CONTENTS Preface ........................................................................................................ix Acknowledgments......................................................................................xv PART ONE Performing Clinical Research 1 Forming Research Questions..............................................................3 2 Conducting Systematic Reviews ......................................................15 3 Finding Information About the Burden of Disease ........................49 4 An Introduction to Performing Therapeutic Trials ..........................59 5 The Tactics of Performing Therapeutic Trials ..................................66 6 The Principles Behind the Tactics of Performing Therapeutic Trials ............................................................................173 7 Testing Quality Improvement Interventions ..................................244 8 Evaluating Diagnostic Tests ............................................................273 9 Determining Prognosis and Creating Clinical Decision Rules......................................................................323 10 Assessing Claims of Causation........................................................356 11 Generating Outcome Measurements, Especially for Quality of Life ............................................................................388 PART TWO Becoming a Clinical Researcher 12 Becoming a Successful Clinician-investigator................................415 13 Preparing a Research Protocol to Improve its Chances for Success ..................................................................429 14 Online Data Collection ....................................................................440 15 Analyzing Data..................................................................................446 16 Preparing Reports for Publication and Responding to Reviewers’ Comments................................................................461 17 Dealing with the Media....................................................................474 Index ........................................................................................................487 Read me....................................................................................................000 vii (cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:9)(cid:10)(cid:11)(cid:1)(cid:12)(cid:5)(cid:1)(cid:13)(cid:14)(cid:1)(cid:15)(cid:6)(cid:7)(cid:16)(cid:17)(cid:18)(cid:16)(cid:5)(cid:17)(cid:19)(cid:20)(cid:21)(cid:1)(cid:22)(cid:16)(cid:23)(cid:7)(cid:24)(cid:1)(cid:25)(cid:14)(cid:1)(cid:26)(cid:16)(cid:27)(cid:28)(cid:19)(cid:10)(cid:10)(cid:21)(cid:1)(cid:29)(cid:3)(cid:6)(cid:24)(cid:3)(cid:17)(cid:1)(cid:18)(cid:14)(cid:1)(cid:29)(cid:30)(cid:5)(cid:16)(cid:10)(cid:10)(cid:21)(cid:1)(cid:16)(cid:17)(cid:24)(cid:1)(cid:31)(cid:19)(cid:10)(cid:19)(cid:6)(cid:1) (cid:30)(cid:8)!(cid:19)""(cid:14) (cid:2)(cid:3)(cid:4)(cid:5)(cid:4)(cid:6)(cid:7)(cid:3)(cid:1)(cid:8)(cid:9)(cid:4)(cid:10)(cid:11)(cid:12)(cid:4)(cid:13)(cid:3)(cid:13)(cid:14)(cid:15)(cid:16)(cid:1)(cid:17)(cid:13)(cid:18)(cid:1)(cid:19)(cid:13)(cid:1)(cid:20)(cid:13)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:4)(cid:6)(cid:7)(cid:3)(cid:1)(cid:21)(cid:22)(cid:7)(cid:6)(cid:19)(cid:4)(cid:6)(cid:11)(cid:1)(cid:23)(cid:11)(cid:24)(cid:11)(cid:7)(cid:22)(cid:6)(cid:25)(cid:14) EVALUATING DIAGNOSTIC TESTS ——— 321 20. RaoPM, FeltmateCM, RheaJT, et al.Helical computed tomography in differentiating appendicitis and acute gynecologic conditions. Obstet Gynecol 1999;93:417–421. 21. Fleiss JL. Measuring agreement between two judges on the presence or absence of a trait. Biometrics 1975;31:651–659. 22. DonnerA, KlarN. The statistical analysis of kappa statistics in multiple samples. J Clin Epidemiol 1996;9:1053–1058. 23. Cook RJ, Farewell VT. Conditional inference for subject-specific and marginal agree- ment: two families of agreement measures. Can J Stat 1995;23:333–344. 24. Meade MO, Cook RJ, Guyatt GH, et al. Interobserver variation in interpreting chest radiographs for the diagnosis of acute respiratory distress syndrome. Am J Respir Crit Care Med 2000;161:85–90. 25. McGinn T, Guyatt G, Cook R. Measuring agreement beyond chance. In: Guyatt G, RennieD, eds. Users’ guides to the medical literature: A manual for evidence-based clinical practice. Chicago, IL: AMA Press, 2002. 26. Dacie SV, Lewis SM. Practical haematology, 4th ed. London: Churchill Livingstone, 1984:107–109. 27. Kemppainen EA, Hedstrom JI, Puolakkainen PA, et al. Rapid measurement of urinary trypsinogen-2 as a screening test for acute pancreatitis. N Engl J Med1997;336:1788–1793. 