Epidemiology: 11) Propensity Scores

Front 1

Propensity score

Back 1

Likelihood that a person will receive a treatment, E+, given potential confounders

Front 2

Utility of propensity scores

Back 2

Subjects in E+ and E- groups with equal propensity scores will tend to have similar distributions in covars used to estimate propensity (eliminates confounding by these covars)

Front 3

5 steps to estimate propensity score

Back 3

1) Identify potential confounders (include if uncertain)
2) Model E+ as a function of confounders using entire cohort (E+ is outcome var for propensity score estimation)
3) DO NOT INCLUDE D+
4) Logistic regression usually used
5) Propensity score = Pr(E+|confounders)

Front 4

three matching techniques for Propensity scores

Back 4

1)Nearest available matching on estimated propensity score
2) Mahalanobis metric matching including propensity score
3)Nearest available Mahalanobis metric matching with calpers defined by propensity score

Front 5

3 parts to Nearest available matching on estimated propensity score

Back 5

1) Select E+ subjects and find E- with closest prop score
2) repeat untill all E+ matched
3) It is easiest method

Front 6

Mahalanobis metric matching including propensity score

Back 6

Mahalanobis distance: measure of distance between to subjects which is a fn of the matching covars as well as propensity score

Front 7

Nearest available Mahalanobis metric matching with calpers defined by propensity score

Back 7

Combo of Nearest available matching on estimated propensity score and Mahalanobis metric matching including propensity score. (preset amount)

Front 8

3 Limits of Propensity scores

Back 8

1) Can only adjust for observed confounding covars, unlike RTC which adjusts for all
2) Work better in large samples in order to attain distributional balance of observed cavars
3) Irrelevant covars reduces efficiency