Abstract
Dynamic treatment regimes are fast becoming an important part of medicine, with the corresponding change in emphasis from treatment of the disease to treatment of the individual patient. Because of the limited number of trials to evaluate personally tailored treatment sequences, inferring optimal treatment regimes from observational data has increased importance. Q-learning is a popular method for estimating the optimal treatment regime, originally in randomized trials but more recently also in observational data. Previous applications of Q-learning have largely been restricted to continuous utility end-points with linear relationships. This paper is the first attempt at both extending the framework to discrete utilities and implementing the modelling of covariates from linear to more flexible modelling using the generalized additive model (GAM) framework. Simulated data results show that the GAM adapted Q-learning typically outperforms Q-learning with linear models and other frequently-used methods based on propensity scores in terms of coverage and bias/MSE. This represents a promising step toward a more fully general Q-learning approach to estimating optimal dynamic treatment regimes.
Similar content being viewed by others
References
Chakraborty B (2011) Dynamic treatment regimes for managing chronic health conditions: A statistical perspective. Am J Publ Health 101(1):40–45
Chakraborty B, Laber EB, Zhao Y (2013) Inference for optimal dynamic treatment regimes using an adaptive m-out-of-n bootstrap scheme (submitted)
Chakraborty B, Moodie EEM (2013) Estimating optimal dynamic treatment regimes with shared decision rules across stages: An extension of Q-learning (submitted)
Chakraborty B, Murphy SA, Strecher V (2010) Inference for non-regular parameters in optimal dynamic treatment regimes. Stat Methods Med Res 19(3):317–343
Fava M, Rush AJ, Trivedi MH, Nierenberg AA, Thase ME, Sackeim HA, Quitkin FM, Wisniewski S, Lavori PW, Rosenbaum JF, Kupfer DJ (2003) Background and rationale for the sequenced treatment alternatives to relieve depression (STAR*D) study. Psychiatr Clin North Am 26(2):457–494
Golub G, Heath M, Wahba G (1979) Generalized cross-validation as a method for choosing a good ridge parameter. Technometrics 21:215–224
Hastie T, Tibshirani R (1986) Generalized additive models. Stat Sci 1(3):297–318
Hastie T, Tibshirani R (1990) Generalized additive models. Chapman & Hall, London
Huang X, Ning J (2012) Analysis of multi-stage treatments for recurrent diseases. Stat Med 31:2805–2821
Li KC (1987) Asymptotic optimality of C p , C L , cross-validation and generalized cross-validation: Discrete index set. Ann Stat 15:958–975
Moodie EEM, Chakraborty B, Kramer MS (2012) Q-learning for estimating optimal dynamic treatment rules from observational data. Can J Stat 40:629–645
Moodie EEM, Richardson TS (2010) Estimating optimal dynamic regimes: Correcting bias under the null. Scand J Stat 37:126–146
Murphy SA, Oslin DW, Rush AJ, Zhu J (2007) Methodological challenges in constructing effective treatment sequences for chronic psychiatric disorders. Neuropsychopharmacology 32:257–262
Murphy SA (2005) A generalization error for Q-learning. J Mach Learn Res 6:1073–1097
Nahum-Shani I, Qian M, Almirall D, Pelham WE, Gnagy B, Fabiano GA, Waxmonsky JG, Yu J, Murphy SA (2012) Q-Learning: A data analysis method for constructing adaptive interventions. Psychol Methods 17:478–494
R Core Team (2012) R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna. ISBN 3-900051-07-0
Robins JM, Hernán MA, Brumback B (2000) Marginal structural models and causal inference in epidemiology. Epidemiology 11:550–560
Robins JM (2004) Optimal structural nested models for optimal sequential decisions. In: Lin DY, Heagerty P (eds) Proceedings of the second Seattle symposium on biostatistics. Springer, New York, pp 189–326
Rosenbaum PR, Rubin DB (1983) The central role of the propensity score in observational studies for causal effects. Biometrika 70:41–55
Rosthoj S, Fullwood C, Henderson R, Stewart S (2006) Estimation of optimal dynamic anticoagulation regimes from observational data: A regret-based approach. Stat Med 25:4197–4215
