Model criticism for evidence synthesis models of infectious disease Anne Presanis MRC Biostatistics Unit in collaboration with David Ohlssen (Novartis) David Spiegelhalter (Cambridge) Daniela De Angelis (MRC BSU) Acknowledgements: Paul Birrell (MRC BSU), Tony Ades (Bristol), Richard Pebody (HPA) UCL Biostatistics Symposium 18 Oct 2012 A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 1 / 26 Introduction Outline 1 Introduction Motivation Evidence synthesis 2 Criticising evidence syntheses Influence & Conflict 3 Conflict diagnostics General method 4 Examples HIV prevalence in women Influenza severity Multivariate eg: rat weights 5 Discussion A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 2 / 26 Introduction Motivation Motivation The implementation and evaluation of public health policies aimed at control and prevention of epidemics rely crucially on knowledge of fundamental aspects of the disease of interest, such as prevalence, incidence and severity. These characteristics are typically not easily measurable as little direct data are available on them. There is plenty of indirect information on functions of these quantities. Estimation through the synthesis of diverse and fragmented sources of evidence is often feasible. A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 3 / 26 Introduction Evidence synthesis Evidence synthesis: long-established idea Methods for combining evidence are not new: The Bayesian paradigm combining prior knowledge with new Meta-analysis combining studies of same type Confidence Profile Method [Eddy et al (1992)] combining information of different types/study designs Multi-parameter evidence synthesis [Spiegelhalter et al (2004), Ades & Sutton (2006)] health technology assessment [Spiegelhalter & Best (2003), Demiris & Sharples (2006)] infectious disease epidemiology [Welton & Ades (2005), Goubar et al (2008), Sweeting et al (2008), De Angelis et al (2009), Presanis et al (2011a), Birrell et al (2011), Albert et al (2011)] environmental epidemiology [Ryan (2008), Jackson et al (2008), Jones et al (2009)] other fields [Henderson et al (2010), Clark et al (2010)] A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 4 / 26 Introduction Evidence synthesis Statistical formulation Interest: estimation of θ = (θ1 , θ2 . . . , θk ) on the basis of a collection of data y = (y1 , y2 . . . , yn ) Each yi provides information on a single component of θ (“direct” data), or a function of one or more components, i.e. on a quantity ψi = f (θ) (“indirect” data) Thus inference is conducted on the basis of both direct and indirect information. Maximum likelihood: L = Qn i=1 Li (yi Bayesian: p(θ | y) ∝ p(θ) × L A. M. Presanis (MRC BSU) | θ) Conflict in complex evidence synthesis UCL, 18 Oct 2012 5 / 26 Introduction Evidence synthesis Graphical representation (DAG) Basic parameters θ1 ... θi θi+1 ... θk Functional parameters ψ1 ... ψj ψj+1 ... ψn Data y1 ... yj yj+1 ... yn A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 6 / 26 Introduction Evidence synthesis Simple example: HIV prevalence HIV prevalence π δ π(1 − δ) Proportion diagnosed Prevalence of undiagnosed infection y1 A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 7 / 26 Introduction Evidence synthesis Simple example: HIV prevalence HIV prevalence π δ π(1 − δ) y1 A. M. Presanis (MRC BSU) Proportion