Randomized, placebo-controlled, parallel group versus crossover study designs for the study of dementia in Parkinson's disease

Abstract

In studies of dementia, crossover designs are controversial, reflecting concerns about temporal stability of disease, confounding of treatment effects with period by treatment interactions and/or carryover effects. Carryover effects are differences in the lingering effect of treatments (placebo) into subsequent periods. In the context of a trial to study the effect of donepezil on dementia in patients with Parkinson's disease, we examine two-sequence crossover studies with two or four periods, and a four-sequence design with two periods. We quantify bias in estimated treatment effects due to carryover effects and explore the use of biased estimators in hypothesis testing. For hypothesis testing, type I error rates are valid if (1) repeated administration of treatment alters the outcome only for effective treatments and (2) carryover effects due to placebo following treatment periods are nonzero only for effective treatments. For crossover and parallel group designs, sample sizes are adjusted for reduced statistical power due to carryover effects and temporal changes in variance. For the proposed clinical study, we estimate that a single-period parallel group design with baselines would require 104 patients and take about 23 months to complete. A two-sequence, four-period parallel group design with baselines would require about 80 patients and about 20 months to complete. We conservatively assume a carryover effect of 50% of the treatment effect for a two-sequence four-period crossover design. The estimated treatment effect for this model may underestimate the true treatment effect by up to 13%. The sample size/study length requirements are 28 patients or 12.4 months, respectively, a substantial saving over either parallel group design. The cost of allowing for carryover in the sample size calculation is about 1.2 months of study time.

Introduction

In addition to motor dysfunction, patients with advanced Parkinson's disease frequently develop dementia. Therapies to treat dementia in patients with Parkinson's disease are in an early stage [1]. Because the physiology of dementia in Alzheimer's and Parkinson's disease is similar, previous work in Alzheimer's disease should offer insights into types of therapeutic agents that may be effective in Parkinson's disease as well as suitable study designs 2, 3, 4. In Alzheimer's disease, cholinesterase inhibition dominates therapeutic strategies for treating dementia [5]. Donepezil is a centrally active, reversible acetylcholinesterase inhibitor approved for Alzheimer's disease that may also be effective with Parkinson's disease [6]. Tacrine, its precursor, was extensively studied for its effects on cognitive function, behavioral disturbance and functional autonomy. A recent meta-analysis and review of tacrine discuss 18 separate studies 5, 7. Of these, ten were crossover (CO) designs, seven were parallel group (PG) and one was a mixed PG/CO.

In PG designs, subjects are randomized to receive either active treatment (T) or placebo (P) for single or multiple periods. In CO designs, some or all subjects receive more than one treatment sequentially. For example, in a two-sequence, two-period CO design (PT:TP), subjects randomized to the first sequence (PT) receive P in the first measurement period and T in the second measurement period, while subjects on the second sequence (TP) receive T followed by P. Both Qizilbash et al. and Conway criticize CO designs, warning that they “harbor methodological problems” and that the most convincing studies demonstrating benefits of tacrine therapy in Alzheimer's disease are large-scale PG designs with long treatment phases 5, 7. Similarly, Leber, writing for the Food and Drug Administration in its draft guidelines for the clinical evaluation of antidementia drugs, recommends PG designs [8].

Recent studies of donepezil show improved cognitive function in Alzheimer's disease patients in both PG and CO studies 9, 10. Here we examine the suitability of PG and CO designs for clinical studies of dementia in patients with Parkinson's disease. In theory, CO designs are suitable for chronic diseases, with outcomes that are relatively stable over the study 11, 12. In particular, subjects cannot be “cured” by the treatments of interest, i.e., treatment cannot permanently alter the normal course of disease. CO designs are efficient because treatment effects are estimated, at least in part, from within-subject contrasts, typically requiring fewer subjects than PG designs where treatment effects reflect between-subject contrasts. However, treatment effects in CO designs are potentially confounded with nuisance parameters, such as period effects, or period by treatment interactions. For example, period effects are systematic changes in outcome that apply to all patients, due perhaps to temporal changes in disease or to the measurement instrument [13]. Period by treatment interactions occur when efficacy varies across periods, e.g., treatment is effective in the first but not the second period. Carryover effects, differences in the lingering effect of a treatment administered in one period into the subsequent period, can bias the estimated treatment effect. Thus, PT:TP designs may be recommended only when carryover effects can be eliminated, e.g., by using a washout period between the two treatments 11, 14.

Here we compare several randomized PG and CO designs. Our motivation, illustrated in the proposed study section, is a placebo-controlled trial examining effects of donepezil therapy on cognitive impairment in patients suffering from dementia. The study design and statistical model section introduces the statistical model, uses several designs to illustrate period effects, treatment×period interactions and carryover effects and discusses a two-stage method of analysis. The section on extensions to other designs examines an alternative, but potentially biased, approach and illustrates sample size calculations that are adjusted for losses in power due to carryover, treatment×period interactions and temporal changes in variance.