28. Begg CB, Greenes RA. Assessment of diagnostic tests when disease verification is sub- ject to selection bias. Biometrics 1983;39:207–215. 29. Gray R, Begg CB, Greenes RA. Construction of receiver operating characteristic curves when disease verification is subject to selection bias. Med Decis Making1984;4:151–164. 30. Ransohoff DF, Feinstein AR. Problems of spectrum and bias in evaluating the efficacy of diagnostic tests. N Engl J Med 1978;299:926–930. 31. ChoiBCK. Sensitivity and specificity of a single diagnostic test in the presence of work- up bias. J Clin Epidemiol 1992;45:581–586. 32. The PIOPED Investigators. Value of ventilation/perfusion scan in acute pulmonary embolism. Results of the prospective investigation of pulmonary embolism diagnosis (PIOPED). JAMA 1990;263:2753–2759. 33. PungliaRS, D’AmicoAV, CatalonaWJ, et al.Effect of verification bias on screening for prostate cancer by measurement of prostate-specific antigen. N Engl J Med 2003;349: 335–342. 34. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982;143:29–36. 35. HanleyJA, McNeilBJ. A method of comparing the areas under receiver operating char- acteristic curves derived from the same cases. Radiology 1983;148:839–843. 36. Albert A. On the use and computation of likelihood ratios in clinical chemistry. Clin Chem 1982;28:1113–1119. 37. PattersonC, GuyattGH, SingerJ, et al.Iron deficiency anemia in the elderly: the diag- nostic process. Can Med Assoc J1991;144:435–440. 38. Dales RE, Stark RM, Sankaranarayanan R. Computed tomography to stage lung can- cer. Am Rev Respir Dis 1990;141:1096–1101. 39. GuyattGH, TugwellP, FeenyDH, et al.A framework for clinical evaluation of diagnos- tic technologies. Can Med Assoc J1986;134:587–594. 40. Committee of Principal Investigators. WHO cooperative trial on primary prevention of ischemic heart disease using clofibrate to lower serum cholesterol: mortality follow-up. Lancet 1980;2:370–385. 41. Echt DS, Liebson PR, Mitchell LB, et al. Mortality and morbidity in patients receiving encainide, flecainide, or placebo. The cardiac arrhythmia suppression trial. N Engl J Med 1991;324:781–788. 42. Brown VA, Sawers RS, Parsons RJ, et al. The value of antenatal cardiotocography in the management of high-risk pregnancy: a randomized controlled trial. Br J Obstet Gynaecol 1982;89:716–722. Copyright© by R. BrianHaynes, David L. Sackett, Gordon H. Guyatt, and Peter Tugwell. Clinical Epidemiology: How to Do Clinical Practice Research. 322 ——— PERFORMING CLINICAL RESEARCH 43. Flynn AM, Kelly J, Mansfield H, et al. A randomized controlled trial of non-stress an- tepartum cardiotocography. Br J Obstet Gynaecol 1982;89:427–433. 44. BarrattA, IrwigL, GlasziouP, et al.Recommendations about screening. In: GuyattG, RennieD, eds. Users’ guides to the medical literature: A manual for evidence-based clinical practice. Chicago, IL: AMA Press, 2002. 45. Sox HC Jr, Margulies I, Sox CH. Psychologically mediated effects of diagnostic tests. Ann Intern Med 1981;95:680–685. 46. McDonald IG, Daly J, Jelinek VM, et al. Opening pandora’s box: the unpredictability of reassurance by a normal test result. BMJ 1996;313:329–332. Copyright© by R. BrianHaynes, David L. Sackett, Gordon H. Guyatt, and Peter Tugwell. Clinical Epidemiology: How to Do Clinical Practice Research. 9 DETERMINING PROGNOSIS AND CREATING CLINICAL DECISION RULES Gordon Guyatt Chapter Outline 9.1 Why study prognosis? 9.2 Prognosis versus diagnosis 9.3 Clinical prediction rules 9.4 Basic principles of conducting prognostic studies CLINICAL RESEARCH SCENARIO Critically ill patients managed in intensive care units (ICUs) are at in- creased risk of gastrointestinal (GI) bleeding. This is true even for pa- tients without previous problems with gastric or duodenal ulcers, or with other GI pathology. In the late 1980s, stress ulcer prophylaxis with histamine H -receptor antagonists (drugs that reduce acid secretion in 2 the stomach) was widely used in ICUs to prevent GI bleeding in criti- cally ill patients. Clinicians were increasingly aware that the incidence of serious bleeding had apparently decreased markedly and suspected that it might be low enough that prophylaxis was no longer warranted for all patients. As a result, they raised questions about the advisability of stress ulcer prophylaxis. Specifically, clinicians questioned what sub- groups of critically ill patients (if any) still had a risk of bleeding that was sufficiently great that they should receive stress ulcer prophylaxis. We set out to address this issue. 