Optimal Dynamic Treatment Regimes

Slides:



Advertisements
Similar presentations
Controlling for Time Dependent Confounding Using Marginal Structural Models in the Case of a Continuous Treatment O Wang 1, T McMullan 2 1 Amgen, Thousand.
Advertisements

A workshop introducing doubly robust estimation of treatment effects
2 – In previous chapters: – We could design an optimal classifier if we knew the prior probabilities P(wi) and the class- conditional probabilities P(x|wi)
Lecture 4 Linear random coefficients models. Rats example 30 young rats, weights measured weekly for five weeks Dependent variable (Y ij ) is weight for.
Dynamic Treatment Regimes, STAR*D & Voting D. Lizotte, E. Laber & S. Murphy LSU ---- Geaux Tigers! April 2009.
Dynamic Treatment Regimes, STAR*D & Voting D. Lizotte, E. Laber & S. Murphy ENAR March 2009.
Dynamic Treatment Regimes, STAR*D & Voting D. Lizotte, E. Laber & S. Murphy Psychiatric Biostatistics Symposium May 2009.
Statistical Methods For Engineers ChE 477 (UO Lab) Larry Baxter & Stan Harding Brigham Young University.
Advanced Statistics for Interventional Cardiologists.
Poisson Random Variable Provides model for data that represent the number of occurrences of a specified event in a given unit of time X represents the.
Biostatistics Case Studies 2005 Peter D. Christenson Biostatistician Session 5: Classification Trees: An Alternative to Logistic.
Review of Chapters 1- 5 We review some important themes from the first 5 chapters 1.Introduction Statistics- Set of methods for collecting/analyzing data.
Chapter 6 Lecture 3 Sections: 6.4 – 6.5.
Propensity Score Matching for Causal Inference: Possibilities, Limitations, and an Example sean f. reardon MAPSS colloquium March 6, 2007.
Lecture 12: Cox Proportional Hazards Model
Single-Factor Studies KNNL – Chapter 16. Single-Factor Models Independent Variable can be qualitative or quantitative If Quantitative, we typically assume.
Categorical Independent Variables STA302 Fall 2013.
Chapter 6 Lecture 3 Sections: 6.4 – 6.5. Sampling Distributions and Estimators What we want to do is find out the sampling distribution of a statistic.
1 Hester van Eeren Erasmus Medical Centre, Rotterdam Halsteren, August 23, 2010.
© 2010 Pearson Prentice Hall. All rights reserved 7-1.
Parameter Estimation. Statistics Probability specified inferred Steam engine pump “prediction” “estimation”
Lecture 3 (Chapter 4). Linear Models for Longitudinal Data Linear Regression Model (Review) Ordinary Least Squares (OLS) Maximum Likelihood Estimation.
1 Ka-fu Wong University of Hong Kong A Brief Review of Probability, Statistics, and Regression for Forecasting.
A Study on Speaker Adaptation of Continuous Density HMM Parameters By Chin-Hui Lee, Chih-Heng Lin, and Biing-Hwang Juang Presented by: 陳亮宇 1990 ICASSP/IEEE.
Unit 3: Probability.  You will need to be able to describe how you will perform a simulation  Create a correspondence between random numbers and outcomes.
Research and Evaluation Methodology Program College of Education A comparison of methods for imputation of missing covariate data prior to propensity score.
Matching methods for estimating causal effects Danilo Fusco Rome, October 15, 2012.
Kelci J. Miclaus, PhD Advanced Analytics R&D Manager JMP Life Sciences
Eastern Michigan University
Survival-time inverse-probability- weighted regression adjustment
Analysis for Designs with Assignment of Both Clusters and Individuals
The comparative self-controlled case series (CSCCS)
Harvard T.H. Chan School of Public Health
Constructing Propensity score weighted and matched Samples Stacey L
Math a Discrete Random Variables
Probability Theory and Parameter Estimation I
Relating Reinforcement Learning Performance to Classification performance Presenter: Hui Li Sept.11, 2006.
Chapter 5 Sampling Distributions
Analytics in Higher Education: Methods Overview
Chapter 5 Sampling Distributions
The Nature of Probability and Statistics
Impact evaluation: The quantitative methods with applications
Matching Methods & Propensity Scores
Matching Methods & Propensity Scores
Methods of Economic Investigation Lecture 12
Single-Factor Studies
Review of Hypothesis Testing
Single-Factor Studies
Chapter 5 Sampling Distributions
Chapter 5 Sampling Distributions
Comparing Populations
Pattern Classification All materials in these slides were taken from Pattern Classification (2nd ed) by R. O. Duda, P. E. Hart and D. G. Stork, John.
The Nature of Probability and Statistics
Matching Methods & Propensity Scores
Optimal scaling for a logistic regression model with ordinal covariates Sanne JW Willems, Marta Fiocco, and Jacqueline J Meulman Leiden University & Stanford.
Clinical outcome after SVR: ANRS CO22 HEPATHER
Additional notes on random variables
Least-Squares Regression
Additional notes on random variables
Statistics II: An Overview of Statistics
Analysing RWE for HTA: Challenges, methods and critique
Parametric Methods Berlin Chen, 2005 References:
An Introductory Tutorial
Ch 3. Linear Models for Regression (2/2) Pattern Recognition and Machine Learning, C. M. Bishop, Previously summarized by Yung-Kyun Noh Updated.
Inferential Statistics
New Techniques and Technologies for Statistics 2017  Estimation of Response Propensities and Indicators of Representative Response Using Population-Level.
Chapter 5 Sampling Distributions
Counterfactual models Time dependent confounding
Analysis of Covariance
N. Ganesh, Adrijo Chakraborty, Vicki Pineau, and J. Michael Dennis
Presentation transcript:

