Statistical Methods in Clinical Trials Ziad Taib Biostatistics AstraZeneca February 22, 2012 Types of Data Continuous Blood pressure Time to event Categorical sex quantitative qualitative Discrete No of relapses Ordered Categorical Pain level Types of data analysis (Inference) Parametric Vs Non parametric Model based Vs Data driven Frequentist Vs Bayesian Steps in data analysis and presentation of results • • • • Data cleaning. Descriptive analyses. Statistical Inference. Reporting and presenting the results. Continuous Data Part I Classical Inference Part II Bayesian Inference Part I Classical Inference 1. 2. 3. 4. One sample Paired data Two samples Many samples » One way anova » Two way anova » Ancova One sample From a population of patients we draw a sample of size n of patients and measure some response on each of these. We want to test whether the average response in this population is equal to (or is larger or smaller than) some predefined value μ0 (e.g. the corresponding value in the healthy population). Estimation • Assume is a sample (i.i.d) of size n from some distribution F with mean μ and standard deviation σ. Then and • Are (i) unbiased and (ii) consistent estimators of μ and σ² i.e. (i) (ii) (LLN) Estimation in general • The maximum likelihood method • The moment method • Uniformly Minimum Variance Unbiased estimators • Least squares • etc Hypothesis testing We can e.g. test the null hypothesis against using the test statistic when n is large, T follows, under the null hypothesis, the standard normal distribution (CLT). Hypothesis testing For moderate values of n assuming follow a normal distribution, then follows, under the null hypothesis, the t (student) distribution with (n-1) degrees of freedom. The distribution of the average when n is large The true distribution among the patients (exponential with mean 0.5). Confidence intervals • A (1-α)100% confidence interval for μ has the form (i) when n is large (ii) when n moderate but sample normally distributed. A hypothesis (or a confidence interval) can be one sided or two sided 95% confidence interval The normal distribution 1-α=0.95 The t distribution 1-α=0.95 Correspondence Theorem Hypothesis testing and confidence intervals are two sides of the same coin! Reject null hypothesis (α) Parameter value according to null hypothesis not included in confidence interval (α) Paired data • Sometimes we measure a response at baseline and at the end of the treatment. A suitable analysis can be to consider the endpoint changes from baseline. • and to use the (one sample) test statistic Two independent samples From two populations/distributions we draw two samples of sizes n1 and n2 of patients and measure some response on each of these. We want to test whether the average responses, μ1 and μ2, in these two populations are equal. The population variances are denoted by σ1² and σ2² respectively. H 0 (null) 1 2 H A (alternative ) 1 2 1 2 1 2 1 2 1 2 Two samples (continued) • When both sample sizes are large we can use the central limit theorem according to which the test statistic below follows the standard normal distribution • When the sample sizes are moderate but normal with equal variances we can use the t-statistic (having n1+n2-2 d.f.) and the pooled variance. • The remaining case can be handled using a method known as Satterwaite’s confidence intervals. Many samples One way analysis of variance • Here we want to compare several (k) groups (treatments) with respect to the average response in each group. The following model is assumed Anova1 Anova1 • Main point: individuals only differ with respect to one criterion: group belonging. • Intuitively: Looking at the pictures we see that the variation says something about the difference between the cases where the groups are different and when they are similar. Anova1 continued • The null hypothesis and the alternative hypothesis can be formulated as • The deviation of an individual observation from the overall mean can be described by SST/SST SSA/SSB SSE/SSW Anova 1 continued • Manipulating these gives: MSE MSA MSA • When the null hypothesis is false we expect the ratio MSA/MSE to be large. To