9.1 WHY STUDY PROGNOSIS? There are two fundamental reasons why clinicians need to know a patient’s prognosis. One is that, irrespective of management decisions, patients often are interested in knowing how well, or badly, they are likely to fare. Certain diagnoses, such as a diagnosis of cancer, carry with them almost inevitable 323 Copyright© by R. BrianHaynes, David L. Sackett, Gordon H. Guyatt, and Peter Tugwell. Clinical Epidemiology: How to Do Clinical Practice Research. 324 ——— PERFORMING CLINICAL RESEARCH questions about how long a patient is going to live. Clinicians must be pre- pared to address patients’ questions about their fates, or prognosis. The second fundamental reason why clinicians need to know a pa- tients’ prognosis is that it may alter their recommended management plan. If a patient’s risk of the adverse outcome that the treatment is designed to prevent is sufficiently low, a therapeutic intervention may not be war- ranted. The reason is that all interventions carry with them a burden of cost, inconvenience, and adverse effects. Furthermore, the opportunity for benefits varies from patient to patient, depending on their prognosis. The research scenario that introduced this chapter provides an example: In the late 1980s, clinicians began to suspect that the incidence of stress ulcera- tion in patients with GI bleeding was sufficiently low that, at least in some critically ill patients, the cost, inconvenience, and side effects of stress ulcer prophylaxis were too great to warrant prophylaxis. As the risk of the adverse outcomes that treatment is designed to pre- vent decreases, and the risk of treatment-induced adverse outcomes increases (I’ll call such events “treatment toxicity”), patients become progressively less enthusiastic about a potential therapeutic intervention. When the probability of adverse outcomes becomes sufficiently low, or when the treatment toxic- ity becomes sufficiently great, patients begin to cross a threshold beyond which they would decline even effective therapy. Thresholds differ in indi- viduals, but as the incidence of adverse events in untreated patients decreases further, or as toxicity rises further, more and more patients would decline therapy designed to lower their risk. Eventually, even those who value the benefits highly and are less concerned about adverse effects no longer find the benefits worth the risks. 9.2 PROGNOSIS VERSUS DIAGNOSIS In what ways are studies of prognosis and diagnosis similar? Both types of studies often have the same purpose: to identify patients who have a higher or lower probability of a particular condition (diagnosis), outcome (prog- nosis), or responsiveness to therapy. For instance, in patients with carotid stenosis who benefit from carotid endarterectomy, as the number of car- diovascular risk factors [age more than 70 years; systolic blood pressure greater than 160 or diastolic less than 90; stroke or transient ischemic at- tacks (TIAs) in the last 30 days; stroke rather than TIA; carotid stenosis more than 80%; ulceration on angiography; history of hypertension, heart failure, myocardial infarction (MI), dyslipidemia, claudication, or smoking] increases, the risk of a subsequent stroke increases. As the risk of stroke increases, the potential benefits of endarterectomy increase. Table 9–1 pre- sents data from a randomized trial of carotid endarterectomy about the re- lation between the number of risk factors, the magnitude of benefit in terms of the absolute risk reduction that patients can expect from endarterec- tomy, and the number of patients one needs to treat with endarterectomy to prevent a stroke (1). Copyright© by R. BrianHaynes, David L. Sackett, Gordon H. Guyatt, and Peter Tugwell. Clinical Epidemiology: How to Do Clinical Practice Research. DETERMINING PROGNOSIS AND CREATING CLINICAL DECISION RULES ——— 325 TABLE 9–1 Relation between Risk Factors and Risk Difference with Carotid Endarterectomy in Patients with Cerebrovascular Disease (1) Number of Number of Number Needed Risk Factors Patients in Trial Risk Difference (%) to Treat 0–5 305 3.8 27 6 153 9.8 10 7 or more 201 19.4 5 From North American Symptomatic Carotid Endarterectomy Trial Collaborators. Beneficial effect of carotid endarterectomy in symptomatic patients with high-grade carotid stenosis. N Engl J Med 1991;325:445–453, with permission. In certain clinical situations, the distinction between prognosis and diagnosis blurs completely. For instance, when a clinician asks “does this patient with chest pain in the emergency room have a myocardial infarc- tion,” the real, underlying question of interest may be “does this patient have a risk of adverse events sufficiently great that the patient requires ad- mission to the hospital.” It isn’t difficult to think of other instances in which distinctions be- tween