Nicola Cooper Centre for Biostatistics and Genetic Epidemiology,

Evidence synthesis of competing interventions

when there is inconsistency in how effectiveness

outcomes are measured across studies Nicola Cooper

Centre for Biostatistics and Genetic Epidemiology, Department of Health Sciences, University of Leicester,

UK.http://www.hs.le.ac.uk/group/bge/

Acknowledgements: Tony Ades, Guobing Lu, Alex Sutton & Nicky Welton

                                     MULTIPLE EVENT OUTCOMES

• Often the main clinical outcome differs across trials

Inconsistent reporting (e.g. mean, median)

Change in outcomes used over time

• Possible to use the available data to inform the estimation of others

                                     EXAMPLE: Anti-viral for influenza• Three antiviral treatments for influenza

• Amantadine

• Oseltamivir

• Zanamivir

• No direct comparisons of the different antiviral treatments

• All trials compare antiviral to standard care

• Different outcome measures

• Time to alleviation of fever

• Time to alleviation of ALL symptoms

• Different summary statistics reported

• Median time to event

• Mean time to event

Time to alleviation of:

Number of trials Fever Symptoms

Amantadine vs. standard care 6 Oseltamivir vs. standard care 8 Zanamivir vs. standard care 5

• Oseltamivir and Zanamivir trials report median time to event of interest

• Amantadine trials report mean time to event of interest

• No direct comparison trials of all antivirals & standard care. Important to preserve within-trial randomised treatment comparison of each trial whilst combining all available comparisons between treatments (i.e. maintain randomisation).

DATA AVAILABLE

• In clinical studies with time to event data as the principal outcome, median time to event usually reported.

• However, for economic evaluations the statistic of interest is the mean => Area under survival curve

(i.e. provides best estimate of expected time to an event)

• Often mean time to an event canNOT be determined from observed data alone due to right-censoring

(i.e. actual time to an event for some individuals unknown either due to loss of follow-up or event not incurred by end of study)

                                     BACKGROUND

Time to symptoms alleviated (hours)

pro

po

rtio

n w

ith s

ymp

tom

s

0 100 200 300 400 500

0.0

0.2

0.4

0.6

0.8

1.0

control75mg bid

                                     PROBLEM: Mean

undefined

Last observation

censored => mean

undefined

Trt 1

Trt 2

                                     CALCULATING MEAN FROM MEDIAN TIME

• Simplest approach is to assume an Exponential distribution for time with influenza, thus assuming a constant hazard function over time,

• Probability of still having influenza at time t,

• At median time,

)texp()t(S

medmedmed t

)2ln(

t

)5.0ln()texp(5.0

• Mean time to event = 1/ & its corresponding variance = 1/2r, where r = number of events incurred during the study period

THREE STATE MARKOV MODEL

1 = transition rate (hazard) from influenza onset to alleviation of fever

2 = transition rate from alleviation of fever to alleviation of symptoms

1/1 = expected time from influenza onset to alleviation of fever

1/2 = expected time from alleviation of fever to alleviation of symptoms

(1/1 + 1/2) = expected time from influenza onset to alleviation of symptoms

Influenza Alleviation of fever

Alleviation of symptoms

h1 h2 Alleviation of ALL symptoms

Alleviation of ALL symptoms

1 2

Assumptions:

• Equal treatment effects in each period, jk

• Baseline hazard during second period same as in first period plus an additional random effect term, j

EVIDENCE SYNTHESIS MODEL

ln(i1) = j + jk = -ln(1) # i1, flu to fev

alleviated

ln(i2) = j + j + jk = -ln(2) # i2, fev to sym alleviated

i3 = i1 + i2 = 1/1 + 1/2 # i3, flu to sym alleviated

jk ~ Normal(dk , 2) # log hazard ratio

j ~ Normal(g , Vg) # random effect

where i = trial arms, j = trials, k = treatments.

