Design of large case
Transcription
Design of large case
W. Dana Flanders Emory University 1. 2. 3. 4. 5. Study Goal Cohort Studies Cases and Controls Validity Issues Spatial comparison ` Before studying something we ought to define what we wish to study. ` Obvious. A bit like some Yogisms “It ain’t over til it’s over.” 4 ` Study Goal ◦ Presumed; explicit ` Estimate Causal Risk Ratio ◦ Risk in exposed group divided by what the risk in the exposed group would have been, had they been unexposed ◦ Counterfactual ` Causal Risk Ratio ≈ Causal Rate Ratio ◦ rare disease (e.g. Cancer for short periods) ` Occasionally in case-control studies (population known), Causal Risk Difference, others ` To design a case-control study, first conceptually design a cohort study ` ` To design a case-control study, first conceptually design a cohort study Why? ` ` To design a case-control study, first conceptually design a cohort study Because, then, one can argue rigorously that the odds ratio has meaning in terms of risk ratios, or other meaningful measures (e.g. rate ratio) ` To design a case-control study, first conceptually design a cohort study ◦ Population (inclusion, exclusion criteria) ` To design a case-control study, first conceptually design a cohort study ◦ Population (inclusion, exclusion criteria) ◦ Exposure, covariates (source, collection) ` To design a case-control study, first conceptually design a cohort study ◦ Population (inclusion, exclusion criteria) ◦ Exposure, covariates (source, collection) ◦ Follow-up (identification of incident cases) ` To design a case-control study, first conceptually design a cohort study ◦ Population (inclusion, exclusion criteria) x What is the underlying source population for cases ◦ Exposure, covariates (source, collection) x Which variables, how to collect ◦ Follow-up (identification of incident cases) x Means of identifying cases during observation period, and link to underlying source population ` ` To design a case-control study, first conceptually design a cohort study From another viewpoint, to identify casesone must observe, sometimes incompletely, some cohort (the source population) Recruiting Controls Identifying Cases Source population ` Ideally, will identify and include all, or nearly all, cases in a defined cohort of interest (just like in a cohort study) ` ` ` Ideally, will identify and include all, or nearly all, cases in a defined cohort of interest (just as in a cohort study) Controls are then a random sample from this source population Information on exposures and confounders is available in both cases and controls ` ` Design for selecting controls is fairly straightforward, once know the cohort Some options and general principles next ` General Principles ◦ Purpose of controls – provide estimate of exposure frequency in source population ◦ good control group: random sample from underlying source, perhaps some special eligibility criteria ◦ corollary - selection is independent of exposure (conditional on stratification, if any) ` General Principles ◦ Some specific types, and key advantage x Risk set sampling: sample from source population, just when each case develops – allows estimation of rate ratio, no rare disease (person-time) x Density-sampling – sample controls to estimate p-t, e.g. dynamic cohort x Sampling from the source population at start of follow-up - allows estimation of risk ratio, no rare disease (cohort) x Matching – allows efficient control of confounding by the matching factors ` General Principles ◦ Some common sources x RDD – population based, but account for cell phones x Neighborhood – often similar SES, so control efficient, but probably out here x Hospital- very difficult x Insurance, other rosters ◦ But whatever it is: a good rule is try to have the exposure in controls be representative of that in the source population ` ` ` Key threats are the usual selection bias, confounding, misclassification Apply to both cohort and case-control Special care in case-control study to ◦ Assure temporality (exposure precedes disease) ` Case-control studies often more subject to selection bias ◦ Cases and controls may not be from same cohort ◦ Participation occurs after disease, and exposure known: can depend on both leading to bias ` Observational Study of radiation exposure ◦ Exposure varies with space and time, so primarily spatial, perhaps temporal-spatial comparisons x Risk in one area (exposure level) compared with that in another area (different exposure) ` Opens door for confounding by risk factors that vary spatially, or in time and space ◦ ◦ ◦ ◦ ◦ SES Lifestyle, smoking, alcohol, nutrition, diet Genetics (neighborhoods?) Occupation Access to care ` Observational Study of radiation exposure ◦ Exposure varies with space and time, so primarily spatial, perhaps temporal-spatial comparisons x Risk in one area (exposure level) compared with that in another area (different exposure) ` Opens door for confounding by risk factors that vary spatially, or in time and space ◦ ◦ ◦ ◦ ◦ ` SES Lifestyle, smoking, alcohol, nutrition, diet Genetics (neighborhoods?) Occupation Access to care Issues the same in Cohort and Case-control ` ` ` ` ` Think cohort first, then case-control Case-control similar to cohort, but with sampling to save time and money Case-control can yield valid estimates: Rate or Risk Ratios, sometimes even differences Extra challenges – participation, assuring cases and controls come from same cohort Common challenge to both- confounding “It ain’t over til its over.” The first part of this presentation is over, but time remains for questions. 27 ` Assumptions ◦ The distribution of residences with exposures X and specular exposure Y is same as distribution of residences with exposures Y and specular exposure X (symmetry of actual and specular exposure distributions) ◦ The actual and specular residences are similar w/r/t confounders