Obtaining Robust Causal Evidence from Observational Studies: Can Genetic Epidemiology Help?

Observational epidemiological studies have clearly made important contributions to understanding the determinants of population health. However, there have been high-profile problems with this approach, highlighted by apparently contradictory findings emerging from observational studies and from randomized controlled trials (RCTs) of the same issue. These situations, of which the best known probably relates to the use of hormone-replacement therapy (HRT) in coronary heart disease (CHD) prevention, have been discussed elsewhere (Davey Smith & Ebrahim, 2002) . The HRT controversy is covered elsewhere in this volume (see Chapter 5). Here, I will discuss two examples. First, consider the use of vitamin E supplements and CHD risk. Several observational studies have suggested that the use of vitamin E supplements is associated with a reduced risk of CHD, two of the most influential being the Health Professionals Follow-Up Study (Rimm et al., 1993) and the Nurses’ Health Study (Stampfer et al., 1993), both published in the New England Journal of Medicine in 1993. Findings from one of these studies are presented in Figure 9.1, where it can be seen that even short-term use of vitamin E supplements was associated with reduced CHD risk, which persisted after adjustment for confounding factors. demonstrates that nearly half of U.S. adults are taking either vitamin E supplements or multivitamin/multimineral supplements that generally contain vitamin E (Radimer et al., 2004). presents data from three available time points, where there appears to have been a particular increase in vitamin E use following 1993 (Millen, Dodd, & Subar, 2004), possibly consequent upon the publication of the two observational studies already mentioned, which have received nearly 3,000 citations between them since publication. The apparently strong observational evidence with respect to vitamin E and reduced CHD risk, which may have influenced the very high current use of vitamin E supplements in developed countries, was unfortunately not realized in RCTs, in which no benefit from vitamin E supplementation use is seen.

General Approaches to Analysis of Course: Applying Growth Mixture Modeling to Randomized Trials of Depression Medication

This chapter discusses the assessment of treatment effects in longitudinal randomized trials using growth mixture modeling (GMM) (Muthén & Shedden, 1999; Muthén & Muthén, 2000; Muthén et al., 2002; Muthén & Asparouhov, 2009). GMM is a generalization of conventional repeated measurement mixed-effects (multilevel) modeling. It captures unobserved subject heterogeneity in trajectories not only by random effects but also by latent classes corresponding to qualitatively different types of trajectories. It can be seen as a combination of conventional mixed-effects modeling and cluster analysis, also allowing prediction of class membership and estimation of each individual’s most likely class membership. GMM has particularly strong potential for analyses of randomized trials because it responds to the need to investigate for whom a treatment is effective by allowing for different treatment effects in different trajectory classes. The chapter is motivated by a University of California, Los Angeles study of depression medication (Leuchter, Cook, Witte, Morgan, & Abrams, 2002). Data on 94 subjects are drawn from a combination of three studies carried out with the same design, using three different types of medications: fluoxetine (n = 14), venlafaxine IR (n = 17), and venlafaxine XR (n = 18). Subjects were measured at baseline and again after a 1-week placebo lead-in phase. In the subsequent double-blind phase of the study, the subjects were randomized into medication (n = 49) and placebo (n = 45) groups. After randomization, subjects were measured at nine occasions: at 48 hours and at weeks 1–8. The current analyses consider the Hamilton Depression Rating Scale. Several predictors of course of the Hamilton scale trajectory are available, including gender, treatment history, and a baseline measure of central cordance hypothesized to influence tendency to respond to treatment. The results of studies of this kind are often characterized in terms of an end point analysis where the outcome at the end of the study, here at 8 weeks, is considered for the placebo group and for the medication group.

The Need for Dimensional Approaches in Discerning the Origins of Psychopathology

The 2008 meeting of the American Psychopathological Association was framed by a very challenging topic: causality. Indeed, setting aside any possible application in understanding psychopathology, causality is a deep concept—a fact that has kept philosophers gainfully employed for some time now. One thing is clear, however, at least in the behavioral sciences: If one wants to make credible causal claims, it helps to be able to directly manipulate the variables of interest. Indeed, some would go so far as to say that causality cannot be inferred without this kind of experimental manipulation. Through manipulation, one can systematically vary a variable of interest, while holding others constant, including the observational conditions. Consider, for example, how this is conveyed to new students in the behavioral sciences in a very useful text by Stanovich (2007). Stanovich (2007) first reviews the classic observation that simply knowing that two things (A and B) tend to occur together more often than one would expect by chance (a correlation) is not enough evidence to conclude that those two things have some sort of causal relationship (e.g., A causes B). To really claim that A causes B, ‘‘the investigator manipulates the variable hypothesized to be the cause and looks for an effect on the variable hypothesized to be the effect while holding all other variables constant by control and randomization’’ (p. 102). The implications of this experimental perspective on causality for psychopathology research are readily apparent: The situation is nearly hopeless, at least in terms of getting at the original, antecedent, distal causes of psychopathology. It is axiomatically unethical to manipulate variables to enhance the likelihood of psychopathology; we cannot directly manipulate things to create psychopathology in persons who do not already suffer from psychopathology. This is not to say that, once psychopathology is present, experimental designs are not fundamentally helpful in understanding the mechanisms underlying its expression. Indeed, the discipline of experimental psychopathology is founded on this premise, involving comparisons of the behaviors of persons with psychopathology and persons without psychopathology under precisely controlled conditions.

