SlideShare une entreprise Scribd logo
1  sur  45
What is your question?
Causal and predictive purposes of clinical
trials and the implications for analysis
Stephen Senn
Edinburgh
(c) Stephen Senn 1
stephen@senns.uk
http://www.senns.uk/Blogs.html
Acknowledgements
Many thanks for the invitation
This work was partly supported by the European Union’s 7th
Framework Programme for research, technological development and
demonstration under grant agreement no. 602552. “IDEAL”
(c) Stephen Senn 2
(c) Stephen Senn 2
Outline
•Frameworks and purposes
•Randomisation, design and analysis
•Personalised medicine
•Wrapping it all up
(c) Stephen Senn 3
Frameworks and Purposes
Questions before answers
(c) Stephen Senn 4
Statistical frameworks
Frameworks
Framework Notes
Random
sampling
Data are generated from a
population of interest using
probability sampling
Randomisation Data are generated by random
assignment of treatment to units
Linear model Conditional on predictors, treated
as fixed, outcomes are assumed to
have random errors
Multivariate
analysis
Variates vary together and we study
certain conditional distributions
Problems and issues
• Random sampling is most often invoked
in elementary courses
• It has little if any relevance to clinical
trials
• Randomisation is used not random
sampling
• In practice linear models are very often
used
• and incorrectly described as being
multivariate
• This is all rather confused
• Sometimes it does not matter but
sometimes it does
(c) Stephen Senn 5
Causal versus predictive inference
An important further issue
• Clinical trials can be used to try and answer a number of very
different questions
• Two examples are
• Did the treatment have an effect in these patients?
• A causal purpose
• What will the effect be in future patients?
• A predictive purpose
• Unfortunately, in practice, an answer is produced without stating
what the question was
• Given certain assumptions these questions can be answered using the
same analysis but the assumptions are strong and rarely stated
(c) Stephen Senn 6
Two purposes
Predictive
• The population is taken to be ‘patients in
general’
• Of course this really means future patients
• They are the ones to whom the treatment
will be applied
• We treat the patients in the trial as an
appropriate selection from this population
• This does not require them to be typical
but it does require additivity of the
treatment effect
• Additivity implies low heterogeneity of
effect on the scale used
Causal
• We take the patients as fixed
• We want to know what the effect was for
them
• Unfortunately there are missing
counterfactuals
• What would have happened to control
patients given intervention and vice-versa
• The population is the population of all
possible allocations to the patients studied
(c) Stephen Senn 7
Coverage probabilities for two questions
Heterogeneity large
Predictive Causal
(c) Stephen Senn
Coverage probabilities for two questions
Heterogeneity small
Predictive Causal
(c) Stephen Senn
Implications
General
• It is the question that is important
• Given certain assumptions the
same answer can apply to different
questions but this is not always the
case
• We should be careful in
interpreting our answers
predictively
Application
• Whether you run a fixed effects or
random effects meta-analysis is a
consequence of the question you
wish to answer
• It is not a matter of whether you
found heterogeneity or not
• That just governs the extent to
which the answers differ
• However it is the question that
dictates the approach
(c) Stephen Senn 10
Randomisation, design and analysis
Designing to analyse and analysing according to design
(c) Stephen Senn 11
A famous experiment
The experiment
• Hills-Armitage cross-over trial in
enuresis
• Two sequences
• Drug-placebo and placebo-drug
• 58 patients
• Measured over 14 days
• Number of dry nights measured
What I am trying to do
• Compare parametric and
randomisation based analyses
• Show that for a causal purpose it
makes little difference which we
use
• Provided that you reflect the
randomisation!
• This gives a justification for linear
models (which are flexible
parametric approaches)
(c) Stephen Senn 12
Cross-over trial in
Eneuresis
Two treatment periods of
14 days each
1. Hills, M, Armitage, P. The two-period
cross-over clinical trial, British Journal of Clinical
Pharmacology 1979; 8: 7-20.
13
(c) Stephen Senn
0.7
4
0.5
2
0.3
0
0.1
-2
-4
0.6
0.2
0.4
0.0
Density
Permutatedtreatment effect
Blue diamond shows
treatment effect whether or
not we condition on patient
as a factor.
It is identical because the
trial is balanced by patient.
However the permutation
distribution is quite different
and our inferences are
different whether we
condition (red) or not
(black) and clearly
balancing the randomisation
by patient and not
conditioning the analysis by
patient is wrong
14
(c) Stephen Senn
The two permutation* distributions summarised
Summary statistics for Permuted difference no
blocking
Number of observations = 10000
Mean = 0.00561
Median = 0.0345
Minimum = -3.828
Maximum = 3.621
Lower quartile = -0.655
Upper quartile = 0.655
P-value for observed difference 0.0340
*Strictly speaking randomisation distributions
Summary statistics for Permuted difference
blocking
Number of observations = 10000
Mean = 0.00330
Median = 0.0345
Minimum = -2.379
Maximum = 2.517
Lower quartile = -0.517
Upper quartile = 0.517
P-value for observed difference 0.0014
15
(c) Stephen Senn
Two Parametric Approaches
Not fitting patient effect
Estimate s.e. t(56) t pr.
2.172 0.964 2.25 0.0282
(P-value for permutation is 0.034)
Fitting patient effect
Estimate s.e. t(28) t pr
.
2.172 0.616 3.53 0.00147
(P-value for Permutation is 0.0014)
16
(c) Stephen Senn
Lessons
specific and general
Specific
• Because each subject acted as his
own control , the design controlled for
• 20,500 genes
• All life history until the start of the trial
• A ton of epigenetics
• The analysis that reflects this is
correct
• The analysis that doesn’t is wrong
• The point estimate is the same
• Standard errors matter
General
• Contrary to what is commonly claimed
the standard analysis of parallel group
trials does not rely on all factors being
balanced
• If they were the analysis would be wrong
• That’s because the analysis allows for
their being imbalanced
• It relies on their not being balanced
• It makes an allowance for them
• If your causal analysis doesn’t
consider standard errors it is
misleading at best and wrong at worst
(c) Stephen Senn 17
Thinking smaller
• So far we have been looking at determining causation and making
predictions based on a collection of subjects
• Average group causal effects
• Predictions based on data on groups
• What are the implications for individual subjects?
• We still need
• To consider counterfactuals
• To use controls
• To pay attention to variation
• Unfortunately there is much slack thinking
(c) Stephen Senn 18
Personalised medicine
The hope and the hype
(c) Stephen Senn 19
Zombie statistics 1
Percentage of non-responders
