|
Article Excerpt
The recent World Health Report (World Health Organization, 2002) again confirms that alcohol consumption is one of the major risk factors for morbidity and mortality worldwide. In established market economies, 9.2% of the disease burden in 2000 was attributable to alcohol, surpassed only by the disease burden associated with tobacco (12.2%) and blood pressure (10.9%). In emerging societies with low child and adult mortality rates, alcohol was the most important out of the 27 risk factors examined, accounting for 6.2% of the burden of disease. Only in poorer developing countries, where risk factors such as malnutrition, poor water conditions and unsafe sex are highly prevalent, is alcohol not among the 10 leading risk factors. Given the importance of alcohol as a risk factor, a valid and reliable measurement of alcohol consumption is needed. As can be inferred from earlier reviews (Alanko, 1984; Rehm, 1998b; Room, 1990; Sobell and Sobell, 1995a), there is not a single widely accepted, "optimal" instrument to measure alcohol consumption. The present paper attempts to give an overview of the measurement instruments available; to provide guidelines on which measures, under which circumstances, may be more useful than others; and to give examples of how to model the association between alcohol consumption and related outcomes.
The measurement of alcohol consumption cannot be seen independently from its different foci. Following Beaton (1994), four approaches to data analysis can be distinguished in alcohol research:
1. Estimation of means for groups or comparison of means across groups.
2. Correlation and regression based on continuous data.
3. Distributional analysis--e.g., establishing the distribution of average alcohol consumption in a population or determining thresholds for individuals with excessive intake.
4. Categorical analysis, classifying individuals by intake and looking at covariation with occurrence of disease, as in epidemiological studies (e.g., classifying individuals as abstainers, 0.1-20 g/day; 20.1-40 g/day: > 40 g/day).
With these four approaches four aspects have to be considered: a) whether alcohol consumption is the dependent or the independent variable, b) the level of data aggregation, c) whether between-individuals and within-individual variation in consumption is of major interest or concern, and d) how strongly systematic biases of measurement influence the analysis. One argument for using alcohol measurements despite potential systematic bias--e.g., because of the underestimation of alcohol consumption owing to denial or concealment by respondents in surveys--is that rank order is preserved. This means that although individuals may underreport their consumption, it occurs either linearly or proportionately related to the respondents' "true" consumption, so "true" heavier consumers still report higher consumption than "true" moderate consumers. Such systematic errors are acceptable in correlational analysis or in mean comparisons if rank order is preserved, which is often the case (Willett, 1998). Correlational analysis or regression analysis may, for example, look at the individual level association between alcohol consumption (independent variable) and some outcome (dependent variable)--e.g., disease occurrence, social well-being, or aggression. Mean comparisons may look at group level differences in alcohol consumption (dependent variable) across, for example, sex or socioeconomic status. In such analyses, the precise level of alcohol consumption is not as important as whether there are significant differences across groups in their (mean) alcohol consumption or whether there are significant associations, in that with increasing consumption the occurrence of the outcome increases or decreases. The level of aggregation may be different, and concern the population as a whole. For example, one question could be whether a change in alcohol pricing was accompanied by a change in per capita consumption at the population level, or whether alcohol consumption co-varies with liver cirrhosis rates. The latter question could be analyzed by looking at the temporal covariation of alcohol sales data and recorded statistics of liver disease. A drawback of sales statistics is that although they may give a more precise estimate of a population's overall volume of consumption, they cannot be used to measure the number of abstainers or mean consumption in different subgroups--e.g., by socioeconomic status or even sex. Data based on individually derived measures are usually necessary in this respect.
Distributional and categorical analysis attempts to classify individuals according to their drinking status. The focus of such research may be, for example, to determine the proportion of abstainers or heavy drinkers according to some given threshold. Another goal may be to determine the threshold of low-risk drinking limits for all-cause mortality (Rehm et al., 2001), or the threshold of alcohol consumption at which drinkers are at elevated risks for a certain disease, if there is not a monotonically increasing (e.g., linear) risk relation. The analysis is therefore different from simple regression analysis, which can be based on correlations and where systematic underreporting by everybody would not distort results. For distributional and categorical analysis, the exact level of consumption is important to know and makes a difference. One needs to know that a certain risk is associated with a certain drinking level--e.g., 50 g pure alcohol--and this is different from a simple association stating that any increase in consumption is associated with a corresponding increase in risk. Categorical analysis would be used, for example, to create sensible drinking guidelines (Bondy et al., 1999). It is also different from group mean comparisons, because the "exact" number of consumers at each level of the distribution is needed, which depends on aspects of the measurement instrument. To give an example, the measurement of yesterday's consumption, if projected to annual consumption, may provide an unbiased estimate of a group's mean annual consumption, but it will certainly overestimate the number of abstainers if all yesterday-abstainers are taken as people abstaining in the past 12 months. Thus, the recall period for measuring alcohol consumption is important (see Rehm, 1998b, for an example, and the detailed explanation below). In summary, distributional and categorical analysis requires the most detailed and most precise information because "true" actual intake is needed (de Vries et al., 1999).
