EPI201_2018_07_Effect Modification (1).pdf - Fall2018...

Unformatted text preview: Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Key Concepts Effect measure modification □ additive □ multiplicative Lecture 7 Effect Measure Modification Impact of baseline risk Confounding vs. effect measure modification Reciprocity of effect measure modification Epidemiologic Methods 1 Murray A. Mittleman Department of Epidemiology, Harvard School of Public Health 1 2 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Additive and Multiplicative Effect Measure Modification Effect Measure Modification In the presence of effect measure modification, the magnitude of the association between exposure and disease varies according to the value of (across strata of) a third factor, which is called an effect modifier. Presence or absence, as well as magnitude, of effect measure modification is usually determined by the scale chosen for measuring departures from homogeneity across strata Additive scale: Strata have different Effect measure modification is an intrinsic phenomenon and cannot be eliminated from a study through clever design □ Cumulative incidence difference (CID) □ Incidence rate difference (IRD) Multiplicative scale: Strata have different Effect measure modification is a finding to be reported rather than a bias to be avoided □ Cumulative incidence ratio (CIR) □ Incidence rate ratio (IRR) □ Odds ratio (OR) Synonyms: interaction, synergy, antagonism, heterogeneity of treatment effect Issue of external validity therefore limits generalizability (transportability) Evaluating effect measure modification on each scale provide different information regarding the association of interest 3 4 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Ratio vs. Difference Measures Angiosarcoma of the Liver The rate ratio for angiosarcoma of the liver from occupational vinyl chloride exposure is near 1000. Some authors claim that difference measures are measures of the public health and clinically relevant effect of exposure and ratio measures are measures of the “scientific effect”. However, this interpretation depends on the assumptions that are made in the model of causation. The incidence rate for angiosarcoma of the liver is 10 to 20 cases per year in the United States ≈ 3.33/108 PY. Epidemiologists commonly use multiplicative statistical models that by default assume the ratio measure is constant across all strata. In these models, ratio measures are independent of the baseline rate (or risk) of disease occurrence. The rate difference is then [1000*3.3/108 PY] ‐ [3.3/108 PY] = 3.297/105 PY. In the sufficient‐component cause model, 2 component causes will have additive effects when they do not occur in the same sufficient cause (have independent causal action). This is an extremely small difference. This suggests that from a public health point of view, the removal of vinyl chloride from the occupational environment will prevent few cases of angiosarcoma of the liver since the baseline incidence is so low. 5 6 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Cigarette Smoking and Lung Cancer Cigarette Smoking and Coronary Heart Disease (CHD) CHD Mortality Lung Cancer Mortality Non‐Smokers IRD 55‐64 969/105 PY 501/105 PY 468/105 PY 65‐74 1,710/105 PY 1,015/105 PY 695/105 PY 1800 1600 1400 IRR Age 21‐39 cig/day Non‐Smokers IRD IRR 1.93 55‐64 400/105 PY 40/105 PY 360/105 PY 10 1.68 65‐74 720/105 PY 80/105 PY 640/105 PY 9 Incidence Rate/105 Person‐Years 21‐39 cig/day Incidence Rate/105 Person‐Years Age Non‐Smoker Smoker 1200 1000 1710 800 600 969 400 200 1015 501 0 800 700 600 500 400 65‐74 720 300 400 200 100 0 55‐64 Non‐Smoker Smoker 40 55‐64 7 80 65‐74 8 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Confounding Confounding vs. Effect Measure Modification Scale Dependence Effect Measure Modification Elaborated Bias that we try to description of