Epidemiology Kept Simple Chapter 12 11/11/2018 Epidemiology Kept Simple Chapter 12 Error in Epidemiologic Research 11/11/2018 Error in Epidemiologic Research

§12.1 Introduction View analytic studies as exercises in measuring association between exposure & disease All such measurements are prone to error We must understand the nature of measurement error if we are to avoid train wrecks* * a disaster or failure, especially one that is unstoppable or unavoidable 11/11/2018

Parameters and Estimates Understanding measurement error begins by distinguishing between parameters and estimates Parameters Error-free Hypothetical Measure effect Statistical estimates Error-prone Calculated Measure association Epidemiologic statistics are mere estimates of the parameters they wish to estimate. 11/11/2018

Target Analogy Let Φ represent the true relative risk for an exposure and disease in a population Replicate the study many times How closely does an estimate from a given study reflect the true relative risk (parameter)? 11/11/2018

Target Analogy 11/11/2018

Random Error (Imprecision) Balanced “scattering” To address random error, use confidence intervals hypothesis tests of statistical significance 11/11/2018

Confidence Intervals (CI) Surround estimate with margin of error Locates parameter with given level of confidence (e.g., 95% confidence) CI width quantifies precision of the estimate (narrow CI  precise estimate) 11/11/2018

Mammagraphic Screening and Breast Cancer Mortality Exposure ≡ mammographic screening Disease ≡ Breast cancer mortality First series (studies 1 – 10): women in their 40s Second series (studies 11 – 15): women in their 50s 11/11/2018

Hypothesis Tests Start with this null hypothesis H0: RR = 1 (no association) Use data to calculate a test statistic Use the test statistic to derive a P-value Definition: P-value ≡ probability of the data, or data more extreme, assuming H0 is true Interpret P-value as a measure of evidence against H0 (small P gives one reason to doubt H0) R.A. Fisher 11/11/2018

Conventions for Interpretation Small P-value provide strong evidence against H0 (with respect to random error explanations only!) By convention P ≤ .10 is considered marginally significant evidence against H0 P ≤ .01 is considered highly significant evidence against H0 Do NOT use arbitrary cutoffs (“surely, god loves P = .06”) R.A. Fisher 11/11/2018

Childhood SES and Stroke Factor RR P-value Crowding (persons/room) < 1.5 = 0.4 1.5 – 2.49 = 1.0 (ref.) 2.5 – 3.49 = 0.6  3.5 = 1.0 trend P = 0.53 Tap water 0.73 P = 0.53 Toilet type flush/not shared = 1.3 flush/shared = 1.0 (referent) no flush = 1.0 trend P = 0.67 Ventilation good = 1.0 (ref.) fair = 1.7 poor = 1.7 trend P = 0.08 Cleanliness good= 1.1 fair = 1.0 (ref.) poor = 0.5 trend P = 0.07 Nonsignif. evidence against H0 Marginally significant evidence against H0 11/11/2018 Source: Galobardes et al., Epidemiologic Reviews, 2004, p. 14

Systematic Error (Bias) Bias = systematic error in inference (not an imputation of prejudice) Amount of bias Direction of bias Toward the null (under-estimates risks or benefits) Away from null (over-estimates risks or benefits) 11/11/2018

Categories of Bias Selection bias: participants selected in a such a way as to favor a certain outcome Information bias: misinformation favoring a particular outcome Confounding: extraneous factors unbalanced in groups, favoring a certain outcome 11/11/2018

Examples of Selection Bias Berkson’s bias Prevalence-incidence bias Publicity bias Healthy worker effect Convenience sample bias Read about it: pp. 229 – 231 11/11/2018

Examples of Information Bias Recall bias Diagnostic suspicion bias Obsequiousness bias Clever Hans effect Read about it on page 231 11/11/2018

Memory Bias Memory is constructed rather than played back Example of a memory bias is The Misinformation Effect ≡ a memory bias that occurs when misinformation affects people's reports of their own memory False presuppositions, such as "Did the car stop at the stop sign?" when in fact it was a yield sign, will cause people to misreport actually happended Loftus, E. F. & Palmer, J. C. (1974). Reconstruction of automobile destruction. Journal of Verbal Learning and Verbal Behaviour, 13, 585-589. 11/11/2018

Recall Bias in Case-Control Studies 11/11/2018

Differential & Nondifferential Misclassification Non-differential misclassification: groups equally misclassified Differential misclassification: groups misclassified unequally 11/11/2018

Nondifferential and Differential Misclassification Illustrations 11/11/2018

Confounding A distortion brought about by extraneous variables From the Latin meaning “to mix together” The effects of the exposure gets mixed with the effects of extraneous determinants 11/11/2018

Properties of a Confounding Variable Associated with the exposure An independent risk factor Not in causal pathway 11/11/2018

Example of Confounding E ≡ Evacuation method D ≡ Death following auto accident Died Survived Total Helicopter 64 136 200 Road 260 840 1100 R1 = 64 / 200 = .3200 R0 = 260 / 1100 = .2364 RR = .3200 / .2364 = 1.35 Does helicopter evacuation really increase the risk of death by 35%? Surely you’re joking! Or is the RR of 1.35 confounded? 11/11/2018

Serious Accidents Only (Stratum 1) Died Survived Total Helicopter 48 52 100 Road 60 40 R1 = 48 / 100 = .4800 R0 = 60 / 100 = .6000 RR1 = .48 / .60 = 0.80 Helicopter evacuation cut down on risk of death by 20% among serious accidents 11/11/2018

Minor Accidents Only (Stratum 2) Died Survived Total Helicopter 16 84 100 Road 200 800 1000 R1 = 16 / 100 = .16 R0 = 200 / 1000 = .20 RR2 =.16 / .20 = 0.80 Helicopter evacuation cut down on risk of death by 20% among minor accidents 11/11/2018

Accident Evacuation Properties of Confounding Seriousness of accident (Confounder) associated with method of evacuation (Exposure) Seriousness of accident (Confounder) is an independent risk factor for death (Disease) Seriousness of accident (Confounder) not in causal pathway between helicopter evaluation (Exposure) and death (Disease) 11/11/2018