What the FDA says Where they got it
Paving the way for personalized
medicine, FDA Oct 2013
Spear, Heath-Chiozzi & Huff, Trends in
Molecular Medicine, May 2001
20
(c) Stephen Senn
Zombie statistics 2
Where they got it Where those who got it got it
Spear, Heath-Chiozzi & Huff, Trends in
Molecular Medicine, May 2001 21
(c) Stephen Senn
The Real Truth
• These are zombie statistics
• They refuse to die
• Not only is the FDA’s claim not right, it’s not even wrong
• It’s impossible to establish what it might mean even if it were true
22
(c) Stephen Senn
88.2% of all statistics are made up
on the spot
Vic Reeves
23
(c) Stephen Senn
Painful comparison
Cochrane Collaboration meta-
analysis 2016
• Meta-analysis of placebo-controlled
trials of paracetamol in tension
headache
• 23 studies
• 6000 patients in total
• Outcome measure:
– Pain free by 2 hours
Baayen (2016) Significance article
• Explanation of Novartis’s MCP-Mod
dose-finding approach using a trial
run by Merck
• 7 doses + placebo
• 517 patients in total
• Outcome measure
– Pain free by 2 hours
24
(c) Stephen Senn
In both cases
• The patients were only studied once
• A dichotomy of a continuous measure was made
• Patients were labelled as responders and non-responders
• A causal conclusion was drawn that went beyond simply
comparing proportions
– Baayen talked about the proportion of patients who would respond
– Cochrane talked about the proportion of patients to whom it would make a
difference in terms of response
25
(c) Stephen Senn
What I propose to do
• Create a simple statistical model to mimic the Cochrane result
– In terms of time to pain resolution every patient will have the same
proportional benefit
• In fact I shall be using a form of proportional hazards model
– The dichotomy will classify patients as responders or non-responders
– We shall be tempted to conclude that some don’t benefit and some do and
that this is a permanent feature of each patient
26
(c) Stephen Senn
The Numerical Recipe
• I shall generate pain duration times for 6000 headaches treated with placebo
– This will be done using an exponential distribution with a mean of just under 3 hours
(2.97 hrs to be exact)
– Each such duration will then be multiplied by just over ¾ (0.755 to be exact) to create
6000 durations under paracetamol
• I shall then take the 6000 pairs and randomly erase one member of the pair
to leave 3000 unpaired placebo values and 3000 unpaired paracetamol
values
• I shall then analyse the data
27
(c) Stephen Senn
Why this recipe?
• The exponential distribution with mean
2.970 is chosen so that the probability of
response in less than two hours is 0.49
– This is the placebo distribution
• Rescaling these figures by 0.755 produces
another exponential distribution with a
probability of response in under two hours
of 0.59
– This is the paracetamol distribution
28
(c) Stephen Senn
29
(c) Stephen Senn
However
• So far I have only gone half way in my simulation recipe
• I have simulated a placebo headache and a corresponding
paracetamol headache
• However I can’t treat the same headache twice
• One of the two is counterfactual
• I now need to get rid of one member of each
factual/counterfactual pair
30
(c) Stephen Senn
31
Note log scale
(c) Stephen Senn
32
(c) Stephen Senn
To sum up
• The results reported are perfectly consistent with paracetamol having the
same effect not just for every patient but on every single headache
• This does not have to be the case but we don’t know that it isn’t
• The combination of dichotomies and responder analysis has great potential
to mislead
• Researchers are assuming that because some patients ‘responded’ in terms
of an arbitrary dichotomy there is scope for personalised medicine
33
(c) Stephen Senn
Implications
• This suggests that we could use designs in which patients are
repeatedly studied as a means of teasing out personal response
• Such designs are sometimes referred to as n-of-1 trials
• An example is a double cross-over
• Patients randomised to two sequences AB or BA in cycle 1 and then again in
cycle 2
• A simulated example follows
(c) Stephen Senn 34
35
(c) Stephen Senn
Results 1
Second Cross-over
Responder Non-
responder
Total
First
Cross-Over
Responder 827 49 876
Non-
responder
38 86 124
Total 865 135 1000
36
correlation coefficient is 0.8 Conditional probability of “response”
in 2nd cross-over
827/876=0.94 if responded1st
38/124=0.31 if did not respond 1st
(c) Stephen Senn
37
(c) Stephen Senn
Results 2
Second Cycle
Responder Non-
responder
Total
First
Cycle
Responder 797 95 892
Non-
responder
93 15 108
Total 865 135 1000
38
correlation coefficient is 0.0.2 Conditional probability of “response”
in 2nd cross-over
797/892=0.89 if responded1st
93/108=0.86 if did not respond 1st
(c) Stephen Senn
Statistical pitfalls of personalized medicine
• We have been promised a
personalised medicine revolution
for over 25 years now
• The revolution, if it is coming, still
seems to be a long way ahead
• In this paper I argue that the scope
for personalized medicine has been
greatly exaggerated
• One of the reasons is naive
interpretation of observed
variation
• N-of-1 trials is one way we could
do better
(c) Stephen Senn 39
Senn, Nature, 29 November 2018
Limitations
• Of course, this is only possible for chronic diseases
• For many diseases replicating treatments is not possible
• Some progress may be made by
• Using repeated measures
• Requires strong assumptions
• Studying Variances
• Comparing variances in treatment and control groups
• Looking at subgroups
(c) Stephen Senn 40
Wrapping it all up
It’s not just Christo and Jeanne-Claude who do this
(c) Stephen Senn 41
Modelling
• The statistician should look for scales for analysis that are additive
• These are not necessarily clinically relevant
• However they can be translated into clinically relevant predictions for
individual patients
• Lubsen & Thyssen 1989
• May require real world data
(c) Stephen Senn 42
A lesson from quality control/assurance
• It is the duty of every manager to understand the variation in the
system
• Intervening naively to try and deal with observed variation can
increase variation
• We need smart statistical design
• We need much more emphasis on components of variation
• Biostatisticians should spend more time discussing these issues with
life scientist colleagues
(c) Stephen Senn 43
Lessons
• Paradoxically embracing random variation can lead to more precise results
• Statistics abounds with examples
• Recovering information in incomplete blocks experiments
• Being Bayesian
• Not overinterpreting data
• Penalised regression
• We should be wary of looking for bedrock causal explanations of everything
• By doing this we may actually increase variation by intervening where doing
something useful is impossible
• Much so-called personalised medicine has overlooked this
• Components of variation matter
• Standard errors matter
• So do questions
(c) Stephen Senn 44
Thank you for your attention
(c) Stephen Senn 45
stephen@senns.uk
http://www.senns.uk/Blogs.html

