ICT Clinical Trial Management Sample Sizing Alan Phillips of ICON Clinical Research investigates strategies for reducing patient numbers in clinical trials during the design phase With increased pressure to design more efficient and effective clinical trials in order to speed up the drug development process, it is imperative that strategies for reducing the number of patients are considered during the design phase of a clinical trial. A large amount of research (1-7) has been undertaken on the problem of sample size estimation for different trial designs. However, in practice the methods are rarely used and sample sizes are usually based on simple formulae derived for the comparison of two means or proportions. After reviewing how sample sizes are derived, strategies for reducing the number of patients needed in clinical studies will be discussed; for example: N N N N N Improving the precision of estimates without increasing patient numbers Expanding the treatment effect to be detected Implementing an adaptive sample size re-estimation strategy The use of one-sided tests in early phase development Incorporating study design features and data structures when estimating patient numbers This is where σ is the anticipated standard deviation of the outcome variable which is assumed to be equal in both populations, and f (α, β) is a function of the type I and II errors. For a two-sided five per cent significance level and type II level of 0.2, the formula approximates to: 16σ 2 n = -----δ2 Referring to the example, 64 patients per group are needed to detect a difference of 5mm Hg, assuming a standard deviation of 10. When the outcome is nominal, δ and σ are replaced by functions of π1 and π2 – the proportions expected to respond under the two treatments. MINIMISING THE NUMBER OF PATIENTS Now we know how the number of patients is determined, we can consider strategies for reducing the number required. These are discussed in more detail below. Use of Quantitative Measurements It will be shown that matching the sample size estimation process to the design and analysis strategy often reduces the number of patients in a clinical trial. BASICS OF SAMPLE SIZE ESTIMATION Sample size estimates for a clinical trial depend on several factors, including: δ : The clinically meaningful difference to be detected α : The type I error or probability of detecting a significant difference when the treatments are equally effective β : The type II error or the probability of not detecting a significant difference when there is a difference of magnitude δ and given level of statistical significance Consider a randomised, parallel group, double-blind study to compare a new anti-hypertensive agent against placebo in hypertensive patients. Suppose that the primary efficacy variable is the change in diastolic blood pressure, which is assumed to be continuous and normally distributed. Then using a two-tailed test, the required number of patients per treatment is given by: 2σ 2 n = ------ x f (α , β) δ2 10 ICT l www.samedanltd.com As discussed by Pocock (6), there are basically three types of response data in clinical trials: 1. 2. 3. Quantitative Qualitative Time to response Statistical analysis methods that incorporate the use of quantitative measurements such as diastolic blood pressure are more powerful in mathematical terms than methods for qualitative data, given the same sample size. Subsequently, quantitative measurements should be used whenever possible in order to minimise the sample size. For example, simply to classify hypertensive patients into responders and nonresponders according to diastolic blood pressure < or ≥ 90mm Hg does not fully utilise the data, and consequently requires more patients compared to analysing diastolic blood pressure values. In some clinical trials, the primary endpoint is set in terms of time to an event such as death. In such trials it is not uncommon to find that not all patients have experienced the event and so one can not use time-to-death as a quantitative measurement. However methods are available that account for the censored nature of the data. These methods tend to be more powerful than converting the data to a qualitative variable such as ‘yes/no’. Improve the Precision of Estimates Expand the Treatment Effect to be Detected In some types of clinical studies, precision can be improved by simply repeating the measurement and incorporating the additional data into the analysis strategy; for example, multiple blood pressure readings at baseline. Frison and Pocock (5) investigated how the sample size can be reduced when the measurements are repeated multiple times at baseline and at follow up assessments. Assuming measurements are equally correlated, they showed that the number of patients can be reduced by a factor of: This is where: From the perspective of minimising the number of patients, the most important parameter when determining the number of patients is the effect size to be detected. In most