Polonius asks Hamlet, “What do you read, my lord?” and hears in response “Words, words, words…” (William Shakespeare, Hamlet, 1601). The March Hare says to Alice, “You should say what you mean,” to which she replies, “I do…at least I mean what I say—that’s the same thing, you know.” (Lewis Carroll, Alice’s Adventures in Wonderland, 1865). The Hatter wins the gambit by pointing out that it is “not the same thing a bit! Why, you might just as well say that ‘I see what I eat’ is the same thing as ‘I eat what I see’!”
Words, what they mean, and how they are used are obviously important. Their careless use can distort thoughts and perceptions.
Surrogates and Sense
A salient example of linguistic erosion has crept into research pertinent to pharmacological agents and devices and its interpretation—namely, the use of the term “surrogate.” We are told that cholesterol lowering, blood pressure lowering, and restoration of coronary patency through the use of thrombolytic drugs are surrogates for mortality in studies of cholesterol-lowering drugs, antihypertensive agents, and thrombolytic agents. Are they? Certainly not. “Surrogate” means substitute in the sense of “deputy” as applied to people or, more generally, in the sense of a stand-in. Does ambiguity attendant on the misuse of words such as “surrogate” amount to anything more than semantics? The answer is certainly yes.
In studies of antihypertensive agents, we determine what happens to systolic and diastolic blood pressure, cardiac output, peripheral vascular resistance, and other pathophysiological variables in the context, of course, of patients stratified with respect to the etiology, duration, and severity of hypertension and target organ involvement. In mechanistic studies, the dependent variable of interest is a biological phenomenon (ie, in this case, blood pressure), which is, in turn, dependent on specific physiological determinants such as cardiac output and peripheral vascular resistance. If hypotheses that elevated blood pressure causes vasculopathy and that vasculopathy causes mortality are valid, inferences regarding long-term outcomes that are predicated on the basis of the antihypertensive effects of a given drug will be valid. If, however, vasculopathy is largely determined by covariates of blood pressure rather than by blood pressure itself, by a common ancestral phenomenon that influences blood pressure and vasculopathy in parallel yet independently, or by factors that are unrelated to blood pressure, inferences about the effects of the antihypertensive agent on outcome that are based on observations of its effect on blood pressure will be weak. In the first situation, we might use blood pressure not only as a primary end point in mechanistic research but also as a criterion for valid inference regarding outcome. In the second situation, the absence of causal contiguity makes such reasoning illogical.1 That “truth” is often elusive in medicine is a dilemma.
Clarity regarding the biological consequences of a given intervention delineated in mechanistic research and the clinical outcomes is essential because both are important. Inferences relative to one class of end points can be made on the basis of observations pertinent to the other, but their strengths can vary markedly. Cholesterol lowering can be induced by the use of statins, and clinical benefit accompanies cholesterol lowering. However, if cholesterol lowering is a result of metastatic cancer or starvation, an improved outcome (ie, enhanced survival) will not parallel cholesterol lowering. Does this tell us anything about (1) whether statins lower cholesterol or (2) whether cholesterol lowering is beneficial in decreasing the risk of coronary events? Of course not. What it tells us is that misguided thinking about surrogates—in this case, cholesterol lowering as a “stand-in” for reduction of mortality, is replete with risk.
Undue reliance on one type of end point as a substitute (surrogate) for another was unmasked by the CAST investigators.2 The suppression of premature ventricular complexes (PVCs) by antiarrhythmic agents did not lead to an anticipated mortality reduction in the CAST trial. The uncoupling occurred because PVCs and mortality may not be reflections of the same underlying electrophysiological phenomena, but it occurred in particular because the number of PVCs (a pathophysiological variable) is not a surrogate for mortality (a clinical event variable).
Multiple Diverse Effects of Pharmacological Agents
An additional conundrum results from the multiple actions of any given drug and the diverse effects of a given drug in the setting of diverse manifestations of a single-system or multisystem disease in a given individual. Typically, agents used for treatment of the same condition have different actions. The treatment of a given condition often involves the use of several alternative drugs, each of which has multiple actions. Each may effect the same change in one dependent variable, such as blood pressure. However, the other actions of each drug, whether overt or occult, may effect beneficial or deleterious changes in other dependent variables, regardless of whether they are assessed. As extreme examples, consider arsenic, cyanide, endotoxin, and snake venom, all of which lower blood pressure. Obviously, none qualifies as a therapeutic agent. Occult as well as overt multiple and diverse actions of drugs make it incumbent on clinicians to recognize that the validity of inferences about the clinical consequences of the use of any given agent to induce a specific pharmacologic effect will depend not only on the extent to which it affects the targeted biological phenomenon but also on the extent to which all of the actions of the agent have been defined and the extent to which all affect the entire organism, alone and in concert. This lesson has been learned painfully in, for example, studies of phosphodiesterase inhibitors as agents with positive inotropic effects. Their use ameliorated symptoms in keeping with augmented cardiac output, but with long-term use, mortality increased unexpectedly and paradoxically.3 The ultimate test of the usefulness of a drug or device depends on the determination of outcomes, ideally in randomized, controlled clinical trials of sufficient scope and duration.