Contenu connexe

Tendances

The Rothamsted school meets Lord's paradox
The Rothamsted school meets Lord's paradoxThe Rothamsted school meets Lord's paradox
The Rothamsted school meets Lord's paradox
Stephen Senn
 
De Finetti meets Popper
De Finetti meets PopperDe Finetti meets Popper
De Finetti meets Popper
Stephen Senn
 

Tendances (20)

The Rothamsted school meets Lord's paradox
The Rothamsted school meets Lord's paradoxThe Rothamsted school meets Lord's paradox
The Rothamsted school meets Lord's paradox
 
The Seven Habits of Highly Effective Statisticians
The Seven Habits of Highly Effective StatisticiansThe Seven Habits of Highly Effective Statisticians
The Seven Habits of Highly Effective Statisticians
 
Thinking statistically v3
Thinking statistically v3Thinking statistically v3
Thinking statistically v3
 
Has modelling killed randomisation inference frankfurt
Has modelling killed randomisation inference frankfurtHas modelling killed randomisation inference frankfurt
Has modelling killed randomisation inference frankfurt
 
First in man tokyo
First in man tokyoFirst in man tokyo
First in man tokyo
 
Minimally important differences
Minimally important differencesMinimally important differences
Minimally important differences
 
NNTs, responder analysis & overlap measures
NNTs, responder analysis & overlap measuresNNTs, responder analysis & overlap measures
NNTs, responder analysis & overlap measures
 
On being Bayesian
On being BayesianOn being Bayesian
On being Bayesian
 
The revenge of RA Fisher
The revenge of RA Fisher The revenge of RA Fisher
The revenge of RA Fisher
 
The challenge of small data
The challenge of small dataThe challenge of small data
The challenge of small data
 
Seventy years of RCTs
Seventy years of RCTsSeventy years of RCTs
Seventy years of RCTs
 
In search of the lost loss function
In search of the lost loss function In search of the lost loss function
In search of the lost loss function
 
Why I hate minimisation
Why I hate minimisationWhy I hate minimisation
Why I hate minimisation
 
To infinity and beyond v2
To infinity and beyond v2To infinity and beyond v2
To infinity and beyond v2
 
Understanding randomisation
Understanding randomisationUnderstanding randomisation
Understanding randomisation
 
Numbers needed to mislead
Numbers needed to misleadNumbers needed to mislead
Numbers needed to mislead
 
To infinity and beyond
To infinity and beyond To infinity and beyond
To infinity and beyond
 
Real world modified
Real world modifiedReal world modified
Real world modified
 
De Finetti meets Popper
De Finetti meets PopperDe Finetti meets Popper
De Finetti meets Popper
 
