International Guidelines for the Estimation of the Avoidable Costs of Substance Abuse

3 Avoidable costs: the Feasible Minimum

3.1 Introduction

Where a particular event or medical condition can have more than a single cause it is necessary to have estimates of drug-attributable fractions. For instance, alcohol is not the only cause of road accidents nor is smoking the only cause of fire deaths. If the smoking-attributable fraction for lung cancer were estimated to be 0.84, it would then be known that 84 per cent of lung cancer cases were caused by smoking, the remaining 16 per cent of cases being attributable to other causes. In the absence of the relevant attributable fraction it is impossible to attribute the correct proportion of the total harm to substance abuse. In almost all cases the use of direct measures involves knowledge of attributable fractions.

This requirement can represent, in some areas of harm, a major difficulty in cost estimation. Calculation of attributable fractions (also known as aetiological fractions or attributable proportions) requires two fundamental pieces of information - the relative risk (measuring the causal relationship between exposure to the risk behaviour and the condition being studied) and prevalence (measuring the proportion of the relevant population engaging in the risky activity). For some types of harm the relative risk can be assumed to be similar for genetically and economically similar populations. Applying the estimated prevalence for each population to the relevant relative risk will yield attributable fractions which can be used to estimate harm in various populations (countries).

In practice, the attributable fractions in cost studies are derived from estimates of relative risks derived from research studies across a range of comparable countries. Generally, the calculations of the attributable fractions for all countries have in the past derived from studies conducted in a large set of countries with similarly high levels of economic development.

However, for some types of harm it would be quite unsafe to assume similar relative risks even in countries at similar levels of economic development. The most obvious type of harm in this context is crime where, for a range of cultural, social, legal and other reasons, relative risks can vary greatly even between countries with similarly high per capita incomes. A good example of the different rates of this type of harm is alcohol-attributable violence where experience varies greatly, even among Western European countries.

The manner in which attributable fractions are derived allows us to estimate how many deaths and hospitalizations are attributable to substance abuse. If there had been no substance abuse, then all of these deaths and hospitalizations would not have occurred. But that does not mean that all of those deaths and hospitalizations were avoidable. There are at least three major difficulties with considering all deaths and hospitalizations caused by substance abuse to be avoidable:

1. Even if substance abuse were to end immediately, many deaths and hospitalizations (and other adverse consequences which lead to costs to the economy) would continue due to the lagged effects of past substance abuse.
   
   
2. When a risk factor (like substance abuse) causes a death or hospitalization, in a sense it may prevent another risk factor from causing a death or hospitalization. The obvious example is the fact that mortality is inevitable: everyone dies eventually from something or another, and if it is not due to substance abuse, it will be due to another cause.
   
   
3. There would necessarily be huge policy costs to end all substance abuse (in the unlikely event that this were possible), and it virtually impossible to estimate these costs as such policies have not even been identified, let alone costed.

The extent of the first problem is very difficult (and in practical terms, virtually impossible) to estimate. Given a set of assumptions concerning time intervals between different levels of use or risky behaviours and the onset of disease or death (information that is generally lacking), one could attempt to make estimates of lagged effects for some causes. But even if this could be done (which is highly doubtful), it would be limited to only a small subset of diagnoses. And in any case, the exercise would be of little use due to the second and third problems. Even with perfect information on the aetiology of substance-related diseases and accidents, researchers must still confront the other two problems.

It could be considered that the term substance abuse is misleading, because a significant portion of the burden attributable to drugs is caused by use only, i.e. by individuals who do not fall under the gold standard definitions of substance dependence or abuse (Rehm, 2003). For example, a road accident may be attributable to the intoxication (as defined by legislated maximum permissible blood alcohol content) of a driver who, nevertheless, would not fulfil any criterion of dependence or abuse in the International Classification of Diseases, Tenth Revision (ICD-10) (WHO, 1992-1994) or the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV-R) (American Psychiatric Association, 1994). Such alcohol use is often labelled as alcohol misuse. On the other hand, much alcohol use can, by any definition, clearly be categorised as abuse

Since social cost studies are essentially economic studies, they require a definition of substance abuse which is meaningful in economic terms. To economists, substance abuse exists when substance use involves the imposition of social costs additional to the resource costs of the provision of that drug. Abuse occurs if society, including the substance user, incurs extra costs as a result of the drug use. Any use of illicit substances is deemed to be abuse since if use is deemed to be illegal it is clearly considered by society to be abuse

For the purposes of simplicity of expression in these guidelines the term substance abuse is used throughout and is defined as consisting of tobacco abuse, illicit drug abuse, and alcohol abuse and misuse.

