1. Methodology Research Group
Evaluation of moderation
and mediation in the
development of personalised therapies
(stratified medicine)
MHRN conference, London, 20 March 2013
Sabine Landau, Institute of Psychiatry, King’s
College London
&
Graham Dunn, Institute of Population
Health, University of Manchester
2. Outline
Methodology Research Group
1. Introduction to key concepts
• What is personalised therapy/
stratified medicine?
Sabine
• Causal effects, confounding and RCTs
• Treatment effect moderation
• Treatment effect mediation
2. Recap and development of ideas
• Correct and incorrect approaches to
treatment effect moderation (stratification)
• Using moderator (predictive marker) by Graham
treatment interactions as instruments for
mediation investigations
3. Research Programme:
Efficacy and Mechanisms
Evaluation Methodology Research Group
Funded by MRC Methodology Research Programme
• Design and methods of explanatory (causal) analysis for randomised trials of complex
interventions in mental health (2006-2009)
– Graham Dunn (PI), Linda Davies, Jonathan Green, Andrew Pickles, Chris Roberts,
Ian White & Frank Windmeijer.
• Estimation of causal effects of complex interventions in longitudinal studies with
intermediate variables (2009-2012)
– Richard Emsley (MRC Fellow), Graham Dunn.
• Designs and analysis for the evaluation and validation of social and psychological
markers in randomised trials of complex interventions in mental health (2010-12)
– Graham Dunn (PI), Richard Emsley, Linda Davies, Jonathan Green, Andrew
Pickles, Chris Roberts, Ian White & Frank Windmeijer with Hanhua Liu.
• Developing methods for understanding mechanism in complex interventions (2013-16)
– Sabine Landau (PI), Richard Emsley, Graham Dunn, Ian White, Paul Clarke,
Andrew Pickles & Til Wykes.
4. Aims of Session 1
Methodology Research Group
• To provide an introduction to causal inference using
potential outcomes (counterfactuals).
• To show that the concepts of stratified medicine and
treatment effect moderation are intrinsically linked to
treatment effect heterogeneity.
• To describe some standard approaches to evaluating
treatment-effect mechanisms including the key
assumptions, and highlight some of the potential
problems with this.
• To briefly describe some newer approaches to
mechanism evaluation so that you are familiar with
these concepts and their potential.
5. Example 1: efficacy and
mechanisms evaluation and
personalised medicine Methodology Research Group
• Parenting training may be effective at improving conduct of
children with behavioural problems, but its effect might be
greater in some children than in others.
• Similarly, the training is likely to improve aspects of
parenting and, again, its effect on such parent outcomes is
likely to vary from one patient to another.
• We might expect that if one parent‟s parenting has been
improved considerably more than that of another parent
then the conduct of the first parent‟s child has been
improved more than that of the second parent‟s child.
– Who are parenting training programmes effective
for?
– What proportion of the training programme effect
on child conduct is explained by its effect on
parenting practice?
6. Example 2: efficacy and
mechanisms evaluation and
personalised medicine Methodology Research Group
• A recent large-scale randomised controlled trial (RCT)
provided evidence for the effectiveness of augmentation of
antidepressant medication with cognitive behavioural
therapy (CBT) as a next-step for patients whose depression
has not responded to pharmacotherapy (Wiles et al, 2012).
• Thus the treatment (CBT) was shown to work for a
subpopulation who were identified as “non-responders to
antidepressants”.
• CBT is supposed to work by changing the way how people
think about themself, the world and other people.
– Who does CBT work for?
– What proportion of the CBT effect on depressive
symptoms is explained by its effect on cognition?
7. General principle of causal
inference Methodology Research Group
• Effect size estimates (correlations, regression coefficients,
odds ratios etc.) can only tell us about association between
two variables (say X and Y).
• The aim of causal inference is to infer whether this
association can be given a causal interpretation (e.g. X
causes Y) by:
– defining the causal parameters,
– being explicit about the assumptions made when using a
respective estimators,
– thinking about other possible explanations for observed
effects, especially confounding.
8. Ideas of causality
(Cox and Wermuth, 2001) Methodology Research Group
• Causality as a stable association
– An observed association that cannot be accounted for by
any postulated confounder(s)
» (but, on its own, this says nothing about the direction of the
causal effect)
• Bradford Hill‟s criteria
– A series of conditions which make the hypothesis of
causality more convincing
» (but none are either necessary or sufficient to prove causality)
• Causality as an effect of an intervention
– Potential Outcomes/Counterfactuals (Neyman, Rubin, etc.)