Statistical Methodology for a SMART Design in the Development of Adaptive Treatment Strategies

The past two decades have brought new pharmacotherapies as well as behavioral therapies to the field of drug-addiction treatment (Carroll & Onken, 2005; Carroll, 2005; Ling & Smith, 2002; Fiellin, Kleber, Trumble-Hejduk, McLellan, & Kosten, 2004). Despite this progress, the treatment of addiction in clinical practice often remains a matter of trial and error. Some reasons for this difficulty are as follows. First, to date, no one treatment has been found that works well for most patients; that is, patients are heterogeneous in response to any specific treatment. Second, as many authors have pointed out (McLellan, 2002; McLellan, Lewis, O’Brien, & Kleber, 2000), addiction is often a chronic condition, with symptoms waxing and waning over time. Third, relapse is common. Therefore, the clinician is faced with, first, finding a sequence of treatments that works initially to stabilize the patient and, next, deciding which types of treatments will prevent relapse in the longer term. To inform this sequential clinical decision making, adaptive treatment strategies, that is, treatment strategies shaped by individual patient characteristics or patient responses to prior treatments, have been proposed (Greenhouse, Stangl, Kupfer, & Prien, 1991; Murphy, 2003, 2005; Murphy, Lynch, Oslin, McKay, & Tenhave, 2006; Murphy, Oslin, Rush, & Zhu, 2007; Lavori & Dawson, 2000; Lavori, Dawson, & Rush, 2000; Dawson & Lavori, 2003). Here is an example of an adaptive treatment strategy for prescription opioid dependence, modeled with modifications after a trial currently in progress within the Clinical Trials Network of the National Institute on Drug Abuse (Weiss, Sharpe, & Ling, 2010). . . . Example . . . . . . First, provide all patients with a 4-week course of buprenorphine/naloxone (Bup/Nx) plus medical management (MM) plus individual drug counseling (IDC) (Fiellin, Pantalon, Schottenfeld, Gordon, & O’Connor, 1999), culminating in a taper of the Bup/Nx. If at any time during these 4 weeks the patient meets the criterion for nonresponse, a second, longer treatment with Bup/Nx (12 weeks) is provided, accompanied by MM and cognitive behavior therapy (CBT). However, if the patient remains abstinent from opioid use during those 4 weeks, that is, responds to initial treatment, provide 12 additional weeks of relapse prevention therapy (RPT). . . .

Understanding the Effects of Menopausal Hormone Therapy: Using the Women’s Health Initiative Randomized Trials and Observational Study to Improve Inference

Over the last decade, several large-scale randomized trials have reported results that disagreed substantially with the motivating observational studies on the value of various chronic disease–prevention strategies. One high-profile example of these discrepancies was related to postmenopausal hormone therapy (HT) use and its effects on cardiovascular disease and cancer. The Women’s Health Initiative (WHI), a National Heart, Lung, and Blood Institute–sponsored program, was designed to test three interventions for the prevention of chronic diseases in postmenopausal women, each of which was motivated by a decade or more of analytic epidemiology. Specifically, the trials were testing the potential for HT to prevent coronary heart disease (CHD), a low-fat eating pattern to reduce breast and colorectal cancer incidence, and calcium and vitamin D supplements to prevent hip fractures. Over 68,000 postmenopausal women were randomized to one, two, or all three randomized clinical trial (CT) components between 1993 and 1998 at 40 U.S. clinical centers (Anderson et al., 2003a). The HT component consisted of two parallel trials testing the effects of conjugated equine estrogens alone (E-alone) among women with prior hysterectomy and the effect of combined estrogen plus progestin therapy (E+P), in this case conjugated equine estrogens plus medroxyprogesterone acetate, among women with an intact uterus, on the incidence of CHD and overall health. In 2002, the randomized trial of E+P was stopped early, based on an assessment of risks exceeding benefits for chronic disease prevention, raising concerns among millions of menopausal women and their care providers about their use of these medicines. The trial confirmed the benefit of HT for fracture-risk reduction but the expected benefit for CHD, the primary study end point, was not observed. Rather, the trial results documented increased risks of CHD, stroke, venous thromboembolism (VTE), and breast cancer with combined hormones (Writing Group for the Women’s Health Initiative Investigators, 2002). Approximately 18 months later, the E-alone trial was also stopped, based on the finding of an adverse effect on stroke rates and the likelihood that the study would not confirm the CHD-prevention hypothesis.