3.2 The concept of avoidability

The social costs of substance abuse worldwide are high (see, for example, Single et al, 1998; Harwood et al, 1998; Collins and Lapsley, 2002; and Andlin-Sobocki and Rehm, 2005). These costs are mainly related to the costs of health care, crime and law enforcement, and losses in productivity. The substance abuse-related burden of disease is composed of the mortality and morbidity attributable to the abuse of alcohol, tobacco, and illegal drugs and is one of the main underlying components in substance abuse-related social costs. Knowing the overall attributable costs of substance abuse or the overall substance-abuse attributable disease burden has been found unsatisfactory, as it is not clear which proportion of the costs or the burden could in principle be changed. This changeable part has been labelled the avoidable cost or avoidable burden (WHO, 2002). Estimating avoidable burden is an important element in the process of estimating avoidable costs.

The first step in estimating avoidable burden is to conceptualise the attributable burden of disease; that is, the burden of a given disease in a given population that is identified as due to a specific exposure to a risk factor or multiple risk factors. Consequently, that portion of disease burden could, in principle, be reduced or eliminated if the causative exposure is reduced or eliminated. Attributable burden is conceptualized regardless of whether such a reduction is achievable in practice or not.

Based on the conceptualization of attributable burden, it is then possible to introduce the term avoidable burden of disease. The latter term denotes the proportion of disease burden that can be reduced by changing the current exposure distribution to an alternative, more favoured, exposure distribution. Clearly, the size of the avoidable burden caused by a given risk factor will always be smaller than or, at most, equal to the burden attributable to that risk factor. Little has been written on the problems of estimating avoidable burdens/costs and so a discussion is presented below of a method to estimate avoidable burden specifically for substance use as a risk factor, with special emphasis on methodological problems and potential solutions.

3.3 Avoidability, optimality and zero tolerance

The avoidable burden/costs discussed here should be contrasted with the economist's concept of the optimal level of substance abuse. Economists argue that the optimal level of drug consumption is reached when the incremental cost to the community as a whole of achieving a given reduction in consumption is exactly matched by the incremental benefit to the community of that reduction. If the incremental benefit is greater than the incremental cost, achieving optimality would require a further reduction in consumption. If the cost exceeds the benefit, then consumption has been reduced to sub-optimal levels.

The concepts of avoidability and optimality can lead to quite different outcomes. It is perfectly possible that optimal levels of consumption may not be achievable. For example, optimal levels of tobacco consumption may be well below the levels which are achievable in a real world in which severe constraints exist on the public resources available to achieve reductions in smoking prevalence. On the other hand, it may be technically possible to reduce tobacco consumption below levels which economists judge to be optimal. In any event, there are likely in practice to be severe informational problems in determining optimal consumption levels. These guidelines concentrate on issues relating to the estimation of the avoidable costs of substance abuse, not the optimal levels of substance abuse.

Some public health advocates consider that the target for public health interventions should be a zero level of substance abuse. This might be called a zero tolerance approach. There are, for the purposes of these guidelines, problems with this approach from an economic perspective.

Economists would argue, as explained above, that in most situations the optimal outcome is not zero substance abuse but a level at which the additional costs of reducing abuse further are matched by the additional social benefits of that reduction. Even if zero substance abuse were achievable, in most instances the costs of achieving that outcome would exceed the benefits. In other words, the resources could be used more productively elsewhere.

In practice, zero abuse is not likely to be an achievable outcome. The concept of avoidable costs relates to what is achievable in the real world, not what would be desirable in an ideal world of unlimited resources.