Personalised medicine a sceptical view
Personalised medicine a sceptical viewPersonalised medicine a sceptical view
Personalised medicine a sceptical view
 

Similaire à What is your question

Critical appraisal of randomized clinical trials
Critical appraisal of randomized clinical trialsCritical appraisal of randomized clinical trials
Critical appraisal of randomized clinical trials
Samir Haffar
 
2010 smg training_cardiff_day1_session4_harbord
2010 smg training_cardiff_day1_session4_harbord2010 smg training_cardiff_day1_session4_harbord
2010 smg training_cardiff_day1_session4_harbord
rgveroniki
 
biostatists presentation
biostatists presentationbiostatists presentation
biostatists presentation
Anil kumar
 
Study design2 6_07
Study design2 6_07Study design2 6_07
Study design2 6_07
Dan Fisher
 

Similaire à What is your question (20)

The Rothamsted School & The analysis of designed experiments
The Rothamsted School  & The analysis of designed experimentsThe Rothamsted School  & The analysis of designed experiments
The Rothamsted School & The analysis of designed experiments
 
Clinical trials are about comparability not generalisability V2.pptx
Clinical trials are about comparability not generalisability V2.pptxClinical trials are about comparability not generalisability V2.pptx
Clinical trials are about comparability not generalisability V2.pptx
 
Clinical trials are about comparability not generalisability V2.pptx
Clinical trials are about comparability not generalisability V2.pptxClinical trials are about comparability not generalisability V2.pptx
Clinical trials are about comparability not generalisability V2.pptx
 
scope and need of biostatics
scope and need of  biostaticsscope and need of  biostatics
scope and need of biostatics
 
What is the point of point estimates
What is the point of point estimates What is the point of point estimates
What is the point of point estimates
 
Oac guidelines
Oac guidelinesOac guidelines
Oac guidelines
 
Depersonalising medicine
Depersonalising medicineDepersonalising medicine
Depersonalising medicine
 
Evidence based psychiatry
Evidence based psychiatryEvidence based psychiatry
Evidence based psychiatry
 
Adjusting for treatment switching in randomised controlled trials
Adjusting for treatment switching in randomised controlled trialsAdjusting for treatment switching in randomised controlled trials
Adjusting for treatment switching in randomised controlled trials
 
Advanced Biostatistics and Data Analysis abdul ghafoor sajjad
Advanced Biostatistics and Data Analysis abdul ghafoor sajjadAdvanced Biostatistics and Data Analysis abdul ghafoor sajjad
Advanced Biostatistics and Data Analysis abdul ghafoor sajjad
 
Clinical trials: three statistical traps for the unwary
Clinical trials: three statistical traps for the unwaryClinical trials: three statistical traps for the unwary
Clinical trials: three statistical traps for the unwary
 
Critical appraisal of randomized clinical trials
Critical appraisal of randomized clinical trialsCritical appraisal of randomized clinical trials
Critical appraisal of randomized clinical trials
 
2010 smg training_cardiff_day1_session4_harbord
2010 smg training_cardiff_day1_session4_harbord2010 smg training_cardiff_day1_session4_harbord
2010 smg training_cardiff_day1_session4_harbord
 
biostatists presentation
biostatists presentationbiostatists presentation
biostatists presentation
 
Research Design for health care students
Research Design for health care studentsResearch Design for health care students
Research Design for health care students
 
Basics of Statistics.pptx
Basics of Statistics.pptxBasics of Statistics.pptx
Basics of Statistics.pptx
 
Techniques in clinical epidemiology
Techniques in clinical epidemiologyTechniques in clinical epidemiology
Techniques in clinical epidemiology
 
Study design2 6_07
Study design2 6_07Study design2 6_07
Study design2 6_07
 
Clinical trial SSKM EXEC FOR EDUCATIONS
Clinical trial SSKM EXEC  FOR EDUCATIONSClinical trial SSKM EXEC  FOR EDUCATIONS
Clinical trial SSKM EXEC FOR EDUCATIONS
 
Testing of hypothesis and Goodness of fit
Testing of hypothesis and Goodness of fitTesting of hypothesis and Goodness of fit
Testing of hypothesis and Goodness of fit
 

Dernier

Jodhpur Park | Call Girls in Kolkata Phone No 8005736733 Elite Escort Service...
Jodhpur Park | Call Girls in Kolkata Phone No 8005736733 Elite Escort Service...Jodhpur Park | Call Girls in Kolkata Phone No 8005736733 Elite Escort Service...
Jodhpur Park | Call Girls in Kolkata Phone No 8005736733 Elite Escort Service...
HyderabadDolls
 
In Riyadh ((+919101817206)) Cytotec kit @ Abortion Pills Saudi Arabia
In Riyadh ((+919101817206)) Cytotec kit @ Abortion Pills Saudi ArabiaIn Riyadh ((+919101817206)) Cytotec kit @ Abortion Pills Saudi Arabia
In Riyadh ((+919101817206)) Cytotec kit @ Abortion Pills Saudi Arabia
ahmedjiabur940
 
怎样办理圣地亚哥州立大学毕业证(SDSU毕业证书)成绩单学校原版复制
怎样办理圣地亚哥州立大学毕业证(SDSU毕业证书)成绩单学校原版复制怎样办理圣地亚哥州立大学毕业证(SDSU毕业证书)成绩单学校原版复制
怎样办理圣地亚哥州立大学毕业证(SDSU毕业证书)成绩单学校原版复制
vexqp
 
