Are clinicians Bayesian: Can a Bayesian approach help clinical decision making? Martin and Laura Green Focus • Evidence required by decision makers – – – – – In this example, clinicians Applies to Stakeholders Policy makers Us! Aim • What motivates clinicians take up new results from research • How data are viewed and interpreted – Summary of a clinical trial – Assessment of likely uptake given clinicians vary – Results from elicitation on use of antibiotics to ‘cure’ infection • Conclude with usefulness of Bayesian framework for decision making as a natural approach The example of cattle mastitis Mastitis in dairy cows • • • • Bacterial infection in the udder Udder, 4 quarters (mammary glands) Somatic cell count increases >200K 5 - 6 common bacterial species (Staphs, Streps, E coli) many other possible causes – Infection enters via teat – Risks • other infected cows, transmission via milking equipment • Environment, E. coli, Streps via contaminated ground Common cause of disease • Current incidence rate of clinical mastitis (IRCM) ~50-60 cases per 100 cow-years • Unchanged for decades • Approx 1 in 20 cows culled pa because of mastitis = twice number culled for bTB each year = total cull for cattle FMD 2001 every five years Huge economic and welfare importance An intervention study testing mastitis control in dairy herds in England and Wales 2004 - 2005 M J Green, K A Leach, J E Breen, L E Green, A J Bradley Aim of research study • Assess the effectiveness of a highly specified mastitis control plan – 52 randomly selected dairy herds in England and Wales – Representative of a target population • > 35 clinical cases per 100 cow-years • Funded by the Milk Development Council Mastitis diagnosis and control plan (MDCP) • Targeted plan – But addressed many putative risks • Developed using literature – Rare that separate risk factors tested in controlled field trials in livestock • A clear, structured process – could subsequently be employed by vets Data collected on all aspects of farm management • Sections: • • • • • • • • • • • • • General farm issues Management between milking Pre-milking management Milking routine Milking machine maintenance Post-milking management Dry cow management Calving cow management Treatment strategies Biosecurity Young stock management Monitoring and recording Nutritional management Farm specific diagnosis of mastitis • Characterised patterns of mastitis • Differentiate time of infection • Dry period vs lactation • Rate of clinical mastitis during lactation • SCC patterns • Season • Parity (including maiden heifers) • Recurrence rates Heterogeneity in post calving (30 day) incidence rates of clinical mastitis Incidence rate of clinical mastitis per 100 cows per 30 calved days Proportion of Cows with Clinical Mastitis 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 Farms inAscending ascending order of incidence rate Farms in order of Clinical Mastitis Incidence Survival to clinical mastitis Heterogeneity in rates of clinical mastitis during lactation 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 Weeks after calving Hierarchy of ‘importance’ of management controls “Diagnosis” Must Should Could 10-20 Action Points COMPLIANCE Clinical trial design Time –12 mths Time 0 Time + 12 mths Gp 2: 26 control farms 52 Outcome: Change as proportion of initial level Gp 1: 26 intervention farms Mastitis Plan Initial analysis Yi = α + β1 Ii + βnCi + ei Yi = (mastitis yrt - mastitis yrt-1) / mastitis yrt-1 on farm i I = Indicator for intervention farm Covatiates… – Mean cow annual milk yield – Mean herd size – Year 1 IRCM /SCC Model fit – standardised residuals plots – assessment of homoscedasticity – investigation of influence & leverage Proportional change in incidence rate of cows affected Results: one year follow up 97.5 %ile * *p=0.01 Rank by proportional change Control Farms Intervention Farms 3 2 Proportional change in Proportional change in IRCA affected cows rate of incidence 1.4 1.2 1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 3 3 3 2 3 2 2 1 1 1 3 3 2 2 2 3 Compliance group of intervention farms 1 2 1 1 1 1 1 2 Proportional change in incidence rate of cows affected Rank - degree of compliance Conclusion • Highly specified approach – Pattern assessment – Targeted control strategies – Can reduce IRCM..... if follow plan • Clinically important... – Welfare – Finances • Would the plan be adopted by vet practitioners? Spiegelhalter et al, 2004 • Four advantages for evaluating medical information in a Bayesian framework – Flexible – Particularly for more complex ‘real life’ situations – Efficient – Can use all available evidence – Ethical – Past experiences can be exploited – Useful – Transparency of uncertain elements – Natural platform to provide predictions that include full uncertainty • Framework to take into account clinical beliefs Investigation of veterinary beliefs on clinical decision making M J Green A J Bradley, W J Browne, J E Breen, K A Leach, L E Green, G F Medley Making a clinical decision • Clinicians want to know what works, not disprove what does not work • Have to be definite in a reasonably