With RJ Niewoehner and Yong Xia (PhD student)
Last Update (Jun '26): Submitted a MAJOR REVISION at Manufacturing & Service Operations Management
Working Paper: SSRN Link to Working Paper
Problem Definition: Despite emergency care guidelines that emphasize prompt diagnostic ordering, emergency physicians often evaluate multiple patients before submitting diagnostic orders. We call this practice batch ordering. Batch ordering conflicts with prompt-ordering norms because it delays diagnostic initiation for earlier patients in the sequence, but it may also reduce order-entry setup costs and synchronize diagnostic work. We study why a cross-patient sequencing choice that appears to delay care can improve patient flow.
Methodology/Results: We use detailed order-level data from more than 272,000 patient encounters and estimate batching effects using an instrumental-variables design that leverages within-physician persistence in prior-shift batching intensity. The central result is paradoxical: batch ordering increases turnaround time for individual diagnostic orders, yet reduces overall ED service time. We explain this pattern through two channels. First, batching changes placement timing: physicians place larger initial diagnostic bundles and fewer later orders, moving diagnostic work earlier within the visit. Second, batching improves synchronization: diagnostic results arrive in a tighter window, shortening the time to actionable information. A fork-join model shows why these forces can reduce patient service time even when individual diagnostic branches are slower. Benefits concentrate under manageable congestion and attenuate under high congestion. Results remain stable across complementary designs, alternative batching definitions, restricted samples, and alternative instruments.
Managerial Implications: Diagnostic ordering is both a cross-patient sequencing problem and a within-patient clinical decision. ED leaders should avoid treating batch ordering as uniformly harmful or uniformly efficient. Workflow and EHR design can preserve the coordination benefits of coalesced ordering by reducing avoidable order-entry setup costs and supporting targeted batching when waiting costs are low, while discouraging batching for time-sensitive patients or under severe congestion.
With Luxi Shen
Last Update (July '26): Submitted a MAJOR REVISION at Management Science
Working Paper: SSRN Link to Working Paper
Emergency departments face persistent pre-treatment waiting, but whether it shapes physician decisions after care begins is poorly understood. We analyze electronic medical record and cost data from two ED systems with different intake rules: a U.S. acuity-prioritized system and a Chinese first-come-first-served system. We instrument wait with prior-hour walk-in arrival counts and use a U.S. ransomware diversion shock as complementary evidence. Longer waits increase first-round diagnostic ordering in both settings (1-2.5 orders per hour in the U.S.; 1-6 across Chinese departments), and encounter-level ordering also rises, showing that the response is not merely earlier order entry. In the U.S. hospital's order-content analysis, added first-round orders concentrate in low-yield rather than guideline-central tests. The wait-induced increase in time-to-workup initiation appears when capacity allows but attenuates under front-end constraints, while ordering continues to rise. Longer waits do not consistently improve clinical outcomes; instead, they raise post-wait service time and encounter-level expense. A preregistered hospital-physician vignette shows that patient-conveyed waiting shifts intended first-contact ordering under fixed clinical facts. Field diagnostics show that the ordering response is stronger among higher-pain patients and does not increase with physician workload. The evidence supports a cross-agent spillover: an upstream operational delay changes physician decisions after care begins and propagates into downstream resource use. Pre-treatment wait is a managerial lever on both patient flow and diagnostic demand. These findings make reducing avoidable delay a priority and motivate hospitals to test communication during the wait, capacity protection at first contact, and diagnostic-backlog visibility at order entry.
With Yongyi Guo, Hongyu Shan, Zhengyuan Zhou
Last Update (Oct '25): Submitted a MAJOR REVISION at Manufacturing & Service Operations Management
Working Paper: Upon Request
Problem Definition: In the past three years, a particularly important development in the digital advertising industry is the shift from second-price auctions to first-price auctions for online display ads. This shift immediately motivated the intellectually challenging question of how to bid in first-price auctions, because unlike in second-price auctions, bidding one's private value truthfully is no longer optimal. In this paper, we study how to adaptively bid in repeated first-price auctions under binary feedback currently used by almost all ad exchanges-where the bidder only sees whether he wins or loses the bid after each auction.
Methodology/Results: We assume others' highest bid can be modelled by $m_t = \langle \theta^*, x_t \rangle + \epsilon_t$, where $x_t \in \mathbf{R}^d$ is a context representing all the relevant features about this auction (e.g. the user features and the impression opportunity features) $\theta^*$ is an unknown vector of parameters and $\epsilon_t$ is a mean-zero random variable that has an unknown cumulative distribution function (CDF). We develop a semi-nonparametric adaptive bidding policy that achieves $\tilde{O}((d^2T)^{\frac{2\nu+1}{4\nu-1}})$ regret, when the CDF of $\epsilon_t$ is $\nu$-times continuously differentiable ($\nu \ge 2$).
