- Ph.D. Candidate in Operations, Information, and Techonlogy
- Stanford Graduate School of Business
- Email contact: joelgoh AT stanford DOT edu
- Mailing address:
- Graduate School of Business
- Ph.D. Program
- E146 Faculty Building East
- 655 Knight Way
- Stanford, CA 94305
- Research Interests:
- Healthcare Applications
- Inventory Theory / Supply Chain Management
- Robust Optimization
- Curriculum Vitae
"Active Postmarketing Drug Surveillance for Multiple Adverse Events" [PDF]
, with Margret V. Bjarnadottir, Mohsen Bayati and Stefanos A. Zenios
Active postmarketing drug surveillance is important for consumer safety. However, existing methods have limitations that prevent their direct use for active drug surveillance. One of the important consideration that has been absent thus far is the modeling of multiple adverse events and their interactions. In this paper, we propose a method to monitor the effect of a single drug on multiple adverse events, which explicitly captures interdependence between events. Our method uses a sequential hypothesis testing paradigm, and employsan intuitive test-statistic. Stopping boundaries for the test-statistic are designed by asymptotic analysis and by reducing the design problem to a convex optimization problem. We apply our method to a dynamic version of Cox's proportional hazards model, and show both analytically and numerically how our method can be used as a test for the hazard ratio of the drug. Our numerical studies further verify that our method delivers Type I/II errors that are below pre-specified levels and is robust to distributional assumptions and parameter values.
"The Relationship Between Workplace Practices and Mortality and Health Costs in the United States" [PDF]
, with Jeffrey Pfeffer and Stefanos A. Zenios
Even though epidemiological evidence links specific workplace practices to health outcomes, the aggregate contribution of these factors to overall mortality and health spending in the United States is not known. In this paper, we build a model to estimate the excess mortality and incremental health expenditures associated with the following ten workplace exposures: Unemployment, lack of health insurance, exposure to shift work, long working hours, job insecurity, work-family conflict, low job control, high job demands, low social support at work, and low organizational justice. Our model uses input parameters obtained from publicly-accessible data sources. We estimated health spending from the Medical Panel Expenditure Survey, joint probabilities of workplace exposures from the General Social Survey, and we conducted an extensive meta-analysis of the epidemiological literature to estimate the relative risks of poor health outcomes associated with facing these workplace exposures. The model was designed to overcome inherent data limitations, which center around the correlated nature of workplace exposures and the potential double-counting of their effects on health. The model separately derives optimistic and conservative estimates of the effect of multiple workplace exposures on health, and uses optimization to calculate upper and lower bounds around each estimate, which accounts for the correlation between exposures. We find that more than 120,000 deaths per year and approximately 5-8% of annual healthcare costs are associated with and may be attributable to how U.S. companies manage their work force. Our results suggest that more attention should be paid to management practices as important contributors to health outcomes and costs in the U.S..
"Using Uncertain Markov Chains to Assess Cost-effectiveness of Medical Innovations" [PDF]
, with Mohsen Bayati and Stefanos A. Zenios
Cost-effectiveness studies of medical innovations often suffer from data inadequacy. When Markov chains are used as a computational tool for such studies, a possible consequence of this lack of data is that the transition matrix of the chain cannot be estimated precisely. In this paper, we study how to compute maximal and minimal values for the discounted value of the chain (with respect to a vector of state-wise costs or rewards) as these uncertain transition parameters jointly vary within a given uncertainty set. We show that these problems are computationally tractable if the uncertainty set has a row-wise structure. Conversely, we prove that if the row-wise structure is even relaxed slightly, the problems become computationally intractable (NP-hard). We apply our model to assess the cost-effectiveness of fecal immunochemical testing (FIT), a new screening method for colorectal cancer (CRC). Our results show that despite the large uncertainty in FIT's performance, it is a cost-effective alternative screening modality to colonoscopy.
Inventory Theory / Supply Chain Management
"Multi-Echelon Inventory Management with Advance Order Commitments", with Evan L. Porteus
We extend the Clark-Scarf serial multi-echelon inventory model to include advance commitment discounts. In our base model, each facility has the option to process at two different cost rates; the cheaper rate can only be utilized up to the amount committed in the previous period. We characterize the structure of the optimal dynamic inventory and advance commitment policies, which are stated in terms of Clark-Scarf echelon inventory levels and analogously defined echelon capacities. The optimal cost function is additively separable, there exists an optimal inventory policy that a generalized finite base-stock policy, and there exists an optimal advance commitment policy that is a base-stock policy. Extensions and limitations are discussed.