Optimal Dynamic Treatment Regimes Julia Wang, Kwan Lee, and Xiwu Lin RWE Analytics at Janssen Quantitative Sciences Sept. 7th, 2017

What is Optimal DTR? Dynamic Treatment Regimes (DTR) are individually tailored treatment sequences based on subject-level information. A DTR algorithm provides a mapping from the space of subject-level information (including baseline, treatments, and responses) to the treatment action space. Optimal DTRs maximize overall average treatment effects for a population following such regimes.

bmiData from r Package iqLearn > head(bmiData) gender race parent_BMI baseline_BMI month4_BMI month12_BMI A1 A2 1 0 1 31.59683 35.84005 34.22717 34.27263 CD MR 2 1 0 30.17564 37.30396 36.38014 36.38401 CD MR 3 1 0 30.27918 36.83889 34.42168 34.41447 MR CD 4 1 0 27.49256 36.70679 32.52011 32.52397 CD CD 5 1 1 26.42350 34.84207 33.72922 33.73546 CD CD 6 0 0 29.30970 36.68640 32.06622 32.15977 MR MR

Baseline Characteristics

Baseline Characteristics Observations On-study Baseline Characteristics

Baseline Characteristics Observations On-study Baseline Characteristics Randomized Groups

Two-Stage Randomized Treatment Set-up MR or CD Stage 2: MR or CD Baseline

Subject-level information at Stage 1: Gender, Race, Baseline BMI, and Parents BMI Subject-level information at Stage 2: Gender, Race, Baseline BMI, Parents BMI and Month 4 BMI Stage 1 Baseline BMI vs. Parent's BMI Stage 2 Month 4 BMI vs. Parent's BMI 50 50 45 45 Baseline BMI Month 4 BMI 40 40 35 35 30 30 25 25 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

Randomized Regime Stage 1 Baseline BMI vs. Parent's BMI Stage 2 Month 4 BMI vs. Parent's BMI MR CD 50 50 45 45 Baseline BMI Month 4 BMI 40 40 35 35 30 30 MR CD 25 25 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

A partition? Optimized Regime Stage 1 Baseline BMI vs. Parent's BMI Month 4 BMI vs. Parent's BMI MR CD 50 50 45 45 Baseline BMI Month 4 BMI 40 40 35 35 30 30 MR CD 25 25 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

Data Generation Process Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space Data Generation Process Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD Y Baseline Characteristics This process can be either randomized or observational, how do we estimate the optimal DTR from the data obtained?

