1 / 50

Alternatives to Randomized Trials for Estimating Treatment Efficacy (or Harm)

Alternatives to Randomized Trials for Estimating Treatment Efficacy (or Harm). Thomas B. Newman, MD, MPH Professor of Epidemiology and Biostatistics and Pediatrics, UCSF. AltToRcts2012. Lecture Outline. Announcements Background Instrumental variables and natural experiments

Download Presentation

Alternatives to Randomized Trials for Estimating Treatment Efficacy (or Harm)

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Alternatives to Randomized Trials for Estimating Treatment Efficacy (or Harm) Thomas B. Newman, MD, MPH Professor of Epidemiology and Biostatistics and Pediatrics, UCSF AltToRcts2012

  2. Lecture Outline • Announcements • Background • Instrumental variables and natural experiments • Measuring alternate variables to estimate bias • Propensity scores • Practice problems

  3. Announcements • Exam question due in section Thursday, 11/15, PREFER EARLIER! • Collaboration OK; see new instructions on web • 2 pages maximum • Take-home final will be posted/handed out in section 11/29, discussed in lecture 12/6

  4. Background • Why do randomized blinded trials? • Randomize: to assemble comparable groups (avoid confounding) • Blinding to avoid placebo effect, cointerventions, and bias in measuring outcome variable • Observational studies • May be able to assemble comparable groups or use statistical adjustment • Won’t be blinded

  5. Why is it hard to assemble comparable groups without randomizing? • People who get treated differ from those who don’t • Important differences are with respect risk of the outcome • Treated people often at higher risk (confounding by indication for treatment); “suppression” of benefit • Treated people may be at lower risk (selection bias, healthy complier effect)

  6. Pre-test: Audience Participation Observational studies can never establish causation. Proof of causation requires randomized trials. True or false? • True • False

  7. Do you believe there is a causal relationship between… • Electric shock and recovery from ventricular fibrillation? • Acetaminophen overdose and liver failure? • Wearing glasses for refractive errors and improved vision? • Land mine explosions and limb injuries?

  8. When is causal inference from observational studies easy? • Outcomes • Notrelated to indications for treatment • Rarely if ever occurs spontaneously • Highly localized in time or space • Strongly associated with treatment • Treatment • well-understood biologically • very rapidly acting

  9. When it’s hard: • Outcomes are related to indications or selection for treatment, are delayed, non specific, weak or not well understood • Learning disabilities in children treated with anticonvulsants • Suicide in users of antidepressants • Exclusive breast feeding and lower risk of sudden infant death syndrome

  10. Natural Experiments and Instrumental Variables • Natural experiments: Find a times or places when/where receipt of treatment was unlikely to be related to prognosis/risk of outcome (except through known covariates) • E.g., time-series analyses where something changed (e.g. new intervention became available) • Instrumental variables (IV): measurable factors that influence probability oftreatment (or exposure to the factor of interest) that are not otherwise associated with outcome

  11. Use of large databases • Allows use of (weak) surrogate measures for actual predictor • Biased towards null • Achieve statistical significance with large sample size • Algebraically reverse bias towards null (with various assumptions)

  12. Delayed Effects of the Military Draft on Mortality • Origin of study: Agent Orange concern • Design: “Randomized natural experiment” using the draft lottery • Data source: computerized death certificate registries, CA and PA • Predictor variable of interest: military service Hearst N, Newman TB, Hulley SB. NEJM 1986; 314:620-24

  13. Why not compare outcomes according to the predictor variable of interest? • Biased comparison – those who serve in the military start out healthier • “Healthy warrior effect” • Similar to “healthy worker effect” and “healthy vaccinee effect.”

  14. Delayed Effects of the Military Draft on Mortality • The instrumental variable measured: draft lottery number below cutoff (based on date of birth) • Instrumental variable associated with predictor variable of interest, not independently associated with outcome

  15. BUT: Having an eligible number was a poor measure of military service:

  16. Draft Lottery Study: Results

  17. Long-term Survival Following Partial vs Radical Nephrectomy Among Older Patients With Early-Stage Kidney Cancer* • Retrospective cohort study • Subjects: Medicare beneficiaries with early kidney cancer treated with partial or radical nephrectomy • Partial: remove the part of the kidney with tumor. (Newer treatment) • Radical: remove entire kidney, collecting system, adrenals and lymph nodes • Outcome: long term survival *Tan et al. JAMA 2012;307(15):1629-1635

  18. Audience response • What would be the main reason to do an RCT to compare survival? • Need for blinding for outcome measure • Need for randomization to avoid confounding by indication • Need to avoid stage-migration bias • Need to avoid slippery linkage bias

  19. Long-term Survival Following Partial vs Radical Nephrectomy Among Older Patients With Early-Stage Kidney Cancer* • Instrument: “Differential distance” to a partial nephrectomy surgeon • Outcomes: Total mortality and kidney cancer-specific mortality *Tan et al. JAMA 2012;307(15):1629-1635