– The idea of fixing (setting) the values of the explanatory
variables (Pearl)
• Causality as an explanation of a process
– This is where science comes in…
9. How can we formally define
a causal treatment effect? Methodology Research Group
• The potential outcomes/counterfactual approach.
• It is a comparison between what is and what might
have been.
• We wish to estimate the difference between a patient‟s
observed outcome and the outcome that would have
been observed if, contrary to fact, the patient‟s
treatment or care had been different (Neyman, 1923;
Rubin, 1974).
• Without the possibility of comparison the treatment
effect is not well defined e.g. gender as a cause.
10. Individual treatment
effects (ITEs) Methodology Research Group
• For a given individual, the effect of treatment
is the difference:
ITE=Outcometreatment - Outcomecontrol
We can never observe this!
11. Causal inference using
counterfactuals Methodology Research Group
Receive treatment Receive control
Measure outcome Measure outcome
Comparison of outcomes gives an
individual treatment effect
12. Causal inference using
counterfactuals Methodology Research Group
Receive treatment Receive control
Measure outcome Measure outcome
Comparison of outcomes will not give an
individual treatment effect
13. Average treatment effect
(ATE) Methodology Research Group
• The average treatment effect ATE is:
Average[ITE] = Average[Outcometreatment - Outcomecontrol]
• If the selection of treatment options is purely random (as in
a perfect RCT) then:
Ave[Outcometreatment - Outcomecontrol]
= Ave[Outcometreatment|treatment] - Ave[Outcomecontrol|Control]
= Ave[Outcome|treatment] - Ave[Outcome|Control]
• ATE defines the efficacy of the treatment w. r. t. to
control.
14. Causal inference using
counterfactuals Methodology Research Group
Receive treatment Receive control
Measure outcome Measure outcome
Comparison of average outcomes gives an
average treatment effect
15. Problem of confounding
Methodology Research Group
U
Exposure Outcomes
• Observed variables in squares, unobserved (latent) variables in
circles.
• An arrow (directed link) between variables represents a causal effect.
• We are interested in the causal effect of Exposure on Outcome (black
path)
• U is an unmeasured confounder (=cause of Exposure and Outcome).
• The confounder provides a backdoor path connecting Exposure and
Outcome (red path)
16. Why randomisation?
Methodology Research Group
• The strength of randomisation is that it ensures that there are
no variables (both observed or unobserved) that drive
treatment allocation.
• In terms of a causal graph, there are no arrows into randomi-
sation from any other variable, observed or unobserved:
– Random treatment group is not a descendent of any other
variable.
– It is exogenous in the model with response=Outcome and
covariate=Random treatment group.
• This means that any comparison between randomisation
groups (e.g. mean difference) estimates a (total) causal
effect…
– …provided the trial has been well designed and executed.
17. Mendelian randomisation
(from Davey-Smith 2011) Methodology Research Group
• “The principle of Mendelian randomization relies on the
basic (but approximate) laws of Mendelian genetics. If the
probability that a postmeiotic germ cell, that has received
any particular allele at segregation, contributes to a viable
conception is independent of environment (following from
Mendel‟s first law), and if genetic variants sort
independently (following from Mendel‟s second law), then
at a population level these variants will not be associated
with the confounding factors that generally distort
conventional observational studies.”
• Basically, genotypes are entirely derived from parents but
can be considered randomly allocated,
– e.g. if both parents are type AB, then genotype could
be AA (probability .25), AB (0.50) or BB (0.25).
18. Mendelian randomisation
(from Davey-Smith 2011) Methodology Research Group
• Genotypes are equivalent to randomisation…
• As before, in causal graph terms, there are no arrows into
genes from any other variable, observed or unobserved:
– Gene is not a descendent of any other variable.
– It is exogenous in the model with response=Outcome
and covariate=Gene.
• This means that any comparison between genes (e.g.
mean difference) estimates a (total) causal effect.
19. Treatment effect
heterogeneity Methodology Research Group
• Importantly the definition of a causal parameter, the average
causal effect (ATE) does not require that the ITEs are equal for
everyone.
Positive effect Detrimental effect
Receive
treatment
Receive
control
20. Personalised medicine and
treatment effect heterogeneity Methodology Research Group
• The existence of variation in individual treatment effects
(ITEs) is the foundation of personalised medicine.
– Stratified medicine
– Predictive medicine
– Genomic medicine
• If we are to pursue the idea of stratified medicine then we
must believe in treatment effect heterogeneity.
• We should therefore use statistical methodology that
explicitly accounts for such causal effect heterogeneity.