如何办理英国诺森比亚大学毕业证(NU毕业证书)成绩单原件一模一样
如何办理英国诺森比亚大学毕业证(NU毕业证书)成绩单原件一模一样如何办理英国诺森比亚大学毕业证(NU毕业证书)成绩单原件一模一样
如何办理英国诺森比亚大学毕业证(NU毕业证书)成绩单原件一模一样
wsppdmt
 
Gartner's Data Analytics Maturity Model.pptx
Gartner's Data Analytics Maturity Model.pptxGartner's Data Analytics Maturity Model.pptx
Gartner's Data Analytics Maturity Model.pptx
chadhar227
 
Top profile Call Girls In Bihar Sharif [ 7014168258 ] Call Me For Genuine Mod...
Top profile Call Girls In Bihar Sharif [ 7014168258 ] Call Me For Genuine Mod...Top profile Call Girls In Bihar Sharif [ 7014168258 ] Call Me For Genuine Mod...
Top profile Call Girls In Bihar Sharif [ 7014168258 ] Call Me For Genuine Mod...
nirzagarg
 
Top profile Call Girls In Satna [ 7014168258 ] Call Me For Genuine Models We ...
Top profile Call Girls In Satna [ 7014168258 ] Call Me For Genuine Models We ...Top profile Call Girls In Satna [ 7014168258 ] Call Me For Genuine Models We ...
Top profile Call Girls In Satna [ 7014168258 ] Call Me For Genuine Models We ...
nirzagarg
 

Dernier (20)

Jodhpur Park | Call Girls in Kolkata Phone No 8005736733 Elite Escort Service...
Jodhpur Park | Call Girls in Kolkata Phone No 8005736733 Elite Escort Service...Jodhpur Park | Call Girls in Kolkata Phone No 8005736733 Elite Escort Service...
Jodhpur Park | Call Girls in Kolkata Phone No 8005736733 Elite Escort Service...
 
In Riyadh ((+919101817206)) Cytotec kit @ Abortion Pills Saudi Arabia
In Riyadh ((+919101817206)) Cytotec kit @ Abortion Pills Saudi ArabiaIn Riyadh ((+919101817206)) Cytotec kit @ Abortion Pills Saudi Arabia
In Riyadh ((+919101817206)) Cytotec kit @ Abortion Pills Saudi Arabia
 
Top Call Girls in Balaghat 9332606886Call Girls Advance Cash On Delivery Ser...
Top Call Girls in Balaghat  9332606886Call Girls Advance Cash On Delivery Ser...Top Call Girls in Balaghat  9332606886Call Girls Advance Cash On Delivery Ser...
Top Call Girls in Balaghat 9332606886Call Girls Advance Cash On Delivery Ser...
 
怎样办理圣地亚哥州立大学毕业证(SDSU毕业证书)成绩单学校原版复制
怎样办理圣地亚哥州立大学毕业证(SDSU毕业证书)成绩单学校原版复制怎样办理圣地亚哥州立大学毕业证(SDSU毕业证书)成绩单学校原版复制
怎样办理圣地亚哥州立大学毕业证(SDSU毕业证书)成绩单学校原版复制
 
Vadodara 💋 Call Girl 7737669865 Call Girls in Vadodara Escort service book now
Vadodara 💋 Call Girl 7737669865 Call Girls in Vadodara Escort service book nowVadodara 💋 Call Girl 7737669865 Call Girls in Vadodara Escort service book now
Vadodara 💋 Call Girl 7737669865 Call Girls in Vadodara Escort service book now
 
Predicting HDB Resale Prices - Conducting Linear Regression Analysis With Orange
Predicting HDB Resale Prices - Conducting Linear Regression Analysis With OrangePredicting HDB Resale Prices - Conducting Linear Regression Analysis With Orange
Predicting HDB Resale Prices - Conducting Linear Regression Analysis With Orange
 
20240412-SmartCityIndex-2024-Full-Report.pdf
20240412-SmartCityIndex-2024-Full-Report.pdf20240412-SmartCityIndex-2024-Full-Report.pdf
20240412-SmartCityIndex-2024-Full-Report.pdf
 
Dubai Call Girls Peeing O525547819 Call Girls Dubai
Dubai Call Girls Peeing O525547819 Call Girls DubaiDubai Call Girls Peeing O525547819 Call Girls Dubai
Dubai Call Girls Peeing O525547819 Call Girls Dubai
 
Aspirational Block Program Block Syaldey District - Almora
Aspirational Block Program Block Syaldey District - AlmoraAspirational Block Program Block Syaldey District - Almora
Aspirational Block Program Block Syaldey District - Almora
 
High Profile Call Girls Service in Jalore { 9332606886 } VVIP NISHA Call Girl...
High Profile Call Girls Service in Jalore { 9332606886 } VVIP NISHA Call Girl...High Profile Call Girls Service in Jalore { 9332606886 } VVIP NISHA Call Girl...
High Profile Call Girls Service in Jalore { 9332606886 } VVIP NISHA Call Girl...
 
