in clustered data Dmitry Tumin PhD Department of Anesthesiology and Pain Medicine NCH Department of Pediatrics OSUCOM Surgery in the US gt50 million inpatient procedures year 15 mortality rate ID: 1035903
Download Presentation The PPT/PDF document "Analysis of surgical outcomes" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1. Analysis ofsurgical outcomesin clustered dataDmitry Tumin, PhDDepartment of Anesthesiology and Pain Medicine, NCHDepartment of Pediatrics, OSUCOM
2. Surgery in the US>50 million inpatient procedures / year~1-5% mortality rate for inpatientsOR time costs $36 / minute
3. Surgical outcomes dataHospital recordsClinical registriesInsurance claims
4. ClusteringPatientSurgeonHospitalGeography
5. When things go horribly wrong“Broken Hearts”Journal retractspaper over clusteredstandard errorsAnesth Analg 2016;122:1231-3
6. “The rate of death or major complications among patients undergoing coronary artery bypass surgery varies markedly across anesthesiologists.”The article demonstrated something we havealways thought: some anesthesiologists are better than others.The problem was subtle, the difference between “robust cluster (anes)” and “robust” in the Stata program code. It seemed to me and the reviewers that the analysis should be clustered, [but] “when fixed effects and clustering are specified at the same level, tests that involve the fixed effects themselves are inadvisable (the standard errors on fixed effects are likely to be substantially underestimated, though this will not affect the other variance estimates).”
7. RetractedarticleCorrectedarticle
8. A slight misunderstanding…Why should we adjust for clustering How should we do it What can we conclude
9. Scenario 1Cases are not independently sampledCluster robust standard errorsProblem: does not address biased coefficients
10. Scenario 2Confounding by cluster, and differences between clusters are ignorableOnly care about correctly estimatingassociation between patient factorand patient outcomeFixed effects models
11. ExampleLung transplantcold ischemiaLonger time =worse outcomes?But, hospitals willing to take lungs fromfarther away tend to be more experienced
12. Stratified Cox = comparing ischemiatime among patientstransplanted at thesame hospitalIncreased HRHayes et al. Am J Transplant 2017
13. Scenario 3Cluster differences NOT ignorableAlso care about identifying“better/worse” clustersDo we know what makes a clusterbetter or worse?
14. Example: median household income in geographic area (neighborhood, county)
15. What if...Cluster-level measuredoes not predict outcomes?Other possible factors at thecluster level cannot be measured?
16. Scenario 4we do NOT know why some clusters have better/worse outcomes
17. Some options…Cluster ID as covariateBayesian analysisRandom effects analysis, if overall variation among clusters is of interest
18. Example
19. Variance of shared frailty termHow different are clusters of patientsafter controlling for observed covariates?H0 = no residual difference between clustersIf fail to reject H0,No reason for testing specific cluster measuresNo reason for including clustering level (county)in risk-adjustment models
20. Heart transplant outcomes by countyAfter controlling for county SES:Same variation in mortality across countiesNo improvement in model fit
21. Visualizing shared frailty varianceRarely done in surgical outcomes researchVariance estimate is often not even reportedBut, easy to do using invgammap + gammaden
22. // HR for 80th percentile vs. mediandi invgammap(1/`theta’, (0.8))*`theta’// Density plottw function y=exp(gammaden(1/`theta1',`theta1',0,x)
23. SummaryClustered observations are common insurgical outcomes researchNeed to account for confounding by cluster,not just lack of independence in samplingAnalytic approach—“single out” outliers,or capture overall variation among clusters?Depends on clinical/administrative need
24. Questions / commentsDmitry.Tumin@nationwidechildrens.org