Alcohol and the so-called prevention paradox: how does it look today?


Sometimes, high-risk individuals may contribute quite substantively to the overall rate of certain problems in society, in spite of the low prevalence of such individuals. However, in other cases, the major share may come from high-prevalence groups with low or moderate risk. What will actually be observed in each particular case will vary according to the type of health or social problem we are studying, and it will depend on the relative risk and the relative size of the groups in question.

The observation that groups with low and moderate risk may be responsible for the majority of certain types of alcohol problems was probably made first by Moore & Gerstein (1981). The observation was brought to the attention of a wider audience in Kreitman's analysis a few years later (Kreitman 1986). The fact that a fancy name—the ‘prevention paradox’ (borrowed from Rose 1981)—was attributed to the observation, may have contributed to its rise to fame. Even in science, what things are called is not irrelevant. However, there is nothing truly paradoxical about this observation. Perhaps it came as a surprise to some, the result not being what they intuitively expected, which only shows that intuitions can be highly unreliable.

The validity of the ‘prevention paradox’ for alcohol problems seems to vary across the spectrum of such problems. In the case of somatic health problems with strongly convex risk functions, such as cirrhosis of the liver, the very heavy drinkers are responsible for the majority of the problems (Skog 1985, 1999a; Norström 1995). However, when it comes to acute health and social problems due to alcohol, drinkers at ‘normal’ levels seem to be responsible for the majority of the problems (Skog 1985, 1999b; Norström 1995; Stockwell et al. 1996; Gmel et al. 2001; Gruenewald et al. 2003; Rossow & Romelsjö 2006).

However, the result also depends on how the risk groups are delineated. In the studies mentioned above, drinkers were classified according to annual intake. Obviously, other criteria could be used. In their critique of Kreitman's contribution, Stockwell et al. (1996, p. 7) claim that the ‘preventive paradox disappears when consideration is given to the amount of alcohol consumed either (i) the day of highest alcohol intake out of the last four, or (ii) the day on which acute alcohol-related harm occurred’. They find the majority of the problems among those who engage in binge drinking. This result is quite natural, as the problems under study are related to drunkenness, and the finding has been confirmed in several of the above-mentioned studies.

Furthermore, it has been suggested that the two observations—(i) the majority of acute social problems are found among consumers who drink moderately in terms of annual intake and (ii) the majority of such problems occur in the high-risk group, when defined in terms of amount per occasion—are easily reconciled (Skog 1999a). It simply means that most of the binge drinking is found among consumers with a moderate annual consumption level. This so-called second-order ‘prevention paradox’ has been confirmed by Gmel et al. (2001) and Rossow & Romelsjö 2006).

The above summary is probably not very controversial. The following three issues still need clarification: (1) Can the observations be trusted, or are we up against serious methodological problems here? (2) How sensitive are the results to cultural context? (3) What are the implications of these observations for prevention?


The evidence is derived mainly from population surveys, and we all know that we are faced with severe methodological problems in this area. First, we have measurement problems and potential biases in the samples. Secondly, the causal attribution of problems to drinking may be problematic. Neither of these problems have been seriously addressed in the ‘prevention paradox’ literature.

  • • In most surveys, response rates are far below 100%. If non-response is random, this problem is not serious, but only the most dedicated survey-analyst would believe that it is. Occasionally, serious attempts are made at evaluating selective non-response, for instance by comparing response rates for identified problem drinkers and the general population, but most often only trivial controls are made in terms of demographic variables. The serious analyses suggest that samples can indeed be severely biased. For instance, Nilsson & Svensson (1971) demonstrated that non-respondents in a Swedish population survey had a public record of alcohol abuse three times as often as the respondents, and Hauge & Nordli (1983) demonstrated that the response rate among Norwegian youth who had a police record for (mostly minor) drug offences was only about half the response rate in the youth population at large. Hence, it is very likely that individuals with drinking problems are underrepresented in surveys, and this clearly distorts the results. Underestimation of the share of problems due to heavy consumption (however defined) would be the result, i.e. a bias in the direction of the ‘prevention paradox’.
  • • Those actually responding typically offer less than perfectly accurate information about their own drinking and the harm experienced. Even if these errors were entirely unsystematic, they would tend to bias results. It is well known that unsystematic error in the independent variable (in casu, consumption) will result in underestimation of the effect of this variable on the dependent variable (Johnston 1991). The result would be overestimation of the moderate and light drinkers’ share of the problems. On the other hand, systematic errors could, in general, bias results in both directions. Of particular concern, however, is the tendency of some respondents to report quite trivial incidents as drinking related problems—incidents that other respondents would not count as a problem. This would tend to bias results in the direction of the ‘prevention paradox’. The effect would be the same when heavy drinkers with real problems conceal or downsize their problems. More generally, the ‘problems’ reported typically in surveys might represent a highly mixed lot, and we certainly need better data on the dependent variable (in casu, harm). A simple procedure would consist of obtaining information from identified victims of different types of damage, both regarding the causes of the incident and the drinking habits of the victim/perpetrator. Emergency-room surveys and similar approaches should offer a good opportunity for an alternative testing of the ‘prevention paradox’. It should be more reliable regarding the dependent variable—although not necessarily regarding the independent variable.
  • • The survey data reported in evaluations of the ‘prevention paradox’ report typically only cross-tabulations of consumption level (however defined) and alleged drinking-related problems. The causal status of these associations is seldom analysed more thoroughly. However, we should not take for granted that the alcohol-related incidents reported by, for instance, the moderate consumers, are in fact caused by their drinking—and the same applies to the heavier drinkers. The respondent's causal interpretation may be less reliable than his or her description of the event. If individuals with certain attributes (e.g. low self-control, impulsivity, sensation-seeking) experience many problems because of the way they behave, and if they also drink more than average they may have a tendency to blame their misbehaviour on alcohol. This particular mechanism would exaggerate the importance of the heavy drinking group, but one could also imagine mechanism with the opposite effect. Students of the ‘prevention paradox’ still have some work waiting to be conducted.


