Are we really including all relevant evidence? Considerations for network meta-analysis Beth Woods Centre for Health Economics, University of York with thanks to: David Scott, Neil Hawkins
Jun 19, 2015
Are we really including all relevant evidence? Considerations for
network meta-analysis
Beth Woods
Centre for Health Economics, University of York
with thanks to: David Scott, Neil Hawkins
What do we need for economic evaluation?
• A set of consistent relative treatment effects on the endpoint(s) of interest for all comparators of interest
• Network meta-analysis (NMA) or indirect comparison is typically used to estimate these effects
Including all relevant evidence
• To improve precision and avoid bias
• Demarcation of relevance i.e. what forms the relevant evidence space?
– Study set (comparators only)
– Endpoint space
Include Decision Comparators
• Three decision comparators (DC1, DC2, DC3)
• Direct comparisons only
Expand to Synthesis Comparators
• Second-order comparisons
• Third-order comparisons
All synthesis comparators here are secondary comparators
How far should we go?
• Connected network
• Increased precision
• Increased heterogeneity and inconsistency
– Scope for addressing these issues may increase
• Ability of decision makers to scrutinise NMA
Iterative searching
• At design stage knowledge of network may be limited – Value of a wider network unclear
– Can’t search for third-order comparisons unless you know the secondary comparators
• Iterative searching (Hawkins et al 2009) – Searching sequentially for primary comparators,
then secondary etc.
– Each iteration uncovers all direct, second-order, etc. comparisons
Endpoint space
Endpoint of interest
Statistic 1
Time point 1 Time point 2 Time point n
Statistic 2
Time point 1 Time point 2 Time point n
Statistic n
Time point 1 Time point 2 Time point n
Endpoint space
Endpoint space
Endpoint of interest
Statistic 1
Time point 1 Time point 2 Time point n
Statistic 2
Time point 1 Time point 2 Time point n
Statistic n
Time point 1 Time point 2 Time point n
Example (Woods et al 2010)
• Treatments for chronic obstructive pulmonary disease (COPD)
• Trial endpoints historically focused on measures of lung function and exacerbations
• Interest in the potential treatment effect on mortality
Available statistics
• Majority of trials reported mortality data only in adverse event reporting – Binary data
• New trials started analysing mortality endpoint using standard survival analysis methods – hazard ratios from Cox proportional hazards model
– includes additional information on time to event and censoring
• Challenge – incorporate all evidence, selecting preferred statistic where reported
BUGS shared parameter model
• Separate “loops” required to incorporate hazard ratio and binary data
• Treatment effect estimates (β’s) are the “shared parameters”
• Fixed effects model shown
Model for hazard ratio data (two arm trials)
• Normal likelihood 𝑥 𝑠,𝑘,𝑏 ~ 𝑁(ln (ℎ𝑟𝑠,𝑘,𝑏), 𝑠𝑒𝑠,𝑘,𝑏)
• Treatment effect model ln ℎ𝑟𝑠,𝑘,𝑏 = β𝑘 − β𝑏
β1 = 0
Model for hazard ratio data (multi-arm trials)
• Correlation in contrast data ln ℎ𝑟𝑠,𝑘,𝑏 = ln (ℎ𝑠,𝑘)−ln (ℎ𝑠,𝑏)
• Could model data as multivariate normal
• Or convert contrast data to arm level data
Converting contrast to arm-level data
• Set ℎ𝑠,𝑏 = 0
• Compute 𝑠𝑒(ℎ𝑠,𝑏) – Use fact that 𝑣𝑎𝑟(ℎ𝑟𝑠,𝑘,𝑏) = 𝑣𝑎𝑟 ℎ𝑠,𝑘 + 𝑣𝑎𝑟(ℎ𝑠,𝑏) – If have variance for hazard ratios comparing 2 vs. 1, 3
vs. 1 and 2 vs. 3 then can solve for 𝑣𝑎𝑟(ℎ𝑠,𝑏)
– If not assume common 𝑠𝑑(ℎ𝑠,𝑖)
• Normal likelihood 𝑥 𝑠,𝑘 ~ (ln (ℎ𝑠,𝑘), 𝑠𝑒𝑠,𝑘)
• Treatment effect model ln ℎ𝑠,𝑘 = 𝛼𝑠 + β𝑘 − β𝑏
Model for binary data
• Binomial likelihood 𝑟𝑠,𝑘~𝐵𝑖𝑛(𝐹𝑠,𝑘 , 𝑛𝑠,𝑘)
• Derive arm-specific log cumulative hazard
ln 𝐻𝑠,𝑘 = ln (− ln 1 − 𝐹𝑠,𝑘 )
• If we assume proportional hazards the ratio of cumulative hazards must equal the ratio of instantaneous hazards
• Treatment effect model ln 𝐻𝑠,𝑘 = 𝛼𝑠 + 𝛽𝑘 − 𝛽𝑏
Data # Data set descriptors
list(LnObs = 5, BnObs = 2, nTx = 4, nStudies = 3)
# Log hazard ratio data
# Binary data
Lstudy[] Ltx[] Lbase[] Lmean[] Lse[] multi[]
1 1 1 0 0.066 1
1 2 1 0.055 0.063 1
1 3 1 -0.154 0.070 1
1 4 1 -0.209 0.072 1
2 2 1 -0.276 0.203 0
Bstudy[] Btx[] Bbase[] Br[] Bn[]
3 3 1 1 229
3 1 1 1 227
BUGS code - model
#For hazard ratio reporting studies
for(ii in 1:LnObs ){
Lmean[ii] ~ dnorm(Lmu[ii],Lprec[ii])
Lprec[ii] < - 1/pow(Lse[ii],2)
Lmu[ii] < - alpha[Lstudy[ii]]*multi[ii] + beta[Ltx[ii]] - beta[Lbase[ii]] }
#For binary data reporting studies
for(ss in 1:BnObs){
Br[ss] ~ dbin(cumFail[ss], Bn[ss])
cumFail[ss] < - 1-exp(-1*exp(logCumHaz[ss]))
logCumHaz[ss] < - alpha[Bstudy[ss]] + beta[Btx[ss]] - beta[Bbase[ss]] }
Shared parameters
Shared parameter models
• Important when different statistics are reported for a specific endpoint – Median, mean and % event free at 21/28 days
(Welton et al 2008)
– Mean events per week, patients achieving ≥1 event by week 10
– Mean values of continuous outcome X, proportion achieving ≥X*
• Additional assumptions required
• Bayesian approach required?