A Conditional Expectation Property The Goal: Maximize E ( Y ) E ( Y ) = E ( E ( Y | S1 )) = E ( E ( E ( Y | S1 , S2 ))) = E ( E ( E ( E ( Y | S1 , S2, S3 )))

A Conditional Expectation Property The Goal: Maximize E ( Y ) E ( Y ) = E ( E ( Y | S1 )) = E ( E ( E ( Y | S1 , S2 ))) = E ( E ( E ( E ( Y | S1 , S2, S3 ))) Maximize by partition the subject-level space defined by S1, S2, S3

A Conditional Expectation Property The Goal: Maximize E ( Y ) E ( Y ) = E ( E ( Y | S1 )) = E ( E ( E ( Y | S1 , S2 ))) = E ( E ( E ( E ( Y | S1 , S2, S3 ))) M-EY Maximize by partition the subject-level space defined by S1, S2, S3

A Conditional Expectation Property The Goal: Maximize E ( Y ) E ( Y ) = E ( E ( Y | S1 )) = E ( E ( E ( Y | S1 , S2 ))) = E ( E ( E ( E ( Y | S1 , S2, S3 ))) M-EY M-EY Maximize by partition the subject-level space defined by S1, S2, S3

A Conditional Expectation Property The Goal: Maximize E ( Y ) E ( Y ) = E ( E ( Y | S1 )) = E ( E ( E ( Y | S1 , S2 ))) = E ( E ( E ( E ( Y | S1 , S2, S3 ))) M-EY

A Conditional Expectation Property The Goal: Maximize E ( Y ) E ( Y ) = E ( E ( Y | S1 )) = E ( E ( E ( Y | S1 , S2 ))) = E ( E ( E ( E ( Y | S1 , S2, S3 ))) M-EY Maximize by partition the subject-level space defined by S1 and S2

A Conditional Expectation Property The Goal: Maximize E ( Y ) E ( Y ) = E ( E ( Y | S1 )) = E ( E ( E ( Y | S1 , S2 ))) = E ( E ( E ( E ( Y | S1 , S2, S3 ))) MM-EY M-EY Maximize by partition the subject-level space defined by S1 and S2

A Conditional Expectation Property The Goal: Maximize E ( Y ) E ( Y ) = E ( E ( Y | S1 )) = E ( E ( E ( Y | S1 , S2 ))) = E ( E ( E ( E ( Y | S1 , S2, S3 ))) MM-EY MM-EY M-EY Maximize by partition the subject-level space defined by S1, and S2

A Conditional Expectation Property The Goal: Maximize E ( Y ) E ( Y ) = E ( E ( Y | S1 )) = E ( E ( E ( Y | S1 , S2 ))) = E ( E ( E ( E ( Y | S1 , S2, S3 ))) MM-EY MM-EY M-EY

A Conditional Expectation Property The Goal: Maximize E ( Y ) E ( Y ) = E ( E ( Y | S1 )) = E ( E ( E ( Y | S1 , S2 ))) = E ( E ( E ( E ( Y | S1 , S2, S3 ))) MM-EY MM-EY M-EY Maximize by partition the subject-level space defined by S1

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD Y Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD Y Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD Y Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD Y Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD M-EY Y Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD M-EY Y Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD M-EY Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD M-EY Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD M-EY Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD MM-EY M-EY Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD MM-EY Baseline Characteristics

Estimation of DTR as a Recursive Partitioning of the Subject-level Information Space History Stage 1 TRT OBS MR CD Stage 2 TRT OBS MR CD Stage 3 TRT OBS MR CD Stage 4 TRT OBS MR CD MMM-EY Baseline Characteristics We need to be able to link the outcome Y with the history space. One such method of optimal DTR estimation is Q-learning