diagnosed Prevalence of undiagnosed infection y2 Conflict in complex evidence synthesis UCL, 18 Oct 2012 7 / 26 Introduction Evidence synthesis Simple example: HIV prevalence HIV prevalence π δ Prevalence of undiagnosed infection π(1 − δ) y1 A. M. Presanis (MRC BSU) y2 Proportion diagnosed y3 Conflict in complex evidence synthesis UCL, 18 Oct 2012 7 / 26 Introduction Evidence synthesis Motivation II Evidence synthesis leads to complex probabilistic models Combination of all available relevant data sources ideally should lead to more precise estimates Multiple sources informing a single parameter ⇒ potential for conflicting evidence Sparsity of data ⇒ parameters unidentifiable without further model constraints, e.g. exchangeability We can already fit such models The next step? How do we assess and criticise them? A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 8 / 26 Criticising evidence syntheses Influence & Conflict What are the components of inference? Model criticism requires understanding the role of three components: the model structure (part of prior? [Box (1980), Efron (2010)]); the prior; and the likelihood. Influence & Conflict Detection & quantification: When does heterogeneity become conflict? Cross-validatory / posterior / mixed predictive methods [Rubin (1984), Gelman et al (1996), Bayarri & Berger (1999), Marshall & Spiegelhalter (2007)] Conflict diagnostics, including prior-predictive [Box (1980), O’Hagan (2003), Evans & Moshonov (2006), Gasemyr & Natvig (2009)] Resolution: Drop suspect/biased data Expand the model to account for biased data May require evidence expansion (external evidence) for identifiability A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 9 / 26 Conflict diagnostics General method Defining/measuring conflict Existing methods can be generalised to the question: How do we compare two sets of evidence? prior-data [prior-predictive methods] posterior-data [posterior-predictive methods] data-data [conflict diagnostics] More generally: in a DAG, consider node-splitting to carry out posterior-posterior comparison for two disjoint sets of evidence (data & priors) informing a single node [Marshall & Spiegelhalter (2007), Dias et al (2010), Ohlssen & Spiegelhalter, Presanis et al (to be submitted)] A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 10 / 26 Conflict diagnostics General method Simple example: prior-likelihood conflict Informative prior Reference prior Informative prior θ θ1 y y A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis θ2 UCL, 18 Oct 2012 11 / 26 Conflict diagnostics General method More generally, conflict at any node θ pa(θ) cp(θ) θ si(θ) ch(θ) pa(θ) = Parents ch(θ) = Children si(θ) = Siblings cp(θ) = Co-parents A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 12 / 26 Conflict diagnostics General method More generally, conflict at any node θ pa(θ) cp(θ) θ si(θ) pa(θ) ch(θ) pa(θ) = Parents ch(θ) = Children si(θ) = Siblings cp(θ) = Co-parents A. M. Presanis (MRC BSU) cp1(θ) θ1 θ2 si(θ) ch1(θ) ch2(θ) cp2(θ) Conflict in complex evidence synthesis UCL, 18 Oct 2012 12 / 26 Conflict diagnostics General method A general conflict measure c θ1 θ2 0 δ = g(θ1) − g(θ2) For general ‘separator’ node(s) θ, split into θ1 based on data y1 and θ2 based on data y2 . Define δ = g (θ1 ) − g (θ2 ) where g is a function such that it is reasonable