21. Baseline predictors
Methodology Research Group
• How does stratified medicine exploit treatment effect
heterogeneity?
• We are interested in knowing in advance of treatment
allocation/decisions to treat who a treatment is most
effective for.
• For personalised medicine we need access to pre-
treatment (baseline) characteristics that predict
treatment-effect heterogeneity
– We don‟t just want to predict outcome
22. Moderators of treatment
Methodology Research Group
Baseline (pre-treatment) characteristics that
influence the effect of treatment on outcome
Random
allocation Outcomes
Marker
Note this path diagram is no longer a causal graph.
We call such baseline variables a “marker” – for more see Section 2.
23. Moderation assessment
in trials Methodology Research Group
• The ability of a baseline variable to act as a treatment
moderator (also referred to as treatment effect modifier)
can be investigated by assessing the interaction between
treatment and the moderator variable in terms of the
outcome.
• When the treatment has been randomised then the causal
effect of the treatment (its efficacy) within subpopulations
defined by the level of the moderator can be estimated.
• (In particular, randomisation within strata defined by the
levels of the moderator maximises the power of this
assessment.)
24. Moderation assessment in
treated cohorts Methodology Research Group
• Often investigators look for outcome heterogeneity in a
cohort of people who received the treatment and interpret
such heterogeneity as evidence for moderation
– E.g. for schizophrenics receiving a psychological
therapy compare functioning between SCZ subtypes
• This approach does not address the moderation
question!
• The approach assesses whether a baseline variable is
predictive of outcome but NOT whether it is predictive of
treatment effects.
25. Prognostic baseline
variables Methodology Research Group
• Cohort studies of treated patients can only provide
assessments of the ability of baseline variables to be
predictive of the outcome;
– That is whether they are prognostic variables.
• They cannot say anything about the ability of baseline
variables to predict treatment effects;
– That is whether they are predictive (moderator)
variables.
• In personalised medicine we are after investigating
moderators.
• However, we may make use of prognostic variables to do
this in a more powerful way (see Session 2).
26. Treatment effect mediation
Methodology Research Group
• The aim of efficacy and mechanism investigations is to go
beyond evaluating whether an intervention is effective and
to explain why it might be efficacious:
– What are the putative mechanisms through
which the treatment acts?
• Usual analysis methods dominated by decomposing total
effects into direct and indirect effects:
– Mental health and psychology has been concerned with
this idea for decades.
– Widely cited Baron and Kenny paper for mediation
analysis in social sciences.
– Makes implicit assumptions which are unlikely to hold.
27. Simple mediation diagram
Methodology Research Group
Mediator
Exposure Outcomes
Total effect = direct effect + indirect effect
28. Confounded mediation
assessment in epidemiology Methodology Research Group
U
U
Mediator
Exposure Outcomes
U
If treatment is not randomised then there is likely to be even more
unmeasured confounding.
29. How does randomisation
help? Methodology Research Group
U
U
Mediator
Random
Outcomes
allocation
U
“Blocked” by
randomisation
30. Mediation in trials
Methodology Research Group
U – the unmeasured confounders
error
U
Mediator
Random error
Outcomes
allocation
Covariates
31. Mediation in genetic
epidemiology Methodology Research Group
U – the unmeasured confounders
error
U
Mediator
Gene Outcomes error
Covariates
32. Possible solutions
Methodology Research Group
• There are basically two ways by which we can ensure that we
can estimate causal parameters of interest in mechanisms
investigations (direct and indirect treatment effects):
– Measure and adjust for potential confounders (sounds
obvious, not always done) …
» so that there remains no hidden confounding and traditional
Baron and Kenny mediation analysis approaches can be
applied
– Use estimators that can consistently estimate mediation
parameters in the presence of hidden confounding …
» a class of estimators called instrumental variables estimators
allows for this
» however, these also require assumptions (see below)
33. Measuring confounders
Methodology Research Group
• This can be difficult when knowledge about underlying
processes is only patchy.
• However, when the putative confounder(s) are known it
might be possible to obtain measures and thus enable causal
mediation assessments even for only partly observed
mediators.
• Example
– Immunology (Follman, 2006):
» Trial to compare vaccination with HIV vaccine against
controls
» Putative mediator= immune response (only observed in the
vaccinated group)
» Interested in whether the vaccination effect on infection rate
is mediated by the immune response
34. Vaccine trials
Methodology Research Group
• It is easy to demonstrate that immune response is a
correlate of protection in the vaccinated arm: the higher
the response, the lower the infection rate.
