Original articleAn algorithm for prospective individual matching in a non-randomized clinical trial
Introduction
Susceptibility bias in a clinical trial arises when treatment groups differ with respect to the probability of developing the outcome under study [1]. Random assignment to treatment groups has gained widespread acceptance as the favored strategy to minimize susceptibility bias and to achieve balanced allocation of subjects in treatment groups. In certain circumstances, variations on simple randomization may be desirable. For instance, if the set of prognostic or risk factors is small, randomizations may be carried out within strata defined by the joint distribution of the risk factor levels [2]. If there are many factors, an adaptive randomization procedure [3], or a minimization procedure 4, 5 might be more suitable. The goal of all these methods is to make treatment assignments in such a way as to maximize the comparability of treatment groups with respect to known (and, it is hoped, unknown or unmeasured) prognostic or risk factors.
In some situations, however, random assignment to treatment group is impossible and/or unethical, and the rigor of the trial is compromised. A few examples of these situations include: assessing treatment efficacy where all patients in a center or clinic will receive the new treatment (e.g., changing the standard of care); evaluating the effect of a new widespread public health program or policy; studying the effects of structural or policy changes in health care delivery systems (e.g., closing of a hospital, new managed care or Medicaid system). In order to minimize susceptibility bias and thus preserve the ability of non-randomized clinical trials to generate valid estimates of the effect of the treatment or intervention, alternative strategies to achieve balanced allocation are clearly and urgently needed.
In a recent clinical trial of a multicomponent intervention to prevent incident delirium in hospitalized elderly patients—the Delirium Prevention Trial—randomization was not feasible, but the investigators achieved successful balanced allocation of subjects to treatment groups by using a novel method: prospective individual matching. In the Delirium Prevention Trial, the intervention was performed on one floor in the hospital, with “usual care” being delivered on two other control floors. At the time of the study, the hospital functioned at 120% patient capacity on the medical service; thus, medical patients were assigned frequently to non-medical units such as surgery or obstetrics. During a 2-month pilot period, beds on the intended study floors were often unavailable and only four patients could be randomly assigned to the appropriate study floor. Thus, enrollment of the targeted sample size of at least 800 subjects during a 3-year period would not have been possible using the classic randomized trial design. The floor assignments in the clinical trial were beyond the control of the investigators and were determined strictly by hospital admission procedures, which are driven by bed availability and are independent of patient characteristics. The investigators therefore needed to develop a method to create balanced treatment groups after treatment assignments had already taken place.
Section snippets
Specific aims
The specific aims of this report are to: (1) describe a method for achieving balanced allocation in a non-randomized clinical trial; (2) discuss the implementation of the method in the Delirium Prevention Trial; (3) evaluate the success of this method in achieving balance of risk factors; and (4) recommend strategies for implementing the method in other clinical trials.
Description of the method
In our method, balance is accomplished by selecting appropriate control patients, rather than by randomly allocating treatment assignments. The basic strategy is to enroll every possible intervention patient, and then to select and enroll controls, using a prospective individual matching strategy. The controls are selected from a reasonably large pool of candidates in such a way as to make the distributions of prognostic or risk factors as similar as possible between the intervention and
Implementation of prospective individual matching in the Delirium Prevention Trial
The Delirium Prevention Trial is a study of the impact of a multiple risk factor intervention strategy on preventing incident delirium during hospital stay [7]. The intervention was managed by a specially trained and coordinated staff of nurses, specially trained staff dedicated to the program, physicians and volunteers, and involved modifications to patient care routines as well as to environmental factors. Because many interventions were unit-wide, it was not feasible to target this protocol
Evaluation of the matching algorithm
Fig. 2 illustrates the patient flow, from screening eligibility through matching, for the Delirium Prevention Trial. Over a 35-month screening and enrollment phase, 564 intervention floor patients qualified for enrollment into the Delirium Prevention Trial, and 756 qualified for enrollment from the control floors. The ratio of available controls per intervention patient was therefore 1.34. The intervention floor was relatively larger (32 beds) than the two control units (27 and 24 beds,
Implementation considerations
The success of our method in achieving a reasonable pair match rate depends primarily on two conditions: selection of a proper set of matching factors, and the availability of an adequate number of matching controls to ensure a reasonable pair match rate. Ideally, for this study design, the investigators should match on a small number of factors (three or fewer) that are strong prognostic determinants for the study outcome in order to achieve matches more efficiently. Limiting the number of
