IJE Advance Access published online on December 4, 2008
International Journal of Epidemiology, doi:10.1093/ije/dyn259
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Rose's Strategy of Preventive Medicine. Geoffrey Rose with commentary by Kay-Tee Khaw and Michael Marmot.
E-mail: sam.harper{at}mcgill.ca
Rose's Strategy of Preventive Medicine. Geoffrey Rose with commentary by Kay-Tee Khaw and Michael Marmot. Oxford University Press, 2008, pp. 192. £19.95 (Paperback). ISBN: 978-0-19-263097-1.
Since its publication in 1992, Geoffrey Rose's The Strategy of Preventive Medicine has been required reading for those interested in population health, social epidemiology, or preventive medicine. Given the explosion of research on the social determinants of health that has taken place in the ensuing years, interested readers are likely to welcome a new edition of this classic text. The fact that Rose died shortly after its original publication, however, may also make readers curious as to what exactly is new in this new edition. The honest answer is nothing, if one's interest is in new material from Rose. There is, however, a new addition to the text in the form of a lengthy commentary by Kay-Tee Khaw and Michael Marmot, both of whom studied under Rose, which attempts to broaden considerably the scope of Rose's ideas.
Khaw and Marmot begin by briefly recapping the main tenets of Rose's book. One of the major themes he explored was the distribution of health risks in populations, which gave rise to the notion that when risk is widely diffused a large share of the population burden of disease may come from many individuals exposed to low or moderate risks, i.e. those at levels below the threshold typically considered ‘high risk’. Rose mainly took his examples from univariate distributions of cardiovascular risk factors, such as blood pressure and cholesterol, noting that many, if not most, incident cases of coronary heart disease (CHD) came from those with risk factor levels below what was typically considered high risk. The implication for prevention was that traditional clinical approaches targeting those with high levels of risk factors would do little to affect population levels of disease because this missed the major source of future cases. Population strategies aimed at shifting entire distributions of risk were therefore necessary, as were efforts to understand the causes of the population distribution and why it varies across societies. Khaw and Marmot provide additional examples where many ‘deviant’ cases come from those at moderate-to-low risk—intraocular pressure and glaucoma risk, HbA1c and diabetes risk, bone heel ultrasound and the risk of fractures—as further evidence backing the population approach. I found some of these examples less convincing than others. For example, in the intraocular pressure and glaucoma example, Khaw and Marmot note that two-thirds of glaucoma cases come from those at moderate-to-low levels of intraocular pressure. This is true, but they also note that fully one-third of cases come from the extreme tail of the distribution making up just 4% of the population. That such a disproportionate number of cases comes from such a small population group strikes me as a potential case where an effective intervention targeting this high-risk 4% would produce significant public health gains at perhaps less cost—yet Khaw and Marmot suggest this clearly favours Rose's population approach.
A good deal has been said since the initial publication of Rose's book about population vs high-risk strategies—largely, though not exclusively, in support of Rose's arguments. However, Khaw and Marmot argue that a less-discussed aspect of Rose's work, one which they say has profound implications, is the relationship between the population mean and the prevalence of ‘deviant’ individuals—what they refer to as the ‘single distribution theory’.
The notion that the population mean predicts the prevalence of ‘deviants’ (i.e. those at extreme values of the distribution) is due to Rose and Day's cross-country analyses of the Intersalt study,1 which found strong correlations across countries between population mean values of blood pressure, body weight, alcohol consumption and sodium intake and the respective prevalence of hypertension, overweight, heavy drinking and excess sodium consumption. Thus, one might argue that there is a general (or ‘single’) population distribution of these risk factors and countries differ in the prevalence of bad outcomes because variation in the mean of this distribution will lead to different percentages of the population being above some fixed ‘high-risk’ threshold [e.g. a body mass index (BMI) 
25 kg/m2]. The strength of this correlation led Rose to argue that the most effective way to reduce the prevalence of deviants was to shift the entire population distribution rather than by trying to eliminate the high-risk tail. Khaw and Marmot provide additional evidence of this phenomenon from a cross-national survey of educational achievement by showing that mean mathematics and science test scores are correlated with the prevalence of high- and low-achievers (though it is not clear how the achievement thresholds are defined). Like Rose, they conclude from this that the only reliable way to alter the fraction of extreme cases is by shifting the population mean. And while Rose largely limited his prior work to health-related outcomes, Khaw and Marmot are eager to extend this notion across a much wider spectrum of population-level outcomes not traditionally associated with health, such as gambling, intelligence, athletic prowess and artistry. In essence, they argue that Rose's idea that the mean value of some population attribute ‘predicts’ the number of deviants can—and should—be applied much more widely. From the perspective of population health (and other population measures) the implication is clear: we should aim to move population averages if we want to achieve real gains in reducing the number of extreme cases.