clinical trials the effect size is set to be the clinically meaningful difference (δ). However, in some scenarios, using the clinically meaningful difference can be unnecessarily small. Significant savings in patient numbers can be achieved by expanding the difference to be detected. Such approaches can be justified when the expected therapeutic benefit of the drug under investigation is greater than the clinically meaningful difference or in preliminary studies when the objective is to justify a more ambitious study in a larger number of patients with a smaller meaningful difference. p = Number of baseline measures Implement an Adaptive Sample Size Re-estimation Strategy 1 + (r - 1) ρ [ ---------------r - p ρ2 ------------------- ] 1 + (p - 1) ρ r = Number of follow up measures ρ = Correlation between pairs of assessments Based on a survey of a variety of clinical trials, Frison and Pocock suggest ρ = 0.65 is unlikely to lead to major errors. Figure 1 shows the percentage reduction in patient numbers that can be achieved by repeating follow up measurements when no baseline measures are taken, assuming ρ = 0.65, α = 0.05 and β = 0.2. Figure 2 shows the percentage reduction in patient numbers that can be achieved by including additional baseline assessments compared with a trial with a single baseline and follow up measure. As you can see from the figures, repeating measurements can have an effect on the sample size. The relative reduction in sample size reduces rapidly beyond about two or three measures. Figure 1: Percentage reduction in patient numbers Percentage reduction 30 25 20 15 10 5 0 1 2 3 Percentage reduction 18 16 14 12 10 8 6 4 2 0 2 3 4 12 ICT l www.samedanltd.com In drug development, two-tailed tests are commonly performed whatever the comparator being studied. That is, it is not uncommon to find that a clinical trial is designed to test a hypothesis of no difference between the treatments under investigation against an alternative hypothesis that the treatment of interest has some effect, either positive or negative. In some studies, especially during the early phases of drug development, there is little value is investigating whether a drug exhibits negative effects. Thus the real alternative hypotheses of interest should be that the treatment of interest has some positive effect, which leads to one-tailed tests. Using one-sided testing can reduce the sample sizes by around 20 per cent for the commonly used type II error values of 0.2 and 0.1. However, the use of one-sided tests should be justified prospectively in the protocol. Even though formulae exist for different types of trial design, they are rarely used in practice. In most trials, sample size estimates are usually based on simple formulae derived for the comparison of two means or two binominal proportions. However, significant reductions can be achieved in sample size estimates by taking into account the design of the trial and the structure of the data being collected. Figure 2: Percentage reduction in patient numbers Number of baseline assessments One-Sided Tests Use of Study Design Features and Data Structure 4 Number of follow up assessments 1 An alternative strategy is to employ an adaptive design; that is, include an interim analysis where the primary purpose is sample size-re estimation. Although very controversial in confirmatory trials, sample size re-estimation looking at the observed treatment effect may be an appropriate strategy in early phase development. If employed, the study may need to include appropriate stopping rules. Although methods are available that protect the type I error (8-10), they do not seem to be universally accepted. Even taking into account the controversies, when planned correctly sample size re-estimation remains a potential strategy for reducing the number of patients in many studies. 5 To illustrate the percentage savings that can be achieved, consider a randomised, double-blind, parallel group clinical trial to evaluate a new anti-inflammatory against a standard treatment in patients with degenerative joint disease of the knee. Suppose that the primary end-point is the investigator’s clinical improvement assessment, rated using a five-point scale (2 = Much better; 1 = Better; 0 = No change; -1 = Worse; -2 = Much worse). Often in Probabilities such studies, the sample size is based on categorising the Figure 3: Expected response probabilities data into a binary response. If the data are to be analysed by the proportional odds model of McCullagh (11), then 0.4 0.35 for such data a more appropriate method for determining 0.3 0.25 the number of patients needed per group was derived by Comparator 0.2 New drug Whitehead (12). Suppose that the expected response 0.15 0.1 probabilities for each treatment is given in Figure 3. 