"Periodic-Review Inventory Management with Dynamic Forecasts", with Evan L. Porteus
We consider a periodic-review inventory model of a single retailer that provides dynamic forecasts of its future orders to a single supplier. The retailer's forecasted order quantity for each future period represents its commitment to procuring the forecasted quantity in that period, and it is given a unit discount for early commitment. The retailer may revise its forecasts periodically, but revisions carry some financial consequences. These revisions must also be made within certain limits of its previous forecast. We characterize the optimal policies for the retailer and show that there exists an optimal inventory and forecast policy that are generalized finite base-stock policies.
- "Total Cost Control in Project Management via Satisficing" [link], with Nicholas G. Hall, 2013. Management Science 59(6), pp. 1354-1372
We consider projects with uncertain activity times and the possibility of expediting, or crashing, them. Activity times come from a partially specified distribution within a family of distributions. This family is described by one or more of the following details about the uncertainties: support, mean, and covariance. We allow correlation between past and future activity time performance across activities. Our objective considers total completion time penalty plus crashing and overhead costs. We develop a robust optimization model that uses a conditional value-at-risk satisficing measure. We develop linear and piecewise-linear decision rules for activity start time and crashing decisions. These rules are designed to perform robustly against all possible scenarios of activity time uncertainty, when implemented in either static or rolling horizon mode. We compare our procedures against the previously available Program Evaluation and Review Technique and Monte Carlo simulation procedures. Our computational studies show that, relative to previous approaches, our crashing policies provide both a higher level of performance, i.e., higher success rates and lower budget overruns, and substantial robustness to activity time distributions. The relative advantages and information requirements of the static and rolling horizon implementations are discussed.
- "Portfolio Value-at-Risk Optimization for Asymmetrically Distributed Asset Returns." [link], with Kian Guan Lim, Melvyn Sim, and Weina Zhang, 2012. European Journal of Operations Research 221(2), pp. 397-406
We propose a new approach to portfolio optimization by separating asset return distributions into positive and negative half-spaces. The approach minimizes a newly-defined Partitioned Value-at-Risk (PVaR) risk measure by using half-space statistical information. Using simulated data, the PVaR approach always generates better risk-return tradeoffs in the optimal portfolios when compared to traditional Markowitz mean-variance approach. When using real financial data, our approach also outperforms the Markowitz approach in the risk-return tradeoff. Given that the PVaR measure is also a robust risk measure, our new approach can be very useful for optimal portfolio allocations when asset return distributions are asymmetrical.
- "Robust Optimization Made Easy with ROME" [link], with Melvyn Sim, 2011. Operations Research 59(4), pp. 973-985
We introduce ROME, an algebraic modeling toolbox for a class of robust optimization problems. ROME serves as an intermediate layer between the modeler and optimization solver engines, allowing modelers to express robust optimization problems in a mathematically meaningful way. In this paper, we discuss how ROME can be used to model (1) a service-constrained robust inventory management problem, (2) a project-crashing problem, and (3) a robust portfolio optimization problem. Through these modeling examples, we highlight the key features of ROME that allow it to expedite the modeling and subsequent numerical analysis of robust optimization problems. ROME is freely distributed for academic use at http://www.robustopt.com.
- "Distributionally Robust Optimization and its Tractable Approximations." [link], with Melvyn Sim, 2010. Operations Research 58(4 part 1), pp. 902-917
In this paper we focus on a linear optimization problem with uncertainties, having expectations in the objective and in the set of constraints. We present a modular framework to obtain an approximate solution to the problem that is distributionally robust and more flexible than the standard technique of using linear rules. Our framework begins by first affinely extending the set of primitive uncertainties to generate new linear decision rules of larger dimensions and is therefore more flexible. Next, we develop new piecewise-linear decision rules that allow a more flexible reformulation of the original problem. The reformulated problem will generally contain terms with expectations on the positive parts of the recourse variables. Finally, we convert the uncertain linear program into a deterministic convex program by constructing distributionally robust bounds on these expectations. These bounds are constructed by first using different pieces of information on the distribution of the underlying uncertainties to develop separate bounds and next integrating them into a combined bound that is better than each of the individual bounds.