• Unfortunately, this correlation does not necessarily imply a
causal effect.
– Protection to infection specifically induced by the HIV
vaccine is confounded with underlying levels of
protection in the absence of vaccination.
– Someone capable of producing a large immune
response would be more resistant to infection, even in
the absence of vaccination.
35. “Strange result”
Methodology Research Group
• Confounding explained the strange result:
– Immune response observed after HIV vaccination.
» …though really what is being observed here is the
combination of protection due to general and specific (HIV
vaccine) factors
– Antibody response to the HIV vaccination was strongly
associated with infection risk in the vaccine group.
» … though that could just be protection due to general
factors correlating with infection risk
– But NO effect of HIV vaccination on infection rate (large
trial of approx. 5000 participants).
• A correlate of protection is not necessarily a treatment-effect
mediator, let alone a valid surrogate outcome.
36. A hypothetical HIV
vaccine trial (Follmann, 2006) Methodology Research Group
• Vaccinate everyone before randomisation with an irrelevant
vaccine (against Rabies, for example).
• Measure the immune response to the Rabies vaccine (a
proxy of protection due to general factors).
• Randomly allocate participants to receive HIV vaccine or
Placebo.
• Measure immune response in the HIV vaccinated group.
• Use response to the Rabies vaccine to (multiply) impute the
missing HIV vaccine response in the Placebo participants.
• Carry out a Baron and Kenny analysis on the imputed data
which controls for the now observed confounder.
37. Why do we need
instrumental variables? Methodology Research Group
• All available statistical methods we usually use (for any
standard analysis), including:
– Stratification
– Regression
– Matching
– etc.
require the one unverifiable condition we identified previously:
NO UNMEASURED CONFOUNDING
• Instrumental variables allow us to relax this assumption.
38. Instrumental variables
Methodology Research Group
• For mediation assessment in a trial we are looking for a
variable that is:
1. (Strongly) predictive of the intermediate variable;
2. Has no direct effect on the outcome, except through the
intermediate variable;
3. Does not share common causes with the outcome.
• If these conditions hold, in addition to one further
assumption (no interactions or monotonicity), then such a
variable can be used as an instrumental variable (IV).
• Randomisation, where available, satisfies criteria 1 and 3.
• If we consider this when designing the trial, we can measure
variables that MIGHT meet these requirements.
39. Mediation diagram
with instrumental variables
Methodology Research Group
error
U
Instruments
Mediator
Random error
allocation Outcomes
Covariates
40. Possible instruments
Methodology Research Group
• The following variables might serve as instrumental
variables to enable mediation investigations in trials:
– Baseline variable x randomisation interactions (see
Section 2)
» E.g. Mother mental health x training programme interaction
in parenting example
– Trial x randomisation interaction in meta-analysis of trials
– Randomly allocated non-standardised aspects of
interventions
» E.g. how and high intensity versions of therapy
– Genes
» An application of Mendelian randomisation where it is
assumed that a gene determining the intermediate
phenotype only affects the distal phenotype by changing the
intermediate
42. Assumptions for instrumental
variables Methodology Research Group
• IV methods require FOUR assumptions
• The first 3 assumptions are from the definition:
– The association between instrument and mediator.
– No direct effect of the instrument on outcome.
– No unmeasured confounding for the instrument and
outcome.
• There are a wide variety of fourth assumptions and
different assumptions result in the estimation of
different causal effects:
– E.g. no interactions, monotonicity (no defiers).
43. Instrumental variables:
pros and cons Methodology Research Group
Advantages Disadvantages
1. Can allow for unmeasured 1. It is impossible to verify that a
confounding; variable is an instrument and
using a non-instrument
2. Can allow for measurement introduces additional bias.
error;
2. A weak instrument increases
the bias over that of ordinary
3. Randomisation often meets regression (for finite samples).
the definition so is an ideal
instrument.
3. Instruments by themselves are
actually insufficient to
estimate causal effects and we
require additional
assumptions.
See Hernán and Robins (2006), Epidemiology for further details
44. Assumption trade-off
Methodology Research Group
• IV methods replace one unverifiable assumption of no
unmeasured confounding between the intermediate variable
and the outcome by other unverifiable assumptions
– no unmeasured confounding for the instruments, or
– no direct effect of the instruments.
• We need to decide which assumptions are more likely to
hold in our analysis.
• An IV analysis will also decrease the precision of our
estimates because of allowing for the unmeasured
confounding.
45. In the next session…
Methodology Research Group
• Combining all these ideas:
– Using baseline moderator variables (predictive
markers) for evaluation of treatment effect
mechanisms.