true prevent/address association No Yes Validity Concern Internal External Statistical Test No Yes Arises from non‐exchangeability between the exposed and unexposed Confounding Concept Not scale dependent Primarily an issue of internal validity □ estimates that lack internal validity (biased estimates) should not be generalized Confounding and effect measure modification □ not mutually exclusive □ can co‐occur, but it is also possible to have one without the other 9 10 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Perioperative Beta‐Blocker Therapy and Mortality Perioperative beta‐blocker therapy and mortality after major noncardiac surgery by Revised Cardiac Risk Index (RCRI) score Question… RCRI Score 0 1.36 (1.27 ‐ 1.45) RCRI Score 1 RCRI Score 2 1.09 (1.01 ‐ 1.19) 0.88 (0.80 ‐ 0.98) RCRI Score 3 0.71 (0.63 ‐ 0.80) RCRI Score >=4 0.58 (0.50 ‐ 0.67) 0.99 (0.95 – 1.04) Overall Adjusted 0.40 0.60 0.80 1.00 2.00 Odds Ratio (95% Confidence Interval) (log scale) *The Revised Cardiac Risk Index is a risk assessment tool developed to predict a patient’s risk of having cardiac complications associated with non‐cardiac surgery; scores range from 0 (low risk) to ≥ 4 (very high risk). Lindenauer PK, Pekow P, Wang K, Mamidi DK, Gutierrez B, Benjamin EM. N Engl J Med. 2005;353(4):349‐61. 11 12 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Perioperative Beta‐Blocker Therapy and Mortality Alcohol Intake and Coronary Heart Disease Perioperative beta‐blocker therapy and mortality after major noncardiac surgery by Revised Cardiac Risk Index (RCRI) score BACKGROUND: Light to moderate alcohol consumption is associated with a reduced risk of coronary heart disease. This protective effect of alcohol, however, may be confined to middle‐aged or older individuals. Coronary heart disease incidence is low in men <40 years of age and in women <50 years of age; for this reason, study cohorts rarely have the power to investigate the effects of alcohol on coronary heart disease risk in younger adults. This study examined whether the beneficial effect of alcohol on coronary heart disease depends on age. RCRI Score 0 1.36 (1.27 ‐ 1.45) RCRI Score 1 RCRI Score 2 1.09 (1.01 ‐ 1.19) 0.88 (0.80 ‐ 0.98) METHODS AND RESULTS: In this pooled analysis of 8 prospective studies from North America and Europe including 192,067 women and 74,919 men free of cardiovascular diseases, diabetes, and cancers at baseline, average daily alcohol intake was assessed at baseline with a food frequency or diet history questionnaire. An RCRI Score 3 0.71 (0.63 ‐ 0.80) RCRI Score >=4 0.58 (0.50 ‐ 0.67) 0.99 (0.95 – 1.04) inverse association between alcohol and risk of coronary heart disease was observed in all age groups; hazard ratios among moderately drinking men (5.0 to 29.9 g/d) 39 to 50, 50 to 59, and >or=60 years of age were 0.58 (95% confidence interval [CI], 0.36 to 0.93), 0.72 (95% CI, 0.60 to 0.86), and 0.85 (95% CI, 0.75 to 0.97) compared with abstainers. However, the analyses indicated a smaller incidence rate difference between abstainers and moderate consumers in younger adults (incidence rate difference, 45 per 100,000; 90% CI, 8 to 84) than in middle‐aged (incidence rate difference, 64 per 100,000; 90% CI, 24 to 102) and older (incidence rate difference, 89 per 100,000; 90% CI, 44 to 140) adults. Similar results were observed in women. Overall Adjusted 0.40 0.60 0.80 1.00 2.00 Odds Ratio (95% Confidence Interval) (log scale) *The Revised Cardiac Risk Index is a risk assessment tool developed to predict a patient’s risk of having cardiac complications associated with non‐cardiac surgery; scores range from 0 (low risk) to ≥ 4 (very high risk). CONCLUSIONS: Alcohol is also associated with a decreased risk of coronary heart disease in younger adults; however, the absolute risk was small compared with middle‐aged and older adults. Hvidtfeldt UA, Tolstrup JS, Jakobsen MU et. al. Circulation. 