0.05 0 Then, in this instance, at least a 25 per cent reduction in Much better Better No change Worse Much worse patient numbers can be achieved by taking into account the trial design and expected response probabilities, employed such as focusing on higher risk patients, using a run-in compared with basing the sample size on the comparison of two phase before randomisation to reduce variability, running trials binominal proportions. for longer periods and so on. The method of calculating the required sample size should always take into account the final Dose-ranging trials is another example where the number of design of the trial. It is often the case that matching the sample patients needed can often be reduced because the study design size estimation process to the design and analysis strategy often features are not included when estimating the sample size. In reduces the number of patients in a clinical trial. As always, such trials the objective of the study is usually to assess the seeking professional advice from qualified statisticians can help relationship between several doses of the same treatment. to ensure success and reduce patient numbers. N When more than two treatment groups are being compared, it is important that the hypotheses of interest are clearly stated; References for example, when three treatments are compared, different sample sizes are needed, say, to detect a difference between 1. Armitage P and Berry G, Statistical methods in medical Research, 3rd ed Blackwell, Oxford, p195, 1994 each pair-wise comparison than to show that one of the treatments differ from the others. Although formulae do not 2. Bland JM, An Introduction to medical Statistics, 2nd ed exist for determining the sample size for a large number of the Oxford medical Publications, Oxford, pp331-341, 1995 statistical methods used for analysing dose ranging data such 3. Machin D and Campbell MJ, Statistical tables for the as William’s test (13,14) or Joncheere’s test (15), simulation design of clinical trials, Blackwell Scientific Publications, Oxford, 1987 can be used to determine the sample size. Simulation also allows exploration of different drop-out rates, for example. 4. Fleiss JL, The design and analysis of clinical experiments, Wiley, Chichester, pp371-376, 1986 CONCLUSION In a large number of clinical trials, sample size estimates are usually based on simple formulae derived for the comparison of two means or binomial proportions. Consequently, in some clinical trials more patients are recruited than necessary. With the increased pressure to design more efficient and effective clinical trials in order to speed up the drug development process, it is imperative that strategies for reducing the number of patients are considered during the design phase of a clinical trial. As well as the strategies discussed in this article, other strategies can be About the author Alan Phillips is Senior Director Biostatistics, Europe and the Rest of the World, at ICON Clinical Research. Prior to joining ICON, Alan worked at Wyeth Pharmaceuticals, where his last position was Senior Director Biostatistics at Wyeth Research’s worldwide headquarters in the US. His statistical interests include biostatistical guidelines for regulatory submissions, clinical trials and adaptive designs. Alan was a member of the Statisticians in Pharmaceutical Industry main committee from 1993 to 1996, the scientific sub-committee from 1995 to 1996, and the training sub-committee from 1991 to 1994. He has belonged to the Association of British Pharmaceutical Industry Statistics Task Force/PSI Regulatory sub-committee since 1998. Email: phillipsa@iconuk.com 14 ICT l www.samedanltd.com 5. Frison L and Pocock S, Repeated measures in clinical trials: analysis using mean summary statistics and its implications for design, Statistics in Medicine, 11, pp1,685-1,704, 1992 6. Pocock SJ, Clinical Trials: A Practical Approach, Wiley, Chichester, pp128-229, 1983 7. Campbell MJ, Julious SAJ and Altman DG, Estimating sample sizes for binary, ordered categorical, and continuous outcomes in two group comparisons, Br Med J 311, pp1,145-1,148, 1995 8. Bauer P and Kohne K, Evaluation of experiments with adaptive interim analyses, Biometrics, 50, pp1,029-1,041, 1994 (Correction Biometrics; 52: p380, 1996) 9. Proschan MA and Hunsberger SA, Designed extension of studies based on conditional power, Biometrics, 51, pp1,315-1,324, 1995 10. Li G, Weichung JS and Tailiang X, A sample size adjustment procedure based on conditional power, Biostatistics, 3, pp277-287, 2002 11. McCullagh P, Regression models for ordinal data, J Roy Statist Soc B, 43, pp109-142, 1980 12. Whitehead J, Sample size calculations for ordered categorical data, Statistics in Medicine, 12, pp2,255-2,271, 1993 13. William DA, A test for differences between treatment means when several dose levels are compared with zero dose control, Biometrics, 27, pp103-177, 1971 14. Williams DA, The comparison of several dose levels with a zero dose control, Biometrics, 28, pp519-531, 1972 15. Jonckheere AR, A distribution-free K sample test against ordered alternatives, Biometrika, 41, pp133-145, 1954
© Copyright 2024