Biostatistics Case Studies Peter D. Christenson Biostatistician Session 6: Control of Confounding Bias Using Propensity Scoring.

Slides:



Advertisements
Similar presentations
The Application of Propensity Score Analysis to Non-randomized Medical Device Clinical Studies: A Regulatory Perspective Lilly Yue, Ph.D.* CDRH, FDA,
Advertisements

How would you explain the smoking paradox. Smokers fair better after an infarction in hospital than non-smokers. This apparently disagrees with the view.
If we use a logistic model, we do not have the problem of suggesting risks greater than 1 or less than 0 for some values of X: E[1{outcome = 1} ] = exp(a+bX)/
Controlling for Time Dependent Confounding Using Marginal Structural Models in the Case of a Continuous Treatment O Wang 1, T McMullan 2 1 Amgen, Thousand.
V.: 9/7/2007 AC Submit1 Statistical Review of the Observational Studies of Aprotinin Safety Part I: Methods, Mangano and Karkouti Studies CRDAC and DSaRM.
Departments of Medicine and Biostatistics
Observational Studies Based on Rosenbaum (2002) David Madigan Rosenbaum, P.R. (2002). Observational Studies (2 nd edition). Springer.
Estimation and Reporting of Heterogeneity of Treatment Effects in Observational Comparative Effectiveness Research Prepared for: Agency for Healthcare.
Chapter 11 Survival Analysis Part 2. 2 Survival Analysis and Regression Combine lots of information Combine lots of information Look at several variables.
Basic Business Statistics, 10e © 2006 Prentice-Hall, Inc. Chap 9-1 Chapter 9 Fundamentals of Hypothesis Testing: One-Sample Tests Basic Business Statistics.
Review for Exam 2 Some important themes from Chapters 6-9 Chap. 6. Significance Tests Chap. 7: Comparing Two Groups Chap. 8: Contingency Tables (Categorical.
EVIDENCE BASED MEDICINE
Cohort Studies Hanna E. Bloomfield, MD, MPH Professor of Medicine Associate Chief of Staff, Research Minneapolis VA Medical Center.
Assessing Survival: Cox Proportional Hazards Model Peter T. Donnan Professor of Epidemiology and Biostatistics Statistics for Health Research.
Marshall University School of Medicine Department of Biochemistry and Microbiology BMS 617 Lecture 12: Multiple and Logistic Regression Marshall University.
Survival analysis Brian Healy, PhD. Previous classes Regression Regression –Linear regression –Multiple regression –Logistic regression.
Business Statistics, A First Course (4e) © 2006 Prentice-Hall, Inc. Chap 9-1 Chapter 9 Fundamentals of Hypothesis Testing: One-Sample Tests Business Statistics,
Advanced Statistics for Interventional Cardiologists.
Simple Linear Regression
Dr Laura Bonnett Department of Biostatistics. UNDERSTANDING SURVIVAL ANALYSIS.
Biostatistics Case Studies 2005 Peter D. Christenson Biostatistician Session 4: Taking Risks and Playing the Odds: OR vs.
Biostatistics in Clinical Research Peter D. Christenson Biostatistician January 12, 2005IMSD U*STAR RISE.
Evidence-Based Medicine 3 More Knowledge and Skills for Critical Reading Karen E. Schetzina, MD, MPH.
Biostatistics Case Studies Peter D. Christenson Biostatistician Session 5: Analysis Issues in Large Observational Studies.
Understanding the Variability of Your Data: Dependent Variable Two "Sources" of Variability in DV (Response Variable) –Independent (Predictor/Explanatory)
Biostatistics Case Studies 2005 Peter D. Christenson Biostatistician Session 5: Classification Trees: An Alternative to Logistic.
Statistical Bootstrapping Peter D. Christenson Biostatistician January 20, 2005.
Assessing Survival: Cox Proportional Hazards Model
Influence of Comorbid Depression and Antidepressant Treatment on Mortality for Medicare Beneficiaries with Chronic Obstructive Pulmonary Disease by SSDI-eligibility.
Biostatistics in Practice Peter D. Christenson Biostatistician LABioMed.org /Biostat Session 6: Case Study.
Article Review Cara Carty 09-Mar-06. “Confounding by indication in non-experimental evaluation of vaccine effectiveness: the example of prevention of.
Biostatistics Case Studies 2008 Peter D. Christenson Biostatistician Session 5: Choices for Longitudinal Data Analysis.