  20. Differential Distance Instrument* • Distance to closest surgeon performing partial nephrectomy minus distance to closest surgeon performing radical nephrectomy *Tan et al. JAMA 2012;307(15):1629-1635

  21. Long-term Survival Following Partial vs Radical Nephrectomy Among Older Patients With Early-Stage Kidney Cancer* • Lower mortality for partial nephrectomy • Details: Biostat 215 (Spring Qtr) *Tan et al.,JAMA 2012;307(15):1629-1635

  22. RCT as an Instrumental Variable: Health effects of exclusive breast feeding • Can’t do RCT of exclusive breast-feeding • Can do RCT of breast-feeding PROMOTION • Assignment to BF promotion group should be associated with exclusive breast feeding, but not independently associated with outcome • Need very large sample size • Algebraic correction

  23. Promotion of Breastfeeding Intervention Trial (PROBIT) • Cluster-randomized trial at 31 sites in Belarus • Subjects 17,046 term singleton infants with birth weight > 2500 g, initially breastfed • Intervention: WHO/UNICEF “Baby Friendly Hospital Initiative” • Outcomes: • Breastfeeding at 3,6,9,12 months • Allergic, gastrointestinal and respiratory disease • F/U to 12 months on 16,491 (96.7%) Kramer MS, et al. JAMA 2001;285:413-20.

  24. PROBIT, RQ #1 • Does a “Baby Friendly Hospital” increase exclusive breastfeeding? • Predictor = Group assignment • Outcome = Exclusive breast feeding • Intention-to-treat (ITT) analysis is fine • Exclusive BF at 3 months (rounded) 40% vs 5%; P < 0.001

  25. Probit RQ #2 • Does exclusive breastfeeding reduce the risk of eczema in the infant? • If the only effect of intervention related to eczema is increasing exclusive BF, then • Predictor = Group assignment • Outcome = Eczema • ITT analysis: biased towards null; informative mainly if study positive • Eczema risk 3.3% vs 6.3%; adjusted OR = 0.54 (95% CI 0.31-.95 based on GLIMMIX; P = 0.03)

  26. PROBIT, RQ #3 • How much does exclusive breastfeeding reduce the risk of eczema in the infant? (What is the NNEBF*? ) • Predictor = Group assignment • Outcome = Eczema • ITT won’t work -- too much misclassification. (Gives the number needed to be exposed to the intervention, not the NNEBF.) *Number Needed to Exclusively Breast Feed

  27. Algebraic correction • If all of the difference in eczema is due to the difference in exclusive breast feeding, it can be shown* that the ARR is: • In other words, the observed risk difference is divided by the difference in proportions exposed to the treatment of interest. * See speaker notes for this slide.

  28. NNEBF and caveat • Since ARR = 8.6%, NNEBF to prevent 1 case of eczema is about 1/.086 = 12 • Caveats: • Results are for the effect of breastfeeding in response to the intervention • Assumes the only effect of the Baby Friendly Hospital is via difference in exclusive breastfeeding • Similarly, effects of draft lottery only apply to those who served as a result of the lottery.

  29. Summary/other examples • If variables known NOT to be associated with outcome are associated with treatment of interest, consider this approach. • Generalizes to many “natural experiments.” • E.g., an intervention is intermittently available, or only available to certain groups. -- different outcome by day of the week, etc.

  30. More natural experiments: • Costs of discontinuity of care: increased laboratory test ordering in patients transferred to a different team the next morning* • Effect of increased ED copayment: decrease in ED visits without an increase in hospitalizations or ICU admissions** • Aircraft cabin air recirculation and symptoms of the common cold: no difference by type of air recirculation in aircraft *** * Lofgren, RO. J Gen Intern Med. 1990;5:501-5 **Hsu J, et al. Health Services Research 2006;41:1801-20 *** Zitter JN et al. JAMA 2002;288:483-6

  31. Alternate variables to estimate bias or confounding • Measure an outcome that WOULD be affected by bias, but not by treatment of interest (and see if it is) • Measure a predictor that WOULD cause the same bias as the predictor of interest (and see if it does)

  32. Alternate outcome: Observational study of screening sigmoidoscopy • Possible bias: patients who agree to sigmoidoscopy are likely to be different • Solution: measure an outcome that would be similarly affected by bias • Results: • Decreased deaths from cancers within the reach of the sigmoidoscope (OR= 0.41) • No effect on deaths from more proximal cancers (OR= 0.96). Selby et al, NEJM 1992;326:653-7

  33. Alternate outcome: Admission day of the week as an instrument to study staffing ratios • Higher mortality on weekends for diagnoses thought to be related to staffing (ruptured abdominal aortic aneurysms, acute epiglottitis and pulmonary embolism) • No difference in mortality for control conditions (myocardial infarction, intracerebral hemorrhage, acute hip fracture) Bell CM, Redelmeier DA (2001). "Mortality among patients admitted to hospitals on weekends as compared with weekdays." N Engl J Med345(9): 663-8.