calculate exact significance levels we use the fact that it F-distributed with (k-1,N-k) d.f. Anova table Source of variation Sum of d.f. squares Mean F/p-value squares Treatment SSA K-1 MSA Error SSE N-k MSE Total SST N-1 F0=MSA/MSE /p • Example: 3 groups with the same mean 20 and the same standard deviation 2. We reject for large values of F0. F is F-distributed and F0 the value of MSA/MSE We cannot reject Source of variation Sum of squares d.f. Mean squares F0/p-value Treatment 9.56 2 4.78 1.22/ 0.29 Error 4671.1 1197 3.9 Total 4680.65 1199 • Example: 3 groups with means 20, 20 and 20.5 and the same standard deviation 3. We can reject Source of variation Sum of squares d.f. Mean squares F/pvalue Treatmen t 59.4 2 29.7 3.25/ 0.039 Error 10942.8 1197 9.14 Total 11002.3 1199 We can reject Multiple comparisons • When we reject the null hypothesis we only know that the groups are not equal but some of them might still be. To find out more we have to consider all the pairwise comparisons between the groups. With k groups this gives m=k(k-1)/2 such comparisons. How do we do this and still have a reasonable overall significance level? The simplest way to deal with this is using Bonferroni’s inequality. This implies that when performing m tests if each test is at the 1-α/m level then the tests taken simultaneously will be on the 1- α level. We will deal with problem in detail later on. Two way anova • Here we assume that individuals can differ with respect to two factors (two drugs, treatment and centre). The following model is usually suitable for this situation • As before the deviation of an individual observation from the overall average can be decomposed into terms related to the effects of treatment, center, treatment by center as well as to a pure random component. Anova2 • We use a similar notation as before • Tests of treatment effect, center effect or treatment by center interaction can be performed by using the appropriate ratio between mean square values. Anova2 • As an example assume we want to test if there is a treatment effect: • This can be tested using the test statistic • Which has under the null hypothesis an F distribution with (a-1) and ab(n-1) degrees of freedom. • The various tests are summarized in the following table Two-way anova Analysis of covariance • Here we assume that individuals can differ with respect baseline values. It is sometimes desirable to adjust the model for the endpoint measures Y so baseline values X are taken into account. The following model can then be useful The analysis uses an F distribution based on Mean sums of squares (cf. p.319) Various forms of models and relation between them Classical statistics (Observations are random, parameters are unknown constants) Mixed models, both random and constant parameters LM: Assumptions: 1. independence, 2. normality, 3. constant parameters LMM: Assumptions 1) and 3) are modified GLM: assumption 2) Exponential family Maximum likelihood LM - Linear model Repeated measures: Assumptions 1) and 3) are modified GLMM: Assumption 2) Exponential family and assumptions 1) and 3) are modified Longitudinal data Non-linear models GLM - Generalised linear model LMM - Linear mixed model GLMM - Generalised linear mixed model Bayesian statistics, parameters are random 39 The Linear Mixed-effects Model • The linear mixed effects model is quite flexible and does not need balance, independence etc. Usually some version of maximum l likelihood is used for the inference Average evolution Subject specific 40 Part II Bayesian Inference What is a probability? 1. the limit of a relative frequency Problem: Not all events all repeatable! 2. the degree of plausibility Or of belief Problem: subjective? P=? P=1/6 Thomas Bayes (1702 to 1761) Bayes in brief • Bayes’ conclusions were accepted by Laplace in 1781 • Rediscovered by Condorcet, and remained unchallenged until • Boole questioned them. • Since then Bayes' techniques have been subject to controversy. Is Classical Inference Really Classical? • Only in the sense that it started a realtively long time ago with Fisher around 1920. • Bayesian Statistics is even more classical since it was dominating until then. • Classical inference was introduced as a way to introduce objectivity in the scientific process. Simplified scientific Process 1. A theory or a hypothesis needs to be tested 2. We perform an experience and obtain some data 3. Are our data in agreement with our theory? 