Fun all Day Call Girls in Jaipur 9332606886 High Profile Call Girls You Ca...
Fun all Day Call Girls in Jaipur   9332606886  High Profile Call Girls You Ca...Fun all Day Call Girls in Jaipur   9332606886  High Profile Call Girls You Ca...
Fun all Day Call Girls in Jaipur 9332606886 High Profile Call Girls You Ca...
 
如何办理英国诺森比亚大学毕业证(NU毕业证书)成绩单原件一模一样
如何办理英国诺森比亚大学毕业证(NU毕业证书)成绩单原件一模一样如何办理英国诺森比亚大学毕业证(NU毕业证书)成绩单原件一模一样
如何办理英国诺森比亚大学毕业证(NU毕业证书)成绩单原件一模一样
 
Gartner's Data Analytics Maturity Model.pptx
Gartner's Data Analytics Maturity Model.pptxGartner's Data Analytics Maturity Model.pptx
Gartner's Data Analytics Maturity Model.pptx
 
Top profile Call Girls In Bihar Sharif [ 7014168258 ] Call Me For Genuine Mod...
Top profile Call Girls In Bihar Sharif [ 7014168258 ] Call Me For Genuine Mod...Top profile Call Girls In Bihar Sharif [ 7014168258 ] Call Me For Genuine Mod...
Top profile Call Girls In Bihar Sharif [ 7014168258 ] Call Me For Genuine Mod...
 
Digital Advertising Lecture for Advanced Digital & Social Media Strategy at U...
Digital Advertising Lecture for Advanced Digital & Social Media Strategy at U...Digital Advertising Lecture for Advanced Digital & Social Media Strategy at U...
Digital Advertising Lecture for Advanced Digital & Social Media Strategy at U...
 
TrafficWave Generator Will Instantly drive targeted and engaging traffic back...
TrafficWave Generator Will Instantly drive targeted and engaging traffic back...TrafficWave Generator Will Instantly drive targeted and engaging traffic back...
TrafficWave Generator Will Instantly drive targeted and engaging traffic back...
 
Top profile Call Girls In Satna [ 7014168258 ] Call Me For Genuine Models We ...
Top profile Call Girls In Satna [ 7014168258 ] Call Me For Genuine Models We ...Top profile Call Girls In Satna [ 7014168258 ] Call Me For Genuine Models We ...
Top profile Call Girls In Satna [ 7014168258 ] Call Me For Genuine Models We ...
 
7. Epi of Chronic respiratory diseases.ppt
7. Epi of Chronic respiratory diseases.ppt7. Epi of Chronic respiratory diseases.ppt
7. Epi of Chronic respiratory diseases.ppt
 
DATA SUMMIT 24 Building Real-Time Pipelines With FLaNK
DATA SUMMIT 24  Building Real-Time Pipelines With FLaNKDATA SUMMIT 24  Building Real-Time Pipelines With FLaNK
DATA SUMMIT 24 Building Real-Time Pipelines With FLaNK
 
Charbagh + Female Escorts Service in Lucknow | Starting ₹,5K To @25k with A/C...
Charbagh + Female Escorts Service in Lucknow | Starting ₹,5K To @25k with A/C...Charbagh + Female Escorts Service in Lucknow | Starting ₹,5K To @25k with A/C...
Charbagh + Female Escorts Service in Lucknow | Starting ₹,5K To @25k with A/C...
 

