What can we learn from data ? A comparison of direct, indirect and observational evidence on treatment efficacy 2 nd Workshop 'Consensus working group.

Slides:



Advertisements
Similar presentations
A small taste of inferential statistics
Advertisements

Hypothesis Testing Goal: Make statement(s) regarding unknown population parameter values based on sample data Elements of a hypothesis test: Null hypothesis.
How would you explain the smoking paradox. Smokers fair better after an infarction in hospital than non-smokers. This apparently disagrees with the view.
Properties of Least Squares Regression Coefficients
Evidence synthesis of competing interventions when there is inconsistency in how effectiveness outcomes are measured across studies Nicola Cooper Centre.
The potential role of mixed treatment comparisons Deborah Caldwell Tony Ades MRC HSRC University of Bristol.
Previous Lecture: Distributions. Introduction to Biostatistics and Bioinformatics Estimation I This Lecture By Judy Zhong Assistant Professor Division.
Irwin/McGraw-Hill © Andrew F. Siegel, 1997 and l Chapter 12 l Multiple Regression: Predicting One Factor from Several Others.
Testing Theories: Three Reasons Why Data Might not Match the Theory.
1 G Lect 2a G Lecture 2a Thinking about variability Samples and variability Null hypothesis testing.
Segment 4 Sampling Distributions - or - Statistics is really just a guessing game George Howard.
Chapter 19 Confidence Intervals for Proportions.
Chapter 4 Multiple Regression.
Evaluating Hypotheses
How Science Works Glossary AS Level. Accuracy An accurate measurement is one which is close to the true value.
Lehrstuhl für Informatik 2 Gabriella Kókai: Maschine Learning 1 Evaluating Hypotheses.
Review for Exam 2 Some important themes from Chapters 6-9 Chap. 6. Significance Tests Chap. 7: Comparing Two Groups Chap. 8: Contingency Tables (Categorical.
Sample Size Determination Ziad Taib March 7, 2014.
DISCUSSION Alex Sutton Centre for Biostatistics & Genetic Epidemiology, University of Leicester.
Standard Error of the Mean
Standard Error and Research Methods
Statistics 11 Hypothesis Testing Discover the relationships that exist between events/things Accomplished by: Asking questions Getting answers In accord.
1 1 Slide Statistical Inference n We have used probability to model the uncertainty observed in real life situations. n We can also the tools of probability.
Inference in practice BPS chapter 16 © 2006 W.H. Freeman and Company.
The problem of sampling error It is often the case—especially when making point predictions—that what we observe differs from what our theory predicts.
Modeling errors in physical activity data Sarah Nusser Department of Statistics and Center for Survey Statistics and Methodology Iowa State University.
1 Today Null and alternative hypotheses 1- and 2-tailed tests Regions of rejection Sampling distributions The Central Limit Theorem Standard errors z-tests.
Testing Theories: Three Reasons Why Data Might not Match the Theory Psych 437.
CHAPTER 16: Inference in Practice. Chapter 16 Concepts 2  Conditions for Inference in Practice  Cautions About Confidence Intervals  Cautions About.
Health and Disease in Populations 2001 Sources of variation (2) Jane Hutton (Paul Burton)
Chapter 8 Introduction to Hypothesis Testing
 Is there a comparison? ◦ Are the groups really comparable?  Are the differences being reported real? ◦ Are they worth reporting? ◦ How much confidence.
Physics 270 – Experimental Physics. Standard Deviation of the Mean (Standard Error) When we report the average value of n measurements, the uncertainty.
Simon Thornley Meta-analysis: pooling study results.
Exam Exam starts two weeks from today. Amusing Statistics Use what you know about normal distributions to evaluate this finding: The study, published.
Biostatistics Case Studies 2008 Peter D. Christenson Biostatistician Session 5: Choices for Longitudinal Data Analysis.
Economic evaluation of health programmes Department of Epidemiology, Biostatistics and Occupational Health Class no. 19: Economic Evaluation using Patient-Level.
Testing Theories: The Problem of Sampling Error. The problem of sampling error It is often the case—especially when making point predictions—that what.
University of Warwick, Department of Sociology, 2014/15 SO 201: SSAASS (Surveys and Statistics) (Richard Lampard) Week 7 Logistic Regression I.
MGS3100_04.ppt/Sep 29, 2015/Page 1 Georgia State University - Confidential MGS 3100 Business Analysis Regression Sep 29 and 30, 2015.
Application 3: Estimating the Effect of Education on Earnings Methods of Economic Investigation Lecture 9 1.
Copyright © 2012 Pearson Education. All rights reserved © 2010 Pearson Education Copyright © 2012 Pearson Education. All rights reserved. Chapter.
Stat 112: Notes 2 Today’s class: Section 3.3. –Full description of simple linear regression model. –Checking the assumptions of the simple linear regression.
Chapter 5 Parameter estimation. What is sample inference? Distinguish between managerial & financial accounting. Understand how managers can use accounting.
통계적 추론 (Statistical Inference) 삼성생명과학연구소 통계지원팀 김선우 1.
Correlation Assume you have two measurements, x and y, on a set of objects, and would like to know if x and y are related. If they are directly related,
1 Comparing multiple tests for separating populations Juliet Popper Shaffer Paper presented at the Fifth International Conference on Multiple Comparisons,
Hypothesis Testing An understanding of the method of hypothesis testing is essential for understanding how both the natural and social sciences advance.
JS Mrunalini Lecturer RAKMHSU Data Collection Considerations: Validity, Reliability, Generalizability, and Ethics.
Simple examples of the Bayesian approach For proportions and means.
Health and Disease in Populations 2002 Sources of variation (1) Paul Burton! Jane Hutton.
Education 793 Class Notes Inference and Hypothesis Testing Using the Normal Distribution 8 October 2003.
Sampling Design and Analysis MTH 494 Lecture-21 Ossam Chohan Assistant Professor CIIT Abbottabad.
Q2010 Special session 34 Data quality and inference under register information Discussion by Carl-Erik Särndal.
Copyright © Cengage Learning. All rights reserved. 5 Joint Probability Distributions and Random Samples.
A Quantitative Overview to Gene Expression Profiling in Animal Genetics Armidale Animal Breeding Summer Course, UNE, Feb Analysis of (cDNA) Microarray.
The inference and accuracy We learned how to estimate the probability that the percentage of some subjects in the sample would be in a given interval by.
The accuracy of averages We learned how to make inference from the sample to the population: Counting the percentages. Here we begin to learn how to make.
Statistical Inference for the Mean Objectives: (Chapter 8&9, DeCoursey) -To understand the terms variance and standard error of a sample mean, Null Hypothesis,
1 The effect of regression towards the mean in assessing crime reduction interventions using non-randomised trials Campbell Collaboration Colloquium London.
1 Basics of Inferential Statistics Mark A. Weaver, PhD Family Health International Office of AIDS Research, NIH ICSSC, FHI Lucknow, India, March 2010.
Statistics in Clinical Trials: Key Concepts
Stat 2411 Statistical Methods
Statistical Data Analysis
PSY 626: Bayesian Statistics for Psychological Science
The Scientific Method Unit 1.
PSY 626: Bayesian Statistics for Psychological Science
Statistical Data Analysis
Chapter 7: The Normality Assumption and Inference with OLS
MGS 3100 Business Analysis Regression Feb 18, 2016
Presentation transcript:

