the lowest education level, this parameter is forced to be 0. The parameter η[gi,si] describes how heavy alcohol consumption affects participation. For this parameter, we need an informative prior. The parameter α2[gi,Yi] describes how age at study affect participation. Finally, α3[ri] is a term for the region. For one of the regions, this parameter is forced to be 0. We selected a model that included important factors affecting participation while ensuring the convergence of the MCMC chains in Bayesian inference. Table A1. Prior distributions Notation Distribution Interpretation Participation model  η Logistic(0, τ = 2.05) How heavy alcohol consumption affects the participation.  α0, α1, α2, α3 N(0, 1000−1) Other parameters. Risk factor model  β0, β1 N(0,1000−1) Other parameters. Survival model  h0,0(25) Unif(0, 20) Hazard for men at age 25–26.  h0,1(25) Unif(0,20) Hazard for women at age 25–26.  h0,0(t), t = 26, 27, … Unif(h0,0(t–1), 20) Hazard for men at age t.  h0,1(t), t = 26, 27, … Unif(h0,1(t–1), 20) Hazard for women at age t.  γ1 N(0, 1000−1) How heavy alcohol consumption affects hazard for men.  γ2 N(0, 1000−1) How heavy alcohol consumption affects hazard for women. Notation Distribution Interpretation Participation model  η Logistic(0, τ = 2.05) How heavy alcohol consumption affects the participation.  α0, α1, α2, α3 N(0, 1000−1) Other parameters. Risk factor model  β0, β1 N(0,1000−1) Other parameters. Survival model  h0,0(25) Unif(0, 20) Hazard for men at age 25–26.  h0,1(25) Unif(0,20) Hazard for women at age 25–26.  h0,0(t), t = 26, 27, … Unif(h0,0(t–1), 20) Hazard for men at age t.  h0,1(t), t = 26, 27, … Unif(h0,1(t–1), 20) Hazard for women at age t.  γ1 N(0, 1000−1) How heavy alcohol consumption affects hazard for men.  γ2 N(0, 1000−1) How heavy alcohol consumption affects hazard for women. View Large Table A1. Prior distributions Notation Distribution Interpretation Participation model  η Logistic(0, τ = 2.05) How heavy alcohol consumption affects the participation.  α0, α1, α2, α3 N(0, 1000−1) Other parameters. Risk factor model  β0, β1 N(0,1000−1) Other parameters. Survival model  h0,0(25) Unif(0, 20) Hazard for men at age 25–26.  h0,1(25) Unif(0,20) Hazard for women at age 25–26.  h0,0(t), t = 26, 27, … Unif(h0,0(t–1), 20) Hazard for men at age t.  h0,1(t), t = 26, 27, … Unif(h0,1(t–1), 20) Hazard for women at age t.  γ1 N(0, 1000−1) How heavy alcohol consumption affects hazard for men.  γ2 N(0, 1000−1) How heavy alcohol consumption affects hazard for women. Notation Distribution Interpretation Participation model  η Logistic(0, τ = 2.05) How heavy alcohol consumption affects the participation.  α0, α1, α2, α3 N(0, 1000−1) Other parameters. Risk factor model  β0, β1 N(0,1000−1) Other parameters. Survival model  h0,0(25) Unif(0, 20) Hazard for men at age 25–26.  h0,1(25) Unif(0,20) Hazard for women at age 25–26.  h0,0(t), t = 26, 27, … Unif(h0,0(t–1), 20) Hazard for men at age t.  h0,1(t), t = 26, 27, … Unif(h0,1(t–1), 20) Hazard for women at age t.  γ1 N(0, 1000−1) How heavy alcohol consumption affects hazard for men.  γ2 N(0, 1000−1) How heavy alcohol consumption affects hazard for women. View Large Risk factor model The model for risk factor (heavy alcohol consumption) is logit(P(Yi=1|Xi))=β0[gi,ri,si,ei]+(si−ai−1938)β1[gi,ri,si,ei]. (2) The risk factor model is stratified by gender gi, region ri, study year si and education ei using similar notation as in (1). The parameter β0[gi,ri,si,ei] is constant for persons born in 1938. The parameter β1[gi,ri,si,ei] determines how the heavy alcohol consumption prevalence changes with the year of birth. Survival model Let dNi(t) be the number of new events (increment) for the individual i at the time t. The increment follows a Poisson distribution with intensity parameters λi(t). The intensity λi(t) is modelled independently for both genders consisting of one-year period piecewise-constant baseline hazard terms h0,0(t) for men and h0,1(t) for women, and heavy alcohol consumption term exp(γ1Yi) and exp(γ2Yi) indicating the effect of heavy alcohol consumption for men and women, respectively dNi(t)∼Poisson(λi(t)) λi(t)={exp(γ1Yi)h0,0(t),giventhatTi≥tandg=0exp(γ2Yi)h0,1(t),giventhatTi≥tandg=10,Ti<t. Prior distributions The prior distributions are specified in Table A1. The prior distributions for piecewise constant hazard terms h0,0(t) and h0,1(t) are specified such that the hazard becomes increasing function with respect to t. APPENDIX B The study questions in 1987 Consumption of Alcohol 1. Do you use any alcoholic drinks, even occasionally (f. ex. beer, wine or spirits)? 