short time • Decision – May be clear cut • Individual animal, known condition – Often very complex • Population level (farm or national) • Always a degree of uncertainty • Always context dependent • Uptake of new knowledge dependent on evidence and individual’s characteristics The decision process NEWNEW INFORMATION DATA (Research or CPD) (Research or CPD) PREVIOUSPRIOR INFORMATION (Research or or (Research Experiences) Experiences) POSTERIOR CURRENT VIEW (with (with some some uncertainty) uncertainty) DELIVER DECISION (Simplified?, politicised?, justified?, ?probabilistic) LOCAL CONTEXT (Animal, population, interested parties, ethics, real constraints etc) How do clinicians do this?? What they might do What time to do? • Read research and undertake CPD • Use of mental strategies (heuristics) • Use all available information in all subject areas • New and complex knowledge cannot be simplified – therefore not adopted • May not keep abreast of all current knowledge – Time • Simplify the decision process, ignore complexity of research – – Assume generalisability Ignore consequences of variability e.g. 2% – 80% Evidence-based medicine Constraints -Fascilities -Business -etc Approach to Economics Mastitis Control Vet New Science? Farmer -Ability -Personality -Goals -Experience -Psychology -Peers -etc -etc A Bayesian approach Our question… • Having published a clinical trial on a mastitis control scheme:– Farm specific – Control measures weighted – Found a mean reduction in clinical mastitis of 20% • How might this be interpreted by vets? Vets vary sceptic...enthusiast • Own context/experiences • Context of other science Current context - prior beliefs • Attempt to capture actual beliefs by formal elicitation • Investigate theoretically using a “community of priors” to represent a reasonable spectrum of possible views {Kass and Greenhouse (1989)} Name of Prior Vague Description of View Represented No view or ability to make a choice as to what the likely parameter values could be. Very Sceptic A mean effect size of 0 with a 2.5% probability that the effect size could be a reduction in mastitis more than - 0.10. Sceptic A mean effect size of 0 with a 2.5% probability that the effect size could be a reduction in mastitis more than 0.20 Mid Sceptical A mean effect size of -0.10 with a 15% probability that Enthusiastic the effect size could be a reduction in mastitis greater than -0.20 or less than 0. Enthusiastic A mean effect size of -0.20 with a 2.5% probability that the effect size could be a reduction in mastitis less than 0. Very Enthusiastic A mean effect size of -0.30 with a 2.5% probability that the effect size could be a reduction in mastitis less than 0.20. “Community of Priors” - Spectrum of reasonable clinical viewpoints Mean 0 0 0 -0.1 -0.2 -0.3 9 8 Probability Density 7 6 5 4 3 2 1 0 -0.6 -0.4 -0.2 0.0 0.2 0.4 Prior Distributions (Proportional Reduction in Clinical Mastitis) StDev 100 0.05 0.1 0.1 0.1 0.05 Posteriors of interest ‘Expected’ reduction in clinical mastitis using this control scheme – Each prior updated using trial data Predictions of expected cost saving (room for investment) from using the control scheme Financial estimates • Cost of a case of clinical mastitis was based on literature (Mean £212, Var £30.3) • Assumed incidence rate of clinical mastitis of 0.5 cases per cow per year (approximate mean value for UK farms) Results Prior Distributions Mean StDev Vague 0 100 Very Sceptic 0 0.05 0.1 Sceptic 0 -0.1 0.1 Mid Sceptic-Enthusiastic -0.2 0.1 Enthusiastic -0.3 0.05 Very Enthusiastic 9 8 Probability Density Density Probability 7 6 5 4 3 2 1 0 -0.6 -0.6 -0.4 -0.4 -0.2 -0.2 0.0 0.0 0.2 0.4 0.4 Posterior Distributions(Proportional (ProportionalReduction Reduction in Clinical Prior Distributions ClinicalMastitis) Mastitis) Probability Minimum Expected Gain gain expected minimum of of Probability Financial gains from the control plan, anticipated by clinicians with different prior beliefs Sceptic Very Sceptic Vague Enthusiastic Mid Sceptic-Enthusiast Cautious Sceptic Very Enthusiastic 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 5 10 15 20 25 30 35 Minimum Expected Gain (£s per cow in herd per year) Minimum expected gain (UK£s /cow /yr) 40 45 50 Update views with 3 further ‘identical RCT’ Theoretical prior ... posterior Prior … posterior Prior … posterior Prior … posterior Anticipated financial gains from implementing the control plan, three further equivalent clinical trials Posterior after original trial with sceptic prior Posterior after 1 simulated trial Posterior after 2 simulated trials Posterior after 3 simulated trials Enthusiastic prior at original trial Vague prior at original trial Probability of Minimum Expected Gain 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 5 10 15 20 25 30 35 40 Minimum Expected Gain (£s per cow in herd per year) 45 50 The relevance of prior uncertainty… Mean Prior = 0 Mean Prior = -0.05 Mean Prior = -0.1 Mean Prior = -0.2 Mean Prior = -0.25 Mean Prior = -0.3 Mean Prior = -0.15 0 Mean of Posterior -0.05 -0.1 -0.15 -0.2 -0.25 -0.3 -0.35 0 0.2 0.4 0.6 0.8 Standard Deviation of Prior 