to assume a Uniform prior for g (θ1 ). Then define c = Pr {pδ (δ|y1 , y2 ) < pδ (0|y1 , y2 )} where pδ is the posterior density of δ. A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 13 / 26 Conflict diagnostics General method Defining c Univariate node-split For symmetric unimodal distributions p(δ|y1 , y2 ), define tail area c = 2 min(p(δ > 0|y1 , y2 ), 1 − p(δ > 0|y1 , y2 )) For skewed and/or multi-modal distributions, we may take a kernel density estimate to obtain c. Multivariate node-splits: p(δ|y1 , y2 ) multivariate normal Define the standardised discrepancy measure ∆ = Ep (δ)T Covp (δ)−1 Ep (δ) and compare to a X 2 distribution: c = 1 − Pr Xk2 ≤ ∆ . A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 14 / 26 Conflict diagnostics General method Defining c Multivariate node-splits: p(δ|y1 , y2 ) symmetric/uni-modal Samples δ i from posterior, calculate Mahalanobis distance from mean ∆i = {δ i − Ep (δ)}T Covp (δ)−1 {δ i − Ep (δ)} Then define c = Pr {∆i > ∆}, the proportion of points that are further away from the mean than is 0. Multivariate node-splits: p(δ|y1 , y2 ) skew/multi-modal Obtain a kernel density estimate of the posterior to calculate the probability that the posterior density at δ is less than (at a lower contour than) at 0: c = Pr {p(δ|y1 , y2 ) < p(0|y1 , y2 )} A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 14 / 26 Examples HIV prevalence in women Example: HIV prevalence in women Ades & Cliffe, Medical Decision Making (2002) Already diagnosed? HIV infection? Risk group? SSA a Yes Yes c No 1−c Yes IDU b d Rest 1−f Yes g No 1−g No 1−d (1 − a − b) f No Yes e No Yes h No 1−h 1−e A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 15 / 26 Examples HIV prevalence in women Example: HIV prevalence in women Ades & Cliffe, Medical Decision Making (2002) Description of data Parameter Proportion born SSA p1 = a Proportion IDU last 5 years p2 = b HIV prevalence, women born in SSA p3 = c HIV prevalence in female IDUs p4 = d HIV prevalence, women not born in SSA p5 = [db + e(1 − a − b)]/(1 − a) Overall HIV seroprevalence in pregnant women p6 = ca + db + e(1 − a − b) Diagnosed HIV in SSA women, out of all diagnosed HIV p7 = fca/[fca + gdb + he(1 − a − b)] Diagnosed HIV in IDUs, out of non-SSA diagnosed HIV p8 = gdb/[gdb + he(1 − a − b)] Overall proportion of HIV diagnosed p9 = [fca + gdb + he(1 − a − b)]/[ca + db + e(1 − a − b)] Proportion of infected IDUs diagnosed p10 = g Proportion of serotype B in infected women from SSA p11 = w Proportion of serotype B in infected women not from SSA p12 = [db + we(1 − a − b)]/[db + e(1 − a − b)] A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 15 / 26 Examples HIV prevalence in women HIV eg: two types of node-split a b c d e ... p1 p2 p3 p4 p5 p6 n1 y1 n2 y2 A. M. Presanis (MRC BSU) n3 y3 n4 y4 n5 y5 Conflict in complex evidence synthesis ... ... n6 y6 ... UCL, 18 Oct 2012 16 / 26 Examples HIV prevalence in women HIV eg: two types of node-split (a) basic parameter level split a1 a2 p1 b ... p2 n1 y1 A. M. Presanis (MRC BSU) n2 y2 ... ... ... p5 p6 ... ... ... n5 y5 Conflict in complex evidence synthesis n6 y6 ... ... UCL, 18 Oct 2012 16 / 26 Examples HIV prevalence in women HIV eg: two types of node-split (b) data level split p1,1 a b p1,2 p2 n1 y1 A. M. Presanis (MRC BSU) ... n2 y2 ... ... ... p5 p6 ... ... ... n5 y5 Conflict