Randomized Trial of Ea rly S urgery Versus Conventional Treatment for Infective E ndocarditis (EASE) Duk-Hyun Kang, MD, PhD on behalf of The EASE Trial.
Estimating Causal Effects from Large Data Sets Using Propensity Scores Hal V. Barron, MD TICR 5/06.
Introduction to Survival Analysis Utah State University January 28, 2008 Bill Welbourn.
Randomized Trial of Preoperative Chemoradiation Versus Surgery Alone in Patients with Locoregional Esophageal Carcinoma, Ursa et al. Statistical Methods:
Surgical outcome of native valve infective endocarditis in srinagarind hospital
Adaptive randomization
1 THE ROLE OF COVARIATES IN CLINICAL TRIALS ANALYSES Ralph B. D’Agostino, Sr., PhD Boston University FDA ODAC March 13, 2006.
Empirical Efficiency Maximization: Locally Efficient Covariate Adjustment in Randomized Experiments Daniel B. Rubin Joint work with Mark J. van der Laan.
Statistics for Managers Using Microsoft Excel, 4e © 2004 Prentice-Hall, Inc. Chap 8-1 Chapter 8 Fundamentals of Hypothesis Testing: One-Sample Tests Statistics.
Biostatistics Case Studies 2005 Peter D. Christenson Biostatistician Session 6: “Number Needed to Treat” to Prevent One Case.
Lecture 9: Analysis of intervention studies Randomized trial - categorical outcome Measures of risk: –incidence rate of an adverse event (death, etc) It.
Medical Statistics as a science
Biostatistics Case Studies 2006 Peter D. Christenson Biostatistician Session 4: An Alternative to Last-Observation-Carried-Forward:
1 Lecture 6: Descriptive follow-up studies Natural history of disease and prognosis Survival analysis: Kaplan-Meier survival curves Cox proportional hazards.
Chap 8-1 Fundamentals of Hypothesis Testing: One-Sample Tests.
Biostatistics in Practice Peter D. Christenson Biostatistician Session 4: Study Size for Precision or Power.
01/20151 EPI 5344: Survival Analysis in Epidemiology Actuarial and Kaplan-Meier methods February 24, 2015 Dr. N. Birkett, School of Epidemiology, Public.
Biostatistics Case Studies 2006 Peter D. Christenson Biostatistician Session 2: Correlation of Time Courses of Simultaneous.
IMPORTANCE OF STATISTICS MR.CHITHRAVEL.V ASST.PROFESSOR ACN.
Using Propensity Score Matching in Observational Services Research Neal Wallace, Ph.D. Portland State University February
Biostatistics Case Studies 2014 Youngju Pak Biostatistician Session 5: Survival Analysis Fundamentals.
Multilevel and multifrailty models. Overview  Multifrailty versus multilevel Only one cluster, two frailties in cluster e.g., prognostic index (PI) analysis,
Biostatistics Case Studies 2007 Peter D. Christenson Biostatistician Session 2: Aging and Survival.
Biostatistics Case Studies 2006 Peter D. Christenson Biostatistician Session 3: An Alternative to Last-Observation-Carried-Forward:
Matching. Objectives Discuss methods of matching Discuss advantages and disadvantages of matching Discuss applications of matching Confounding residual.
POPLHLTH 304 Regression (modelling) in Epidemiology Simon Thornley (Slides adapted from Assoc. Prof. Roger Marshall)
Chapter 9: Introduction to the t statistic. The t Statistic The t statistic allows researchers to use sample data to test hypotheses about an unknown.
Biostatistics Case Studies 2006 Peter D. Christenson Biostatistician Session 1: Demonstrating Equivalence of Active Treatments:
Marshall University School of Medicine Department of Biochemistry and Microbiology BMS 617 Lecture 13: Multiple, Logistic and Proportional Hazards Regression.
Date of download: 9/18/2016 Copyright © 2016 American Medical Association. All rights reserved. From: Effect of Structured Physical Activity on Prevention.
Early Surgery versus Conventional Treatment for Infective Endocarditis
Unit 5: Hypothesis Testing
Sec 9C – Logistic Regression and Propensity scores
Table 1 Characteristics of study population, by pneumococcal vaccination status. From: Prior Pneumococcal Vaccination Is Associated with Reduced Death,
Comparing 2 Groups Most Research is Interested in Comparing 2 (or more) Groups (Populations, Treatments, Conditions) Longitudinal: Same subjects at different.
Comparing Populations
Selecting the Right Predictors
Improving Overlap Farrokh Alemi, Ph.D.
Presentation transcript:

Biostatistics Case Studies Peter D. Christenson Biostatistician Session 6: Control of Confounding Bias Using Propensity Scoring

Case Study Impact of Valve Surgery on 6-Month Mortality in Adults With Complicated, Left-Sided Native Valve Endocarditis A Propensity Analysis Holenarasipur R. Vikram, MD; Joan Buenconsejo, MPH; Rodrigo Hasbun, MD; Vincent J. Quagliarello, MD JAMA. 2003;290:

Subjects N=513 from records at 7 hospitals : Jan Jan Definite (3%) or probable (97%) endocarditis. Left-sided infection of native valve. Have a complication for which surgery is considered: CHF New valvular regurgitation. Refractory infection. Systemic embolism to vital organs. Echocardiographic evidence of vegetation.

Measured Characteristics Exposure: Surgery or no surgery for valve repair or replacement. Outcome: All-cause mortality within 6 months of baseline = date of surgery or date of decision to not operate. Other factors: Hospital, baseline demographics, comorbidity, previous heart disease, symptoms, physical findings, blood cultures, EKG, echo, type of surgery, operative findings.

Analysis #1: Typical Survival Analysis Step 1: Individually determine factors related to mortality with Cox regression (survival analysis). Surgery is one. Step 2: Include factors related to mortality in step 1, and surgery, in a multivariable survival analysis that adjusts the effect of mortality for these other factors.

Analysis #1: Results Table 3 and p top center Step 1: Factors significantly associated with greater mortality are: female, older, immunocompromised, fever, intracardiac abscess, comorbidity on standard scale, CHF, abnormal mental status, bacteria other than viridans strep, elevated serum creatinine, and refractory infection. Surgery: 37/230=16% died, non-surgery: 94/160=33% died in six months. Hazard ratio = 0.43 (95%CI: 0.29 – 0.63). Step 2: Adjustment for “heterogeneity factors” gives a slightly greater surgical effect: Hazard ratio = 0.35 (95%CI: 0.23 – 0.54). But, why are these factors in * of Table 3 not the same as those in step 1?

Analysis #2: Propensity Analysis Step 1: Use logistic regression to estimate the propensity that a subject will have surgery based on baseline factors: e = Prob [surgery] = function (baseline factors). Step 2: Randomize the order of those having surgery and sequentially go through the list of surgery subjects, finding the single non- surgery subject with propensity score closest to the surgery subject. If there is no such match to at least the first decimal place, drop the surgery subject from the analysis. Step 3: Use matched survival analysis to find hazard ratio, ignoring e. Step 4: Use matched survival analysis to find hazard ratio, also adjusting for heterogeneity or confounding factors.

Analysis #2: Results Part 1 Tables 1 and 2 Propensity score is a function of: Hospital, HIV+, comorbidity, CHF, abnormal mental status, symptomatic or disabling emboli, refractory infection, intracardiac abscess, new valve regurgitation, microbial etiology, valve involved, number of embolic events. The propensity score does what it is intended to do: Most of the factors in Table 1 are different between surgery (N=230) and non-surgery (N=283) subjects. All of these same factors in are not different between surgery (N=109) and their matched non-surgery (N=109) subjects in Table 2.

Analysis #2: Results Part 2 Table 3 and Figure 1 Table 3: HR (95% CI)p Propensity-matched only 0.45 ( ) 0.02 Include confounder* adjustment 0.45 ( ) 0.02 Add heterogeneity adjustment0.40 ( ) 0.03 *only for slight differences in propensity score within matched pairs. Figure 1: Does not assume proportional hazards. p=0.01

Logic for Propensity Score Q: Why estimate the probability that a subject receives surgery, when we KNOW which subjects did? A: This creates a “quasi-randomized” experiment. If one surgery subject and one non-surgery subject have the same probability that they would have received surgery, then the fact that one of them eventually did is random, a 50% chance for each. Two issues: (1)We assume that the set of covariates contains all of the characteristics related to both treatment and response, i.e., that all confounders are measured. (2)Even if (1) is true, we don’t know true propensity, but need to estimate it from the data.

