Users' Guides to the Medical Literature: Title and subTitle BreakXIX. Applying Clinical Trial Results
A. How to Use an Article Measuring the Effect of an Intervention on Surrogate End Points

You are a physician seeing a 62-year-old woman with postmenopausal osteoporosis. Her bone mineral density, as measured by dual-energy x-ray absorptiometry, is 2.5 SDs below the mean value in premenopausal women. Although she does not have back pain, a spinal radiograph shows an old vertebral fracture. The patient has not yet experienced problems as a result of her vertebral fracture, but she is disturbed by the prospect that she may end up like her mother whose osteoporotic fractures have resulted in severe, long-term back pain.

The patient has reflux esophagitis and a past endoscopy revealed nonspecific gastritis. A specialist had prescribed alendronate, which the patient had to stop taking after several weeks because of dyspepsia. She searched the Web and discovered a new drug, raloxifene, and wonders whether this drug might be an alternative. You know that this drug has been licensed for the prevention of postmenopausal osteoporosis. You promise to examine the literature and to get back to her.

Using MEDLINE you identify a study of raloxifene for the treatment of osteoporosis demonstrating an effect on bone mineral density.¹ You are wondering whether this warrants administration to lower your patient's risk of osteoporotic fracture.

Ideally, clinicians making treatment decisions should refer to methodologically strong clinical trials examining the impact of therapy on clinically important outcomes. By clinically important outcomes we mean outcomes that are important to patients: health-related quality of life, morbid end points such as stroke or myocardial infarction, or death. Often, however, conducting these trials requires such a large sample size, or long-term patient follow-up, that researchers or drug companies look for alternatives. Substituting surrogate end points for the target event allows conduct of shorter and smaller trials, thus offering an apparent solution to the dilemma.

A surrogate end point may be defined as "a laboratory measurement or a physical sign used as a substitute for a clinically meaningful end point that measures directly how a patient feels, functions or survives."² Surrogate end points include physiologic variables (such as bone mineral density as a surrogate for long-bone fractures, blood pressure for stroke, low-density lipoprotein cholesterol levels for myocardial infarction, and CD4 cell count for acquired immunodeficiency syndrome [AIDS] and AIDS-related mortality) or measures of subclinical disease (such as degree of atherosclerosis on coronary angiography).

The use of surrogate end points is indispensable for drug evaluation in phase 2 and early phase 3 trials geared to establishing a drug's promise of benefit. In many countries, companies may obtain drug approval by demonstrating a positive impact on surrogate end points. The use of surrogate end points for regulatory purposes reflects drug approval decisions that regulators must make in the face of public health exigencies.

Reliance on surrogate end points may be beneficial or harmful. On the one hand, use of the surrogate end point may lead to the rapid and appropriate dissemination of new treatments. For example, the Food and Drug Administration's decision to approve new antiretroviral drugs based on information from trials using surrogate end points recognized the enormous need for effective therapies for patients with human immunodeficiency virus (HIV) infection. Subsequently, several of these drugs have proved effective in randomized trials focusing on clinically important outcomes.^{3 - 6}

On the other hand, reliance on surrogate end points may lead to excess morbidity and mortality. For example, while cardiac inotropes may improve short-term cardiac hemodynamic function in patients with heart failure, randomized clinical trials have demonstrated excess mortality with a number of these agents.⁷ In particular, flosequinan was widely prescribed after its release, but had to be withdrawn after a trial revealed its deleterious effects on survival.⁸

How are clinicians to distinguish between these 2 situations? Surrogate outcome will be consistently reliable only if there is a causal connection between change in surrogate and change in the clinically important outcome. Thus, the surrogate must be in the causal pathway of the disease process and an intervention's entire effect on the clinical outcome of interest should be fully captured by a change in the surrogate. This Users' Guide builds on previous discussions of how one can establish a causal relationship⁹ and presents an approach to critical appraisal of studies using surrogate end points and application of their results to manage individual patients.

As our discussion will make evident, the clinician needs to assess far more than a single study to make the decision about the adequacy of a surrogate. Evaluation may require a comprehensive review of observational studies of the relationship between the surrogate and the target, and of some or all of the randomized trials that have evaluated treatment impact on both the surrogate and the target. While most clinicians would hesitate to conduct such an investigation, our guidelines will allow them to evaluate the arguments made by experts or the pharmaceutical industry for prescribing treatments on the basis of their effect on surrogate end points.