Q-learning Recursive Linear Model Fitting, with treatment by covariates interaction

Q-learning Recursive Linear Model Fitting, with treatment by covariates interaction

Q-learning Recursive Linear Model Fitting, with treatment by covariates interaction MMM-EY MM-EY M-EY

r DTR Packages iqLearn DynTxRegime DTRreg DTR DTRlearn

Illustration 1 using bmiData and Q-learning Stage 1 only and Parent’s BMI only

Month 4 and Month 12 Outcome Variable y = -100*(Month12_BMI – Baseline_BMI)/Baseline_BMI y4 = -100*(Month14_BMI – Baseline_BMI)/Baseline_BMI

Weight Reduction Outcome at Month 4 vs. Parental BMI MR CD 20 20 Negative Percent Change in BMI at 4 Month Negative Percent Change in BMI at 4 Month 10 10 −10 −10 −20 −20 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

Weight Reduction Outcome at Month 4 vs. Parental BMI MR CD 20 20 Negative Percent Change in BMI at 4 Month Negative Percent Change in BMI at 4 Month 10 10 −10 −10 −20 −20 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

Treatment by Parent’s BMI interaction Expected Weight Reduction Outcome at Month 4 vs. Parental BMI Treatment by Parent’s BMI interaction 20 Negative Percent Change in BMI at 4 Month ( 30.95 , 7.96 ) ● 10 −10 MR CD −20 25 30 35 40 45 50 Parent's BMI

●● Expected Weight Reduction Outcome at Month 4 vs. Parental BMI 20 Negative Percent Change in BMI at 4 Month ●●● ●● ● ●●● ●●●●●● ( 30.95 , 7.96 ) 10 ●●●●●● ● ●● ●●●●● ● ● ● ●● ● ●● ●●● ●●●●●● ●● ●●●●●●●●●●●●●●●● ●● ●●●● ●●●● ●● ● ● ●●●●●●●●●● ●● ●● ●●● ●●●●●●●● ●●●● ● ● ● ● ●●●●●● ●●●● ●●● ● ● ● ● ●● ●●● ●● ● ● ● ●●● ● ●● ●●●● ● ● −10 ●● ● MR CD −20 25 30 35 40 45 50 Parent's BMI

●●● Expected Weight Reduction Outcome at Month 4 vs. Parental BMI ● MR CD 20 20 Negative Percent Change in BMI at 4 Month ●● ● Negative Percent Change in BMI at 4 Month ● ● ●● ●●● ●●● 10 ● ●●● 10 ● ●●● ●● ●●● ● ● ● ● ●●●●●●●●●●●● ●●●●●●●●● ●●●●●●●●●● ●●●●●●●●●●●●● ●●●●●● ●●●●●● ●● ●● ● ●● ●● ● ● ●● ● ●● ● ● ●● ●● ● ●● ●● ●● ● ● −10 −10 ● −20 −20 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

●● Expected Weight Reduction Outcome at Month 4 vs. Parental BMI ● MR CD 20 20 ●● Negative Percent Change in BMI at 4 Month ●● ● Negative Percent Change in BMI at 4 Month ● ● ● ●● ● ●● ●● ●●● ● ● ●●● ●● 10 ● ●●● 10 ●● ●● ● ●● ●●● ●● ●● ● ●● ●● ● ●● ● ●●●●● ● ● ●●●● ●●●●●●●●●●●●● ●● ●●●●●● ●● ●●●●●●● ●●●● ● ● ● ● ●●●●●● ●●●●● ● ● ● ● ● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ●● ● ●● ●● ● ● ●● ● ●● ● ● ●● ●● ● ●● ●● ●● ● ● −10 −10 ● −20 −20 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