What is your question

  • 1. What is your question? Causal and predictive purposes of clinical trials and the implications for analysis Stephen Senn Edinburgh (c) Stephen Senn 1 stephen@senns.uk http://www.senns.uk/Blogs.html
  • 2. Acknowledgements Many thanks for the invitation This work was partly supported by the European Union’s 7th Framework Programme for research, technological development and demonstration under grant agreement no. 602552. “IDEAL” (c) Stephen Senn 2 (c) Stephen Senn 2
  • 3. Outline •Frameworks and purposes •Randomisation, design and analysis •Personalised medicine •Wrapping it all up (c) Stephen Senn 3
  • 4. Frameworks and Purposes Questions before answers (c) Stephen Senn 4
  • 5. Statistical frameworks Frameworks Framework Notes Random sampling Data are generated from a population of interest using probability sampling Randomisation Data are generated by random assignment of treatment to units Linear model Conditional on predictors, treated as fixed, outcomes are assumed to have random errors Multivariate analysis Variates vary together and we study certain conditional distributions Problems and issues • Random sampling is most often invoked in elementary courses • It has little if any relevance to clinical trials • Randomisation is used not random sampling • In practice linear models are very often used • and incorrectly described as being multivariate • This is all rather confused • Sometimes it does not matter but sometimes it does (c) Stephen Senn 5
  • 6. Causal versus predictive inference An important further issue • Clinical trials can be used to try and answer a number of very different questions • Two examples are • Did the treatment have an effect in these patients? • A causal purpose • What will the effect be in future patients? • A predictive purpose • Unfortunately, in practice, an answer is produced without stating what the question was • Given certain assumptions these questions can be answered using the same analysis but the assumptions are strong and rarely stated (c) Stephen Senn 6
  • 7. Two purposes Predictive • The population is taken to be ‘patients in general’ • Of course this really means future patients • They are the ones to whom the treatment will be applied • We treat the patients in the trial as an appropriate selection from this population • This does not require them to be typical but it does require additivity of the treatment effect • Additivity implies low heterogeneity of effect on the scale used Causal • We take the patients as fixed • We want to know what the effect was for them • Unfortunately there are missing counterfactuals • What would have happened to control patients given intervention and vice-versa • The population is the population of all possible allocations to the patients studied (c) Stephen Senn 7
  • 8. Coverage probabilities for two questions Heterogeneity large Predictive Causal (c) Stephen Senn
  • 9. Coverage probabilities for two questions Heterogeneity small Predictive Causal (c) Stephen Senn
  • 10. Implications General • It is the question that is important • Given certain assumptions the same answer can apply to different questions but this is not always the case • We should be careful in interpreting our answers predictively Application • Whether you run a fixed effects or random effects meta-analysis is a consequence of the question you wish to answer • It is not a matter of whether you found heterogeneity or not • That just governs the extent to which the answers differ • However it is the question that dictates the approach (c) Stephen Senn 10
  • 11. Randomisation, design and analysis Designing to analyse and analysing according to design (c) Stephen Senn 11
  • 12. A famous experiment The experiment • Hills-Armitage cross-over trial in enuresis • Two sequences • Drug-placebo and placebo-drug • 58 patients • Measured over 14 days • Number of dry nights measured What I am trying to do • Compare parametric and randomisation based analyses • Show that for a causal purpose it makes little difference which we use • Provided that you reflect the randomisation! • This gives a justification for linear models (which are flexible parametric approaches) (c) Stephen Senn 12
  • 13. Cross-over trial in Eneuresis Two treatment periods of 14 days each 1. Hills, M, Armitage, P. The two-period cross-over clinical trial, British Journal of Clinical Pharmacology 1979; 8: 7-20. 13 (c) Stephen Senn
  • 14. 0.7 4 0.5 2 0.3 0 0.1 -2 -4 0.6 0.2 0.4 0.0 Density Permutatedtreatment effect Blue diamond shows treatment effect whether or not we condition on patient as a factor. It is identical because the trial is balanced by patient. However the permutation distribution is quite different and our inferences are different whether we condition (red) or not (black) and clearly balancing the randomisation by patient and not conditioning the analysis by patient is wrong 14 (c) Stephen Senn
  • 15. The two permutation* distributions summarised Summary statistics for Permuted difference no blocking Number of observations = 10000 Mean = 0.00561 Median = 0.0345 Minimum = -3.828 Maximum = 3.621 Lower quartile = -0.655 Upper quartile = 0.655 P-value for observed difference 0.0340 *Strictly speaking randomisation distributions Summary statistics for Permuted difference blocking Number of observations = 10000 Mean = 0.00330 Median = 0.0345 Minimum = -2.379 Maximum = 2.517 Lower quartile = -0.517 Upper quartile = 0.517 P-value for observed difference 0.0014 15 (c) Stephen Senn
  • 16. Two Parametric Approaches Not fitting patient effect Estimate s.e. t(56) t pr. 2.172 0.964 2.25 0.0282 (P-value for permutation is 0.034) Fitting patient effect Estimate s.e. t(28) t pr . 2.172 0.616 3.53 0.00147 (P-value for Permutation is 0.0014) 16 (c) Stephen Senn
  • 17. Lessons specific and general Specific • Because each subject acted as his own control , the design controlled for • 20,500 genes • All life history until the start of the trial • A ton of epigenetics • The analysis that reflects this is correct • The analysis that doesn’t is wrong • The point estimate is the same • Standard errors matter General • Contrary to what is commonly claimed the standard analysis of parallel group trials does not rely on all factors being balanced • If they were the analysis would be wrong • That’s because the analysis allows for their being imbalanced • It relies on their not being balanced • It makes an allowance for them • If your causal analysis doesn’t consider standard errors it is misleading at best and wrong at worst (c) Stephen Senn 17
  • 18. Thinking smaller • So far we have been looking at determining causation and making predictions based on a collection of subjects • Average group causal effects • Predictions based on data on groups • What are the implications for individual subjects? • We still need • To consider counterfactuals • To use controls • To pay attention to variation • Unfortunately there is much slack thinking (c) Stephen Senn 18
  • 19. Personalised medicine The hope and the hype (c) Stephen Senn 19