in complex evidence synthesis n6 y6 ... ... UCL, 18 Oct 2012 16 / 26 Examples HIV prevalence in women HIV eg: direct vs indirect evidence δ = x1 − x2 , x ∈ {a, b, c, d, g , w } 100 250 c = 0.165 bw = 0.004 c = 0.01 bw = 0.001 150 80 c = 0.155 bw = 0.001 200 60 150 100 40 100 50 20 50 0 0 0.00 0.05 0.10 0.15 0.20 0 0.00 0.01 0.02 a 0.03 0.04 0.05 120 0.01 0.02 0.03 0.04 0.05 c 5 15 c = 0.009 bw = 0.002 100 0.00 b c = 0.31 bw = 0.033 c = 0.353 bw = 0.01 4 80 10 3 60 2 40 5 1 20 0 0 0.00 0.01 0.02 0.03 d A. M. Presanis (MRC BSU) 0.04 0.05 0 0.0 0.2 0.4 0.6 0.8 1.0 g Conflict in complex evidence synthesis 0.00 0.10 0.20 0.30 w UCL, 18 Oct 2012 17 / 26 Examples HIV prevalence in women HIV eg: data-level cross-validation δ = px1 − px2 , 50 40 x ∈ 1, . . . , 12 100 c = 0.169 bw = 0.003 c = 0.009 bw = 0.002 80 250 200 30 60 150 20 40 100 10 20 0 50 0 −0.06 −0.02 0.02 0 −0.01 p1 2500 2000 0.01 0.03 −0.010 p2 2500 c = 0.038 bw = 0 60 50 40 30 20 10 0 c = 0.161 bw = 0.001 −0.02 p3 c = 0.249 bw = 0 2000 0.000 4 c = 0.245 bw = 0.034 3 4 3 1500 1000 1000 2 2 500 500 1 1 0 0 0 0e+00 −5e−04 p5 4 5e−04 p6 2 1.0 4 1 0.5 2 0 6 0.0 −0.6 −0.2 0.2 p9 A. M. Presanis (MRC BSU) 0.2 0.6 0.0 p10 0.5 −0.2 0.2 p8 4 c = 0.358 bw = 0.019 3 c = 0.042 bw = 0.032 2 1 0 −0.5 −0.8 p7 8 c = 0.129 bw = 0.032 c = 0.364 bw = 0.047 0 −0.2 2.0 c = 0.314 bw = 0.05 1.5 3 0.02 p4 1500 −1e−03 c = 0.008 bw = 0.003 0 −0.3 −0.1 0.1 p11 Conflict in complex evidence synthesis −0.6 −0.2 0.2 p12 UCL, 18 Oct 2012 18 / 26 Examples Influenza severity Influenza severity CF R = cD|H × cH|S × cS|Inf = P r{D|H} × P r{H|S} × P r{S|Inf } sCHR = cH|S = P r{H|S} ICU Death Hospitalised Symptomatic Infected A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 19 / 26 Examples Influenza severity 2009 pandemic: model for the first wave CFR CIR CHR cD|H cI|H cH|S ND NI NH dD Sero-survey data, 2008 πbaseline cS|Inf NS π IAR NInf Sero-survey data, 2009 NP op dH OD OD:H OI:H OH HPA estimates of NS based on GP consultations/virological testing Deaths reported to CMO/HPA Web-based hospital surveillance Full model [Presanis et al (2011)] A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 20 / 26 Examples Influenza severity 2009 pandemic: model for the first wave Sero-survey data, 2008 πbaseline cS|Inf NS π IAR NInf Sero-survey data, 2009 NP op Parent model A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 20 / 26 Examples Influenza severity 2009 pandemic: model for the first wave cD|H ND cI|H NI cH|S NH dD NS dH OD OD:H OI:H OH HPA estimates of NS based on GP consultations/virological testing Deaths reported to CMO/HPA Web-based hospital surveillance Child model A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 20 / 26 Examples Influenza severity NS : parent-child conflict Child model cD|H ND πbaseline cI|H NI cH|S NS1 NH dD cS|Inf NS2 OD:H OI:H OH 2009 NP op Parent model dH OD π IAR NInf 2008 HPA estimates of NS based on GP consultations/virological testing Deaths reported to CMO/HPA Web-based hospital surveillance A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 21 / 26 Examples Influenza severity Conflict & influence: ln(NS ) <1 6 8 10 12 14 1.0 1−4y 1.0 5−14y 1.0 0.8 0.8 0.8 0.6 0.6 0.6 0.4 0.4 0.4 0.2 0.2 0.2 0.0 0.0 0.0 16 6 8 10 12 14 16 ln(Number symptomatic) 1.0 ln(Number symptomatic) 1.0 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 