Optimal Expected Weight Reduction Outcome at Month 4 vs. Parental BMI MR CD 20 20 ●● Negative Percent Change in BMI at 4 Month ●● ● Negative Percent Change in BMI at 4 Month ● ● ● ●● ● ●● ●● ●●● ● ● ●●● ●● 10 ● ●●● 10 ●● ●● ● ●● ●●● ●● ● ●● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●● ●● ● ● ● −10 −10 −20 −20 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

Optimal Expected Weight Reduction Outcome at Month 4 vs. Parental BMI MR CD 20 20 ●● Negative Percent Change in BMI at 4 Month ●● ● Negative Percent Change in BMI at 4 Month ● ● ● ●● ● ●● ●● ●●● ● ● ●●● ●● 10 ● ●●● 10 ●● ●● ● ●● ●●● ●● ● ●● ●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●● ●● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●● ●● ● ● ● −10 −10 −20 −20 25 35 30 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

A Classification Problem Expected Weight Reduction Outcome at Month 4 vs. Parental BMI MR CD Loss 20 20 ●● Negative Percent Change in BMI at 4 Month ●● ● Negative Percent Change in BMI at 4 Month ● ● ● ●● ● ●● ●● ●●● ● ● ●●● ●● 10 ● ●●● 10 ●● ●● ● ●● ●●● ●● ●● ● ●● ●●●●●● ●● ●●●●●● ●● ●●●●●●● ●●●● ● ● ●● ● ●●●●● ●● ● ● ● ●●●● ●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ● ●●●●●● ●●●●● ● ● ● ● ● ● ●● ●● ● ●● ●● ● ● ●● ● ●● ● ● ●● ●● ● ●● ●● ●● ● ● −10 −10 ● A Classification Problem −20 −20 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

Illustration 2 using bmiData Stage 1 only and Parent’s BMI and Baseline BMI

Model Fitted MR (A1=1) and CD (A1=-1)

Predicted Stage 1 Optimal Treatment based on Month 4 Outcome MR Group Demarcation Line MR better CD better 50 45 Baseline BMI 40 35 30 25 30 35 40 45 50 Parent's BMI

Predicted Stage 1 Optimal Treatment based on Month 4 Outcome CD Group Demarcation Line MR better CD better 50 45 Baseline BMI 40 35 30 25 30 35 40 45 50 Parent's BMI

Expected Weight Reduction Outcome at Month 4 vs. Parental BMI Parent’s BMI only 20 Negative Percent Change in BMI at 4 Month ( 30.95 , 7.96 ) ● 10 −10 MR CD −20 25 30 35 40 45 50 Parent's BMI

Predicted Stage 1 Optimal Treatment based on Month 4 Outcome MR Group Demarcation Line MR better CD better 50 45 Baseline BMI 40 35 30 30.95 25 30 35 40 45 50 Parent's BMI

Predicted Stage 1 Optimal Treatment based on Month 4 Outcome MR Group Demarcation Line MR better CD better 50 45 Baseline BMI 40 35 30 30.95 25 30 35 40 45 50 Parent's BMI

●● Expected Weight Reduction Outcome at Month 4 vs. Parental BMI ● MR CD 20 20 ●● Negative Percent Change in BMI at 4 Month ●● ● Negative Percent Change in BMI at 4 Month ● ● ● ●● ● ●● ●● ●●● ● ● ●●● ●● 10 ● ●●● 10 ●● ●● ● ●● ●●● ●● ●● ● ●● ●● ● ●● ● ●●●●● ● ● ●●●● ●●●●●●●●●●●●● ●● ●●●●●● ●● ●●●●●●● ●●●● ● ● ● ● ●●●●●● ●●●●● ● ● ● ● ● ●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ● ●● ●● ● ●● ●● ● ● ●● ● ●● ● ● ●● ●● ● ●● ●● ●● ● ● −10 −10 ● −20 −20 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

Optimal vs. Predicted Month 4 Outcome MR CD 18 20 Optimal Predicted Month 4 Outcome Optimal Predicted Month 4 Outcome 16 10 14 12 10 −10 8 −20 6 −20 −10 10 20 6 8 10 12 14 16 18 Predicted Month 4 Outcome Predicted Month 4 Outcome