Managerial Implications: Our results reveal an interesting phenomenon: as the non-parametric uncertainty component becomes smoother, the regret performance also improves. In the limit, as the CDF of $\epsilon_t$ becomes infinitely smooth, we obtain the regret bound of $\tilde{O}(d\sqrt{T})$, which is minimax optimal in the time horizon up to log factors.
Low-Acuity Patients Delay High-Acuity Patients in the Emergency Department
With Mohsen Bayati, Erica L. Plambeck, Michael Aratow
Last Update: July 2025
Working Paper: SSRN Link to Working Paper
Problem Definition: Emergency departments face persistent congestion and pre-treatment waiting, yet little is known about how pre-treatment waiting shapes physician behavior at first contact and downstream operations.
Methodology/Results: We analyze electronic medical record and cost data from two hospital systems: a U.S. teaching hospital with acuity-prioritized intake and a Chinese hospital that operates first-come-first-served within departments. Using operational timestamps, we verify intake discipline and align estimation with operational queues. We identify the causal effect of pre-treatment wait using registration-time counts of patients still ahead in the relevant queue (high-priority patients in the U.S.; same-department patients in China) and a ransomware diversion shock that increased congestion in the U.S. system. Across both settings, longer waits increase first-round diagnostic ordering (1-2 additional orders per hour in the U.S.; 1-7 across Chinese departments). Waiting also increases time devoted to the initial evaluation when front-end capacity is slack, but this time response collapses under high physician workload and high bed occupancy, while ordering expansion persists. We find no consistent improvements in disposition accuracy or 30-day revisits, but clear operational costs through longer post-wait service time and higher expenditures. A randomized vignette experiment with practicing physicians reproduces the ordering response, supporting a cross-agent spillover from patient waiting to physician decisions.
Managerial Implications: Our findings suggest that more than an access metric, pre-treatment wait serves as a critical input that shifts utilization and downstream resource use. Practical levers include improving communication during the wait and making downstream diagnostic constraints more visible at order entry so ordering decisions better reflect capacity.
With Erica L. Plambeck
Last Update: February 2026
Working Paper: Upon Request
Abstract: In a hospital that aims to have fewer patients leave the Emergency Department without being seen by a physician (LWBS), we field-tested two approaches for displaying an algorithmic prediction of low-acuity patients' wait time to see a physician. The first approach is the prediction rounded to a multiple of 10 minutes, and the second is an interval designed to communicate that the wait time could be even 20 minutes longer. Relative to the control with no wait time information, both approaches significantly reduce the likelihood of LWBS, with the interval approach being more effective. Improved waiting satisfaction, as indicated by our incentivized satisfaction survey of ED patients, and a higher anticipated wait time with the interval approach, indicated by our online experiment, may contribute to these effects. Consistent with prospect theory, we find that to the extent that patients' actual wait time exceeds the displayed wait time, they have higher likelihood to LWBS. According to emergency medicine literature, many patients with ESI level 4-5 and complaint ``dental pain” or ``medication refill” need not be in the ED. Unfortunately, our intervention is most effective at reducing LWBS by those patients.
With Pooyan Kazemian
Last Update: July 2026
Working Paper: Upon Request
Abstract: Weight loss after the first year of diabetes prevention is a useful but incomplete guide to continuation decisions, because participants with similar weight loss can differ substantially in residual metabolic risk and in the expected benefit of additional support. In a landmark analysis of participants free of diabetes at year 1, we examined whether routine year 1 glycemic measures improve post–year 1 risk stratification and treatment targeting beyond weight loss alone. Participants with similar year 1 weight loss showed substantial heterogeneity in metabolic status and later diabetes risk. Adding fasting glucose to year 1 weight loss improved discrimination for post–year 1 diabetes from 0.728 to 0.801, whereas adding hemoglobin A1c provided little further gain; adding the 2-hour oral glucose tolerance test further increased the concordance index to 0.830. In policy analyses, the expected benefit of continued support varied sharply across year 1 metabolic states. At a 10% intensification budget, a routine reassessment strategy based on year 1 weight loss, fasting glucose, and hemoglobin A1c prevented about 36 diabetes cases per 1000 participants, roughly twice as many as a weight-loss-only rule and roughly 3 times as many as random assignment. The oral glucose tolerance test added little to prioritization when resources were tight but improved action choice as coverage expanded. Together, these findings suggest that weight loss alone is not sufficient for year 1 continuation decisions and that routine year 1 reassessment with standard laboratory measures can materially improve the allocation of scarce follow-up support.