25−44y A. M. Presanis (MRC BSU) 45−64y 6 8 10 12 14 ln(Number symptomatic) 65+ Conflict in complex evidence synthesis 16 15−24y 6 8 10 12 14 16 ln(Number symptomatic) 2: Parent model 3: Child model ● UCL, 18 Oct 2012 22 / 26 Examples Influenza severity Conflict & influence: ln(NS ) <1 6 8 10 12 14 1.0 1−4y 1.0 5−14y 1.0 0.8 0.8 0.8 0.6 0.6 0.6 0.4 0.4 0.4 0.2 0.2 0.2 0.0 0.0 0.0 16 6 8 10 12 14 16 ln(Number symptomatic) 1.0 ln(Number symptomatic) 1.0 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 25−44y A. M. Presanis (MRC BSU) 45−64y 6 8 10 12 14 ln(Number symptomatic) 65+ Conflict in complex evidence synthesis 16 15−24y 6 8 10 12 14 16 ln(Number symptomatic) 1: Full model 2: Parent model ● 3: Child model UCL, 18 Oct 2012 22 / 26 Examples Influenza severity δ = ln(NS1 ) − ln(NS2 ) 1.0 1.0 <1 0.8 <1 c=0 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 6 8 10 12 14 16 −4 −2 0 2 4 2 4 ln(Number symptomatic) 1.0 1.0 5−14y 0.8 5−14y c = 0.02 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 6 8 10 12 14 16 −4 A. M. Presanis (MRC ln(Number BSU) Conflict in complex evidence synthesis symptomatic) −2 0 UCL, 18 Oct 2012 23 / 26 Examples Influenza severity δ = ln(NS1 ) − ln(NS2 ) 0.6 c1 = 0.058 c2 = 0.063 bw = 0.5 0.5 0.4 0.3 0.2 0.1 0.0 −4 −2 0 2 4 Difference function, 65+ A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 23 / 26 Examples Multivariate eg: rat weights Bivariate normal model for rat weights Gelfand et al (1990) random linear growth yij µij ∼ N(µij , σ 2 ) = φia + φib tj φi ∼ MVN2 (β, Ω) A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 24 / 26 Examples Multivariate eg: rat weights Bivariate normal model for rat weights Gelfand et al (1990) random linear growth yij µij φ1ia Ω\i φ\ib φ\ia φ2ia φ1ib ∼ N(µij , σ 2 ) β\i φ2ib = φia + φib tj φi ∼ MVN2 (β, Ω) σ2 yij tj y\ij j A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 \i 24 / 26 Examples Multivariate eg: rat weights Outlier: rat 9 2 0 + + ++ + ++ +++ + + + + + + + + + + + +++ + + + +++ + ++ + + ++ +++ + ++ + + ++ +++++++ ++ + +++ ++ ++++++ ++ + +++ + ++ + + ++++ + ++ + + + + + ++ +++++ + + + + ++ ++ + ++++ ++ ++ + + ++++ +++++ ++ + + + ++ ++ + + ++ + +++ + ++ ++ + +++ ++ +++++ +++++++++ + ++ + + +++ + + ++ + +++ +++ ++ ++ ++ + ++++ + ++++ + + ++ ++ + + + + + phi2 * −2 + −4 Rat 9 c = 0.006 −150 −100 −50 0 50 phi1 A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 25 / 26 Discussion Discussion Importance of understanding what is driving conclusions, particularly when combining multiple sources of evidence (informative priors, data from varied sources) A number of methods available: importance of systematic model criticism. Conflict p-values - useful continuous scale. When does influence lead to conflict/inconsistency? Choice of reference priors? Choice of function g such that δ = g (θ1 ) − g (θ2 ) is symmetric & unimodal? Multiple testing problem. Once conflict detected: why is it occurring? Biases? Incorrect model? Accommodate the inconsistency through the introduction of bias parameters? A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 26 / 26 References Evidence synthesis: general and HTA D.M. Eddy, V. Hasselblad, R. Shachter Meta-Analysis by the Confidence Profile Method. Academic Press, 1992. D.J. Spiegelhalter, N.G. Best Bayesian approaches to multiple sources of evidence and uncertainty