What can we learn from data ? A comparison of direct, indirect and observational evidence on treatment efficacy 2 nd Workshop 'Consensus working group on the use of evidence in economic decision models‘ Department of Health Sciences, University of Leicester September 26, 2005 Tony Ades, Debbi Caldwell MRC Health Services Research Collaboration, Bristol

Outline of presentation Introduction: learning about parameters … Fixed effect models Direct data, Indirect data, Observational data: one new study, a meta-analysis of observational data Random effect models What to learn about: mean, variance, new or old groups ? Direct and indirect data in RE models Observational evidence … And Surrogate end-points

Why might this be useful? 1. A “standard” systematic review is carried out. Has all relevant data been included ?” 2. Data is relevant if it reduces uncertainty ….. how effective might different kinds of data be in reducing uncertainty ? 3. Synthesis agenda  Research prioritisation agenda. Why collect more data if you don’t know what you can learn from it ? 4. …. a scientific basis for “Hierarchy of evidence” ?

Data tells us nothing unless there is a model 1.You must have something to learn about: a parameter 2.If you know what you are going to learn about, you must know how much you already know about it: a prior distribution 3.There must be a relationship between what the data estimates and the parameter: a model

… it’s partly a language thing 1.Need to distinguish between data, parameters and estimates. Terms like ‘Log Odds Ratio’ tend to get used as if these were all the same thing. 2.Meta-analysis gives a “summary”. Summary of what? …data, literature, estimates? No “summary” without a model … 3. “evidence” => MODEL => “medicine”

FIXED EFFECT MODEL LOR Parameter  Its prior distribution  ~ N(  0,  0 2 ) LOR data from an RCT Y, with standard error S The model Y ~ N( , S 2 ) FE : data estimates exactly the parameter we want. Uncertainty in prior Uncertainty in data