High- and low-risk drinkers’ share of the overall problem rates may vary across cultures, in response to differences in drinking patterns. We clearly need data from a much wider range of cultures. Most studies to date are based on data from Northern Europe, North America and Australia. We cannot take for granted that the different risk groups’ shares are the same outside this limited selection of drinking cultures. In cultures where a high level of intoxication is a common phenomenon, the distribution of social problems could be quite different, compared to cultures without this tradition. In cultures where binge drinking is less common, the majority of these drinkers may be very heavy drinkers, in which case the second-order ‘prevention paradox’ would not apply.


The distinction between these inferences has not always been made clear, and this may explain some of the disagreements.

  • • First, a conservative inference would simply be that the prevention of alcohol problems should focus not only on the very heavy drinkers (defined in terms of annual intake). Today this may seem rather trivial, but 30 years ago it was not, and we need to see things in their proper historical perspective. For instance, until the mid-1970s WHO's main interest in alcohol problems was centred around ‘alcoholism’. Towards the end of the 1970s and the early 1980s a shift occurred, due partly to changes in the conception of alcohol dependence (Edwards et al. 1977), but probably also as a result of the so-called ‘purple book’ (Bruun et al. 1975), which advocated a more general public health perspective on alcohol problems. It was claimed that heavy drinking and ‘alcoholism’ cannot be seen in isolation from the rest of the drinking population. The latter book paid little attention to social drinking problems, but this void was filled during the following years. The observation that the majority of social problems are found among ‘normal’ drinkers, rather than among ‘alcoholics’ and ‘heavy’ drinkers, fits neatly into Bruun et al.'s picture.
  • • A less conservative inference would be this: Stockwell et al. (2004, p. 67) argue that the ‘prevention paradox’, if valid, is ‘recommending a universal or whole of populations approach to prevention’—as opposed to a targeted high-risk approach. This interpretation also seems to have been implicit in Kreitman's analysis. However, this may, or may not, be a reasonable interpretation, depending on the definition of terms, as well as the empirical effects of preventive efforts.

On one hand, the second-order ‘prevention paradox’ indicates that there are high-risk groups within low-risk segments: some of those who drink moderate amounts per year are binge drinkers. A ‘whole population’ strategy, if it has any demonstrable effect at all, may also affect this high-risk segment, with the desired outcome. The down-side, of course, is that many ‘innocent’ consumers are also affected. On the other hand a targeted strategy, aimed exclusively at binge drinkers, may have the desired effects, and/or unintended side-effects, or none at all. If it has the desired effect and no unintended side-effects, all is well. The ‘prevention paradox’ is not incompatible with this state of affairs. The empirical question is, however, if sufficiently powerful targeted strategies are available. The Swedish rationing system is an interesting historical example of a highly effective targeted strategy (Norström 1987), although its feasibility today is nil.

Perhaps Room (2001) is right when he suggests that most of the strategies of proven effectiveness are the unpopular ones, where targeted and untargeted segments of the population alike are affected. Perhaps free lunches are as rare in alcohol prevention as in economics. If that is so, ‘draining the ocean to prevent shark attacks’ may turn out to be the only truly effective strategy, to borrow a somewhat exaggerated metaphor from Rehm (1999). On the other hand, one obviously cannot exclude a priori that better diagnostic tools and better intervention strategies may be found some day, allowing effective targeted prevention strategies for high-risk groups. This remains to be seen.

In conclusion, much of the disagreement around the ‘prevention paradox’ seems to derive from the failure to make a clear distinction between the conservative and the non-conservative inference. The non-conservative inference is obviously problematic, because it certainly does not follow from the empirical observations that focused high-risk strategies will fail. The conservative interpretation, namely that one should avoid focusing exclusively on the very heavy drinkers, is unproblematic—but perhaps not extremely useful either. Then again, the non-conservative inference is not very useful either, being unwarranted. The empirical facts underlying the ‘prevention paradox’ never suggested that strategies directed at high-risk drinkers are useless.

So far, the ‘prevention paradox’ literature has been limited in scope and only indirectly relevant for prevention, as the efficacy of preventive measures is not addressed. Even if most of the problems should be found in low-risk groups, this is not very useful if measures aimed at preventing these problems are inefficient in this group. Similarly, even if certain problems derive mainly from high-risk groups, preventive measures for this group could be ineffective. If, on the other hand, the latter measures are effective against moderate and low-risk groups, they could be useful—even if this group contributes less than half of the problems. Consequently, the prevention debate needs to focus more on the differential efficacy of strategies, and not only the distribution of problems across risk groups.

My own reading of the literature suggests that a sensible prevention policy needs to apply both targeted and population strategies. Furthermore, measures ought to be aimed at both drinking pattern (frequency of intoxication) and consumption level (intake per year), as well as drinking contexts. One does not have to choose one or the other. A sensible policy should be a mix of these elements, and the mix will have to vary across drinking cultures, depending on drinking habits and the nature of the problem, as well as on political judgements of what is gained and what is lost.