Schneider LS, Tariot PN, Lyketsos CG, Dagerman KS, Davis KL, Davis S (2001) National institute of mental health clinical antipsychotic trials of intervention effectiveness (CATIE): Alzheimer disease trial methodology. Am J Geriatr Psychiatry 9:346–360
Shortreed SM, Moodie EEM (2012) Estimating the optimal dynamic antipsychotic treatment regime: Evidence from the sequential-multiple assignment randomized CATIE schizophrenia study. J R Stat Soc, Ser B, Stat Methodol 61:577–599
Song R, Wang W, Zeng D, Kosorok MR (2013) Penalized Q-learning for dynamic treatment regimes (submitted)
Sutton RS, Barto AG (1998) Reinforcement learning: An introduction. MIT Press, Cambridge
Thall PF, Millikan RE, Sung HG (2000) Evaluating multiple treatment courses in clinical trials. Stat Med 30:1011–1128
Thall PF, Sung HG, Estey EH (2002) Selecting therapeutic strategies based on efficacy and death in multicourse clinical trials. J Am Stat Assoc 97(457):29–39
Topol E (2012) Creative destruction of medicine: How the digital revolution and personalized medicine will create better health care. Basic Books, New York
Wood SN (2004) Stable and efficient multiple smoothing parameter estimation for generalized additive models. J Am Stat Assoc 99(467):673–686
Wood SN (2006) Generalized additive models: An introduction with R. Chapman & Hall, London
Wood SN (2011) Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear models. J R Stat Soc B 73(1):3–36
Xin J, Chakraborty B, Laber EB (2012) qLearn: Estimation and inference for Q-learning. R package version 1.0
Zhao Y, Kosorok MR, Zeng D (2009) Reinforcement learning design for cancer clinical trials. Stat Med 28:3294–3315
Zhao Y, Zeng D, Socinski MA, Kosorok MR (2011) Reinforcement learning strategies for clinical trials in non-small cell lung cancer. Biometrics 67(4):1422–1433
Acknowledgements
We would like to thank Dr. Bibhas Chakraborty for insightful discussions. This work is supported by Dr. Moodie’s Discovery Grant from Canada’s Natural Sciences and Engineering Research Council (NSERC).
Author information
Authors and Affiliations
Corresponding author
Appendix: Derivation of the True Dynamic Regime Parameters
Appendix: Derivation of the True Dynamic Regime Parameters
In the following, we calculate the true values of the first-interval decision rule parameters ψ 10 and ψ 11 in terms of γ’s and δ’s, the parameters of the generative model following the calculations in [2].
1.1 A.1 Bernoulli Utility
We begin with the derivations for the case of a Bernoulli utility. Let M=γ 0+γ 1 C 1+γ 2 O 1+γ 3 A 1+γ 4 O 1 A 1+γ 5 C 2+f 1(C 1)+f 2(C 2), and define μ=M+γ 6 A 2+γ 7 O 2 A 2+γ 8 A 1 A 2. It follows that
where f 1=γ 6+γ 7+γ 8, f 2=γ 6+γ 7−γ 8, f 3=γ 6−γ 7+γ 8, and f 4=γ 6−γ 7−γ 8. Further,
Therefore,
Furthermore,
Since A 1∈{−1,1}, we have \((1-A_{1}^{2})=0\), (1+A 1)2=2(1+A 1) and (1−A 1)2=2(1−A 1). It may therefore be deduced that
and
Let \(k_{1}=\frac{1}{4}\operatorname{expit}(\delta_{1}+\delta_{2})\), \(k_{2}=\frac{1}{4}\operatorname{expit}(-\delta_{1}+\delta_{2})\), \(k_{3}=\frac{1}{4}\operatorname{expit}(\delta_{1}-\delta_{2})\), \(k_{4}=\frac{1}{4}\operatorname{expit}(-\delta_{1}-\delta_{2})\). Therefore,
Applying a logit transformation to Q 1(H 1,A 1) gives that the coefficient of A 1 in the above expression for \(\operatorname{logit}(Q_{1}(H_{1}, A_{1}))\) is
and the coefficient of O 1 A 1 in the expression for Q 1 is
1.2 A.2 Poisson Utility
In this section, we derive the correspondence between the Q-function model parameters and the parameters from the data-generating models when the utility is given by a Poisson count. As above, let M=γ 0+γ 1 C 1+γ 2 O 1+γ 3 A 1+γ 4 O 1 A 1+γ 5 C 2+f 1(C 1)+f 2(C 2), and μ=M+γ 6 A 2+γ 7 O 2 A 2+γ 8 A 1 A 2. Then
where f 1, f 2, f 3, and f 4 are as defined above. Thus,
where k 1, k 2, k 3, k 4 are as above. Therefore, the coefficients of A 1 and O 1 A 1 in the above expression take the same form as in the case of a Bernoulli utility:
Rights and permissions
About this article
Cite this article
Moodie, E.E.M., Dean, N. & Sun, Y.R. Q-Learning: Flexible Learning About Useful Utilities. Stat Biosci 6, 223–243 (2014). https://doi.org/10.1007/s12561-013-9103-z
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s12561-013-9103-z