– Using prognostic baseline variables (markers) as
confounders or instrumental variables.
– Improved trial designs to evaluate treatment-effect
heterogeneity and corresponding mediational
mechanisms.
• First we will have a short break…
46. Methodology Research Group
Evaluation of moderation
and mediation in the
development of personalised therapies
(stratified medicine)
SESSION 2
47. Aims of Session 2
Methodology Research Group
• Recap main ideas from Session 1.
• Develop these ideas to
– verify correct and incorrect approaches to
assessing treatment effect moderation
(stratification).
• Develop these ideas to
– suggest trial designs and analyses that use
moderator (predictive marker) by treatment
interactions as instruments for mediation
investigations.
48. Recap: treatment effects and
treatment-effect moderation Methodology Research Group
• Potential outcomes & treatment effects
• Average treatment effects
• Treatment-effect heterogeneity (moderation)
• Naïve searches for stratifying factors (moderators)
49. Treatment effects
Methodology Research Group
• Treatment effects do not make sense (are not defined)
without comparison.
• We are comparing the outcome we see after therapy
with the outcome we might have seen had the
individual not received therapy, or therapy of a
different kind to that actually experienced.
• We are comparing potential outcomes or
counterfactuals.
50. Potential outcomes
Methodology Research Group
• Consider just two alternatives for the treatment of
depression: therapy (T) or a control condition (C).
• We have an outcome (the Beck Depression Inventory
score) that could be measured six months after the
decision to start therapy (or not).
• Let these two potential outcomes be BDI(T) and BDI(C)
for the therapy and control conditions, respectively.
51. Comparison of potential outcomes
Methodology Research Group
• The treatment effect for any given individual is the
difference
BDI(T)-BDI(C)
which we would expect to be a negative
number if the treatment is beneficial.
• Unfortunately, we never get to see both potential
outcomes so we can never observe this individual‟s
treatment effect.
52. So-called treatment-response is
not a measure of an effect of therapy
Methodology Research Group
• Let‟s now introduced a measure of depression BDI(0)
that is obtained at the time of the start of therapy.
• The change over time under therapy – i.e.
– BDI(T) – BDI(0) is not the same as BDI(T) –BDI(C).
• BDI(0) is NOT BDI(C)!
53. Randomisation and Average
Treatment Effects Methodology Research Group
• We get round our problem by working with
averages:
Average Treatment Effect = ATE
= Ave[(BDI(T) – BDI(C)]
= Ave[BDI(T)] – Ave[BDI(C)]
• If we have random allocation to treatment, R=T or
C, then
• ATE = Ave[BDI|R=T] – Ave[BDI|R=C]
54. Treatment-effect
heterogeneity
Methodology Research Group
• The treatment effect BDI(T)-BDI(C) is highly likely to
vary from one individual to another.
• We would like to know what background information
moderates (or predicts) the individual‟s treatment
effect. This is the essence of stratification.
• Let‟s say we have a genotypic marker (G=0,1). We‟d
like to look at association between G and BDI(T)-
BDI(C).
55. Again, we look at averages
Methodology Research Group
• We are concerned with the evaluation of the
comparison of
ATE|G=0
with
ATE|G=1
• This can be done by estimating and/or testing a
treatment by genotype interaction in a suitably-
powered RCT.
– (e.g.see the GENPOD trial: Lewis et al. BJPsych,
Vol 198, pp 464-471, 2011).
56. This is not rocket science ….
but what do geneticists usually do?
Methodology Research Group
• Investigators have a cohort of treated individuals.
• They have a measure of treatment outcome, say,
BDI(T), or treatment response, BDI(T)-BDI(0), on all
individuals within the cohort. Often, they label people
as „responders‟ or „non-responders‟.
• They investigate associations between treatment
outcome and genotypic markers (G).
57. A treatment outcome is not
a treatment effect Methodology Research Group
• BDI(T) is not BDI(T)-BDI(C)!
• Let the treatment effect be Δ.
• Then treatment outcome, BDI(T), is equal to
BDI(C) + Δ (to note the obvious!).
58. Confounding of treatment-
effects with prognosis Methodology Research Group
• The genotype (G) may be associated with both the
treatment effect (Δ) and with treatment-free outcome,
BDI(C), i.e. prognosis.
• Associating G with treatment outcome, BDI(T), cannot
distinguish between the two.
• Most importantly, it may be possible for treatment
outcome to be associated with G even when there is no
effect of treatment for anyone in the treated cohort!