prognosis and diagnosis blur. For instance, although most clinicians think of troponin I as useful in the diagnosis of acute MI, it may also pro- vide information about the long-term prognosis of patients presenting with acute coronary syndrome (2). Consider the prognostic value of preopera- tive characteristics in estimating the likelihood of cardiac events after non- cardiac surgery. Clinicians only need to follow patients for a few days to determine whether they have attained the outcome of interest. Moreover, long-term follow-up is often a component of the reference standard in many diagnostic test studies. For instance, investigators have used long-term follow-up to establish a diagnosis of “no pulmonary embolus” in patients who did not undergo a pulmonary angiogram at the time of presentation with possible pulmonary embolus (3). Another similarity lies in the nature of the relation between the pre- dictor variable (the diagnostic test or the prognostic factor) and the out- come (disease present or absent, or outcome present or absent). The association between the diagnostic test and the disease is seldom, if ever, causal. That is, the high troponin, or low serum ferritin is a consequence of the pathological processes of MI and iron deficiency, and not the cause. The same is true when troponin is used to predict outcome, or for other common prognostic variables such as lung or cardiac function, or func- tional capacity. Prognostic variables such as smoking, high blood pressure, or high serum lipids may, on the other hand, have an etiologic link to the outcomes with which they are associated. In what ways do prognostic and diagnostic studies differ? Diagnostic studies geared to establish test properties for clinical practice should restrict Copyright© by R. BrianHaynes, David L. Sackett, Gordon H. Guyatt, and Peter Tugwell. Clinical Epidemiology: How to Do Clinical Practice Research. 326 ——— PERFORMING CLINICAL RESEARCH themselves to populations with an intermediate probability that the target condition is present. If the probability is too low—below a “test threshold”—it may not be worth ordering the test, particularly if it is ex- pensive or invasive. If the probability is too high—above the “treatment threshold,”—the clinician should not bother with further testing but rather go ahead and treat. Thus, there is little point in enrolling patients at very high or very low risk in studies of diagnostic tests unless, of course, one cannot identify characteristics that distinguish between high- and low-risk patients. Prognostic studies, however, may enrol a wider spectrum of pa- tients, including those with very low and very high probabilities of the out- come of interest. Diagnostic test studies typically focus on an individual laboratory or radiological test, or a small number of tests. Prognostic studies typically address a number of prognostic factors, often starting with demographic variables (such as age and sex) and including markers of disease severity (such as physical examination findings and disease staging). However, in- vestigators may focus on the prognostic usefulness of tests ordinarily con- sidered as bearing primarily on diagnosis. Classical diagnostic test studies do not usually address the independent contribution of the diagnostic test to the diagnostic process, whereas prognostic studies often use techniques such as multivariable regression to cast light on the independent contri- bution of each prognostic factor. Diagnostic test studies usually compare the test results to a “gold,” reference, or criterion standard ascertained at the same time as, or shortly after, testing. Prognostic studies, by their nature, involve following pa- tients over time to determine whether the outcome of interest occurs. Diagnostic test studies report results in terms of likelihood ratios (LRs) or (less usefully) sensitivity and specificity. Prognostic studies typically re- port relative risks, odds ratios, or survival curves and their associated hazard ratios. Thus, despite many similarities in basic study design, the issues in prognostic and diagnostic test studies are often different enough that we considered it worth having separate chapters addressing these issues. 