4. If the answer to the above question is no, we reject the theory. Otherwise we cannot reject it! Classical Paradigm vs The Bayesian paradigm • The classical paradigm is based on the consideration of P[ Data | Theory ] (1) • How likely is the data if the theory was to be true? Classical vs Bayesian (cont’d) • The Bayesian paradigm is based on the consideration of P[ Theory | Data] (2) • How much support or belief is there in the theory given the data? Bayes’ Formula States that P [ T | D] P [ D | T ]P [ T ] P [ D] (3) Where T = Theory and D = Data Simple formula with many interesting implications. Implications of Bayes’ Formula 1. (1) and (2) are not equivalent. 2. To work out (2) we have to estimate P[ T ] i.e. we need to put a probability on our belief in the theory we are testing. 3. We cannot make P[ T ] disappear. (Similar to the uncertainty principle in quantum mechanics?) Is subjective always bad? In our everyday life we are far from being always objective. Not even when it comes to scientific issues. Nevertheless we are afraid of incorporating subjectivity into the scientific process. This is the reason why Bayesian inference is perceived as being controversial. How to deal with subjectivity? 1. Sometimes we do have some apriori information (earlier studies, similar drugs, expert opinions etc). 2. But even if we do not have any prior knowledge, we can study the effect of having various degrees of belief . (D. Spiegelhalter) 3. We can also invert the problem and ask what prior belief in the theory is required to make our data coherent with that theory? How to deal … 4. The role of the apriori information becomes negligible as data accumulates. It can be shown mathematically that ultimately all investigators will reach the same conclusion regardless of what prior information they had from the beginning. (for a mathematical argument cf. O’hagan, 1994). In the meanwhile the sceptic will just need more time/convincing/data! A textbook example • Let X be a sample from N(,s) of size n. • Let the apriori distribution p() be N(,) which is the so called conjugate distribution. • Then the aposterior distribution is of the following form p ( | x) P ( x | ) p ( ) p (t )P ( x | t )dt n N (x (1 ) , 2 ) 2 s n 2 n 2 2 s n 2 Comments • There are problems in choosing the apriori distribution. • The importance of the apriori distribution decreases as the sample size increases. Example: p-values • Some toxin is associated with certain symptoms. denotes the toxin level in patients having the symptoms and 0 the toxin level in healthy individuals. We want to test if there is a difference. A test can be based on a sample of size n through Under the null hypothesis H 0: = 0, z is an observation from a t-distribution with n-1 degrees of freedom. If, moreover, the p-value is less than 0.05 it is then customary to consider the result as significant, i.e. we reject the null hypothesis. We return to this in moment! Preliminaries • Consider two Hypothesis H0 and H1 P [ D | H0 ] P [ H0 ] P [ H 0 | D] P [ D] P [ D | H1 ] P [ H1 ] P [ H1 | D] P [ D] • We consider the odds ratio between the two P[ H 0 | D] P[ H 0 ] P[ D | H 0 ] P[ H1 | D] P[ H1 ] P[ D | H1 ] Odds ratio = prior odds x likelihood ratio New old Bayes factor Result 1- P [ H0 ] P [ H 0 | D] 1 P [ H 0 ] BF where P[D| H 0 ] BF P[D| H1 ] 1 Back to the t-test P. M. Lee Bayesian Statistics: An Introduction 2nd Ed. London: Arnold, p. 131 (1997) shows that in the case of the t-est and under quite general conditions: BF e z2 2 Consequences • Assume that there is no agreement on the effect of the toxin so we take P(H0)=0.5. Then: 1 P[H0 | D] 1 -z 2 2 e 1 1 2 -z 1 e 2 • Assume further that the value of z in the experiment turns out to be 2. Since this leads to a p-value of 0.044 we conclude that the result is significant at the 0.05 level. Setting z=2 in the formula above leads to: -2 P[H0 | D] 1 e 2 2 1 0.12 P[ H0| D] Z=2 P[ H0] P[ H0] P[ H0| D] Conclusions • By introducing the element of degree of belief about a theory, we arrive at conclusions that do not agree with those obtained using the frequentist approach, i.e. prior knowledge matters • Prior knowledge is part of the Bayesian approach. Backup slides Bayesian Analysis in Clinical Trials Number of Bayes Articles in Medical Journals In short, Bayesian inference provides a coherent, comprehensive and strikingly intuitive alternative to the flawed frequentist methods of statistical inference. It leads to results that are more easily interpreted , more useful, and which more accurately reflect the way science actually proceeds. It is, moreover, unique in its ability to deal explicitly and reliably with the provably ineluctable presence of subjectivity in science. These features alone should motivate many working scientists to find out more about Bayesian inference in their own research. Robert A.J. Mathhews (1998) The use and abuse of subjectivity in scientific research – Part 2. Working paper for the European Science and Environment Forum What for? • The greatest virtue of frequentist clinical trials is extreme rigor and focus on the experiment at hand. • But they tend to become large and expensive and some patients receive inferior experimental therapies. • The Bayesian approach is gaining terrain and has more and more advocates hoping to address these and other challenges Bayesian Framework • The Advantages: – – – – Formal system for incorporating existing information Natural approach to inference Generally more efficient Well suited for decision making • The Challenges: – – – – Determining appropriate prior probabilities Computational complexity Lack of familiarity Lack of software tools What is the difference? Approach Traditional Bayesian Question How likely are the trial results given there really is no difference among treatments? How likely is it that there is a true difference among treatments given the trial data? Approval based on Pivotal trial (Hypothesis testing) Weight of evidence (revising beliefs in light of new evidence) Design Single stage Adaptive In what settings have Bayesian desings been used? • • • • • • Pharmacokinetics Phase I (CRM) Phase II (Thall/Simon) Phase III (Spiegelhalter, Parmar, others) Meta-analysis Medical device clinical trials (guidelines from 2006) Summary of applications • Methods for incomplete data • Adaptive designs (including adaptive sample size) • Confirmatory trials • Evaluating a modified version of an approaved product • Synthesising data in post market surveillance Recommended Reading 1. 2. 3. 4. Berry DA (2006). Bayesian clinical trials. Nature Reviews: Drug Discovery. 5: 27-36; 2006. Goodman, S.N. Toward Evidence-Based Medical Statistics: The P Value Fallacy. Annals of Internal Medicine. 130:9951021; 1999. Spiegelhalter DJ, Keith R, Abrams KR, Myles JP. Bayesian Approaches to Clinical Trials and Health-Care Evaluation. John Wiley & Sons, Ltd. 2004. Winkler, R.L. Why Bayesian Analysis Hasn’t Caught on in Healthcare Decision Making. International Journal of Technology Assessment in Health Care. 17:1, 56-66; 2001. Questions or Comments? Karl Popper (1902-1994) A theory can be called scientific if it can be falsified! We recognise this pirnciple from the foundation of statistical hypothesis testing . Falsifiability (or refutability or testability) is the logical possibility that an assertion can be shown false by an observation or a physical experiment. For example, "all men are mortal" is unfalsifiable, since no amount of observation could ever demonstrate its falsehood. "All men are immortal," by contrast, is falsifiable, by the presentation of just one dead man A white mute swan, common to Eurasia and North America. Two black swans, native to Australia Proof P [ H 0 | D] P [ H 0 ] P [D | H 0 ] P [ H1 | D] P [ H1 ] P [D | H1 ] P [ H 0 ] P [D | H 0 ] P [ H 0 | D] P [ H1 | D] P [ H1 ] P [D | H1 ] C 1 - P [ H 0 | D]) P [ H 0 | D] C P [ H 0 | D] C C 1 C P [ H 0 | D] 1 C C 1- P [ H0 ] 1 P [D | H 0 ] P [ H0 ] P [D | H ] 1 1 1 1- P [ H0 ] P [ H 0 | D] 1 P [ H 0 ] BF 1 1 1 C 1