1 yes 2 no, but I have not quitted completely 3 no, because I quit using alcohol…… years ago 4 I have never used alcohol If you have quitted alcohol use, please specify, why did you quit? no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 2. Have you during the past year (last 12 months) had any alcohol (beer, wine or spirits)? 1 yes 2 no (for your part, the questions are completed) 3. How often do you usually drink beer (III or IV A)? 1 daily 2 a few times a week 3 about once a week 4 few times a month 5 about once a month 6 about once in a few months 7 3–4 times a year 8 twice a year 9 once a year or more seldom 0 never 4. How much do you usually drink beer at a time? 1 less than one bottle 2 1 bottle 3 2 bottles 4 3 bottles 5 4–5 bottles 6 6–9 bottles 7 10–14 bottles 8 15 bottles or more 9 I do not drink beer 5. How often do you usually drink wine (light or strong, also home made)? 1 daily 2 a few times a week 3 about once a week 4 a few times a month 5 about once a month 6 about once in a few months 7 3–4 times a year 8 twice a year 9 once a year or more seldom 0 never 6. How much do you usually drink wine at a time? 1 half a glass 2 one glass 3 two glasses 4 about half a big bottle 5 a little less than one big bottle 6 about one big bottle 7 from one to two big bottles 8 more than two big bottles 9 I do not drink wine 7. How often do you usually drink spirits? 1 daily 2 a few times a week 3 about once a week 4 a few times a month 5 about once a month 6 about once in a few months 7 3–4 times a year 8 twice a year 9 once a year or more seldom 0 never 8. How much do you usually drink spirits at a time? 1 less than one restaurant measure (less than 4 cl) 2 one restaurant measure (about 4 cl) 3 two restaurant measures (about 8 cl) 4 3–4 restaurant measures 5 5–6 restaurant measures (about quarter liter) 6 7–10 restaurant measures 7 about a half liter bottle 6 more than a half liter bottle 7 I do not drink spirits 9. How often have you during the last 12 months had so much beer, wine or spirits that you have felt intoxicated? 1 a few times a week or more often 2 about once a week 3 a few times a month 4 about once a month 5 about once in two months 6 5 times a year 7 2–3 times a year 8 once a year 9 not even once The changes in questions from 1987 to 1992 The questions 1, 6 and 7 have with changes in text. We have highlighted the removed text with strikeout font (e.g. removed) and added text with italic font (e.g. added). The changes are in comparison with the previous survey. 1. Do you use any alcoholic drinks, even occasionally (f. ex. beer, wine or spirits)? 1 yes yes, at least once a month 2 no, but I have not quitted completelyyes, less than once a month 3 no, because I quit using alcohol …… years ago 4 I have never used alcohol If you have quitted alcohol use, please specify, why did you quit? no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 no yes For health reasons 1 2 For economic reasons 1 2 For other reasons 1 2 6. How much do you usually drink wine at a time? 1 half a glass 2 one glass 3 two glasses 4 about one small bottleabout half a big bottle 5 a little less than one big bottle 6 about one big bottle 7 from one to two big bottles 8 more than two big bottles 9 I do not drink wine 7. How much do you usually drink spirits at a time? 1 less than one restaurant measure (less than 4 cl) 2 one restaurant measure (about 4 cl) 3 two restaurant measures (about 8 cl) 4 3–4 restaurant measures 5 5–6 restaurant measures (about quarter liter) 6 7–10 restaurant measures 7 about a half liter bottle 6 more than a half liter bottle 7 I do not drink spirits The changes in questions from 1992 to 1997 The questions 4 and 6 have with changes in text. We have highlighted the removed text with strikeout font (e.g. removed) and added text with italic font (e.g. added). The changes are in comparison with the previous survey. 4. How much do you usually drink beer at a time?(1 bottle = 1/3 liters.) 