In this guide, we follow the framework of previous articles in the series¹⁰ and ask 3 sorts of questions: are the results valid; what were the results; and will the results help me in caring for my patients? (Table 1). When we consider the validity of a surrogate, we must address 2 issues. First, to be consistently reliable, the surrogate must be in the causal pathway from the intervention to the outcome. Second, in considering a particular intervention, we must be confident that there are no important effects of that intervention on the outcome of interest that are not mediated through, or captured by, the surrogate. Our guides for validity (Table 1) bear directly on these 2 issues.

To provide a valid substitute for an important target outcome, the surrogate must be associated or correlated with that target. In general, researchers choose surrogate end points because they have found a correlation between a surrogate and a target outcome in observational studies, and their understanding of the biology makes it plausible that changes in the surrogate will invariably lead to changes in the important outcome. The stronger the association, the more likely the causal link between the surrogate and the target. The strength of an association is reflected in statistics such as relative risk (RR) or odds ratio. We have presented a full discussion of statistics reflecting the strength of association in another article.¹¹ Many biologically plausible surrogates are only weakly associated with clinically important outcomes. For example, measures of respiratory function in patients with chronic lung disease, or conventional exercise tests in patients with heart and lung disease, are only weakly correlated with capacity to undertake activities of daily living.^{12 - 13} When correlations are low, the surrogate is likely to be a poor substitute for the target outcome.

In addition to the strength of the association, one's confidence in the validity of the association depends on whether it is consistent across different studies and after adjustment for known confounders. For example, ecologic studies such as the Seven Countries Study¹⁴ suggested a strong correlation between serum cholesterol levels and coronary heart disease mortality even after adjusting for other predictors such as age, smoking, and systolic blood pressure. Subsequent cohort studies confirmed this association and suggested that long-term reductions in serum cholesterol levels of 0.6 mmol/L (23 mg/dL) would lower the risk of coronary heart disease by approximately 30%. When a surrogate is associated with an outcome after adjusting for multiple other potential prognostic factors we call the association independent.

Similarly, cohort studies have consistently revealed that a single measurement of plasma viral load predicts the subsequent risk of AIDS or death in patients infected with HIV.^{15 - 20} For example, in 1 study the proportion of patients that progressed to AIDS after 5 years in the lowest through the highest quartiles of viral load was 8%, 26%, 49%, and 62%, respectively.²⁰ Moreover, this association retained its predictive power after adjustment for other potential predictors such as CD4 cell count.^{15 - 19}

Returning to the scenario, you are wondering if you can substitute bone mineral density for fractures or health-related quality of life in considering whether to recommend raloxifene. A large cohort study investigated risk factors for hip fracture.²¹ Postmenopausal women with a calcaneal bone density in the highest third had a hip fracture rate of 9.4/1000 woman-years while women in the middle and lowest third had a fracture rate per 1000 woman-years of 14.7 and 27.3, respectively. Furthermore, after considering other risk factors for osteoporotic hip fractures including maternal history of hip fracture, previous fractures from any site, poor self-rated health, use of long-acting benzodiazepines, impaired visual function, and reduced physical activity, bone mineral density continued to predict the risk of hip fracture.²¹ These findings are consistent across studies looking at the association between bone density and fracture risk.^{22 - 23} Thus, bone mineral density is a moderately strong, independent predictor of fracture, and meets our first criterion for an acceptable surrogate end point.

While meeting this first criterion is necessary, it is not sufficient to support reliance on a surrogate outcome. As we will emphasize below (Table 1), before offering an intervention on the basis of effects on a surrogate outcome, the clinician should note a consistent relationship between surrogate and target in randomized trials; the effect of the intervention on the surrogate must be large, precise, and lasting, and the benefit-risk trade-off must be clear.