9.3 CLINICAL PREDICTION RULES Clinical prediction rules, when properly developed, may take studies of prognosis to a higher level. Clinical prediction rules include simultaneous consideration of several factors in predicting the prognosis of individual patients. They may also be developed for use with patients for whom clin- icians have not yet established a firm diagnosis. In the latter case, they may predict the result of either an intermediate test, such as plain x-ray, or a more definitive test or diagnostic pathway. To illustrate this, consider patients presenting to an emergency de- partment with chest pain suggestive of an acute coronary syndrome. In Copyright© by R. BrianHaynes, David L. Sackett, Gordon H. Guyatt, and Peter Tugwell. Clinical Epidemiology: How to Do Clinical Practice Research. DETERMINING PROGNOSIS AND CREATING CLINICAL DECISION RULES ——— 327 such patients with acute coronary syndrome, the nature of the patients’ pain, the electrocardiogram, and cardiac enzymes all play a major role in deciding on the likelihood of acute MI. Once the clinician has established that a particular patient does indeed have an MI, she must simultaneously consider age, degree of cardiac dysfunction, and presence of arrhythmias in deciding on how likely the patient is to be alive a year from presenta- tion. Thus, a clinical prediction rule can help make a diagnosis or can as- sist in the management of the patient once the diagnosis is made. The task of simultaneously considering a number of variables in de- ciding on the probability of a diagnosis, or the likelihood of a subsequent adverse event, represents a considerable cognitive challenge. In particu- lar, clinicians are at risk of double-counting. That is, they may fail to fully take into account that, to the extent that one variable is correlated with another, a patient’s status on the second variable will add relatively little diagnostic or prognostic information. For instance, when consider- ing whether patients who are anemic are iron deficient, it might be tempting to think that a low mean cell volume suggests iron deficiency and that a low serum ferritin strengthens the case. As it turns out, how- ever, the information is redundant, and mean cell volume adds nothing once you’ve considered the serum ferritin. If clinicians treat the tests as if they are independent, they risk over- or underestimating the likelihood of iron deficiency. Considering the magnitude of the cognitive challenge, the possibility naturally arises that mathematical models that simultaneously consider all relevant variables will do a better job than the clinicians’ intuition. When we say “simultaneously consider,” we mean that they avoid the problem of “double-counting” the results of tests that are correlated with one another. All such models provide an estimate of the likelihood of a diagnosis (such as myocardial infarction or ankle fracture) or of future adverse events (such as a poor outcome in patients presenting with syncope, or the need for reoperation in patients with tibial fracture). Thus, such models have received appellations that include “clinical prediction rules” or “clinical prediction guides.” The term “guides” may be more appropriate for models that simply generate estimates without dictating a course of action. On the other hand, some argue that “rules” is a good term because it emphasizes the necessity for rigorous and replicable application of each component of the diagnostic or prognostic strategy. Some models go one step further than models (by any name) that sim- ply generate estimates: They prescribe a particular course of action. For in- stance, the Ottawa Ankle Rule tells clinicians that, in patients presenting with ankle injuries, if they find no tenderness in specific areas and if the patients can bear their own weight, the likelihood of an underlying frac- ture is very low. The rule does not, however, stop there. It also tells clin- icians that, in such low-risk patients, they needn’t order radiographs. One term for such prescriptive models is “clinical decision rules.” Copyright© by R. BrianHaynes, David L. Sackett, Gordon H. Guyatt, and Peter Tugwell. Clinical Epidemiology: How to Do Clinical Practice Research. 328 ——— PERFORMING CLINICAL RESEARCH 9.4 BASIC PRINCIPLES OF CONDUCTING PROGNOSTIC STUDIES A protocol for any study determining patients’ prognosis must attend to the elements outlined in the following checklist: (cid:2) Conduct your literature review. (cid:2) Pose your research question. (cid:2) Recruit your participants. Recruit a representative sample of patients who do not have the outcome of interest at the time of initial observation and who are, preferably, at an identifiable, common, and early point in their disorder or exposure. (cid:2) Choose what you will measure. Record all patient characteristics that might show substantial associations with the outcome of interest. (cid:2) Use measurement strategies. Use accurate and unbiased measurement of the outcome of interest. (cid:2) Select your statistical procedures. Include sample size and analysis. (cid:2) Select your