Method RCT gives DIRECT information on the parameter of interest. Strategy: how much can Indirect Comparisons, Observational Data, etc tell us about the parameter  … …. COMPARED TO the same amount of direct RCT evidence. Use standard deviation S as a measure of “information”

Scale of the day : Log Odds Ratios 1.LORs for Treatment Effects usually well within the range –1 to +1 … corresponding to Odds Ratios 0.4 to And usually in range -0.5 to +0.5 … corresponding to OR : 0.6 to We need to think of uncertainty on this scale. values of  0 or S > 1 are HIGH, <0.25 LOW

…the more you know, the less there is to learn 1.If prior uncertainty is large (  0 high), posterior uncertainty is dominated by the amount of new data – ie by S 2.If prior uncertainty is already low, only a large amount of new data (S low) will make a difference.

Indirect RCT evidence on parameter  AB Target parameter  AB ~ some prior Model for Indirect evidence: Y AC ~ N(  AC, S AC 2 ), Y BC ~ N(  BC, S BC 2 )  BC =  AC -  AB IF S AC = S BC = S, then indirect evidence is equivalent to direct data with sd =  2S = S

… the weakest link BUT, the contribution of indirect comparisons depends on the weakest link: If S AC is high (weak evidence on AC), the contribution to  AB is small, no matter how much evidence there is on BC (ie no matter how low S BC ) …. don’t do big literature search on BC, if you know there is little evidence on AC (unless you are also interested in treatment C !!!)

Multiple indirect comparisons Debbi Caldwell’s presentation: Contribution to  AB relative to a direct RCT of size S via Y AC, Y BC - ONE indirect comparator => S and Y AD, Y BD TWO S … etc THREE 0.82 S … etc FOUR 0.71 S … etc FIVE 0.63 S

Observational data: one new study Observational data is biased: it does not give us an direct estimate of . Instead : Y OBS ~ N(  + , S OBS 2 ) … in any given case we don’t know how big the bias  is, or its direction …”unpredictability” Let’s have a distribution.. perhaps..  ~ N(  B,  B 2 ) … to describe our views about  ?

Prior distribution for bias  ~ N(  B,  B 2 )….. 1.As a ‘first cut’, suppose  B = 0 is our “best guess” 2.For  B … how small / big might the bias be ? an OR of 1.1 either way seems rather optimistic an OR of 1.6 either way seems rather pessimistic 3. … assume these represent 95% credible limits on the amount of bias … (in a “typical” single Observational study) …for example

In which: bias is on average 0.28 on log scale (either way)

 B =0

Some shortcomings of this analysis 1.Assumes that bias is not related to the true . Maybe larger  => larger bias ? … this could be modelled too.. 2. What is our belief about the “AVERAGE BIAS in OBSERVATIONAL STUDIES” :  B ~ N(M,  Exp- B 2 )..  B =0 would mean: M=0, and  Exp- B = 0 … No !

A more reasonable view of the “average bias”  B ~ N(M,  Exp- B 2 ).. The consensus seems to be that observational studies tend to exaggerate effects, ie M>0 No problem: if we knew M exactly, we could adjust!… The problem is we don’t …ie  Exp- B > 0.

Summary : the single observational study Must include : uncertainty in the study bias, and uncertainty in the expectation of bias effects – and the size of the Obs study: Y OBS ~ N(  + ,, S OBS )  ~ N(  B,  B 2 )  B ~ N(M,  Exp- B 2 ).. => Y OBS ~ N(  + M,, S OBS 2 +  B 2 +  Exp- B 2 )..

Meta-analysis of Observational Studies (1) With ONE observational study  ~ N(  B,  B 2 ) is interpreted as uncertainty in bias  With several studies j = 1,2 … J studies  j ~ N(  B,  B 2 ) is interpreted as between-study variation in bias BUT, the values of  B,  B are the same … ….Variation => predictive uncertainty

Variation and Uncertainty … Uncertainty is a state of mind. It can be reduced – by collecting more data Variation is a fact about objects, people, studies, estimates … It cannot be reduced Predictive uncertainty that arises from variation cannot be reduced

Meta-analysis of Observational Studies (2) A random effect Observational meta-analysis would be Y OBS-j ~ N(  +  j,, S OBS-j ) data from study j  j ~ N(  B,  B 2 ) between-study variation in bias  B ~ N(M,  Exp- B 2 ) uncertainty regarding “expected” bias => Y OBS-j ~ N(  + M, S OBS-j +  B 2 ) the mean of the M-A is a biased estimate of target parameter , biased by M… easily corrected … So M-A (if large!) avoids the large uncertainty  B and replaces it with the smaller uncertainty  Exp- B