2010;121(14):1589‐97. Lindenauer PK, Pekow P, Wang K, Mamidi DK, Gutierrez B, Benjamin EM. N Engl J Med. 2005;353(4):349‐61. 13 14 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Alcohol Intake and Coronary Heart Disease Age Rate Ratio 39‐50 0.58 (95% CI: 0.36‐0.93) 45 (95% CI: 8‐84) 50‐59 >=60 0.72 (95% CI: 0.60‐0.86) 64 (95% CI: 24‐102) 0.85 (95% CI:0.75‐0.97) 89 (95% CI: 44‐140) Reciprocity of Effect Measure Modification Effect measure modification is completely reciprocal: If Z modifies the association between E and D, E modifies the association between Z and D Rate Difference (per 100,000 PY) The choice of Z as opposed to E as the modifier is a function of the hypothesis being evaluated This can be related back to component causes in the same sufficient cause Hvidtfeldt UA, Tolstrup JS, Jakobsen MU et. al. Circulation. 2010;121(14):1589‐97. 15 16 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Question… Epidemiologic Methods 1 Murray A. Mittleman Department of Epidemiology, Harvard School of Public Health 17 18 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Reciprocity of Effect Measure Modification Scale Dependence of Effect Measure Modification Gender, Red Wine and Headaches Does gender modify the association between red wine consumption and headaches? Risk of Headache CIE+ Red Wine CIE‐ No Red Wine CIR CID Male .2 .8 0.2/0.8=0.25 0.2‐0.8=‐0.6 Female .3 .9 0.3/0.9=0.33 0.3‐0.9=‐0.6 If both X and Z have effects and there is no modification (heterogeneity) of the difference measure for one factor by the other factor, there has to be modification of the ratio measure. Does red wine consumption modify the association between gender and headaches? Risk of Headache CIE+ Male CIE‐ Female CIR CID Red Wine .2 .3 0.2/0.3=0.67 0.2‐0.3=‐0.1 No Red Wine .8 .9 0.8/0.9=0.89 0.8‐0.9=‐0.1 Conversely, if X and Z have effects and there is no modification of the ratio measure, there has to be modification of the difference measure. Multiplicative scale? YES _____ Additive scale? NO _____ Rothman Modern Epidemiology 3rd Edition. Chapter 5 19 20 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Effect Measure Modification Effect Measure Modification Graph 2 6 5 5 Disease Incidence Disease Incidence Graph 1 6 4 3 2 1 3 2 1 0 0 Young Difference Ratio 4 Young Old Young Old EMM* 2‐1=1 2/1=2 5‐4=1 5/4=1.25 No Yes Difference Ratio Old Young Old EMM* 2‐1=1 2/1=2 4‐2=2 4/2=2 Yes No *Effect measure modification *Effect measure modification 21 22 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Effect Measure Modification Graph 3 6 Disease Incidence 5 4 3 2 Question… 1 0 Young Difference Ratio Old Young Old EMM* 2‐2=0 2/2=1 5‐5=0 5/5=1 No No *Effect measure modification 23 24 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Effect Measure Modification Graph 4 6 Disease Incidence 5 4 3 2 Question… 1 0 Young Young Old Old EMM* Difference Ratio *Effect measure modification 25 26 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Effect Measure Modification Effect Measure Modification Graph 5 Graph 6 6 6 5 5 Disease Incidence Disease Incidence Harvard School of Public Health Epidemiologic Methods I Fall 2018 4 3 2 1 4 3 2 1 0 Young Young 0 Old Old Young EMM* Difference Ratio Difference Ratio Old Young Old EMM* 4‐2=2 4/2=2 2‐4=‐2 2/4=0.5 Yes Yes *Effect measure modification *Effect measure modification 27 28 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Harvard School of Public Health Epidemiologic Methods I Fall 2018 Effect Measure Modification Key Concepts Graph 7 Disease Incidence 6 Effect measure modification 5 □ additive □ multiplicative 4 3 Impact of baseline risk Confounding vs. effect measure modification Reciprocity of effect measure modification 2 1 0 Young Young Difference Ratio Old Old 2.25‐2=0.25 5.75‐2.75=3 2.25/2=1.1 5.75/2.75=2.1 EMM* Yes Yes *Effect measure modification 29 30 ...
View Full Document