Principles of case control studies Part III Matching

Principles of case control studies Part III • Matching Many slides in this presentation are from the World Health Organization and the European Programme for Intervention Epidemiology Training, thank you. Piyanit Tharmaphornpilas MD, MPH The International Field Epidemiology Training Program, Thailand

Confounding Hypothesis: Sunbathe is a risk factor for being diabetes mellitus Sunbathe Diabetes mellitus Age is confounding factor ! need to be controlled Reality: Age Sunbathe Diabetes mellitus

How to control confounding factors n Randomisation Restriction n Matching n n n Adjustment Mutivariate analysis

Because age is confounding factor, so )In cohort study) Age of exposed and unexposed population should be comparable ! Then, effect of age on the study association will be taken off. )In case-control) age of cases and controls should be comparable ! If a case ages 30, his control should age 30 too. Reality: Age is confounding factor ! need to be controlled Age Sunbathe Diabetes mellitus

Types of matching n Frequency matching Large strata: Controls are selected in proportion to the number of cases in each strata of the matching variable n Individual matching Small strata : For each case one or more controls are selected with the matching characteristics

Frequency matching Controls are selected in proportion (%) to the number of cases in each strata of the matching variable Age 15 -24 25 -34 35 -44 45 -54 >54 Total Cases 30 30 20 10 10 100 Controls 60 60 40 20 20 200 The distribution of cases and controls is similar for age, and controls are no more representative of the not-ill population for age

Individual matching For each case one or more controls are selected with the matching characteristics No. Case Control 1 Control 2 1 age 30 � 5 2 age 20 � 5 3 age 10 � 5 The distribution of cases and controls is similar for age, and controls are no more representative of the not-ill population for age

Matching : analysis If…. control enrolment is done by matching Then…. analysis should be adjusted for it (by strata)

Adjustment by Mantel-Haenszel Using confounding (matching) variable as strata OR M-H= S [(ai. di) / Ti] S [(bi. ci) / Ti]

Frequency matching : analysis • Stratified analysis on the frequency matching variable • Mantel Haenszel weigthed OR Exposure Strata 1 yes no Total Cases Controls Total ai ci bi di L 1 i L 0 i C 1 i C 0 i Ti Strata j. . OR M-H = S [(ai. di) / Ti] S [(bi. ci) / Ti]

Individual matching analysis Controls Exposed Not Exposed C+/Ctr + C+/Ctr - C-/Ctr + C-/Ctr - Cases Not Exposed Pairs of cases and controls

Individual matching analysis Controls Exposed Not Exposed e f g h Cases Not Exposed Pairs of cases and controls

Controls Exposed C Exposed A S Not exposed E Total S Not exposed Total e f a g h c b d T Odds ratio: f/g

Atherosclerosis risk in Communities study Association between CMV infection and Carotid Atherosclerosis Controls CMV+ CMV- 214 65 42 19 Atherosclerosis CMV- Cases and controls individually match paired by Age group, sex, ethnicity, field center and date of exam From: PD Sorlie et al, cytomegalovirus and carotid Atherosclerosis, Journal of Medical Virology, Vo 42, pp 33 -37, 1994

We cannot analyze a matched case-control study by unmatched method Why? ? Because matching process introduce selection bias This selection bias is controllable by stratified analysis

Matching : advantages n n n When there is a potentially strong confounding variable Tends to increase the statistical power Logistically straightforward way to obtain a comparable control group

Matching: disadvantages n n n Difficult to find a matched control Cannot assess the association between matching variables and outcome Implies some tailoring of the selection of study groups to make them comparable (less representativeness) Once is done cannot undone, risk of overmatching No statistical power is gained if the matched variables are weak confounders

Don’t match (too much) End of part III