holger langkabel phuse 2016 · very short introduction to propensity-score matching procedure...

23
Proposal of a general macro framework for propensity-score matching (RW03) Holger Langkabel PhUSE 2016

Upload: others

Post on 04-Aug-2020

12 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Proposal of a general macro framework for

propensity-score matching

(RW03)

Holger Langkabel PhUSE 2016

Page 2: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Introduction

Very Short Introduction to Propensity-Score Matching

Procedure Overview

Suggested Framework

Conclusion

Page 3: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Introduction

Very Short Introduction to Propensity-Score Matching

Procedure Overview

Suggested Framework

Conclusion

Page 4: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

What do you need propensity-score matching

for?

Scenario 1

Randomized Controlled Trial Non-Interventional Study (NIS)

Random treatment assignment Self-selection into treatment and control

group

Subject characteristics balanced across

groups

Between-group imbalances due to self-

selection

Naïve estimate unbiased Naïve estimate potentially biased

Page 5: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

What do you need propensity-score matching

for?

Scenario 2

• Naïve estimate potentially biased

Study 1 Study 2

Pool

TRT 1 TRT 2

Page 6: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Solution to the Evaluation Problem

• Matching can be done on the (full) vector of covariates (high-dimensional!)

• or on the propensity score (one-dimensional!).

Population (in the data)

Analysis Population

TRT 1

TRT 2

Search for similar

observations ...

... and match them

to the treated

Page 7: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

What is the propensity score?

• The propensity score is the conditional probability of being in the treatment

group given a specific set of covariables:

𝑃 𝑇 = 1 𝑋

• Estimated via probit/logit model or other

Page 8: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Introduction

Very Short Introduction to Propensity-Score Matching

Procedure Overview

Suggested Framework

Conclusion

Page 9: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Main Steps of Propensity-Score Matching

(1) Check for imbalances between control and treatment group.

(2) Estimate the propensity model and predict individual propensity scores.

(3) Match controls on treated subjects.

(4) Check if propensity-score matching reduced the imbalances.

(5) Estimate the treatment effect.

Page 10: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Step (1): Pre-Matching Balance Assessment

Available measures to be compared:

• Mean

• Standard deviation

• Median and other quantiles

• Proportions

• Boxplot

• Histogram

• ...

The use of statistical tests is generally discouraged.

Page 11: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Step (2): Model Estimation and Prediction of

Propensity Scores

• Use a binary response model (e.g. probit or logit model).

• Prediction easy from the programmer’s perspective:

OUTPUT OUT=SAS-data-set PREDICTED=name;

Page 12: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Step (3): Matching

Available algorithms:

• 𝑘-nearest-neighbor matching

• Caliper matching

• Block-wise matching

• Kernel matching

• Combinations of the above

• Whatever you might come up with ...

Page 13: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Step (4): Post-Matching Balance Assessment

Same as Step (1)

Page 14: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Step (5): Treatment-Effect Estimation

• Depends on the matching algorithm applied.

• Might be very easy.

• Might be very complicated.

Page 15: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Introduction

Very Short Introduction to Propensity-Score Matching

Procedure Overview

Suggested Framework

Conclusion

Page 16: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Overall Process Flow

Pre-matching balance

assessment (%ps_check)

PS-calculation (%ps_calc)

Matching (%ps_match)

Post-matching balance

assessment (%ps_check)

Treatment-effect estimation (%ps_est/

%ps_postest)

Study data

ADPS

No balance achieved

Page 17: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

%ps_check

Proposed example macro call:

%ps_check(inds = adsl

, vars = age basebmi baseweig baseheig

, statistics = table boxplot histogram );

Page 18: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

%ps_calc

Proposed example macro call:

%ps_calc(inds = adsl

, invars = age basebmi baseweig baseheig

, link = probit

, outds = adps1

, cs_diagnosis = yes );

Page 19: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

%ps_match

Proposed example macro call for 1-to-1-nearest-neighbor matching with

caliper and without replacement:

%ps_match_1nn(inds = adps1

, outds = adps2

, seed = 12345

, caliper = 0.1

, replacement = off );

Page 20: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

ADPS-data structure

Minimally required variables

USUBJID Unique identifier of the observation to which the information

applies

TRT Treatment status of the observation

PS Predicted propensity score

STRATUM [For block-wise matching:] Stratum to which the observation was

matched

MATCHED 0/1-variable indicating whether the observation was used for

matching

ID_MATCH Unique identifier of the observation which has been matched to

the current observation

WEIGHT [If more than 1 control is matched to a treated observation:]

Weight in the set of matched observations

Page 21: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

ADPS-data structure

Example for 1-to-1-nearest-neighbor matching

USUBJID TRT PS STRATUM MATCHED ID_MATCH WEIGHT

1234 A 0.186 1 2345

2345 B 0.259 1 1234

3456 B 0.846 0

... ... ... ... ...

Page 22: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Introduction

Very Short Introduction to Propensity-Score Matching

Procedure Overview

Suggested Framework

Conclusion

Page 23: Holger Langkabel PhUSE 2016 · Very Short Introduction to Propensity-Score Matching Procedure Overview Suggested Framework Conclusion . Main Steps of Propensity-Score Matching (1)

Doing now what patients need next