3.4 Approaches to calculation of the Feasible Minimum

From an economic policy perspective, it is necessary to determine the maximum quantifiable, measurable reduction in substance abuse costs which effective policies can be expected to achieve. The lowest achievable level of substance abuse can be termed the Feasible Minimum, and the four methodologies discussed here can be regarded as ways in which a Feasible Minimum can be calculated in order to provide policy objectives. Of course, such avoidable cost estimates assume that the calculation of total cost estimates will have already been undertaken. Various possible approaches exist for estimation of the Feasible Minimum.

One method of achieving an estimate of a Feasible Minimum is the use of the classic epidemiological approach, deriving the attributable burden from calculations of relative risk and the prevalence. From this calculation of the attributable burden, both past and future risk factor distributions can be estimated, which also provide data to enable the calculation of a Feasible Minimum. This approach can be modelled to demonstrate the difference between the attributable and the avoidable burden.

Armstrong applied this concept in his work on prevention, using a measure which he described as the "Arcadian normal" as his feasible minimum for preventable mortality for a range of conditions. Instead of using epidemiological data from which to calculate the feasible minimum, he used as a comparator the lowest recorded rate (e.g. of lung cancer) which had been achieved by a country which could be considered a reasonable comparison with the study country. From that comparator, the amount which can be prevented can be estimated, and from that the calculation for avoidable costs can be made.

This concept does not in itself suggest that similar policies, regulations, and even health behaviours are necessarily appropriate nor transferable from the comparator country, but simply indicates what amount of burden reduction and costs have been able to be achieved. A disaggregation of effective policies from a comparator country may be useful in developing economic evaluation studies, and provide guidance on resource distribution which could be made to different policy implementations.

Another possible approach to estimation of the Feasible Minimum is to use recently-published WHO data on drug-attributable fractions, morbidity and mortality. These data can be used to identify best performance among countries in sub-regions identified by the WHO as having common characteristics. This approach could be adopted for countries where insufficient domestic data are available. Although the only practical solution for many countries, it should be noted that this method is likely to be highly imprecise, given the multiple layering of assumption underlying the application of attributable fractions from one setting to another.

There may also be circumstances in which evidence about the known effectiveness of specific interventions may be used in avoidable proportions. This constitutes a fourth type of approach.

All four approaches are described in more detail below. Inevitably there are difficulties and complexities both in deriving the attributable fraction and then in calculating the proportion which is avoidable, which indicates the Feasible Minimum.

3.5 From the classical to the distributional approach in epidemiology

Two different approaches to the epidemiology of risk factor attribution can be broadly distinguished: the classical epidemiological approach where, on the basis of a 2x2 table, disease risk can be estimated with respect to exposure and non-exposure of various populations to a single specific risk factor. Based on this categorically derived relative risk and the prevalence, the attributable burden can be derived, resulting in an estimate for a population who have already been exposed, i.e. focussing on the past. The main counterfactual scenario (Maldonado and Greenland, 2002) in this approach asks the question: "What would have happened if no exposure had occurred?"

More modern developments of epidemiology ask the question: "What would happen if risk factor distributions shifted to different counterfactual scenarios?" (Murray and Lopez, 1999). The modern approach not only looks at distributional shifts at one time, but also takes the future time dimension into consideration, and thus is able to predict future developments.

The current contribution focuses on the second approach. It thus conceptualizes the avoidable burden by specifying alternative scenarios for risk factor distributions, including potential future distributions. It is important to look at the impact of a risk factor over time, not just an epidemiological snapshot of today's attributable burden, in order to fully estimate the burden of disease contributions of acute and chronic disease burden which might be avoided in the future. This is an especially important factor for policy. We will also consider potential interactions and competing risks between influencing factors.

Note that the proposed framework is conceptualized independently from the measure for burden of disease used, whether it be mortality, morbidity, or any summary measure (such as the traditionally used Disability Adjusted Life-Year (DALY) (Murray and Lopez, 1997)), so that it can be universally applied.

3.5.1 An epidemiological model to estimate avoidable burden

Figure 1 illustrates the conceptual model of the difference between attributable and avoidable burden, cited from the epidemiological model of Murray et al (2003).