in complex cost-effectiveness. modelling. Statist. Med., 22(23):3687–3709, 2003. D.J. Spiegelhalter, K.R. Abrams, J.P. Myles Bayesian approaches to clinical trials and health-care evaluation. Wiley, 2004. A.E. Ades, A.J. Sutton Multiparameter evidence synthesis in epidemiology and medical decision-making:current approaches. J. R. Stat. Soc. A, 169:5–35, 2006. N. Demiris, L.D. Sharples Bayesian evidence synthesis to extrapolate survival estimates in cost-effectiveness studies. Statist. Med., 25(11):1960–1975, 2006. A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 27 / 26 References Evidence synthesis: epidemiology & ecology N.J. Welton, A.E. Ades A model of toxoplasmosis incidence in the UK: evidence synthesis and consistency of evidence. J. R. Stat. Soc. C, 54(2):385–404, 2005. L. Ryan Combining data from multiple sources, with applications to environmental risk assessment. Statist. Med., 27(5):698–710, 2008. C. Jackson, S. Richardson, N. Best Studying place effects on health by synthesising individual and area-level outcomes. Soc. Sci. & Med., 67(12):1995–2006, 2008. M.J. Sweeting, D. De Angelis, M. Hickman, A.E. Ades Estimating hepatitis C prevalence in England and Wales by synthesizing evidence from multiple data sources. Assessing data conflict and model fit. Biostatistics, 9(4):715–734 2008. D. De Angelis, M.J. Sweeting, A.E. Ades, V. Hope, M. Ramsay An evidence synthesis approach to estimating Hepatitis C prevalence in England and Wales. Stat. Meth. Med. Res., 18(4):361–379, 2009. D.R. Jones, J.L. Peters, L. Rushton, A.J. Sutton, K.R. Abrams Interspecies extrapolation in environmental exposure standard setting: A Bayesian synthesis approach. Regul. Toxicol. & Pharm., 53(3):217–225, 2009. J.S. Clark, D. Bell, C. Chu, B. Courbaud, M. Dietze, M. Hersh, J. HilleRisLambers, I. Ibáñez, S. LaDeau, S. McMahon, J. Metcalf, J. Mohan, E. Moran, L. Pangle, S. Pearson, C. Salk, Z. Shen, D. Valle, P. Wyckoff High-dimensional coexistence based on individual variation: a synthesis of evidence. Ecol. Monog., 80(4):569–608, 2010. D.A. Henderson, R.J. Boys, D.J. Wilkinson Bayesian calibration of a stochastic kinetic computer model using multiple data sources. Biometrics, 66(1):249–256, 2010. I. Albert, E. Espie, H. de Valk, J.-B. B. Denis A Bayesian evidence synthesis for estimating campylobacteriosis prevalence. Risk Analysis, 31(7):1141–1155, 2011. A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 28 / 26 References Model criticism references D.J. Spiegelhalter, N.G. Best, B.P. Carlin, A. van der Linde Bayesian measures of model complexity and fit. J. R. Statist. Soc. B, 64(4):583–639, 2002. G.E.P. Box Sampling and Bayes’ inference in scientific modelling and robustness. J. R. Statist. Soc. A, 143(4):383–430, 1980. B. Efron The future of indirect evidence. Statist. Sci., 25(2):145–157, 2010. D.B. Rubin Bayesianly justifiable and relevant frequency calculations for the applied statistician. Ann. Appl. Statist., 12:1151–1172, 1984. A. Gelman, X.