59. … and evaluating the so-called
treatment-response doesn‟t help!
Methodology Research Group
• Δ = BDI(T)-BDI(C)
• Treatment response = BDT(T) – BDI(0)
= Δ + BDI(C) – BDI(0)
• Still confounded!
– At best, these investigations are identifying
candidates for further (more rigorous)
investigation.
– At worst, they are uncovering artefacts.
60. Our approach to stratified medicine
(personalised therapy) Methodology Research Group
• Predicting outcome after treatment (responders vs.
non-responders) is barely scratching the surface of
stratified medicine.
• Understanding the mechanism underlying the
stratification is the key scientific question, and the
methodological challenge.
61. Our “manifesto”
Methodology Research Group
• Personalised (stratified) medicine and treatment-effect
mechanisms evaluation are inextricably linked and
stratification without a corresponding mechanisms
evaluation lacks credibility;
• In the almost certain presence of mediator-outcome
confounding, mechanisms evaluation is dependent on
stratification for its validity;
• Both stratification and treatment-effect mediation can be
evaluated using a marker stratified trial design together
with detailed baseline measurement of all known prognostic
markers and other prognostic covariates;
62. Our methodological
solution Methodology Research Group
• Direct and indirect (mediated) effects should be
estimated through the use of instrumental variable
methods (the instrumental variable being the
predictive marker by treatment interaction)
together with adjustments for all known prognostic
markers (confounders)
– the latter adjustments contributing to increased
precision (as in a conventional analysis of
treatment effects) rather than bias reduction.
63. A purely prognostic marker
Methodology Research Group
Randomised
Outcome
Treatment
Prognostic
Marker
64. Prognostic Marker
Methodology Research Group
Treated
Outcome Untreated
Treatment effect
Marker Level
65. A prognostic marker as
a confounder Methodology Research Group
Randomised Putative
Treatment Mediator
U
Prognostic Clinical
Marker Outcome
66. Instrumental variables
Methodology Research Group
• If the causal influence of the prognostic marker
on the final outcome can be fully explained by its
influence on the intermediate, then the marker
can be used as an instrumental variable (or
instrument, for short).
• This is the theoretical rationale in the use of so-
called „Mendelian Randomisation‟.
67. An instrumental variable (IV)
Methodology Research Group
Random Treatment
Allocation (IV) Received Outcome
U
68. A prognostic marker as an
instrumental variable Methodology Research Group
Randomised Putative
Treatment Mediator
U
Prognostic Clinical
Marker
No direct link to outcome Outcome
69. Predictive markers
Methodology Research Group
• Although they may have direct predictive effects on
both intermediate and final outcomes, their essential
characteristic is that they moderate (influence)
treatment effects.
• If the treatment-effect moderation on final outcome is
wholly explained by the moderation of the effect of
treatment on the intermediate outcome, then the latter
(i.e. a treatment by marker interaction) can be used as
an instrument.
• A more subtle (and more realistic?) version of
Mendelian Randomisation.
70. Predictive marker
(may also be prognostic) Methodology Research Group
Randomised
Outcome
Treatment
Moderating
effect
Predictive
Marker
(moderator)
71. Predictive Marker
Methodology Research Group
Treated
Outcome
Untreated
Treatment effect
depends on marker
Marker Level
72. Putting it all together: potential
joint roles of predictive and
prognostic markers Methodology Research Group
Intermediate
Outcome U
(Mediator)
Predictive
Marker B
(moderator)
Final
A (Clinical)
Randomised Outcome
C
Treatment
Prognostic
Marker U – unmeasured confounders
(risk factor)
73. Potential roles of prognostic markers:
measured confounder
or instrumental variable Methodology Research Group
Intermediate
Outcome U
(Mediator)
B
Final
A (Clinical)
Randomised Outcome
C
Treatment
Prognostic
Marker U – unmeasured confounders
(risk factor)
Dotted line – pathway we might assume are absent
Alternatively, we might assume that there are no longer any Us
74. Option 1 – use prognostic marker(s) as a
measured confounder(s) and then assume
there is no hidden confounding (U)
Methodology Research Group
Intermediate
Outcome
(Mediator)
B
Final
A (Clinical)
Randomised Outcome
C
Treatment
Prognostic
Marker 1 Prognostic
(confounder)
Marker 2
(confounder)
75. Option 2 – Use as prognostic marker as
an instrumental variable
(Mendelian Randomisation) Methodology Research Group
Intermediate
Outcome U
(Mediator)
B
A
Final
(Clinical)
Randomised Outcome
C
Treatment
Prognostic
Marker U – unmeasured confounders
(instrument)
Using the prognostic marker as an instrumental variable
76. Potential problems with
Mendelian Randomisation
Methodology Research Group
• Assumption that there is no direct effect of the
genetic marker on final outcome frequently difficult
to justify, and practically impossible to verify.