Figure 1 - A conceptual model of attributable and avoidable risk with increasing projected burden

• a. Disease burden at T_o attributable to prior exposure
  
  
• b. Disease burden at T_o not attributable to the risk factor of interest (caused by other factors only). The burden not attributable to the risk factor of interest (grey area) may be decreasing, constant or increasing over time. The constant case is shown in the figure.
  
  
• c. Disease burden avoidable at T_x with a 50% exposure reduction at T_o.
  
  
• d. Remaining disease burden at T_x after a 50% reduction in risk factor exposure

Source: Murray, C.J.; Ezzati, M.; Lopez, A.D.; Rodgers, A.; and VanderHoorn, S (2003). "Comparative quantification of health risks conceptual framework and methodological issues",

Population Health Metrics, 1(1): 1-20.

Consider the disease burden at time T_o (that portion which is attributable to prior exposure), denoted by the letter a in Figure 1. This is all the burden of disease which can be attributed to prior exposure of the risk factor under consideration before T_o.

In the example, a general situation is given where the background burden, i.e. the burden due to other factors except the risk factor under consideration, is constant over time. This background burden is denoted by the letter b at time T_o and, because it is constant over time in this specific example, the burden is the same size at all time points. Of course, in other situations the burden attributable to other factors than the risk factor under consideration may fluctuate.

The burden attributable to the risk factor under consideration in the example is increasing constantly over time until time T_o. Then different scenarios are shown. Let us discuss three of them. If nothing (e.g. intervention, changes in cultural acceptance) happens at time T_o, the attributable burden continues to increase linearly. On the other hand, if the exposure is reduced completely, then the attributable burden is decreasing until time T_x, when it reaches zero.

Consider tobacco use as an example in a society where prevalence rates have been increasing and would continue to increase without any intervention. If some drastic intervention could reduce smoking completely at a certain time point, smoking-related disease burden would not be zero the moment thereafter. Instead, some burden of disease would persist, e.g. burden of disease due to existing tobacco-related lung cancer, and some people may even develop new lung cancer based on their past exposure.

Now consider a reduction of exposure to the risk factor by 50 per cent at time T_o. In the example, such a reduction would mean that a constant attributable burden would result. This burden is, of course, a mixture of the impact of past exposure (i.e. prior to T_o) plus the impact of the exposure after T_o. Three main components have to be known to estimate this model. They are:

1. The relationship between exposure (i.e. the risk factor under consideration) and disease burden (attributable risk) and its temporal trend
   
   
2. The amount of burden caused by other factors than the risk factor under consideration and the temporal stability of this burden.
   
   
3. The specific trajectory of burden reduction after exposure reduction

Methods to estimate the relationship between a certain exposure distribution and disease burden are well established (e.g., odds ratios, relative risks; see Rothman and Greenland, 1998). Similarly, methods to estimate the proportion of burden attributable to the distribution of a certain risk factor were developed some decades ago and were first described by Miettinen (1972) in the early 1970s and Walter a little while later (Walter, 1976; 1980). They have since been used to generate estimates of the attributable burden for substance use and other risk factors around the world (Ezzati and Lopez, 2000). The key concept used herewith is that of an attributable fraction (AF; also called the aetiologic fraction). It depends on the exposure distribution (or, in the discrete case, on the prevalence of different exposure categories) and on the relative risk for burden related to the respective exposure levels.

Below are the two formulas for the case of continuous exposure levels and the discrete case, both of which are fully developed and described elsewhere (Walter, 1976; 1980; Eide and Heuch, 2001; Murray et al., 2003).

The contribution of a risk factor to disease can be estimated by comparing the burden due to the observed exposure distribution in a population with that from another distribution (rather than a single reference level such as non-exposed) as described by the generalized equation shown as Equation 1.

Equation 1

where PIF is the "potential impact fraction", a generalized form of the attributable fraction; RR(x) is the relative risk at exposure level x, P(x) is the population distribution of exposure, P' (x) is the counterfactual distribution of exposure, and m the maximum exposure level. As Murray et al. (2003) have noted, this formula can be further generalized to deal with a situation, where the relative risks change in the counterfactual scenario.

The corresponding relationship when exposure is described as a discrete variable with k levels is given by Equation 2.