Integrating Causal Analysis into Psychopathology Research

Both in psychopathology research and in clinical practice, causal thinking is natural and productive. In the past decades, important progress has been made in the treatment of disorders ranging from attention-deficit/hyperactivity disorder (e.g., Connor, Glatt, Lopez, Jackson, & Melloni, 2002) to depression (e.g., Dobson, 1989; Hansen, Gartlehner, Lohr, Gaynes, & Carey, 2005) to schizophrenia (Hegarty, Baldessarini, Tohen, & Waternaux, 1994). The treatments for these disorders include pharmacological agents as well as behavioral interventions, which have been subjected to clinical trials and other empirical evaluations. Often, the treatments focus on the reduction or elimination of symptoms, but in other cases the interventions are designed to prevent the disorder itself (Brotman et al., 2008). In both instances, the interventions illustrate the best use of causal thinking to advance both scientific theory and clinical practice. When clinicians understand the causal nature of treatments, they can have confidence that their actions will lead to positive outcomes. Moreover, being able to communicate this confidence tends to increase a patient’s comfort and compliance (Becker & Maiman, 1975). Indeed, there seems to be a basic inclination for humans to engage in causal explanation, and such explanations affect both basic thinking, such as identification of categories (Rehder & Kim, 2006), and emotional functioning (Hareli & Hess, 2008). This inclination may lead some to ascribe causal explanations to mere correlations or coincidences, and many scientific texts warn researchers to be cautious about making causal claims (e.g., Maxwell & Delaney, 2004). These warnings have been taken to heart by editors, reviewers, and scientists themselves; and there is often reluctance regarding the use of causal language in the psychopathology literature. As a result, many articles simply report patterns of association and refer to mechanisms with euphemisms that imply causal thinking without addressing causal issues head-on. Over 35 years ago Rubin (1974) began to talk about strong causal inferences that could be made from experimental and nonexperimental studies using the so-called potential outcomes approach. This approach clarified the nature of the effects of causes A vs. B by asking us to consider what would happen to a given subject under these two conditions.

The Mathematics of Causal Relations

Almost two decades have passed since Paul Holland published his highly cited review paper on the Neyman-Rubin approach to causal inference (Holland, 1986). Our understanding of causal inference has since increased severalfold, due primarily to advances in three areas: 1. Nonparametric structural equations 2. Graphical models Symbiosis between counterfactual and graphical methods 3. These advances are central to the empirical sciences because the research questions that motivate most studies in the health, social, and behavioral sciences are not statistical but causal in nature. For example, what is the efficacy of a given drug in a given population? Can data prove an employer guilty of hiring discrimination? What fraction of past crimes could have been avoided by a given policy? What was the cause of death of a given individual in a specific incident? Remarkably, although much of the conceptual framework and many of the algorithmic tools needed for tackling such problems are now well established, they are hardly known to researchers in the field who could put them into practical use. Why? Solving causal problems mathematically requires certain extensions in the standard mathematical language of statistics, and these extensions are not generally emphasized in the mainstream literature and education. As a result, large segments of the statistical research community find it hard to appreciate and benefit from the many results that causal analysis has produced in the past two decades. This chapter aims at making these advances more accessible to the general research community by, first, contrasting causal analysis with standard statistical analysis and, second, comparing and unifying various approaches to causal analysis. The aim of standard statistical analysis, typified by regression, estimation, and hypothesis-testing techniques, is to assess parameters of a distribution from samples drawn of that distribution. With the help of such parameters, one can infer associations among variables, estimate the likelihood of past and future events, as well as update the likelihood of events in light of new evidence or new measurements.

Causal Thinking for Objective Psychiatric Diagnostic Criteria: A Programmatic Approach in Therapeutic Context