  • 20. Zombie statistics 1 Percentage of non-responders What the FDA says Where they got it Paving the way for personalized medicine, FDA Oct 2013 Spear, Heath-Chiozzi & Huff, Trends in Molecular Medicine, May 2001 20 (c) Stephen Senn
  • 21. Zombie statistics 2 Where they got it Where those who got it got it Spear, Heath-Chiozzi & Huff, Trends in Molecular Medicine, May 2001 21 (c) Stephen Senn
  • 22. The Real Truth • These are zombie statistics • They refuse to die • Not only is the FDA’s claim not right, it’s not even wrong • It’s impossible to establish what it might mean even if it were true 22 (c) Stephen Senn
  • 23. 88.2% of all statistics are made up on the spot Vic Reeves 23 (c) Stephen Senn
  • 24. Painful comparison Cochrane Collaboration meta- analysis 2016 • Meta-analysis of placebo-controlled trials of paracetamol in tension headache • 23 studies • 6000 patients in total • Outcome measure: – Pain free by 2 hours Baayen (2016) Significance article • Explanation of Novartis’s MCP-Mod dose-finding approach using a trial run by Merck • 7 doses + placebo • 517 patients in total • Outcome measure – Pain free by 2 hours 24 (c) Stephen Senn
  • 25. In both cases • The patients were only studied once • A dichotomy of a continuous measure was made • Patients were labelled as responders and non-responders • A causal conclusion was drawn that went beyond simply comparing proportions – Baayen talked about the proportion of patients who would respond – Cochrane talked about the proportion of patients to whom it would make a difference in terms of response 25 (c) Stephen Senn
  • 26. What I propose to do • Create a simple statistical model to mimic the Cochrane result – In terms of time to pain resolution every patient will have the same proportional benefit • In fact I shall be using a form of proportional hazards model – The dichotomy will classify patients as responders or non-responders – We shall be tempted to conclude that some don’t benefit and some do and that this is a permanent feature of each patient 26 (c) Stephen Senn
  • 27. The Numerical Recipe • I shall generate pain duration times for 6000 headaches treated with placebo – This will be done using an exponential distribution with a mean of just under 3 hours (2.97 hrs to be exact) – Each such duration will then be multiplied by just over ¾ (0.755 to be exact) to create 6000 durations under paracetamol • I shall then take the 6000 pairs and randomly erase one member of the pair to leave 3000 unpaired placebo values and 3000 unpaired paracetamol values • I shall then analyse the data 27 (c) Stephen Senn
  • 28. Why this recipe? • The exponential distribution with mean 2.970 is chosen so that the probability of response in less than two hours is 0.49 – This is the placebo distribution • Rescaling these figures by 0.755 produces another exponential distribution with a probability of response in under two hours of 0.59 – This is the paracetamol distribution 28 (c) Stephen Senn
  • 30. However • So far I have only gone half way in my simulation recipe • I have simulated a placebo headache and a corresponding paracetamol headache • However I can’t treat the same headache twice • One of the two is counterfactual • I now need to get rid of one member of each factual/counterfactual pair 30 (c) Stephen Senn
  • 31. 31 Note log scale (c) Stephen Senn
  • 33. To sum up • The results reported are perfectly consistent with paracetamol having the same effect not just for every patient but on every single headache • This does not have to be the case but we don’t know that it isn’t • The combination of dichotomies and responder analysis has great potential to mislead • Researchers are assuming that because some patients ‘responded’ in terms of an arbitrary dichotomy there is scope for personalised medicine 33 (c) Stephen Senn
  • 34. Implications • This suggests that we could use designs in which patients are repeatedly studied as a means of teasing out personal response • Such designs are sometimes referred to as n-of-1 trials • An example is a double cross-over • Patients randomised to two sequences AB or BA in cycle 1 and then again in cycle 2 • A simulated example follows (c) Stephen Senn 34
  • 36. Results 1 Second Cross-over Responder Non- responder Total First Cross-Over Responder 827 49 876 Non- responder 38 86 124 Total 865 135 1000 36 correlation coefficient is 0.8 Conditional probability of “response” in 2nd cross-over 827/876=0.94 if responded1st 38/124=0.31 if did not respond 1st (c) Stephen Senn
  • 38. Results 2 Second Cycle Responder Non- responder Total First Cycle Responder 797 95 892 Non- responder 93 15 108 Total 865 135 1000 38 correlation coefficient is 0.0.2 Conditional probability of “response” in 2nd cross-over 797/892=0.89 if responded1st 93/108=0.86 if did not respond 1st (c) Stephen Senn
  • 39. Statistical pitfalls of personalized medicine • We have been promised a personalised medicine revolution for over 25 years now • The revolution, if it is coming, still seems to be a long way ahead • In this paper I argue that the scope for personalized medicine has been greatly exaggerated • One of the reasons is naive interpretation of observed variation • N-of-1 trials is one way we could do better (c) Stephen Senn 39 Senn, Nature, 29 November 2018
  • 40. Limitations • Of course, this is only possible for chronic diseases • For many diseases replicating treatments is not possible • Some progress may be made by • Using repeated measures • Requires strong assumptions • Studying Variances • Comparing variances in treatment and control groups • Looking at subgroups (c) Stephen Senn 40
  • 41. Wrapping it all up It’s not just Christo and Jeanne-Claude who do this (c) Stephen Senn 41
  • 42. Modelling • The statistician should look for scales for analysis that are additive • These are not necessarily clinically relevant • However they can be translated into clinically relevant predictions for individual patients • Lubsen & Thyssen 1989 • May require real world data (c) Stephen Senn 42
  • 43. A lesson from quality control/assurance • It is the duty of every manager to understand the variation in the system • Intervening naively to try and deal with observed variation can increase variation • We need smart statistical design • We need much more emphasis on components of variation • Biostatisticians should spend more time discussing these issues with life scientist colleagues (c) Stephen Senn 43
  • 44. Lessons • Paradoxically embracing random variation can lead to more precise results • Statistics abounds with examples • Recovering information in incomplete blocks experiments • Being Bayesian • Not overinterpreting data • Penalised regression • We should be wary of looking for bedrock causal explanations of everything • By doing this we may actually increase variation by intervening where doing something useful is impossible • Much so-called personalised medicine has overlooked this • Components of variation matter • Standard errors matter • So do questions (c) Stephen Senn 44
  • 45. Thank you for your attention (c) Stephen Senn 45 stephen@senns.uk http://www.senns.uk/Blogs.html