Equation 2:

where i: exposure level category

RR_i: relative risk at exposure level i

P_i: prevalence of the ith category of exposure

The concept of attributable fraction (or the generalized form of PIF), as defined here, can only describe a snapshot at a specific time. Attributable fractions without including a time dimension are not able to characterize those cases whose occurrence would have been delayed or in part prevented due to exposure reduction (see Greenland and Robins, 1988). As a remedy, Greenland and Robins (1998) recommend the use of aetiologic fractions with a time dimension to account for this shortcoming. Time-based measures are discussed in some detail by Murray et al (2003) in their seminal conceptual framework of the comparative quantification of health risks.

The theoretical minimum risk (see Figure 1) is trickier to define than the exposure-burden relationship. Theoretical minimum risk denotes the exposure distribution that would result in the lowest population burden, irrespective of whether such a distribution is currently attainable in practice (Murray and Lopez, 1999). In the example it was set at zero attributable burden, but this is not necessarily always the case. Consider alcohol, and assume only one relevant exposure dimension in disease aetiology (see below for a discussion of volume and patterns of drinking as a two-dimensional exposure association), such as average volume of drinking. Minimal risk then occurs at zero (i.e. no drinking at all) for most related diseases such as cancer or traffic injury, but not for some whose risk actually decreases at exposures greater than zero e.g. heart disease (Rehm et al, 2003a; Rehm et al, 2003b; Rehm et al, 2004). For composite outcome measures, e.g. all-cause mortality, there is also reason to believe that the level of the exposure associated with minimum burden is greater than zero (i.e. some level of drinking) (Rehm et al, 2001; Gmel et al, 2003). The exact value of exposure associated with the minimal burden will depend on the composite measure used, and the disease distribution in the country or region examined. Thus, the theoretical minimum risk will fluctuate across cultures.

A theoretical minimum risk with an exposure greater than zero has interesting implications. For alcohol in the above example it means that, even if drinking occurs at the theoretical minimum, there will be some attributable disease burden. For instance, in a society with relatively large coronary burden and assuming a theoretical minimum risk for the population occurring at approximately 1 drink/day, there will be disease burden associated with such an exposure level of moderate drinking, e.g., for certain gastrointestinal diseases (Taylor et al, 2005) or accidents (Rehm and Gmel, 2003).

The above example is hypothetical for several reasons:

1. It assumes all people having the same exposure level (1 drink/day for all people) rather than a distribution of exposure, known to exist for all populations drinking alcohol. (Skog, 1985). Note that the above statement only refers to the existence of a distribution and does not specify the exact shape of it (e.g., lognormal vs. Poisson).
   
   
2. It completely disregards the fact that alcohol has at least two dimensions relevant for disease, average volume of alcohol consumption and patterns of drinking (Rehm et al, 2003a).
   
   
3. There is no known intervention, which would result in the above-mentioned distribution of 1 drink/day for all (Babor et al., 2003; Room et al., 2005).

In more general terms, the following points can be made:

1. We should always model shifts in risk factor distributions when estimating avoidable burden of disease.
   
   
2. We should always model all relevant dimensions of a risk factor when estimating avoidable burden of disease.
   
   
3. We need other counterfactual scenarios in addition to the theoretical minimum risk, when estimating avoidable burden of disease. The question will be where to obtain the counterfactual scenarios from.

The trajectory of the burden reduction after changes in exposure is also difficult to define, not only because it should incorporate changes of exposure reduction as well, but because it also has to make estimates for the change of relative risk over time of several disease outcomes, potentially both acute (e.g., alcohol and deaths from drinking and driving) and chronic (e.g. alcohol and chronic pancreatitis). For instance, after a change of smoking status to abstinence, we have to know the relative risk of an ex-smoker after one year after abstaining, two years after abstaining, three year after abstaining, and so on. For acute outcomes, the problem is much easier. Once the prevalence of alcohol, tobacco or illicit drugs is reduced, all the acute outcomes (e.g. injuries) are reduced accordingly. To give an example: while drinking over the past years does affect the cancer risk of people today, even if they started abstaining in between, it does not affect the traffic accident risk.