presentation and interpretation methods. Note that this checklist is almost identical to that for diagnostic test studies, reflecting the very close relation between the goals and methods of these two types of studies. Planning for the study that will provide the focus for this discussion began 15 years ago. My colleague, Dr. Deborah Cook, led all aspects of the work, from inception to completion. Our team submitted the grant in 1988, the granting agency funded the study in 1989, and the journal arti- cle reporting the results appeared in 1994 (4). We have learned a lot since then, and the narrative of the study will often focus on what we would have done differently if we were conducting the study today. (cid:2) Conduct your literature review. In 1987, when we began planning our study, intensivists regularly ad- ministered stress ulcer prophylaxis with H -receptor antagonists to criti- 2 cally ill patients. The results of a number of randomized trials supported this practice, and, in 1991, our group published a systematic review and meta-analysis that we had conducted while our prognostic study was getting underway. The review demonstrated a 50% reduction in patient- important bleeding with prophylaxis (5). Despite this strong evidence in support of prophylaxis, in the late 1980s, clinicians began to question its routine use because of the increas- ing evidence that the incidence of serious bleeding from stress ulceration Copyright© by R. BrianHaynes, David L. Sackett, Gordon H. Guyatt, and Peter Tugwell. Clinical Epidemiology: How to Do Clinical Practice Research. DETERMINING PROGNOSIS AND CREATING CLINICAL DECISION RULES ——— 329 was decreasing. Our review (which was fully systematic or comprehensive for the randomized trials, but not for the observational studies) found an incidence of serious bleeding of 2.6% in 27 randomized trials. We identi- fied only three studies in adults, published between 1984 and 1987, that had prospectively evaluated the independent contribution of a variety of patient characteristics to risks of bleeding. Whereas our search for observational studies was not comprehensive or systematic, we have repeated the search using the sensitive “Clinical Queries” for prognosis in PubMed (http://www.ncbi.nlm.nih.gov/entrez/ query/static/clinical.html), which utilizes the following search string: “inci- dence” [Medical Subject Headings (MeSH)] OR “mortality” (MeSH) OR “fol- low-up studies” (MeSH) OR ‘mortality” (SH) OR prognos* (WORD) OR predict* (WORD) OR course (WORD). This search failed to reveal addi- tional relevant articles. If your goal is to be comprehensive, you can pursue other useful strategies once an initial search has identified most of the rele- vant articles. For each of these articles, you can use Science Citation Index (particularly useful for older articles), to find newer studies and reviews that cite the older studies, or use the “Related Articles” feature of PubMed. (cid:2) Pose your research question. In the report of the study, we described our project this way: “We undertook this prospective study to determine the incidence of clinically important GI bleeding in a heterogeneous group of critically ill patients and to identify patients at sufficiently low risk of bleeding to obviate the need for prophylaxis.” In this book, we are proposing that you make explicit what we felt was implicit in this statement. The statement we made in the published paper (in the shaded box at the beginning of this section) places the ques- tion squarely in the “clinical decision rule” category. That is, if we suc- ceeded in answering the question, we would be providing clinicians with management advice: “In this particular subgroup of patients, you should withhold prophylaxis.” As we proceed with this discussion we will review the particular challenges implicit in this framing (issues that we were not fully aware of at the time and that resulted in limitations in our study de- sign, and the analysis and interpretation of the data). Consideration of clinical decision rules highlights the need for inves- tigators to be vividly clear on the question they are addressing and the al- ternative designs that flow from that question. Clinicians might ask, “Should clinicians prescribe H -receptor antagonists in particular sub- 2 groups of critically ill patients, in relation to their risk of GI bleeding?” This question, however, is too vague. There are two ways one can shed light on the optimal management of critically ill patients who may be at risk of GI Copyright© by R. BrianHaynes, David L. Sackett, Gordon H. Guyatt, and Peter Tugwell. Clinical Epidemiology: How to Do Clinical Practice Research.