  34. Alternate predictor: Pioglitazone, Rosiglitazone and Bladder Cancer • Nested case-control study using the UK General Practice Research Dtabase • Subjects enrolled at time of new prescription of an antidiabetes drug • Pioglitazone associated with subsequent bladder cancer incidence (OR = 1.83; 95% CI 1.10 to 3.05 • No association with Rosiglitazone (OR = 1.14 (95% CI 0.78 to 1.68) Azoulay et al BMJ 2012;344:e3645

  35. Alternate Predictor: Suicide Risk in Bipolar Disorder During Treatment With Lithium and Divalproex • Retrospective cohort study of Kaiser Permanente and Group Health patients with bipolar disorder • Compared with no treatment, patients treated with Valproex at 2.1 times suicide risk • Concern: confounding by indication • Results: Suicides per 1000 person/years • 31.3 for treatment with divalproex • 15 for no treatment (P<0.001) • 10.8 for Lithium (P<0.001) • If confounding by indication, expect same bias for Lithium Goodwin et al. JAMA. 2003;290:1467-1473

  36. Initial Mood Stabilizer Prescription by Year of Initial Diagnosis Goodwin et al. JAMA. 2003;290:1467-1473

  37. Estimating biases: Cautionary Tale • Nurses’ Health Study • Vitamin E assoc. with decreased risk of CHD (RR ~0.6) • No significant effect of multiple vitamins • Health Professionals Study • Vitamin E assoc. with decreased risk of CHD (RR ~0.6) • No significant effect of Vitamin C • TN began taking Vitamin E * N Engl J Med. 1993;328:1444-9 and 1450-6

  38. Meta-analysis: high-dosage vitamin E supplementation may increase all-cause mortality Miller ER et al. Ann Intern Med. 2005 Jan 4;142(1):37-46

  39. Propensity Scores -1 Big picture: want to know if association between treatment and outcome is CAUSAL Recall competing explanation = confounding by indication for treatment: Factor must be associated with outcome Factor must be associated with treatment Traditional approach: adjust for factors associated with outcome

  40. Propensity Scores -2 Alternative approach: Create a new variable, propensity to be treated with the intervention Then match, stratify, or include it in multivariable analyses Advantages: Better power to control for covariables (because receipt of the intervention may be much more common than occurrence of the outcome) You can more easily tell when treated and untreated groups are not comparable

  41. Propensity to receive treatment Propensity to receive treatment Propensity to receive treatment 1 1 1 C 0 0 0 Treated Treated Not Treated Not Treated Treated Not Treated www.chrp.org How Much Overlap In The Propensity Scores Do We Want? A B

  42. Example: Therapeutic patient education and all-cause mortality in patients with chronic heart failure: A propensity analysis* RQ: Does therapeutic patient education reduce all-cause mortality in patients with chronic heart failure? Design: Cohort study Subjects: 3237 consecutive patients with chronic heart failure in a French cohort study Predictor: Receipt of patient education Outcome: All-cause mortality Crude result: 17% vs 31% mortality after mean 27 months *Juillier et al. Int J Cardiol, in press

  43. Analysis using Propensity Scores Two multivariable analyses: Predictors of receiving education Predictors of death Predictors of receiving education turned into a propensity score Educated and noneducated matched on education propensity score Compare mortality in matched groups

  44. Survival in Full Cohort Recall total N=6174

  45. Survival in Propensity-Matched Patients Recall total N=3237

  46. Propensity score caveats • Can only compare subjects whose propensity scores overlap • Can only generalize to subjects who could have received either treatment • Limitations similar to exclusions from clinical trials • Important variables may be missing from your model • Most convincing when RR gets farther from 1 with propensity analysis

  47. Propensity score question • Are there examples where: • 1. Propensity score analyses gave substantially different results from traditional multivariate methods (e.g. logistic regression) and • 2. The propensity score answer was closer to the right answer? • TN is still looking!

  48. Problem 10.1 • RQ: Do small doses of mercury in the vaccine preservative thimerosal cause autism? • Data source: electronic data from a large HMO (very high N) • Opportunity: • Rhogam (Rh Immune Globulin) is given to Rh-negative women during pregnancy (if they get good prenatal care) • It had 25 mcg of Hg per dose until 2001 • You have data from an electronic record 1990 to the present with blood types, Rhogam, diagnoses of autism • Assume: • You are interested in effects of 25 mcg of Hg given prenataly • The incidence of autism has been increasing • Blood type is not associated with autism • What could you use as an instrument to study the effect of thermosal? • What groups would you compare?

  49. Problem 10.5 • Does perioperative use of lipid-lowering drugs decrease mortality following cardiac surgery? • Why are error limits on the first bar so much wider? • Should we encourage everyone to get this treatment?

  50. Questions?

More Related