– Dependent on prior knowledge.
• The marker is likely to be a rather weak instrument
(i.e. it‟s influence on the intermediate outcome is not
strong enough).
– This can lead to problems (see Session 1.)
• Probably wiser to use available prognostic markers
as observed confounders.
77. Potential role of
predictive markers Methodology Research Group
Intermediate
Outcome U
(Mediator)
Predictive
Marker B
(moderator)
Final
A (Clinical)
Randomised Outcome
C
Treatment
U – unmeasured confounders
Red dotted lines – pathways we might be justified in assuming are absent
78. Stratification & mediational
mechanisms evaluation
Methodology Research Group
Intermediate
Outcome U
(Mediator)
Predictive
Marker B
(moderator)
Final
A (Clinical)
Randomised Outcome
C
Treatment
U – unmeasured confounders
Using the treatment by marker interaction as an instrumental variable
79. Is the treatment by predictive
marker interaction a valid
instrument? Methodology Research Group
• Are we correct in assuming that there is no moderating
effect on pathway B?
• Are we correct in assuming that there is no moderating
effect on pathway C?
• Dependent on prior knowledge of the biology/biochemistry
of the system.
80. Theory-driven stratification
Methodology Research Group
• Prior scientific theory and preliminary evidence
strongly suggests that a given predictive marker has
its influence through a specific mechanism (the
putative mediator).
• No reason to expect that the moderating effect of the
predictive marker works via a pathway not associated
with the above mechanism (i.e. we assume that the
treatment by marker interaction – moderation – is a
valid instrument).
81. Using strong theory and all
available prognostic marker
information Methodology Research Group
Intermediate
Outcome U
(Mediator)
Predictive
Marker B
(moderator)
Final
A (Clinical)
Randomised Outcome
C
Treatment
Prognostic
Marker(s)
as Confounder(s)
U – unmeasured confounders
Using the treatment by marker interaction as an instrumental variable
82. Complicated but Viable!!
Methodology Research Group
• Statistical methods widely available to estimate the
pathways of this model (we won‟t worry about the
technical details).
• Health Warning!!
• This model is pretty complex and is dependent on a
lot of assumptions. Are these assumptions – i.e. the
theory - defensible? Invalid assumptions lead to
invalid solutions.
83. Real examples –
We don‟t have any! Methodology Research Group
• We know of no existing examples of the use of this
design – we are presently writing it up for publication.
• Examples from our mental health trials involve
retrospective analyses of archived data.
• Four funded EME trials are under way:
– Ketamine ECT in depression (Ian Anderson et al.);
– Minocycline and negative symptoms (Bill Deakin et
al.);
– Worry Intervention Trial (Freeman et al.);
– DBT for depression (Lynch et al.);
– but none fully utilise biomarker information as
described here.
84. A computer-simulated
example Methodology Research Group
• Trial with 1000 participants
– (500 treated, 500 controls).
– Quantitative outcome, y.
• Binary predictive marker (x10):
– Treatment effect on mediator (m) in its absence is
10 units; in its presence 60 units.
– Moderating effect of x10 on outcome solely through
the mediator (x10 known to be an IV).
– Variants of x10 equally probable (50:50).
• Nine prognostic uncorrelated binary markers x1-x9.
– All nine are confounders.
– Details of their creation of no consequence, here.
85. The true model (mediator)
Methodology Research Group
Mediator (m):
m=5*x1+5*x2+5*x3+5*x4+5*x5+5*x6+5
*x7+5*x8+5*x9+5*x10+10*treat+50*x
11+e12
Where x11 = treat*x10
(i.e. The treatment by marker interaction)
e12 is a random „error‟ term
“ * ” is a multiplication sign.
86. The true models (outcome)
Methodology Research Group
Outcome (y):
y=5*x1+5*x2+5*x3+5*x4+5*x5+5*x6+5
*x7+5*x8+5*x9+5*x10+2*m+10*treat
+e13
e13 is a random „error‟ term (uncorrelated
with e12).
There is no x11 (interaction) in this model.
