1 A Rorschach Test
S. Stanley Young, NISS Jessie Q. Xia, NISS Banff, Canada Dec 15, 2011 Variable Importance in Environmental Studies
Current Challenges in Statistical Learning 1. Statistical methods 2. Data quality 3. Invalid claims a. Multiple testing b. Multiple modeling c. Bias
Great Smog of '52 or Big Smoke 12,000 estimated deaths
Pope et al. 2009
6 Studied Variables Life Expectancy life-table methods Per capita income (in thousands of $) Lung Cancer (Age standardized death rate) COPD (Age standardized death rate) High-school graduates (proportion of population) PM2.5 (μg/m3) Black population (proportion of population) Population (in hundreds of thousands) 5-Year in-migration (proportion of population) Hispanic population (proportion of population) Urban residence (proportion of population)
7 First Analysis, Regression VariableSS FirstSS Last Income Lung Cancer COPD High School Population PM Hispanic Black Urban Migration
8 Recursive Partitioning
9 Variable Importance VariableRegressionRP Income COPD Lung Cancer PM High School %Black Pop Density %Hispanic Migration Urban
10 East versus West Krewski et al Health Effects In. Enstrom 2005 Inhalation Toxicology Bell et al Env Health Pers Smith et al Inhalation Toxicology Jerrett 2010 CARB workshop
11 Fine particles and Mortality Pope co-author, 2000.
12 Ozone and Mortality
13 Variable Importance Regression Recursive Partitioning
14 Longevity versus PM2.5 East : Gray West : Red
15 Longevity versus Income
16 Hans Rosling's 200 Countries, 200 Years
17 Summary to this point Income is very important. PM2.5 is 4th or 5th in importance. PM2.5 is not important in West. Pope knew or should have known the East/West heterogeneity.
18 E1: Breakfast cereal and boy babies
19 P-value plot
E2 : Peto, NEJM, statins and cancer Hypothesis: The (SEAS) trial has raised the hypothesis that adding ezetimibe to statin therapy might increase the incidence of cancer.
The claim fails to replicate. The relative risk is wide (95% CI, 1.13 to 2.12; 99% CI, 1.02 to 2.33; uncorrected P = before any allowance is made for this being the hypothesis-generating result. NB: 16 x = SEAS New Studies
E3: A multiple testing and modeling train wreck chemicals medical outcomes demographic covariates 275 x 32 = 8800 x 2^10 = ~9 million A CDC “systems” train wreck in progress! JAMA
Author Interpretation There exists an increased risk of myocardial infarction in patients exposed to abacavir and didanosine within the preceding 6 months. E4 : Bias Example: Lancet DAD study
First drug use (Text, page 1422, and Table 3)
25 E4 : BMJ versus JAMA (1) Conclusion: The risk of oesophageal cancer increased with 10 or more prescriptions for oral bisphosphonates and with prescriptions over about a five year period. BMJ 2010; 341:c4444
26 E4: BMJ versus JAMA (2) Conclusion: Oral bisphosphonates was not significantly associated with incident of esophageal or gastric cancer. JAMA 2010; 304(6): 657‐663
27 A Rorschach Test With large, complex data sets, there is enough flexibility to get what you want/need.
28 Consumer Wishes Honest science Valid claims Claims in context + and – of data and methods
29 What do we have? (Deming) A systems failure. Essentially no process control. Journals operating by “quality by inspection”. Workers are happy. Management failure.
30 What to do? Funding agencies need to require data access on publication. Editors need to give up quality by inspection require split sample strategy require number of claims at issue.
31 Statisticians Eventually society will figure it out; Scientific claims are (most) often wrong. Essentially all claims are supported by statistics. Society will ask, “Where were the statisticians?”
32 Contact Stan Young