1 of 13

Propensity Score [weighting]

within complex survey

ehsan.karim@ubc.ca

Oct 7, 2020

SPPH 504/007

2 of 13

Propensity score

Weighting

(ATE + ATT)

2

3 of 13

IPW (inverse probability weighting)

How to conduct propensity score weighting?

​

Step 1: Specify PS & fit model

Step 2: Match subjects by PS Convert PS to IPW

Step 3: Covariate balance in matched weighted sample

Step 4: Estimate treatment effect

​

​

For the purposes of illustration, we will first assume that our data was collected via SRS.

3

Exposure model (RA)

Outcome model (MI)

4 of 13

IPW

Step 1: Fit PS model

A~L

Step 2: Convert PS = IPW(ATE)

​

​

Step 3: Check balance

SMD in IPW-weighted data

Step 4: Outcome model with

Weight = IPW

4

IPW = 1/ps, if A = 1

IPW = 1/(1-ps), if A = 0

5 of 13

IPW in complex survey

Step 1: Fit PS model

A~L (survey-weights as design variable / covariate)

Step 2: Convert PS = IPW(ATE)

​

​

Step 3: Check balance

SMD (data weighted by w = IPW * survey-weights)

Step 4: Outcome model with

Weight = IPW * survey-weights

5

IPW = 1/ps, if A = 1

IPW = 1/(1-ps), if A = 0

“sampling weights in the propensity score estimation stage (as weights, not as a covariate)”

6 of 13

IPW

Step 1: Fit PS model

A~L

Step 2: Convert PS = IPW(ATT)

​

​

Step 3: Check balance

SMD in IPW-weighted data

Step 4: Outcome model with

Weight = IPW

6

IPW = 1, if A = 1

IPW = ps/(1-ps), if A = 0

7 of 13

IPW in complex survey (ATT)

Step 1: Fit PS model

A~L (survey-weights as design variable / covariate)

Step 2: Convert PS = IPW(ATT)

​

​

Step 3: Check balance

SMD (data weighted by w = IPW * survey-weights)

Step 4: Outcome model with

Weight = IPW * survey-weights

7

IPW = 1, if A = 1

IPW = ps/(1-ps), if A = 0

8 of 13

Reasonable approach (my summary)

  • PS model: (population-level)
    • use design variables (cluster + strata + weight) to estimate ps (not as covariate)
    • Combined weight = ipw * survey weight
  • Outcome model: (population-level)
    • use design features (strata+psu as well as combined weight) to get population level estimates

8

9 of 13

Estimates and conclusion

Adult patients with RA are at increased risk for MI in US (based on 2007-08 data)?

9

50% increased risk of CVD death in patients with RA

10 of 13

Estimates from NHANES (2007-08) and conclusion

OR: population-based estimates, sample-based not shown

​

​

​

​

* Also conditional estimates if further adjustment made; SE / CI width is a function of n.

10

​

Adjusted Regression

Matching (Zanutto)

Matching (DuGoff)

Matching (design in both stages)

Weighting

(Ridgeway)

Weighting

(DuGoff)

PATT

​

1.87

(0.86 4.07)

1.26

(0.55, 2.88)

1.66

(0.65, 4.28)

1.38

(0.71, 2.71)

1.37

(0.71, 2.67)

PATE

1.66*

(0.71, 3.89)

​

​

​

1.51

(0.68, 3.35)

1.43

(0.62, 3.28)

11 of 13

NHANES vs. CCHS

  • In the public release data, NHANES provides
    • masked variance pseudo-PSUs, and
    • masked variance pseudo-stratum

to account for the complex survey design.

  • CCHS public use microdata file (PUMF) does not contain PSU / Stratum information. Any SE calculation assumes SRS even if weights are used. RDC provides access to master data with these necessary information.

11

12 of 13

Short Reference and Textbook List

12

13 of 13

Thanks!

ehsan.karim@ubc.ca

www.ehsankarim.com

13