Prior distributions specified for dk , g , 2, Vg

Exp(dk) is the ratio of hazards of recovery at any time for an individual on treatment k relative to an individual on the standard treatment

If exp(dk) < 1 then treatment k is superior

If exp(dk) > 1 then standard treatment is superior

Model fitted in WinBUGS and evaluated using MCMC simulation

HAZARD RATIO / RELATIVE HAZARD

caterpillar plot: e.d

0.5 0.6 0.7 0.8 0.9 1.0

                                     CATERPILLAR PLOT OF HAZARD RATIOS

Hazard Ratio

Improvement in rate of recovery

Oseltamivir

Amantadine

Zanamivir

                                     RANKING TREATMENTSBest 0% Best 33%

Best 67% Best 1%

standard care

rank

1 2 3 4

0.0

0.5

1.0

Oseltamivir

rank

1 2 3 4

0.0

0.5

1.0

Amantadine

rank

1 2 3 4

0.0

0.5

1.0

Zanamivir

rank

1 2 3 4

0.0

0.5

1.0

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30

Days

Pro

po

rtio

n o

f in

div

idu

als

                                     PROPORTION OF INDIVIDUALS IN EACH STATE

Standard careOseltamivirAmantadineZanamivir

InfluenzaAll symptoms alleviated

Fever alleviated

Ose

lta m

ivir

tria

ls 2

1 da

ys (

8 -22

%)

Zan

amiv

ir tr

ials

28

days

(7-

25%

)

                                     WEIBULL MODEL• Relaxes assumption of constant hazard ( = shape)

Time (t)

Ha

zard

= 1

0 < < 1

= 2

Exponential

                                     WEIBULL MODEL (cont.)

• Probability of still having influenza at time t,

S(t)=e -(t/) >0 (shape) >0 (scale)

• At median time,

0.5=e -(tmed/) tmed= (ln(2))1/

• If r out of n individuals still had symptoms at X days (i.e. end of trial), the proportion of censored individuals can be expressed as:

S(t(X)) =r/n=e -(t/)

Time to symptoms alleviated (hours)

pro

po

rtio

n w

ith s

ymp

tom

s

0 100 200 300 400 500

0.0

0.2

0.4

0.6

0.8

1.0

control75mg bid

                                     AVAILABLE TRIAL DATA

Trt 1

Trt 20.5

tmed X

r/n

                                     WEIBULL MODEL (cont.)

• Important to calculate E(S(t(X))|tmed) to take

account of the correlation between median time

(tmed) to alleviation of illness and proportion of

participants (r/n) still ill at X days as they are from the same trial dataset.

• Mean time to event (i.e. statistic of interest)

(1+1/)

                                     EVIDENCE SYNTHESIS MODEL

Assumptions:

• Equal treatment effects in each period, jk

• Baseline hazard during second period same as in first period plus an additional random effect term, j

ln(i1) = j + jk = ln(i1) + ln( (1+1/1)) # i1, flu to fev alleviated

ln(i2) = j + j + jk # i2, fev to sym alleviated

i3 = i1 + i2 = i3 (1+1/3) # i3, flu to sym alleviated

jk ~ Normal(dk, 2) # log hazard ratio

j ~ Normal(g, Vg) # random effect

where i = trial arms, j = trials, k = treatments.

Prior distributions specified for dk , g , 2, Vg

                                     EVIDENCE SYNTHESIS MODEL (cont.)

• and are the shape and scale parameters of a Weibull distribution respectively

• Model assumes the shape parameters are the same for time to alleviation of fever, 1, regardless of antiviral treatment & similarly for time to alleviation of symptoms, 3

• Due to lack of data, to estimate 1 set equal to 1 (i.e. exponential distribution). Could set constraint to ensure proportion still with fever at X days (i.e. end of trial) proportion still with symptoms

caterpillar plot: e.d

0.4 0.6 0.8 1.0

                                     CATERPILLAR PLOT OF HAZARD RATIOS

Improvement in rate of recovery

Zanamivir

Oseltamivir

Amantadine

                                     RANKING TREATMENTS

Best 0%

Best 1%

Best 20%

Best 79%

standard care

rank

1 2 3 4

0.0

0.5

1.0

zanamivir

rank

1 2 3 4

0.0

0.5

1.0

oseltamivir

rank

1 2 3 4

0.0

0.5

1.0

amantadine

rank

1 2 3 4

0.0

0.5

1.0

                                     CONCLUSIONS• Although amantadine is ranked “best” it does have serious

side effects (e.g. gastrointestinal symptoms, central nervous system) which are not taken into account in this analysis

• This type of model could inform the effectiveness parameters of a cost effectiveness decision model

• Allows multiple outcomes & indirect comparisons to be modelled within a single framework

• Appropriate model for nested outcomes (e.g. progression-free survival & overall survival)

• If non-nested outcomes, then a multivariate meta-analysis model, as developed for surrogate outcomes, more appropriate