This is a plausible and intriguing notion, but one wonders just how widely the ‘single distribution theory’ may apply. For risk factors that have a bell-shaped normal curve, it is not surprising that the mean and the fraction above some fixed threshold are highly correlated, but not all risk factor distributions are normal and the prevalence of deviants is also affected by the variance (i.e. the spread around the mean).2 Rose's strategy applies most assuredly when the population mean changes as a result of shifting the entire distribution of a risk factor, rather than its variance. Thus, shifting population means is likely to be effective in the absence of changes in variance, but it is not clear that when overall distributions shift they do so uniformly. Take, for example, population increases in average BMI that have occurred in many countries. Much has been made of the rapid increase in obesity, particularly in the United States, but the increase is far larger than what one would have predicted based on the change in average BMI because the distribution widened considerably due to a fattening of the rightmost tail. As a result, US government's goal of reducing the obesity rate to 1970s levels (
15%) cannot be achieved by simply shifting mean BMI back to 1970s levels because the distribution today is far more right skewed.3,4 There is no doubt that returning average BMI to 1970s levels would reduce obesity rates, but given the widening of the distribution it seems hard to say whether such a reduction would be more beneficial for population health than would an intervention to reduce the rightward skewness without changing mean BMI.
One can tell a similar story about cigarette smoking. After the landmark Surgeon's General's report was published the prevalence of current smoking in the United States declined by roughly one-third between 1965 and 1985, yet the fraction of those smoking 25 or more cigarettes per day increased by 25% as the less-addicted were initially more likely to quit.5 As one might expect, the mean number of cigarettes smoked among daily smokers actually increased during this period of declining prevalence, suggesting a widening of the distribution of smoking. The distribution has since narrowed, but the lesson here is that using the ‘single distribution theory’ as a guide to prevention requires careful consideration of both the mean and the variance of population risk factor distributions.
Khaw and Marmot also say little about the shape of the risk function relating risk factors to disease, but the case for the population strategy is also harder to make when the exposure-risk curve is exponential. For example, in the absence of solid data Rose speculated that a large part of alcohol-related traffic accidents could conceivably be due to the accumulated risks of large numbers of people driving after consuming small amounts of alcohol. Subsequent research has shown, however, that this is decidedly not the case, at least in the United States. The distribution of alcohol-involved fatal accidents in the United States in 2006 shows that 84% of drivers had a blood alcohol content (BAC) of 0.08 or higher, and fully 55% were in excess of 0.15, a level indicating severe impairment.6 A more extreme case is found during periods of heightened risk—weekend evenings—where nearly 72% of driver fatalities occur among those with BACs in excess of 0.10, a group accounting for only 2.7% of the population.7 Thus while it is true that accident risk increases with any alcohol consumption, it does so exponentially, making the contribution of light drinkers to alcohol-involved crashes minor despite their greater representation in the population.
Similar examples where population-based strategies may be inefficient occur in cases where risk factor distributions are decidedly non-normal. Malcom Gladwell recently wrote about societal problems, such as homelessness and auto emissions, which tend to follow a power-law distribution in which a large burden of the problem is generated by a small percentage of the population.8 In such cases population-based strategies may be less efficient, depending on the potential interventions available. This is case, for example, with medical care expenditures in the United States, where 5% of the population generates nearly half of the total spending.9 Similarly, the distribution of US alcohol consumption is highly skewed, where 10% of drinkers are responsible for nearly 60% of the total volume of alcohol consumed.10 These would seem to be cases where trying to shift the population average may require a specific, targeted focus on smaller populations that generate the bulk of the problem.