THERE ARE NO UNMEASURED COMMON
CAUSES
(i.e. x1-x9, and x10, are all measured)
87. Simple summaries
Methodology Research Group
---------------------------------------------------------------------------
-> treat = 0
Variable | Obs Mean Std. Dev. Min Max
------------+-------------------------------------------------------------
m | 500 74.83 7.58 55.22 97.78
y | 500 174.47 22.27 116.21 247.78
--------------------------------------------------------------------------
-> treat = 1
Variable | Obs Mean Std. Dev. Min Max
------------+-------------------------------------------------------------
m | 500 108.92 28.42 55.66 159.70
y | 500 252.91 61.10 124.59 372.49
Note lack of homogeneity of standard deviations across the groups.
TREATMENT GROUP MUCH MORE VARIABLE (AS WE MIGHT EXPECT).
88. Naïve analysis methods
Methodology Research Group
• I won‟t bother to describe these in detail (but see
below).
• In the psychological and social science literature they
will be dominated by approaches similar to those
advocated by Baron & Kenny (about 17000 citations!)
• At the more hi-tech end of medicine they‟ve rarely got
round to using the naive methods!
89. Let‟s pretend we‟ve not
measured x1-x9:
Methodology Research Group
i.e. there are indeed „unmeasured‟
common causes
An instrumental variable regression in Stata:
ivregress 2sls y treat x10 (m = x11), first
This is a two-stage least-squares procedure which
simultaneously estimates the effect of treatment on
m (the first-stage regression), the effect of m on y,
and direct effect of treatment on y (the second
stage).
90. The first-stage regressions
Methodology Research Group
------------------------------------
m | Coef. Std. Err.
---------+--------------------------
treat | 10.07 0.63
x11 | 50.47 0.90
-------------------------------------
91. The second-stage
regressions Methodology Research Group
-------------------------------------
y | Coef. Std. Err.
---------+---------------------------
m | 2.00 0.02
treat | 10.39 0.87
-------------------------------------
92. Naïve methods:
the 2nd-stage regression
Methodology Research Group
Use ordinary least-squares to regress y on x10,
m and treat
regress y m x10 treat
------------------------------------------
y | Coef. Std. Err
-------------+----------------------------
m | 2.19 0.02
treat | 3.67 0.75
------------------------------------------
DIRECT EFFECT OF TREATMENT SEVERERLY BIASED.
93. Now use all available data
Methodology Research Group
ivregress 2sls y treat x1 x2 x3 x4 x5 x6 x7 x8 x9 x10
(m = x11), first
1st stage: Coef. Std. Err.
treat | 9.77 0.26
x11 | 50.73 0.37
2nd stage:
m | 2.01 0.01
treat | 10.01 0.55
CONSIDERABLE GAIN IN PRECISION
Measurement of prognostic markers not essential, but it
makes the design more efficient (i.e. get away with a
smaller trial) – perhaps the difference between a viable
trial and one that‟s just not feasible.
94. „Naïve‟ 2nd-stage regression
using all data Methodology Research Group
regress y x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 m treat
-------------------------------------
y | Coef. Std. Err.
-------------+-----------------------
m | 2.00 0.01
treat | 10.05 0.54
If (but only if) we‟ve measured all confounders then this is
valid and it is the most precise method. But ... we never know!
Returning to IV: there‟s a balance between bias and precision.
We don‟t get something for nothing.
95. The Key Ingredients
Methodology Research Group
• Convincing psychological theory concerning the
potential mechanism for mediation.
• Convincing theory to underline the belief that the
treatment by moderator (predictive marker)
interaction is a valid instrument.
• An appropriately powered trial for
– Valid evaluation of treatment-effect moderation –
on the mediator as well as on the outcome.
– Valid use of instrumental variables estimation to
evaluate the treatment-effect mechanisms
(mediation).
96. Design considerations
Methodology Research Group
• How big does the trial have to be? Considerably
larger than a conventional pragmatic trial.
• How strong does the moderating effect on the
mediator have to be?
– Our simulated example used a very strong
moderating effect.
– However, presumably it has to be reasonably
strong to be of any serious interest.
• What does the prevalence of the alleles for the
predictive biomarker have to be?
– We used 50:50 (maximum power).
– More likely to be of the order 90:10.
97. Conclusions
Methodology Research Group
• The scientific evaluation of stratified/personalised
medicines/therapies is inseparable from mechanisms
evaluation.
• So far, progress in trial design for mechanisms
evaluation appears to have been very limited.
– Interestingly, much more progress for the „softer‟
treatments (psychotherapies) than for hi-tech
medicines.
• Good design involves using prior scientific
knowledge/evidence and makes full use of data from
both prognostic and predictive markers.
• The required statistical methods are available and
reasonably straight forward to use.