Endpoint space
Endpoint of interest
Statistic 1
Time point 1 Time point 2 Time point n
Statistic 2
Time point 1 Time point 2 Time point n
Statistic n
Time point 1 Time point 2 Time point n
Why include repeated measures?
• Relative treatment effects for a study s and comparison k,b can be informed by data at more than one time point
• Precision and bias
• Explore temporal changes in treatment effects
– Within trial
– To inform extrapolation
Existing work
• Within-subject correlation generally ignored (e.g. Dakin et al 2011, Ding & Fu 2013)
• Models for time-varying treatment effects
– Treatment-specific time-by-treatment interactions
– Unrelated piecewise (e.g. Dakin et al 2011) or explicit functional form (e.g. Ding & Fu 2013; Jansen 2011)
Models for decision analysis?
• Sparse data and varying follow-up across treatments – Treatment specific time interactions may not be
estimable or may be very uncertain
• Could impose more structure (assumptions): – Exchangeable and related interactions
– Single common interaction
– Likely to make these assumptions in the decision model…
• Could use external data (observational, expert) – Inform functional form or priors on parameters
Bias in repeated measures NMA
• Changing potential for bias over time
– Long term follow-up may depend on study success
– Trial continuation may depend on efficacy / futility
• Changes in confounders over time may not reflect clinical practice
– Cross-over, adherence, comorbidities
Endpoint space
Multiple outcome (N)MA
• Multivariate meta-analysis • Jointly modelling two or more endpoints
– Incorporate within-study correlation between treatment effects
– Incorporate between-study correlation between treatment effects
• Models typically assume simple correlation (i.e. linear relationship)
• Increased precision – Particularly when correlation is high and reporting
incomplete
• Reduced bias if outcome reporting is selective
Additional benefits for economic evaluation
• Quantify surrogate relationships – E.g. across trials does a treatment effect on PFS lead to a
treatment effect on OS?
– Makes better use of evidence base than common decision modelling approaches • Using trial data from an individual study
• Using independent MTCs for PFS and OS
• Using PFS data and making assumptions about PPS | treatment
• More accurate representation of parameter uncertainty – Provides joint distribution of treatment effects on different
endpoints
Could we incorporate a little more structure?
• Models reflect simple linear relationships between treatment effects
– Unlikely to correspond to any underlying biological process
• Could impose model on surrogate-final endpoint relationship
– Test treatment independence of this relationship
• Examples of this exist (Welton et al 2008)
Potential benefits of structure
• Test modelling assumptions/estimate additional model parameters
• Reduced risk of over-parameterisation
– More appropriate estimation of parameter uncertainty
• Easier to incorporate external data
• Careful consideration needs to be given to the study space
• Extensions to synthesis models can help us include additional relevant endpoint data – They may also provide information about
important aspects of model structure
• Avoiding structure in the synthesis model may lead to stronger structural assumptions in decision models
Conclusion
References Hawkins, N., Scott, D. a & Woods, B. How far do you go? Efficient searching for indirect evidence. Med. Decis. Making 29, 273–81 (2009).
Woods, B. S., Hawkins, N. & Scott, D. a. Network meta-analysis on the log-hazard scale, combining count and hazard ratio statistics accounting for multi-arm trials: a tutorial. BMC Med. Res. Methodol. 10, 54 (2010).
Welton, N. J., Cooper, N. J., Ades, A. E., Lu, G. & Sutton, A. J. Mixed treatment comparison with multiple outcomes reported inconsistently across trials : Evaluation of antivirals for treatment of influenza A and B. Stat. Med. 27, 5620–5639 (2008).
Dakin, H. a et al. Mixed treatment comparison of repeated measurements of a continuous endpoint: an example using topical treatments for primary open-angle glaucoma and ocular hypertension. Stat. Med. 30, 2511–2535 (2011).
Ding, Y. & Fu, H. Bayesian indirect and mixed treatment comparisons across longitudinal time points. Stat. Med. 32, 2613–28 (2013).
Jansen, J. P. Network meta-analysis of survival data with fractional polynomials. BMC Med. Res. Methodol. 11, 61 (2011).