Even in the case Rose argued most strenuously and that Khaw and Marmot consider settled—that of CHD—it remains an empirical question as to the right balance between population and high-risk prevention strategies, depending on the distribution of risk, the shape of the exposure-risk curve and the cost and effectiveness of interventions. Manuel et al.11 compared three different cholesterol prevention strategies for preventing CHD deaths in Canada: shifting the entire cholesterol distribution downward (population strategy), treatment for those with high-cholesterol (single risk factor strategy), and treatment for those with high baseline risk of CHD according to the Framingham risk algorithm (high baseline risk strategy). Because the distribution of absolute risk of CHD in Canada is not widely diffuse, they found the high baseline risk strategy would prevent roughly seven times more CHD deaths over 10 years than the population strategy. The lesson here is that improved strategies for classifying risk in the population (i.e. using multiple risk factors) may lead to different conclusions regarding the balance of prevention strategies.12
In the end, Khaw and Marmot do a fine job of distilling the central tenants of Rose's book down to a manageable commentary. In particular, they conclude with an important discussion of the difficulties of generating evidence for the kind of large-scale population-level interventions they (and, presumably, Rose) would favour, and of the trade-offs between protecting public health and individual liberty that must often be made when human behaviours are important determinants of disease. This is the strongest part of their contribution, but overall I found myself wishing their commentary contained more of the subtlety and clarity of Rose's original prose. In particular, Rose was keenly aware that achieving the right balance between population and high-risk approaches depends heavily on both the underlying distribution of risk factors and the shape of the curve-relating risk factor levels to specific events. In cases where exposure-risk curves are J- or even U-shaped, Rose acknowledged that shifting risk factor distributions would lead to both gains and losses in population health, and warned against oversimplifications of the population approach. And while I would encourage readers to engage the new material by Khaw and Marmot, the major reason to recommend this book remains the important ideas Geoffrey Rose elaborated more than 15 years ago.
 |
References |
|---|
 Top References |
|---|
1 Rose G, Day S. The population mean predicts the number of deviant individuals. B Med J (1990) 301:1031–34.
2 Wen SW, Kramer MS, Hanley JA. Mean, deviants, and spread of distribution. Epidemiology (1995) 6:639.[CrossRef][Web of Science][Medline]
3 Jolliffe D. Continuous and robust measures of the overweight epidemic: 1971–2000. Demography (2004) 41:303–14.[CrossRef][Web of Science][Medline]
4 Flegal KM, Troiano RP. Changes in the distribution of body mass index of adults and children in the US population. Int J Obes (2000) 24:807–18.[CrossRef][Web of Science][Medline]
5 Burns DM, Major JM, Shanks TG. Changes in number of cigarettes smoked per day: cross-sectional and birth cohort analyses using NHIS. In: Smoking and Tobacco Control Monograph 15: Those Who Continue to Smoke: Is Achieving Abstinence Harder and Do We Need to Change our Interventions?—Burns DM, ed. (2003) Bethesda, MD: US Department of Health and Human Services, National Cancer Institute. 83–99.
6 US National Highway Traffic Safety Administration. Traffic Safety Facts: Alcohol-Impaired Driving (2008) Washington, DC: US Department of Transportation, National Highway Traffic Safety Administration.
7 Zador PL, Krawchuk SA, Voas RB. Relative Risk of Fatal Crash Involvement by BAC, Age, and Gender (2000) Washington, DC: US National Highway Traffic Safety Administration. DOT HS 809 050.
8 Gladwell M. Million-Dollar Murray. The New Yorker (2006) 81:96–107.
9 Berk ML, Monheit AC. The concentration of health care expenditures, revisited. Health Aff (2001) 20:9–18.
10 Kerr WC, Greenfield TK. Distribution of alcohol consumption and expenditures and the impact of improved measurement on coverage of alcohol sales in the 2000 National Alcohol Survey. Alcohol Clin Exp Res (2007) 31:1714–22.[CrossRef][Web of Science][Medline]
11 Manuel DG, Lim J, Tanuseputro P, et al. Revisiting Rose: strategies for reducing coronary heart disease. B Med J (2006) 332:659–62.
12 Zulman DM, Vijan S, Omenn GS, et al. The relative merits of population-based and targeted prevention strategies. Milbank Q (2008) 86:557–80.[CrossRef][Web of Science][Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||