Given the possibility of effects unrelated to the surrogate end point, pathophysiologic studies, ecological studies, and cohort studies are insufficient to establish that the link between surrogate and clinically important outcomes is ironclad. We can confidently rely on surrogate end points only when long-term randomized trials have consistently demonstrated that modification of the surrogate is associated with concomitant modifications in the target outcome of interest. For example, although ventricular ectopic beats are associated with adverse prognosis in patients with myocardial infarction²⁴ and class 1 antiarrhythmic agents effectively suppress ventricular arrhythmias in animals and humans,²⁵ these drugs have proved to increase mortality when evaluated in randomized trials.²⁶ In this case, reliance on the surrogate end point of suppression of nonlethal arrhythmias led to the deaths of tens of thousands of patients.²⁷

The treatment of heart failure provides another instructive example. Trials of angiotensin-converting enzyme inhibitors in heart failure treatment have demonstrated parallel increases in exercise capacity^{28 - 31} and decreases in mortality,³² suggesting that clinicians may be able to rely on exercise capacity as a valid surrogate. Milrinone³³ and epoprostenol³⁴ have both demonstrated improved exercise tolerance in patients with symptomatic heart failure. However, when these drugs were evaluated in randomized controlled trials both showed an increase in cardiovascular mortality that in one instance was statistically significant,³⁵ and in the second case led to the early termination of the study.³⁶ Thus, exercise tolerance is inconsistent in predicting improved mortality and is therefore an unsatisfactory substitute. Other suggested surrogate end points in heart failure have included ejection fraction, heart-rate variability, and markers of autonomic function.³⁷ The dopaminergic agent ibopamine positively influences all 3 surrogate end points, and yet a randomized trial demonstrated that the drug increases mortality in heart failure.³⁸

An example of a surrogate end point is CD4 cell count, which has been validated in randomized trials. A number of trials comparing different classes of antiretroviral therapies have demonstrated that patients randomized to more potent drug regimens had higher CD4 cell counts and were less likely to progress to AIDS or death.^{6 ,39} While there is no guarantee that the next trial using a different class of drugs will show the same pattern, these results greatly strengthen our inference that if therapy for HIV infection increases the CD4 count, a reduction in AIDS-related mortality will result.

Returning to our scenario, trials of etidronate^{40 - 41} and alendronate⁴² for the prevention of osteoporotic fractures in postmenopausal women have shown parallel increases in bone mineral density and reduced incidences of new vertebral fractures. This would suggest that clinicians might rely on bone density to evaluate new drugs in osteoporosis in making the assumption that if they saw increases in bone density, decreases in fractures would follow.

However, another secondary prevention trial in postmenopausal women using sodium fluoride showed divergent results.⁴³ Although sodium fluoride increased bone mineral density at the lumbar spine by 35% over 5 years, more vertebral and nonvertebral fractures occurred in the intervention group than in the placebo group (163 and 72 in 101 women with sodium fluoride vs 136 and 24 in 101 women with placebo). In another randomized trial, fluoride again showed a large increase in bone density without any change in fracture rate.⁴⁴ Inferences on the basis of unchanged bone density may also be problematic. A study of calcium and vitamin D in the elderly showed virtually no change in bone density, but a reduction in fracture risk of approximately 50%.⁴⁵ Thus, increase in bone mineral density as a surrogate end point has shown an inconsistent relationship to osteoporotic fractures.

Clinicians are in a stronger position to rely on surrogate end points if the new drug they are considering is from a class of drugs in which the relationship between changes in the surrogate and changes in the target has been verified in randomized trials. For instance, thiazide diuretics and β-blockers have both been shown to reduce blood pressure and clinically important outcomes such as stroke in patients with hypertension. Thus, we would be much more comfortable relying on reduction in blood pressure to justify administering a new β-blocker or thiazide diuretic than to justify offering a novel antihypertensive agent from another class.⁴⁶

For example, although 1 dihydropyridine calcium channel blocker has been shown to reduce clinically important outcomes in patients with hypertension,⁴⁷ 4 other trials have shown that these agents are less efficacious than thiazides or angiotensin-converting enzyme inhibitors in preventing hard clinical end points despite exerting similar degrees of blood pressure lowering.^{48 - 51}

We will consider the example of cholesterol reduction as a surrogate for cardiovascular outcomes such as myocardial infarction and death in part B of this Users' Guide.⁵² Briefly, several large trials of primary and secondary prevention of coronary heart disease with statins have consistently shown that these drugs reduce cardiovascular outcomes.⁵³

We could therefore make the assumption that a new statin with a similar low-density lipoprotein cholesterol–lowering potency may also reduce clinically important outcomes. However, we would be reluctant to generalize to another class of lipid-lowering agents since trials of 1 such class (the fibrates) have shown that these drugs reduce the incidence of myocardial infarction but increase the risk of mortality from other causes (with no impact on overall mortality).^{53 - 55}