Analysis considerations
Because our method results in an individually matched cohort design, data should be analyzed using statistical methods developed for prospectively matched data. Our previous investigations suggest that the prospectively matched design, at least under certain conditions, is more efficient than the randomized-block design arising from a traditional randomized clinical trial [9]. Certain matched models, for example the extension of the conditional logistic regression for prospectively sampled
Alternative observational designs
The two major observational designs that would provide viable alternatives to the prospective individual matching design proposed here are: (1) a prospective cohort design enrolling all eligible patients in both treatment groups, and (2) a modified cohort study in which controls are selected randomly. In both cases, differences in baseline treatment group characteristics would need to be controlled in the analyses. Because the study enrollment process entailed considerable costs for control
Discussion
We have described the successful implementation of a method that employs a prospective individual matching strategy to achieve balance between treatment groups with respect to risk factors in a non-randomized intervention trial. The design is analogous to a matched cohort study, and can be analyzed using well established statistical models for matched designs. The application of this method to a controlled clinical trial represents an innovative and important extension of this design. Moreover,
Acknowledgements
This work was supported in part by grants from the National Institute on Aging (#RO1AG12551), the Commonwealth Fund (#95-47 and 94-90) and the Retirement Research Foundation (#94-71), and by in-kind support from the Claude D. Pepper Older Americans Independence Center (#P60AG10469). Dr. Inouye is recipient of a Midcareer Award from the National Institute on Aging (#K24AG00949) and a Donaghue Investigator Award from the Patrick and Catherine Weldon Donaghue Medical Research Foundation
References (13)
The randomization and stratification of patients to clinical trials
J Chron Dis
(1974)Clinical epidemiologythe architecture of clinical research
(1985)An application of an urn model to the design of sequential controlled clinical trials
JASA
(1978)Minimizationa new method of assigning patients to treatment and control groups
Clin Pharmacol Ther
(1974)- et al.
Sequential treatment assignment with balancing for prognostic factors in the controlled clinical trial
Biometrics
(1975) - et al.
Log-linear models for the analysis of matched cohort data
Am J Epidemiol
(1989)
Cited by (12)
Gene expression profiling during intensive cardiovascular lifestyle modification: Relationships with vascular function and weight loss
2015, Genomics DataCitation Excerpt :Exclusion criteria were: 1) < 21 years of age, 2) presence of unstable coronary syndromes, refractory congestive heart failure, uncontrolled arrhythmia, or high-grade uncorrected cardiac conduction abnormalities, 3) significant left main stenosis (> 50%) and ejection fraction < 35% in patients who did not have revascularization or were not candidates for revascularization, 4) hypotensive response to exercise, 5) known history of autoimmune disease or systemic/chronic disease requiring chemotherapy or long term treatment, 6) history of substance abuse (including alcohol) without self-certification of abstinence for at least three months, and 7) physical disabilities or medical conditions that would preclude program adherence. Non-intervention controls were recruited prospectively and matched to program participants based on age (within ± 5 years), gender, and disease status (presence of CAD or diabetes mellitus) [1]. Control subjects received only standard care from their primary care physician, did not receive any advice, counseling, or information regarding healthy lifestyle behaviors, and did not participate in any component of the lifestyle intervention.
Cardiometabolic risk reduction in an intensive cardiovascular health program
2013, Nutrition, Metabolism and Cardiovascular DiseasesCitation Excerpt :Physician approval, motivation to commit to following the guidelines of the program, and successful abstinence from smoking for at least three months prior to enrollment also were part of the acceptance criteria. Controls (n = 76 white men and women) were matched to participants based on gender, age at baseline within five years, and CAD status (overt CAD or risk factors) using a prospective individual matching strategy to achieve a balanced distribution of risk factors between intervention participants and controls in nonrandomized clinical trials [15]. Controls receiving only standard care from their primary physicians underwent identical examinations at baseline, 12 weeks, and 52 weeks, but did not participate in the program or receive healthy lifestyle information.
Adaptive propensity score procedure improves matching in prospective observational trials
2019, BMC Medical Research MethodologyIntensive cardiovascular risk reduction induces sustainable changes in expression of genes and pathways important to vascular function
2014, Circulation: Cardiovascular GeneticsRandomization versus matching method in resistance training research: A simulation study based on maximal isometric squat strength recovery data
2014, Isokinetics and Exercise Science