Basic Theory for Propensity Score The two conditions: (1)The propensity score is a balancing score, meaning that the statistical distribution of the covariates is the same for surgery or non-surgery subjects with the same propensity. (2)Treatment assignment is strongly ignorable, meaning that the probability of treatment assignment and the distribution of response are independent for a given set of the covariates. [This is (1) from the previous slide.] imply that: Average treatment effect can be measured unbiasedly for a specific propensity, as in a randomized experiment. But, again, (2) is an assumption, and we only have an estimate in (1). In practice, approximately unbiased estimates are hoped for, but cannot be verified statistically.

Alternative Analysis A “Standard” adjusted survival analysis using a set of actual confounder values. Advantage: Uses all subjects (provided data is available). Here, we dropped = 295 subjects, although some due to missing data. Disadvantages: 1. Adjusting for many confounders makes model stability difficult: choosing interactions, multicollinearity, overfitting, many categorical levels, etc. See slide after next one, cf. 2. If the range for just one confounder is very different for surgery and non-surgery subjects, the adjustment is at an interpolated value, which may not occur in practice or make clinical sense, as in example on next slide.

Example of Poor Confounder Adjustment From Gerard Dallal, This problem exists in any regression adjustment, but is more likely when adjustment for many factors is made.

Confounder Adjustment Simulations: Set of many confounders vs. Propensity score Cepeda, et al, American J Epidemiology 2003;158: : Uses logistic regression, rather than survival analysis as here, but probably (?) similar conclusions. Compares the two methods on bias, CI coverage, power, and precision. Concludes: Use propensity score if # deaths per # of parameters for confounders is small, less than 7. Next three slides give reasons for this recommendation. In this study, have 113 deaths, so propensity score method is preferable if at least ~ 113/7 ~ 16 confounding parameters. There are 18 parameters for association with surgery (but not all are confounders?), not including interactions.

Confounder Adjustment Simulations: Set of many confounders vs. Propensity score Precision

Confounder Adjustment Simulations: Set of many confounders vs. Propensity score Confidence Interval: True Probability Level

Confounder Adjustment Simulations: Set of many confounders vs. Propensity score Bias

Borrowing from Peter to Pay Paul? Q: Isn’t this two-step process in propensity analysis just removing the problem of “many confounders” in the survival estimation, as in Alternative A, to the estimation of the propensity, and the problem still remains there? A: Using many confounders in the equation for estimation of treatment effect may lead to over-fitting, specific for the data in the sample, and not generalizable, and to imprecision in the treatment effect. We are not as concerned about over-fitting the equation for propensity scores, since the prediction from it will be less biased, and the imprecision is not as serious, as the simulations in the last few slides show.

Alternative Analysis B “Standard” stratified survival analysis with strata defined by a set of actual confounder values. Advantages: 1. Uses all subjects (provided data is available, and strata are not too numerous). 2. Assumptions and fitting difficulties of regression modeling as in alternative A are not issues. Disadvantage: Too many strata will be required. Thus, small Ns in strata, or some subjects dropped due to no non-surgery subjects in a strata.

Alternative Analysis C Use techniques other than the propensity score for choosing a single summary confounding measure. Most popular method: Mahalanobis metric matching. This was the method of choice prior to propensity scores. Similar to clustering analyses. Approximately visualized for just 2 confounders by plotting each confounder as x or y on a plot, for all non-surgery subjects only. Then, for each surgery subject, add him/her to the plot, and choose the closest non-surgery subject as the match. Disadvantage: difficult to find close matches when there are many confounders.

Other Propensity-type Scores Current Development: Drop-out in clinical trials Propensity = Prob[ completing trial] = function of baseline characteristics Concept: Attempt to reduce the bias due to differential drop out between treated and non-treated groups, by matching on propensity to drop out. Difficulty: This still will require the assumption that reasons for drop out are included in the measured characteristics.

Conclusions The two analyses in the paper are extreme in their handling of potential confounding: Analysis (1) ignores confounding, although it adjusts for some covariates. Analysis (2) strongly corrects for confounding. It ignores subjects who don’t have a matched subject with similar confounders receiving the opposite treatment, and may over- match. Similar conclusions, i.e., a 55% to 65% efficacy of surgery for the different analyses, adds robustness to the conclusion.