-L. Meng, H. Stern Posterior predictive assessment of model fitness via realized discrepancies. Statistica Sinica, 6:733–807, 1996. M.J. Bayarri, J.O. Berger Quantifying surprise in the data and model verification. In Bayesian Statistics 6 (eds. J. M. Bernardo, J. O. Berger, A. P. Dawid, A. F. M. Smith). OUP, 53–82, 1999. E.C. Marshall, D.J. Spiegelhalter Identifying outliers in Bayesian hierarchical models: a simulation-based approach. Bayesian Analysis, 2:409–444, 2007. A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 29 / 26 References Model criticism references II A. O’Hagan HSSS model criticism (with discussion). In Highly Structured Stochastic Systems (eds. P.J. Green, N.L. Hjort, S. Richardson). OUP, 2003. M. Evans, H. Moshonov Checking for prior-data conflict. Bayesian Analysis, 1(4):893–914, 2006. J. Gasemyr, B. Natvig Extensions of a conflict measure of inconsistencies in Bayesian hierarchical models. Scand. J. Statist., 36:822–838, 2009. D. Ohlssen, D.J. Spiegelhalter Conflict diagnostics in directed acyclic graphs. Workshop “Explaining the results of a complex probabilistic modelling exercise: conflict, consistency and sensitivity analysis”, 2006. S. Dias, N.J. Welton, D.M. Caldwell, A.E. Ades Checking consistency in mixed treatment comparison meta-analysis. Stat. Med., 29(7):932–944, 2010. A.M. Presanis, D. Ohlssen, D.J. Spiegelhalter, D. De Angelis Conflict diagnostics in directed acyclic graphs, with applications in Bayesian evidence synthesis. To be submitted, 2012. A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 30 / 26 References ’Flu & HIV evidence syntheses A.M. Presanis, D. De Angelis, The New York City Swine Flu Investigation Team, A. Hagy, C. Reed, S. Riley, B.S. Cooper, L. Finelli, P. Biedrzycki, M. Lipsitch The severity of pandemic H1N1 influenza in the United States, from April to July 2009: A Bayesian analysis. PLoS Med, 6(12):e1000207, 2009. A.M. Presanis, R.G. Pebody, B.J. Paterson, B.D.M. Tom, P.J. Birrell, A. Charlett, M. Lipsitch, D. De Angelis Changes in severity of pandemic (H1N1) 2009 influenza in England: a Bayesian evidence synthesis. BMJ, 343:d5408, 2011. P.J. Birrell, G. Ketsetzis, N.J. Gay, B.S. Cooper, A.M. Presanis, R.J. Harris, A. Charlett, X.-S. Zhang, P.J. White, R.G. Pebody, D. De Angelis Unmasking the pandemic: a Bayesian reconstruction of influenza A/H1N1pdm dynamics in London. PNAS, doi:10.1073/pnas.1103002108, 2011. A. Goubar, A.E. Ades, D. De Angelis, C.A. McGarrigle, C. Mercer, P.Tookey, K.Fenton, O.N. Gill Estimates of HIV prevalence and proportion diagnosed based on Bayesian multi-parameter synthesis of surveillance data. J.R. Stat. Soc. A, (with discussion), 171:541–580, 2008. A.M. Presanis, D. De Angelis, D.J. Spiegelhalter, S. Seaman, A. Goubar, A.E. Ades Conflicting evidence in a Bayesian synthesis of surveillance data to estimate HIV prevalence. J. R. Stat. Soc. A, 171:915-937, 2008. A.M. Presanis, O.N. Gill, T.R. Chadborn, C. Hill, V. Hope, L. Logan, B.D. Rice, V. Delpech, A.E. Ades, D. De Angelis Insights into the rise in HIV infections in England and Wales, 2001 to 2008: a Bayesian synthesis of prevalence evidence. AIDS, 24(18):2849–2858, 2010. A.M. Presanis, D. De Angelis, A. Goubar, O.N. Gill, A.E. Ades Bayesian evidence synthesis for a transmission dynamic model for HIV among men who have sex with men. Biostatistics, 12(4):666–681, 2011. S. Conti, A.M. Presanis, M.G. van Veen, M. Xiridou, M.C. Donoghoe, A. Rinder Stengaard, D. De Angelis Modeling of the HIV infection epidemic in the Netherlands: a multi-parameter evidence synthesis approach. Ann. Appl. Statist., 5(4):2359–2384, 2011. A. M. Presanis (MRC BSU) Conflict in complex evidence synthesis UCL, 18 Oct 2012 31 / 26