Matching In Case Control And Cohort Designs (Midterm 3)

Front 1

Frequency Matching

Back 1

By group

Distribution of matching factors same in cases andcontrols (or exposed and unexposed) but not matchedindividually

Front 2

Individual Matching

Back 2

By case-control

By cohort (exposed-unexposed)

Front 3

Distance Matching

Back 3

Individual matching for multiple factors. Useful for situations with continuous or many factors

Front 4

Purpose of Matching

Back 4

Improve efficiency by protecting against confounder distributions dramatically different between cases and controls (or exposed and unexposed)

Control unmeasured confounders

Front 5

Disadvantages of matching

Back 5

• Cost – complicates sampling scheme– Exclusion of cases when no match found - reduces N(pair matching)
  
• Longer study duration
  
• Reduced flexibility in analysis• Cannot estimate association of the matched variable• Requires use of matched pairs analysis and conditionallogistic regression (pair matching)
  
• Improper matching can prevent estimation of effectsof interest—over-matching
  

Front 6

What variables should be matched

Back 6

• strong confounders
• effect of variable is not of scientific interest (i.e., age, race, sex are common)

Front 7

Case-control matching

Back 7

Studies tend to be small so there is concern aboutoverlap in distribution of confounders if the cases ofdisease differ from the study base dramatically onstrong confounder(s) (e.g., age)

Front 8

Cohort Matching

Back 8

• Studies tend to be large and intended to examinemultiple exposures and outcomes so otherapproaches to controlling confounding usuallypreferable
  
• Prognostic cohort studies more often use matching• Smaller and thus concern about overlap in distribution ofconfounders• Focused on one exposure
  

Front 9

Overmatching

Back 9

• Matching that restricts variability of exposure canreduce efficiency
  
• Matching on an intermediate between exposure anddisease, or a factor that is affected by both exposureand disease can lead to bias
  

Front 10

If you mistakenly match on a collider or pathway variable...

Back 10

this induces selection bias

Front 11

Frequency matching: case-control

Back 11

• ORs must be estimated within strata of the matching factor(s)
  
• The relation between the matching factor and the outcome ismeaningless
  
• Stratification and calculation of an adjusted OR achievesestimation within strata
  
• An adjusted (logistic regression) model with the matchingfactors included as covariates also achieves estimation ofOR within strata of the matching factors
  

Front 12

Frequency matching: cohort

Back 12

Confounding is removed by the matching and no need tocontrol for the matching factors

Front 13

Pair Matching

Back 13

Extreme of when you have a set of matching factors with a largenumber of discrete levels, few individuals will share allvalues

Front 14

Pair Matching OR

Back 14

OR = B/C

Front 15

Breaking the match: Cohort

Back 15

• match themeasure of association will be unbiased
  
• variance of the measure of associationwill be incorrect
  

Front 16

Breaking the match: case-control

Back 16

measure of association will be biased

Front 17

Relationship between pair and frequency matching

Back 17

• pair matching can legitimately be convertedto frequency matching as long as all pairs do not havecompletely unique sets of matched factors
  
• Frequency matching allows more flexibility in analysisthan pair matching
  

Front 18

Main purpose of matching

Back 18

• improveefficiency in adjustment for confounding– Avoiding strata with small N– Thus, most common in small studies
  
• useful to control confounders that aredifficult to measure
  

Front 19

Propensity score matching

Back 19

• used primarily in cohort studies
  
• predicted probability of exposure
  
• creates a counterfactual population

Front 20

Region of support

Back 20

• overlap of propensity scores for the exposed and unexposed distributions
  
• match units within this region