The measurement of alcohol consumption to estimate a risk relationship with an outcome is complicated by the fact that average consumption is not the only relevant factor: rather, the way alcohol is consumed--i.e., the drinking pattern--and consumption on a single occasion are also important. It becomes intuitively clear that there can be no single best instrument for this task when one considers that alcohol consumption is assumed to be a risk factor for about 60 different disease consequences alone (Gutjahr and Gmel, 2001; Rehm et al., 2003b). Some of these consequences are assumed to be associated with volume--i.e., the usual or average consumption over time--which would favor measurement instruments best able to address mean consumption over a certain time. It should be noted that in this context not the mean of a group, but mean consumption of individuals is considered. Examples would be the different cancers affected by chronic alcohol use, which are usually found to have monotonically increasing risk functions. The problem is aggravated if an individual's mean consumption over a short time period is no longer of interest, but the accumulated amount over an extended period or even a lifetime is of importance. The length of the period over which respondents must recall their consumption may be related to measurement errors, such as declining precision of recall, owing to memory deficits.
Other consequences are clearly related to consumption "in the event," which means that consumption in a single drinking occasion is associated with the outcome (risky single occasion drinking = RSOD). A typical example would be alcohol consumption while operating a machine or driving a car. For such consequences, the measurement of the most recent drinking occasion will be of most importance. Measures like blood alcohol concentration (BAC) will be more valuable than those that measure the accumulated volume over time. However, even here the association is less than simple. For example, the reanalysis of the Grand Rapids study (Hurst et al., 1994) has shown that the risk for traffic casualties increases more steeply for less frequent drinkers than for frequent drinkers, probably because more experienced drinkers ale more likely to cope better with the impairing cognitive and psychomotor limitations compared with less experienced drinkers or may have developed more tolerance to the effects of alcohol (Midanik et al., 1996). Thus, in addition to measures "in the event," usual alcohol consumption is needed to establish the risk relationship for differently experienced drinkers.
Even more complicated is the association of outcomes for which it has become clear that the interrelation between mean consumption (volume) and RSOD plays a role. For example, there is good evidence that 14 glasses a week when drunk in the form of two glasses every day may be protective for coronary heart disease, but the same mean of glasses drunk in form of seven drinks on only Friday and Saturday will have a detrimental effect on coronary heart disease (Puddey et al., 1999; Rehm et al., 2003c). Thus, to predict coronary heart disease, both volume and drinking patterns--i.e., the way the same amount is consumed--have to be measured.
According to Alanko (1984), there are two fundamental dimensions of alcohol consumption measurement: the occurrence of drinking over time and the varying amounts of alcohol consumed on the drinking occasions. Thus, the task is to estimate the central tendency for both quantities and frequencies and also the within-individual and between-individuals variability. Figure 1a gives an example of intra-individual variability for nine randomly chosen drinkers of a seven-day-diary survey in Switzerland (work in progress). Figure 1b also shows that even after averaging over all individuals in the sample (n=747), between-days variation remains over the week. The higher the individual variability in drinking (measured as within-individual variance), the stronger the increase in weekend drinking compared with drinking on workdays. Measurement instruments differ according to their ability to detect within-individuals" variation in drinkers and therefore different kinds of drinkers (Dawson, 1998).
[FIGURE 1 OMITTED]
Measurement instruments at the individual level
Alcohol measurement instruments at the individual level can be subdivided into subjective measures (self-reports) and objective measures. Individuals' self-reports can be further classified into two broad categories. The respondents are asked either to summarize their drinking pattern/behavior over a predefined period of time or to report their most recent drinking occasions in a detailed way (Room, 1990). In the latter, respondents are asked to list their drinking over a usually short...
|
