Global web ALERT button

Statistical Methods & Confounding in Large Health Care Databases

 Click to view the video in a new window.

Notes

Recorded at the Comparative Effectiveness Research with Population-Based Data Conference, Baker Institute at Rice University, 2012. Presented by M. Alan Brookhart, Professor in the Department of Population Health Sciences at Duke University. Nationally renowned experts discuss strategies for analyzing large population databases to conduct comparative effectiveness research.

 

Learning Module Notes Modules

  1. Introduction (0:00-1:38)
  2. Brookhart’s Lecture intro: Statistical Methods to Address Confounding Data Analysis (1:38-6:25)
    • Methods Discussed: Propensity score methods & instrumental variable methods
    • Methods differ in terms of assumptions they make & treatment effects they estimate
    • How confounding occurs: treatment & outcome of interest (EX: Disease severity, the healthy user effect, confounding by frailty or serious illness)
    • Summary of these dynamics (6:25-8:10)
    • Comparative new user design minimizes these biases 
  3. Example Study: Non-steroidal anti-inflammatory drugs and GI bleeding risk (8:10-15:45)
    • Data
    • Characteristics of cohort
    • Controlling for observed confounders with statistical models: standard multivariate outcome model OR propensity score and IPTW methods
    • Propensity Score: Probability of receiving treatment (X) given confounders
    • Estimating the propensity score
    • Propensity score theory: *if all confounders are measured & model for treatment is correct, treatment assignment does not depend on the confounders given the PS. Among people with the same propensity score, treatment is effectively randomized. 
  4. Hypothetical distribution of propensity scores (15:45-16:58)
    • Significant area of non-overlap
  5. Methods of using PS (16:58-18:05)
    • With an estimated propensity score, how do you estimate the treatment effect?
  6. Matching on the PS: Matching the exposed patient to an unexposed patient with similar Propensity scores (18:05-19:15 )
    • Limitation: may lose many patients, Generalizability
  7. Inverse Probability of Treatment Weighting (IPTW) (19:15-23:01)
    • Each subject was weighted by the inverse of the probability that they received their observed treatment
    • Example
    • IPTW estimates the average effect of treatment in the population, and assumes “positivity” (aka non-zero probability of being treated or not)
    • Challenge: working with poorly defined populations 
  8. Alternative use of Weights: SMR (Standardize Mortality Ratio) Weight (23:01-23:51)
    • Uses treated group as the standard
    • Yield: effect of treatment among the treated 
  9. Comparing IPTW vs SMR (23:51-24:43)
  10.  Example: Thrombolysis & Mortality (24:43-27:04)
  11. NSAID Example: Return (27:04-32:09)
    • Data
    • Results
    • Limitations: unmeasured confounding
    • Proxies in claims
  12. High-dimensional propensity score algorithm (32:09-33:21)
    • Sources of codes – identify frequently occurring codes 
  13. Apply the High-dimensional propensity score algorithm to the NSAID example cohort (33:21-35:20 )
    • Strengths & limitations of the algorithm
  14. NSAID Example: PS Methods do not find evidence of a clear benefit (35:20-37:51)
    • Instrumental variable methods
    • Estimate treatment effects, place bounds of effects, and rely on having an instrumental variable.
    • Define IV Structural assumptions  
  15. Intention to Treat (ITT) Approach (37:51-39:33)
    • In RCTs with non-compliance, as-treated can be biased estimate of the effect of treatment 
    • IV estimator is a rescaled ITT estimator
  16. NSAID Example: What instrument can we use? (39:33-46:47)
    • How do we make physician preference operational?
    • Estimating preferences and changes throughout the course of the study
    • Limitations
    • Interpretations of result  
  17. Conclusions (46:47-50:16)
    • We cannot know which method is correct – use judgment to decide
    • Recap: Strengths & weaknesses of propensity score methods & instrumental variables 
  18. Q&A (50:16-1:01:33)
  19. Concluding remarks for the conference (1:01:33-1:02:21)

 
  Learning Modules