Illustration 3 using bmiData Both Stages Stage 1: Parents BMI and Baseline BMI Stage 2: Parents BMI and Month 4 BMI

Models Fitted Stage 2: y ~ gender + parent_BMI + month4_BMI + A2*(parent_BMI + month4_BMI) A2 coded as 1 and -1 Stage 1: Opt_y~ gender + race + parent_BMI + baseline_BMI + A1*(parent_BMI + baseline_BMI) A1 coded as 1 and -1

Predicted Stage 2 Optimal Treatment based on Month 12 Outcome MR Group Demarcation Line MR better CD better 45 40 Month 4 BMI 35 30 25 30 35 40 45 50 Parent's BMI

Predicted Stage 2 Optimal Treatment based on Month 12 Outcome CD Group Demarcation Line MR better CD better 45 40 Month 4 BMI 35 30 25 30 35 40 45 50 Parent's BMI

Predicted Stage 1 Optimal Treatment based on Optimal Month 12 Outcome MR Group Demarcation Line MR better CD better 50 45 Baseline BMI 40 35 30 25 30 35 40 45 50 Parent's BMI

Predicted Stage 1 Optimal Treatment based on Optimal Month 12 Outcome CD Group Demarcation Line MR better CD better 50 45 Baseline BMI 40 35 30 25 30 35 40 45 50 Parent's BMI

Models Fitted Stage 2: y ~ gender + parent_BMI + month4_BMI + A2*(parent_BMI + month4_BMI) A2 coded as 1 and -1 Contrast Term Stage 1: Opt_y~ gender + race + parent_BMI + baseline_BMI + A1*(parent_BMI + baseline_BMI) A1 coded as 1 and -1

Estimated Contrast Term vs. Parent's BMI in Stage 2 MR (coded as 1) CD (coded as −1) 4 4 2 2 Estimated Contrast Term Estimated Contrast Term −2 −2 −4 −4 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

Switch!!! Estimated Contrast Term vs. Parent's BMI in Stage 2 MR (coded as 1) CD (coded as −1) 4 4 2 2 Estimated Contrast Term Estimated Contrast Term Switch!!! −2 −2 −4 −4 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

Switch!!! Estimated Contrast Term vs. Parent's BMI in Stage 2 MR (coded as 1) CD (coded as −1) 4 4 2 2 Estimated Contrast Term Estimated Contrast Term Switch!!! −2 −2 −4 −4 25 30 35 40 45 50 25 30 35 40 45 50 Parent's BMI Parent's BMI

The Goal of Optima DTR maximize overall average treatment effects for a population following such regimes The outcome of a fixed treatment regime can be estimated from the population of subjects whose observe regime is the same as the optimal regime.

Estimated Mean Y 9.9284

7.9175 8.0631 3.5236 6.2016 Estimated Mean Y 9.9284

What if the data is observational? Use Inverse Probability Weighting, at each stage It creates an approximate pseudo-randomized population at each stage Why?

IPW creates an approximate pseudo-randomized population Assuming there are N subjects in the study, and the probability of being treated is p We will have (on average) N*p treated subjects. If we weight each of these subjects by 1/p, then these N*p subjects will represent N*p*(1/p)=N subjects. The same reasoning applies to the control subjects weighted by 1/(1-p), and they will represent on average N*(1-p)*(1/(1-p))=N subjects.

An IPW Illustration Assuming there are N=90m (m is an integer multiplier) subjects in the study, evenly spread within a single confounder with 9 subgroups; The probability of being treated varies from p = 0.1, 0.2, to 0.9 within each subgroup of the confounder. Within each subgroup, we will have (on average) n1=10m*p treated and n2=10m*(1-p) controls. This leads to: n1*(1/p) = n2* (1/(1-p)) = 10m Conclusion: IPW balances the distribution of the confounders. However, in practice, these probabilities, also called propensity scores, need to be estimated.