1 less than one bottle 2 1 bottle 3 2 bottle 4 3 bottles 5 4–5 bottles 6 6–9 bottles 7 10–14 bottles 8 15 bottles or more 9 I do not drink beer 6. How much do you usually drink wine at a time? 1 half a glass 2 one glass (1 glass = c. 12 cl) 3 two glasses 4 about half a big bottle (1 bottle = 0,75 l) 5 a little less than one big bottle 6 about one big bottle 7 from one to two big bottles 8 more than two big bottles 9 I do not drink wine The changes in questions from 1997 to 2002 In 2002 the questions 3–8 have been replaced with a new question number 3. 3. How often did you drink the following amounts in one day during the last 12 months? Instruction: Start answering from the first row. Mark (x) the most suitable ‘How often?’ alternative. Then continue row at a time down in the same manner. Please mark only one alternative per row. 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses Doses per day Never Once a month or more seldom 2–3 times a month About once a week 2–3 times a week 4–5 times a week 6–7 times a week 15 or more □ □ □ □ □ □ □ 13–14 □ □ □ □ □ □ □ 11–12 □ □ □ □ □ □ □ 9–10 □ □ □ □ □ □ □ 7–8 □ □ □ □ □ □ □ 5–6 □ □ □ □ □ □ □ 3–4 □ □ □ □ □ □ □ 1–2 □ □ □ □ □ □ □ Doses per day Never Once a month or more seldom 2–3 times a month About once a week 2–3 times a week 4–5 times a week 6–7 times a week 15 or more □ □ □ □ □ □ □ 13–14 □ □ □ □ □ □ □ 11–12 □ □ □ □ □ □ □ 9–10 □ □ □ □ □ □ □ 7–8 □ □ □ □ □ □ □ 5–6 □ □ □ □ □ □ □ 3–4 □ □ □ □ □ □ □ 1–2 □ □ □ □ □ □ □ Doses per day Never Once a month or more seldom 2–3 times a month About once a week 2–3 times a week 4–5 times a week 6–7 times a week 15 or more □ □ □ □ □ □ □ 13–14 □ □ □ □ □ □ □ 11–12 □ □ □ □ □ □ □ 9–10 □ □ □ □ □ □ □ 7–8 □ □ □ □ □ □ □ 5–6 □ □ □ □ □ □ □ 3–4 □ □ □ □ □ □ □ 1–2 □ □ □ □ □ □ □ Doses per day Never Once a month or more seldom 2–3 times a month About once a week 2–3 times a week 4–5 times a week 6–7 times a week 15 or more □ □ □ □ □ □ □ 13–14 □ □ □ □ □ □ □ 11–12 □ □ □ □ □ □ □ 9–10 □ □ □ □ □ □ □ 7–8 □ □ □ □ □ □ □ 5–6 □ □ □ □ □ □ □ 3–4 □ □ □ □ □ □ □ 1–2 □ □ □ □ □ □ □ The changes in questions from 2002 to 2007 In 2007 the new question number 4 has been updated with a small change in the instructions and a change in the categories of consumed doses per day. 4. How often did you drink the following amounts in one day during the last 12 months? Instruction: Start answering from the first row. Mark (x) the most suitable ‘How often?’ alternative. Then continue row at a time down in the same manner. Please mark only one alternative per row. 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses 1 dose = bottle (1/3 liter) beer (class III) or a glass (12 cl) of light wine or a glass (8 cl) of strong wine or a glass (4 cl) of spirits or other strong liquor Bottle (0.33 liter) beer (class IV), Gin Long Drink or strong cider = 1.25 doses Large bottle (0.5 liter) beer (class III) = 1.5 doses Large bottle (0.5 liter) beer (class IV) = 2 doses Bottle (0.75 liter) wine = 7 doses Bottle (0.75 liter) strong wine = 10 doses Bottle (0.5 liter) strong alcohol (e.g. Koskenkorva) = 12 doses Doses per day At least 4 times a week 2–3 times a week About once a week 1–2 times a month 3–10 times a year 1–2 times a year Never 18 or more 1 2 3 4 5 6 7 13–17 1 2 3 4 5 6 7 8–12 1 2 3 4 5 6 7 5–7 1 2 3 4 5 6 7 3–4 1 2 3 4 5 6 7 1–2 1 2 3 4 5 6 7 Doses per day At least 4 times a week 2–3 times a week About once a week 1–2 times a month 3–10 times a year 1–2 times a year Never 18 or more 1 2 3 4 5 6 7 13–17 1 2 3 4 5 6 7 8–12 1 2 3 4 5 6 7 5–7 1 2 3 4 5 6 7 3–4 1 2 3 4 5 6 7 1–2 1 2 3 4 5 6 7 Doses per day At least 4 times a week 2–3 times a week About once a week 1–2 times a month 3–10 times a year 1–2 times a year Never 18 or more 1 2 3 4 5 6 7 13–17 1 2 3 4 5 6 7 8–12 1 2 3 4 5 6 7 5–7 1 2 3 4 5 6 7 3–4 1 2 3 4 5 6 7 1–2 1 2 3 4 5 6 7 Doses per day At least 4 times a week 2–3 times a week About once a week 1–2 times a month 3–10 times a year 1–2 times a year Never 18 or more 1 2 3 4 5 6 7 13–17 1 2 3 4 5 6 7 8–12 1 2 3 4 5 6 7 5–7 1 2 3 4 5 6 7 3–4 1 2 3 4 5 6 7 1–2 1 2 3 4 5 6 7 © The Author(s) 2018. Medical Council on Alcohol and Oxford University Press. All rights reserved. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices)

Journal

Alcohol and AlcoholismOxford University Press

Published: Sep 1, 2018

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off