1 1.2 1.4 Thoughts - dairy cow mastitis • We may fail to convince vets to modify their approach to management of mastitis... – unless we understand their current views • Do we know anything about the views of stakeholders? – e.g. vets for the major endemic conditions; BVD, lameness, reproduction etc Quantitative assessment of vet clinical beliefs to improve preventive healthcare for dairy cattle Helen Higgins, Green, Huxley, O’Hagan, Oakley, Browne, Smith, Clough An initial study: Dry cow antibiotic therapy in dairy cows ? Two main aims 1) Evaluate the variation and strength of vets beliefs 2) Given the current beliefs of the vets, investigate the strength of evidence needed to ‘change their minds’? Sampling strategy Target population – vets who “regularly” deal with dairy cattle Study population – 100 mile radius of Nottingham Two-stage cluster sampling, stratified by “certCHP” 24 vets, 5 practices in England (+ 2 clinical academics) Qualified 9 months - 26 years 7 held extra cattle qualifications Probabilistic elicitation Vets interviewed individually Beliefs quantified numerically as probability density functions using “probabilistic elicitation” (SHELF) Capture uncertainty Definition of variables • 1 = overall cure rate with intra-mammary dry cow therapy • 2= overall cure rate with intra-mammary dry cow therapy AND systemic antibiotics Interested in the joint probability 3 = the additional benefit of using systemic antibiotics, given IDCT has failed 2 = 1 + (1- 1)3 Elicited from vets the marginal distributions: cure rate with DCT = f(1) Additional cure with systemic tx if DCT fails = f(3) Fit a distribution from an appropriate parametric family from summary values (min, max IQR, median) – iterate with vet , Practice 2 Cure rate with standard therapy AND systemic antibiotic 2 Cure rate with standard dry cow therapy - 1 Practice 2 Practice 3 standard therapy 0.8 0.6 0.2 0.4 0.6 0.8 probability of cure with treament 1 Practice 4 Practice 5 Cluster 4 0.0 0.2 0.4 0.6 0.8 probability of cure with treament 1 1.0 1.0 0.8 0.6 0.4 0.2 0.0 0.2 0.4 0.6 0.8 probaility of cure with treatment 2 1.0 Cluster 5 0.0 probaility of cure with treatment 2 2 0.4 0.0 AND systemic antibiotic 0.2 0.0 Cure rate with 1.0 Cluster 3 probaility of cure with treatment 2 Practice 1 0.0 0.2 0.4 0.6 0.8 1.0 Cure rate with standard dry cow therapy - 1 probability of cure with treament 1 1.0 • Given the beliefs of the 24 vets… • …Explore what it would take to convince the majority of vets that it is not worth giving systemic antibiotics • Assume – that treatment does not work (not unrealistic) • Assume - that an additional effect on an OR scale (odds θ2:θ1) >1.5 (e.g. from 70% cure to 77% cure) is ‘clinically worthwhile’ • Assume - the vets need to be very (95% credible interval OR < 1.5) certain to change their action (n.b. reduces their income) • vet’s prior belief likelihood of theta based on new data the updated belief in light of the new evidence New data Synthetic, assumed : From a single RCT trial (of different sizes) that showed no additional benefit with systemic antibiotics. Synthetic data for the likelihood were generated for different sized trials; 30, 60, 250, 500 cows in each treatment arm. θ1 and θ2 were set equal so that θ3 =0; i.e. no difference between the two treatments. Posterior of interest was OR θ2:θ1 0 1 2 3 4 5 6 7 8 9 10 11 Odds ratio Vets’ elicited beliefs expressed as 95% credible intervals on an odds ratio scale Systemic ab’s offer a benefit over IDCT alone Systemic ab’s offer no extra benefit Practice 1 Practice 2 Practice 3 Practice 4 Practice 5 Predicted belief (red) after evidence from a single RCT involving 30 infected cows per group Predicted belief (red) after evidence from a single RCT involving 250 infected cows per group Predicted belief (red) after evidence from a single RCT involving 500 infected cows per group 10 11 9 8 7 6 5 4 3 2 1 0 Odds ratio for theta 2 versus theta 1 Odds ratio Example from a meta-analysis: Predicted belief (red) after the evidence of 5 small RCT ( 2 groups of 100 cows per trial) Cluster1 Cluster2 Practice 1 Practice 2 Cluster3 Practice 3 Cluster4 Practice 4 Cluster5 Practice 5 Therefore… Major variations in vets beliefs Reasons for variation? To what extent does variability occur with clinical beliefs in general? Many of these practitioners are “enthusiastic” and/or “certain” about the efficacy of systemic antibiotics Final thoughts If the point of research is generally to effect change (e.g. clinical decisions) • How often do we have a clear understanding of the stakeholders involved? • Current views (degree of scepticism and certainty)? • When will they change their minds? • Should context of stakeholders influence study design and sample size? Final thoughts • Clinicians naturally Bayesian? – Weighting of information + knowledge of uncertainty • Bayesian analyses may be useful for clinicians if – Transparent – Summarise posteriors are on scales that are easily understood • A final ‘but’ – Link ‘what we think’ with ‘what we do’ Acknowledgements Authors , collaborators and colleagues Funders: Vets and farmers InFER