These examples highlight the point we made earlier: confidence in a surrogate outcome depends on the assumption that the treatment captures any relationship between the treatment and the outcome.^{56 - 57} This assumption can be violated in 2 ways. First, treatment may have a beneficial mechanism of effect on the outcome independent of its effect on the surrogate. For instance, 1 explanation for the superior effect of angiotensin-converting enzyme inhibitors vs calcium antagonists on clinically important outcomes is that angiotensin-converting enzyme inhibition has biological effects independent of lowering blood pressure that reduce risk of stroke or death and that calcium antagonists do not share these effects.

Second, treatment may have deleterious effects on the outcome that are not mediated through the surrogate. Mortality-increasing effects of fibrates rather than inability to lower morbidity and mortality through cholesterol reduction probably explain the lack of effect of fibrates on clinically important outcomes. That such additional effects are less likely across classes of drugs than within classes is what makes us more inclined to rely on within-class evidence from surrogate outcomes.

This criterion is complicated by the variable definitions of drug class. A manufacturer of a drug related to a class of agents with a consistently positive association between modification of a surrogate end point and modification of the target (such as a β-blocker) will naturally argue for a broad definition of class. Manufacturers of agents that are related to drugs with known or suspected adverse effects on target events (clofibrate, or some calcium antagonists) are likely to argue, on the other hand, that the chemical or physiological connection is not sufficiently close to consider the new drug to be in the same class as the harmful agent. Part B will address these issues more fully.⁵²

Returning to the scenario, we have established that because of the inconsistent relationship between increase in bone mineral density and fracture reduction we would be reluctant to offer patients a new antiosteoporotic agent solely on the basis of evidence of its effect on the surrogate end point. Raloxifene, the drug we are considering for our patient, is a nonsteroidal benzothiophene, a selective estrogen-receptor modulator representing a new class of drugs for the prevention of osteoporosis-related bone fractures. Thus, it is likely that the mechanisms of action will be considerably different from bisphosphonates and the conclusion that similar reductions in loss of bone density will lead to parallel reductions in clinical fractures is questionable. In Table 2, we apply our validity criteria to a number of controversial examples of the use of surrogate end points.

We are interested not only in whether an intervention alters a surrogate end point, but also in the magnitude, precision, and duration of the effect. If an intervention shows large reductions in the surrogate end point, the 95% confidence intervals (CIs) around those large reductions are narrow, and the effect persists over a sufficiently long period, our confidence that the target outcome will be favorably affected increases. Positive effects that are smaller, with wider CIs, and shorter duration of follow-up leave us less confident.

We have already cited evidence suggesting that CD4 cell counts may be an acceptable surrogate for mortality in patients with HIV infection. A randomized controlled trial of immediate vs delayed zidovudine therapy in HIV-infected asymptomatic individuals declared a positive result for immediate therapy, largely on the basis of a greater proportion of treated patients with CD4 cell counts above 435 × 10⁶/L at a median follow-up of 1.7 years.⁵⁸ Subsequently, the Concorde study addressed the same question in a randomized trial with a median follow-up of 3.3 years.⁵⁹ The Concorde investigators found a continuous decline in CD4 cell counts in both treated and control groups, but the median difference of 30 × 10⁶/L in favor of treated patients at study termination was statistically significant. However, the study showed no effect of zidovudine in terms of reduced progression to AIDS or death. The median CD4 cell count difference was insufficient to have an impact on clinically important outcomes. The Concorde authors made the following conclusion: the small, but highly significant persistent difference in CD4 cell counts between the groups was not translated into a significant clinical benefit and "called into question the uncritical use of CD4 cell counts as a surrogate endpoint." Had the Concorde analysis showed significantly shorter times to reach a CD4 cell count of 350 × 10⁶/L in the control group and been regarded as fundamental, the trial might have been stopped early with a false-positive result.

Returning to our scenario, the trial of raloxifene in women with osteoporosis demonstrated that after 2 years of treatment, raloxifene-treated patients in the group receiving the highest dosage showed an increase in bone mineral density at the lumbar spine of 2.2% (SE, 0.3%) compared with a slight decrease in the control group 0.8% (SE, 0.3%). This difference in change over time was statistically significant (P<.03). Ideally, the investigators would have provided us with a CI around the 3% difference in percentage change in bone mineral density in the treatment and control groups. As we will illustrate when we consider weighing benefits and harms, the magnitude of the effect on the surrogate may (or may not) help us estimate the size of a possible affect on the target outcome.