Distribution of Confonders before IPW ● ● 0.1 0.2 ● 0.3 ● 0.4 0.5 ● 0.6 ● 0.7 0.8 ● 0.9 Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Gr. 6 Gr. 7 Gr. 8 Gr. 9 Subjects Confonder Subgroups

● Distribution of Confonders for Treated Group before IPW Gr. 1 Gr. 2 Subjects Confonder Subgroups

Distribution of Confonders for Treated Group before IPW ● 1/ 0.9 ● 1/ 0.8 ● 1/ 0.7 ● 1/ 0.6 ● 1/ 0.5 ● 1/ 0.4 ● 1/ 0.3 ● 1/ 0.2 ● 1/ 0.1 Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Gr. 6 Gr. 7 Gr. 8 Gr. 9 Subjects Confonder Subgroups

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Distribution of Confonders for Treated Group before IPW ● ● ● ● ● ● ● ● ● ● Subjects ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 1 / 0.1 ● 2 / 0.2 ● 3 / 0.3 ● 4 / 0.4 ● 5 / 0.5 ● 6 / 0.6 ● 7 / 0.7 ● 8 / 0.8 ● 9 / 0.9 Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Gr. 6 Gr. 7 Gr. 8 Gr. 9 Confonder Subgroups

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Distribution of Confonders for Treated Group before IPW ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Subjects ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Gr. 6 Gr. 7 Gr. 8 Gr. 9 Confonder Subgroups

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Distribution of Confonders for Treated Group after IPW ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Subjects ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Gr. 6 Gr. 7 Gr. 8 Gr. 9 Confonder Subgroups

Distribution of Confonders before IPW ● ● 0.1 0.2 ● 0.3 ● 0.4 0.5 ● 0.6 ● 0.7 0.8 ● 0.9 Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Gr. 6 Gr. 7 Gr. 8 Gr. 9 Subjects Confonder Subgroups

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Distribution of Confonders for Control Group before IPW ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Subjects ● ● ● ● ● ● ● ● ● ● Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Gr. 6 Gr. 7 Gr. 8 Gr. 9 Confonder Subgroups

Distribution of Confonders for Control Group before IPW ● 1/ 0.9 ● 1/ 0.8 ● 1/ 0.7 ● 1/ 0.6 ● 1/ 0.5 ● 1/ 0.4 ● 1/ 0.3 ● 1/ 0.2 ● 1/ 0.1 ● 1/ 0.2 Subjects Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Gr. 6 Gr. 7 Gr. 8 Gr. 9 Confonder Subgroups

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Distribution of Confonders for Control Group before IPW ● 9 / 0.9 ● 8 / 0.8 ● 7 / 0.7 ● 6 / 0.6 ● 5 / 0.5 ● 4 / 0.4 ● 3 / 0.3 ● 2 / 0.2 ● 1 / 0.1 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Subjects ● ● ● ● ● ● ● ● ● ● Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Gr. 6 Gr. 7 Gr. 8 Gr. 9 Confonder Subgroups

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Distribution of Confonders for Control Group before IPW ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Subjects ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Gr. 6 Gr. 7 Gr. 8 Gr. 9 Confonder Subgroups

● Distribution of Confonders for Control Group after IPW Gr. 1 Gr. 2 Subjects Confonder Subgroups

● ● Distribution of Confonders after IPW Gr. 1 Gr. 2 Gr. 3 Gr. 4 Gr. 5 Subjects Confonder Subgroups

Propensity Scores The more accurately a propensity score can be estimated, the better it is at reducing confounding and bias. How to estimate propensity scores properly is a very big topic in itself.

7.9175 8.0631 3.5236 6.2016 Estimated Mean Y 9.9284 dwols: 10.0805

Comparing Q-learning vs. dwols using bmiData Stage 2 dwols Q-learning CD MR 94 1 4 111 Stage 1 dwols Q-learning CD MR 111 1 2 96