How uncertain are we about M? Can set some limits on uncertainty regarding “Average Bias” M If studies suggest that, eg, the average bias is an OR of 0.9, how uncertain is this... ? Credible limits 0.75 to 1.05 ? … etc. … or carry out a huge meta-meta-analysis and obtain a posterior distribution for  B ~ N(M,  Exp- B 2 )

“Fixed effect” parameter: Summary 1. Indirect comparisons – “weakest link” effect, but large uncertainty reduction possible with >1 comparator.. 2. Observational data from single study very weak … between-study variation AND uncertainty in “average bias”. 3. The estimated mean from a random effect Observational meta-analysis more useful: ONLY uncertainty in average bias to worry about….

Random Treatment Effect Models Every RCT j=1,2 …J estimates a different parameter  j Y j ~ N(  j, S j 2 ), The studies and their sampling error  j ~ N(  RE,  RE 2 ) Variation in the true effects, from a common RE distribution  RE ~ N(  0,  0 2 ) Uncertainty in the mean  RE ~ ? ? Uncertainty in the between-trials variation

What do we want to learn more about ? (a)The mean effect :  RE (b)between-study variance  RE 2 (c)LOR  j - in patient group / protocol studied before (d) LOR  J+1 - in a new patient group / protocol from same distribution PROBLEM:  RE is an ‘unbiased’ estimate … … but what is it an estimate of ???

What can we learn from one new RCT ? (a)Not much about the mean effect  RE unless we can assume  RE. - between-studies variation - is very low (b) Not much about between-study variance  RE 2 (c) LOR  j : Efficacy in a patient group / protocol studied before.... Then back to a Fixed Effect model for that group/protocol … (split or lump?)

What does an RE model tell us about the parameter of interest Given a RE distribution, ie  RE,  RE 2 we can work out What we can say about efficacy in a new group How much does data on parameter  j tells us about  k (data on one group, but need info on another)

What can we learn from observational studies, given an RE model ? 1.Difficult: is each observational study giving us a biased estimate of some  j, or is it averaging over many  j and estimating a  RE ? …but no guarantee it’s the “same”  RE as an RCT meta-analysis 2. At BEST (if many very large studies) the mean from a Observational MA is an estimate of (“  RE ” + M) Only problem (still) uncertainty  Exp-B about M

What can we learn from Indirect Comparisons in a Random Effect context? MTC RE meta-analysis provides unbiased information on mean treatment effects  AB via  AC and  BC -just as  AB is informed by  AC and  BC inform. -Same “weakest link” effect … added bonus: far more information on  RE

What can we learn from Surrogate end- points 1.“Validated” surrogate end-points are rated high in the hierarchy of evidence …. 2.Validation, however, usually within trial

What can we learn from Surrogate end-points ? Daniels & Hughes model. T j ~ N(  j,S T,j ) Data on True End Point, study j Z j ~ N(  j,S Z,j ) Data on Surrogate End Point  j =  +   j, Regression relates true TEP to true SEP If we knew the regression parameters  and , then information on SEP would be as good as information on TEP … but we DONT (… also, this is a RE model – one study does not say much)

What do we really know about the regression slope ? 1.Regression of T against S in Untreated cohort studies motivated the surrogacy concept – plenty of data on  and  … uncertainty small 2.But people insist we can learn about  and  only from RCTs … back to uncertainty again.. ! 3.BUT, then they also want to assume  and  are the identical regardless of treatment … flip back to unrealistic certainty !

placebo active

Surrogacy summary … 1. What is a realistic level of certainty in projecting from surrogate evidence to clinical end-points ? 2.Careful analyses of data required – in every case – CD4+ cell count, bone mineral density, blood pressure cholesterol, etc ….

A Research Agenda - Observational Studies 1.Models for bias: additive, multiplicative, combined 2.Evidence-based estimates for Between-studies variation in bias  B 3. Evidence based distribution  B needed => values for average bias M and uncertainty  Exp- B 2 4. Do  B and  B depend on Type of Study: case- control, cohort, register; or on condition ?

A Research Agenda - RCTs 1.Evidence-based estimates of Between-studies variation in treatment effect  RE = 0 quite possible, but  RE > 0.25 unlikely => 95% of ORs 0.6 – 1.65 around their median ? 2. The prior distribution  ~ N(  0,  0 2 ) why begin with complete ignorance about ,  0 > 100… … LORs > 3 (either way) are VERY rare,  0 = 0.55

Any data is ‘relevant’ if it reduces uncertainty …. depends on the model that relates it to the parameter of interest Towards a Hierarchy of RELEVANCE ??? … work in progress