The questions clinicians should ask themselves in applying the results are the same ones we have suggested for any issue of therapy or prevention ⁶⁰ and elaborated on in our Users' Guide regarding applicability.⁶¹ These 3 questions have to do with whether the results can be applied to your patient's care, whether all important outcomes were considered, and whether the likely benefits are worth the down sides of treatment.

"Can the results be applied to my patient's care" refers to the extent to which the patient before you is similar to those who participated in the published studies under consideration, and the extent to which the therapy, and the associated technologies for monitoring and responding to complications, are available in your setting. "Were all important outcomes considered" relates to the focus of this Users' Guide, and all the issues we have raised thus far: was the primary outcome really the one in which patients will be interested?

This second criterion also draws issues of adverse intervention effects to our attention. Applying the third criterion, judging whether the benefits are worth the down sides of treatment, presents particular challenges when investigators have focused on surrogate end points, and we will discuss this criterion in some detail.

To know whether to offer a treatment to their patients, clinicians must be able to estimate the magnitude of the likely benefit. When the data available are limited to the effect on a surrogate end point, estimating the extent to which treatment will reduce clinically important outcomes becomes a challenge.

One approach is to extrapolate from 1 or more randomized trials assessing a related intervention in a similar patient population that provides both surrogate end point and clinical outcome data. For example, until recently there were little long-term data on the efficacy of lovastatin in reducing clinically important outcomes. However, one could extrapolate from short-term dose efficacy studies assessing the surrogate end point of cholesterol lowering. Thus, since 40 mg of lovastatin produced a similar degree of lowering of low-density lipoprotein cholesterol as 40 mg of pravastatin (31% vs 34% reduction) in the CURVES Study,⁷⁷ one could theorize that lovastatin would have similar long-term benefits to pravastatin. Subsequently, the AFCAPS/TexCAPS Trial (a 5-year trial assessing the efficacy of lovastatin in the primary prevention of ischemic heart disease)⁷⁸ confirmed that this agent had a beneficial profile similar to pravastatin (as determined by the 5-year, primary prevention WOSCOPS Trial)⁷⁹ : the RR reductions (and 95% CIs) for myocardial infarction were 40% (17%-57%) and 31% (17%-43%), respectively. However, this approach is likely to be seriously flawed when one is extrapolating from trials of another class of drugs.

Returning to our scenario, to estimate the magnitude of the fracture reduction we might expect with raloxifene (in which we have only surrogate end point data), we could (recognizing the limitations of this approach pointed out above) examine the results of randomized controlled trials of alendronate (a drug from a different class for which we have data on the same surrogate end point as well as clinical end points such as fracture reduction). While alendronate appears to improve vertebral bone density by 7.5% over 2 years (vs control),⁴² raloxifene is associated with only a 3.0% improvement over the same time frame. A systematic overview of the alendronate trials⁸⁰ reported a 29% reduction in RR of nonvertebral fracture over 2 years. Only 1 trial looked at symptomatic vertebral fractures in women with decreased bone density and an existing vertebral fracture.⁸¹ This study demonstrated an RR reduction of 55% with alendronate and suggested that our patient's risk over 3 years of a nonvertebral fracture would be approximately 15%; symptomatic vertebral fracture would be about 5%. Given the RR reductions with alendronate, one would need to treat approximately 25 women to prevent a nonvertebral fracture and 40 women to prevent a symptomatic vertebral fracture over a 3-year period.

Since the improvement in bone mineral density with raloxifene is at best 50% of the effect of alendronate, we would anticipate a considerably lower reduction in fracture risk with raloxifene. However, interim analysis of an ongoing raloxifene trial⁶² reported a 46% RR reduction with this therapy (despite less of an increase in bone mineral density than seen with the alendronate trials). This serves to emphasize the dangers of extrapolating results across classes when it is uncertain that the effects on clinically important outcomes are mediated in the same fashion by the 2 comparison drugs.