Saying What We Mean to Avoid Alice’s Solecism
Examples of dependent variables in mechanistic research and outcomes in clinical studies that must be differentiated are shown in the Table⇓. Relationships among categories (eg, among biochemical phenomena, morbidity, and mortality) cannot be established inferentially. Their delineation requires research that addresses the relationship itself and research that addresses each type of end point. When the relationship differs from one that is assumed, a priori, to exist and to be associated with a favorable clinical outcome, inferences drawn from results in mechanistic research will not necessarily be realized, and anticipated favorable clinical outcomes will not necessarily ensue.
If we want to understand, for example, what determines blood pressure, we must measure it and characterize the responses of each of its physiological determinants to a given intervention. If we want to know whether reduction in blood pressure per se affects survival, we must evaluate survival under conditions in which blood pressure is modified by specific interventions. Blood pressure per se is not a surrogate for survival any more than survival is a surrogate for blood pressure.
Implications Pertinent to Policy
Optimal public policy related to drug development and approval of new drugs requires that those responsible make the essential distinction between dependent variables in mechanistic studies and clinical outcomes such as mortality. Both wisdom and the judicious use of words are needed to avoid obfuscation. The indiscriminate use of the term “surrogate” and the implication that one measurement is a substitute for another—when it is not—should be eschewed. Linguistic care is needed not only for crisp communication but also for insightful assessment of cause/consequence relationships and, in particular, for the therapeutic efficacy of drugs and devices under development. The type of variable that supports interpretations and, in particular, judgments that are necessary in forming policy and making decisions regarding the approval of drugs should be referred to as specific types of variables and classified by terms such as those used for categorization in the Table⇑. Wisdom and scientific facts are needed to determine the extent to which a policy decision (eg, approval of a specific drug) can be based on inferences derived from results of mechanistic studies. If the impact of a pathophysiological variable such as blood pressure on a dependent variable of a clinical study, such as mortality, is very powerful, the level of confidence will be high that an agent that effectively lowers elevated blood pressure will enhance survival. Thus, the mechanistic study can provide a result that permits inference about efficacy of treatment. Under such conditions, assessment of the potential usefulness of a novel agent may not require a mortality trial. However, definitive documentation of safety of any treatment modality must be considered. Regrettably, mechanistic trials per se cannot establish safety. Accordingly, assessment of safety must rely on other research approaches, including prospective, randomized clinical trials and observational studies. Because most adverse drug reactions are not known, a priori, at the time of drug approval, a well-structured surveillance system with careful monitoring is essential.
In contrast, if the strength of an association between a pathophysiological variable such as hypertriglyceridemia and mortality is weak, any inference about the benefit conferred by a particular agent based on mechanistic considerations alone (eg, reduction in elevated triglycerides) is much less likely to be valid compared with the situation that occurs when the association is strong. Under these conditions, the impact on outcomes of confounding factors, including multiple drug effects, is likely to be relatively greater. Thus, prospective, controlled clinical trials will be necessary not only because of the need to address safety considerations but also to provide credible evidence that is pertinent to efficacy.
The term “surrogate” should not be used in describing end points. Instead, descriptions of results and interpretations should be formulated in terms that designate the specific nature and category of variable assessed. This practice will reduce confusion in the interpretation of results and their implications. Biochemical, pathophysiological, pharmacodynamic, and morphological variables, among other dependent variables in mechanistic research, should be characterized to answer questions about how drugs work. They cannot be used to confidently predict clinical outcomes. Such outcomes can be delineated definitively only by characterizing dependent variables that are clinical phenomena. Careful distinction between these two fundamentally different types of variables (Table⇑) is essential for referees of peer-reviewed manuscripts, editors, thought leaders, practitioners, and policy makers as they shape medical literature and healthcare delivery. The same careful distinction is necessary for those who serve in regulatory agencies that are responsible for the drug approval process.
- Copyright © 1997 by American Heart Association