Terms such as disorder, illness, disease, dysfunction, and deviance embody the preconceptions of historical development (Klein, 1999). That individuals become ill for no apparent reason, suffering from pain, dizziness, malaise, rash, wasting, etc., has been known since prehistoric days. The recognition of illness led to the social definition of the patient and the development of various treatment institutions (e.g., nursing, medicine, surgery, quacks, and faith healers). Illness is an involuntary affliction that justifies the sick, dependent role (Parsons, 1951). That is, because the sick have involuntarily impaired functioning, it is a reasonable social investment to exempt them (at least temporarily) from normal responsibilities. Illness implies that something has gone wrong. However, gaining exemption from civil or criminal responsibilities is often desired. Therefore, if no objective criteria are available, an illness claim can be viewed skeptically. By affirming involuntary affliction, diagnosis immunizes the patient against charges of exploitative parasitism. Therefore, illness may be considered a hybrid concept, with two components: (1) the necessary inference that something has actually, involuntarily, gone wrong (disease) and (2) the qualification that the result (illness) must be sufficiently major, according to current social values, to ratify the sickness exemption role. The latter component is related to the particular historical stage, cultural traditions, and values. This concept has been exemplified by the phrase ‘‘harmful dysfunction’’ (Wakefield, 1992). However, this does not mean that the illness concept is arbitrary since the inference that something has gone wrong is necessary. Beliefs as to just what has gone wrong (e.g., demon possession, bad air, bacterial infection) as well as the degree of manifested dysfunction that warrants the sick role reflect the somewhat independent levels of scientific and social development (for further reference, see Lewis, 1967). How can we affirm that something has gone wrong if there is no objective evidence? The common statistical definition of abnormality simply is ‘‘unusual.’’ Something is abnormal if it is rare. Although biological variability ensures that someone is at an extreme, there is a strong presumption that something has gone wrong if sufficiently extreme.

Rare Variant Approaches to Understanding the Causes of Complex Neuropsychiatric Disorders

The distinction between genetic variation that is present in more than 5% of the population (defined as common) and genetic variation that does not meet this threshold (defined as rare) is often lost in the discussion of psychiatric genetics. As a general proposition, the field has come to equate the hunt for common variants (or alleles) with the search for genes causing or contributing to psychiatric illness. Indeed, the majority of studies on mood disorders, autism, schizophrenia, obsessive–compulsive disorder, attention-deficit/hyperactivity disorder, and Tourette syndrome have restricted their analyses to the potential contribution of common alleles. Studies focusing on rare genetic mutations have, until quite recently, been viewed as outside the mainstream of efforts aimed at elucidating the biological substrates of serious psychopathology. Both the implicit assumption that common alleles underlie the lion’s share of risk for most common neuropsychiatric conditions and the notion that the most expeditious way to elucidate their biological bases will be to concentrate efforts on common alleles deserve careful scrutiny. Indeed, key findings across all of human genetics, including those within psychiatry, support the following alternative conclusions: (1) for disorders such as autism and schizophrenia, the study of rare variants already holds the most immediate promise for defining the molecular and cellular mechanisms of disease (McClellan, Susser, & King, 2007; O’Roak & State, 2008); (2) common variation will be found to carry much more modest risks than previously anticipated (Altshuler & Daly, 2007; Saxena et al., 2007); and (3) rare variation will account for substantial risk for common complex disorders, particularly for neuropsychiatric conditions with relatively early onset and chronic course. This chapter addresses the rare variant genetic approach specifically with respect to mental illness. It first introduces the distinction between the key characteristics of common and rare genetic variation. It then briefly addresses the methodologies employed to demonstrate a causal or contributory role for genes in complex disease, focusing on how these approaches differ in terms of the ability to detect and confirm the role of rare variation.

Causes of Posttraumatic Stress Disorder

The definition of posttraumatic stress disorder (PTSD) in the Diagnostic and Statistical Manual of Mental Disorders, 3rd edition (DSM-III), and in subsequent DSM editions is based on a conceptual model that brackets traumatic events from other stressful experiences and PTSD from other responses to stress and links the two causally. The connection between traumatic experiences and a specific mental disorder has become part of the general discourse. PTSD provides a cultural template of the human response to war, violence, disaster, or very bad personal experiences. The DSM-III revolutionized American psychiatry. The manual’s editors wanted a symptom-based, descriptive classification and generally rejected any reference to causal theories about mental processes. PTSD was an exception to the rule of creating a classification that is ‘‘atheoretical with regard to etiology or pathophysiological process’’ (American Psychiatric Association, 1980 p. 7), but the exception was not noted anywhere in the manual. Not only did the PTSD definition include an etiological event, but it incorporated a theory, an underlying process, that connects the syndrome’s diagnostic features (McNally, 2003; Young, 1995). In 1994, the American Psychiatric Association published the fourth edition of the DSM. The definition of PTSD, which had already undergone some revisions in DSM-IIIR, maintained the syndrome’s description but changed materially the stressor criterion. The range of events was widened, and the emphasis shifted to the subjective experience of victims. The list of ‘‘typical’’ traumas in the DSM-IV left no doubt that the intent was to enlarge the variety of experiences that can be used to diagnose PTSD beyond the initial conception of directly experienced, life-threatening events such as combat, natural disaster, rape, and other assault. Persons who learned about a threat to the physical integrity of another person or about a traumatic event experienced by a friend could be considered victims. A novel form of PTSD took shape following the 9/11 terrorist attacks, when the entire population of the United States was considered to have been affected by a ‘‘distant’’ trauma, produced chiefly by viewing television coverage.