In deciding whether the likely magnitude of the treatment effect warrants offering patients the intervention, clinicians must consider not only the uncertainty associated with that estimate, but the trade-off with potential toxic effects and costs of therapy. In addition, clinicians must ponder the consequences of not treating, and the available management alternatives. The deadly and usually relentless progression of HIV infection, and the paucity of alternative therapies, has contributed to the readiness of patients, clinicians, and regulatory agencies to accept evidence from surrogate end points in instituting novel therapies in patients infected with HIV. In osteoporosis, in which the consequences of the condition are less immediately devastating, and a variety of agents are available, the case for relying on surrogate end points is far less compelling.

We have found a strong, consistent, independent, and biologically plausible association between bone mineral density and vertebral and nonvertebral fractures. Randomized trials, however, have failed to show a consistent association between increased bone density and reduction in fracture across all drug classes.

Because our patient is at substantial risk of fracture over the short term, the number needed to treat to prevent both nonvertebral and vertebral fractures is moderate, as is the absolute benefit she might expect. Moreover, she is interested in longer-term fracture prevention, and her risk will grow over time. One might offer her alternative interventions, including hormone replacement therapy, calcium and vitamin D, bisphosphonates, or calcitonin.

While there is strong evidence from randomized trials supporting the use of bisphosphonates to decrease osteoporotic fractures, randomized trial data showing fracture reduction in populations similar to our patient with the other agents is limited. Our patient is concerned about her long-term risk. Raloxifene was well tolerated during this 2-year trial but no information is available about long-term adverse effects including cardiovascular disease, venous thromboembolism, breast and endometrial cancer, and menopausal symptoms. While a number of options (including a trial of etidronate, offering hormone replacement therapy, calcium and vitamin D, calcitonin, or suggesting only a balanced diet and exercise) might be reasonable, ideally the clinician would subject these options to the same scrutiny applied to raloxifene.

Data indicating a reduction in fracture rate would greatly strengthen the case for including raloxifene as the preferred option. Just as you are about to see the patient (and, for us, just before this article went to press) you pick up a few of your latest editions of JAMA from the pile in the corner of your office, and find 2 highly relevant randomized trials.^{82 - 83} The results show that, in women like your patient with a prevalent vertebral fracture, raloxifene decreased radiological vertebral fracture risk (for 60 mg: number needed to treat=16 [RR, 0.7; 95% CI, 0.6-0.9]; and for 120 mg: number needed to treat=10 [RR, 0.5; 95% CI, 0.4-0.7]), but did not decrease the incidence of nonvertebral fracture. In helping your patient to decide on the right course of action, you realize you will have to consider other effects of raloxifene: the JAMA articles also show a 76% RR reduction of breast cancer as detected by mammography (number needed to treat, 126), a 3-fold increase in the risk of venous thromboembolism, and an increased incidence of hot flashes, leg cramps, influenzalike syndromes, and peripheral edema.

When we use surrogate end points to make inferences about expected benefit, we are making assumptions regarding the link between the surrogate end point and the target outcome. We have outlined criteria clinicians can use to decide when these assumptions might be appropriate. Even if a surrogate end point meets all of these criteria, inferences about a treatment benefit may still prove misleading. Thus, treatment recommendations based on surrogate outcome effects can never be strong. Furthermore, difficulties in estimating the magnitude of effects on clinically important end points compromises economic analysis examining the cost-effectiveness of alternative management strategies.

These considerations emphasize that waiting for randomized trials investigating the effect of the intervention on outcomes of unequivocal importance to patients is the only ironclad solution to the surrogate outcome dilemma. When clinicians must choose between alternative interventions, trials should make head-to-head comparisons between competing treatments rather than restricting comparisons of treatment to control or placebo. We expand on this issue in Part B of this Users' Guide. However, when patients' risk of serious morbidity or mortality are high, this "wait-and-see" strategy may pose problems for many patients and their physicians.

We encourage clinicians to critically question therapeutic interventions in which the only proof of efficacy is from surrogate end point data. When the surrogate end point meets all our validity criteria, the effect of the intervention on the surrogate end point is large, the patient's risk of the target outcome is high, the patient places a high value on avoiding the target outcome, and there are no satisfactory alternative therapies, clinicians can recommend therapy on the basis of randomized trials evaluating only surrogate end points. In other situations, clinicians must carefully consider the known adverse effects and cost of therapy, and the possibility of unanticipated adverse effects, before recommending an intervention solely on the basis of surrogate end point data.