Previous month:
December 2012
Next month:
February 2013

January 2013

Profiles: William Bonvillian

William Bonvillian is the director of the MIT Washington Office, which he joined in January 2006 after seventeen years as a senior policy advisor and chief counsel to Senator Joseph Lieberman. In that capacity, he “works with Congress and the executive branch to raise understanding of the contributions of higher education and research to the national welfare” (from the press release). He further engages, along with the rest of the MIT DC staff, in “research and development and education efforts throughout government, managing a wide portfolio of related policy issues”, as explained on the MIT Washington Office page.

This includes “communicating the latest developments in these key policy areas”, “facilitat[ing] interactions between campus experts and those in Congress, the Administration, associated federal agencies, and national organizations seeking that expertise,” as well as “engaging and educating MIT’s students in the larger science and technology policy-making process, either through classes taught on campus or via the office’s internship and science policy fellows programs.”

Bonvillian is an adjunct faculty member at Georgetown and Johns Hopkins University, where he teaches courses in science, technology and innovation, especially related to energy. He also teaches the MIT “bootcamp” on science and technology policy during the Independent Activities Period every January. In addition, he is the co-author, along with Georgetown professor Charles Weiss (formally the Distinguished Professor of Science, Technology, and International Affairs at Georgetown University's Walsh School of Foreign Service, and the first Science and Technology Adviser to the World Bank), of Structuring an Energy Technology Revolution, published in 2009 by MIT Press. His bio further states that “[h]e was the recipient of the IEEE Distinguished Public Service Award in 2007 and was elected a Fellow by the American Association for the Advancement of Science in 2011.”

Bonvillian has given numerous speeches and presentations on the topic of science and technology innovation. In particular, he delivered the 2012 D. Allan Bromley Memorial Lecture at the University of Ottawa, where his talk focused on “bringing advanced innovation to the manufacturing sector.” In his abstract, he summarizes the evolution of manufacturing innovation in the US from “innovate here/produce here” to “innovate here/produce there”, and echoes the concerns of innovation experts such as Gary Pisano and Willy Shih, authors of the famous modularity/maturity matrix (a framework to determine whether to outsource manufacturing), when he writes that “many are now rethinking [the latter] doctrine… since integration of R&D/design and production may prove essential to retention of innovation capacity, a fundamental of US economic organization.”

The talk, which was videotaped and posted on Vimeo (the sound is not very good, unfortunately), touched upon “the need to combine new production technologies with new processes and business models” and “the need to examine a series of policy areas, from the increasing integration of production with services, to competitor nation strategies, to workforce and engineering skills, to financing advanced manufacturing, to the innovation organization problem of how to better connect the seams within the innovation and production pipeline.”

Another presentation by Bonvillian that I found particularly insightful is No More Sputnik: The Tripartite Alliance Fifty Years Later (CSIS/ASTRA, October 2007), in which he goes over the history of innovation and R&D funding in the US as well as current trends. A few highlights:

  • Vannevar Bush, who served as the first director of the US Office of Scientific Research and Development during World War II and oversaw significant R&D support for science and technology, laid out the principle of “federal patronage for the advancement of knowledge” in his book Science, the Endless Frontier and also proposed what would become - after much wrangling with President Harry Truman, as Bonvillian explains elsewhere - the National Science Foundation. It is also impossible to talk about US science and technology innovation without mentioning Sputnik, which later “drove a major influx of R&D funding and science talent support.”
  • The innovation model, which used to be national, is now argued by some to be more and more international, but two key innovation drivers still remain overwhelmingly national: “technological and related innovation”, which was identified by Robert Solow, Nobel-Prize-winning Professor of Economics at MIT, and “human capital engaged in research”, which was studied by Paul Romer, Professor of Economics at Stanford at the time of Bonvillian’s writing and now at NYU.
  • The old innovation model of the national tripartite alliance (Department of Defense, industry, academia, all tightly connected) has become outdated because companies are becoming global. Key issues include the fact that basic research can now be done in many places. Also, the US is possibly following a distributed (what Pisano and Shih call modular) manufacturing model.
  • According to the innovation wave theory studied by Carlota Perez of the University of Essex (following up on arguments made by Joseph Schumpeter about the fact that innovation follows wave cycles such as these), we are in the interim stage of a 40/50-year cycle and thus focus less on breakthrough innovations and more on business models, social science integration, and the like.
  • According to Paul Samuelson, another Nobel-Prize-winning economist, who is quoted by Bonvillian, “economic history is replete with stories of capturing comparative advantage.”
  • Finally, Bonvillian touches upon the differences between dispersing research and dispersing innovation and points out the current absence of international funding model. He concludes by providing some additional systemic US advantages.

A paper Bonvillian wrote for a symposium on “21st century innovation systems for Japan and the United States” offers similarly fascinating insights into innovation organization, defined as “the way in which the direct innovation factors of R&D and talent come together, how R&D and talent are joined in an innovation system”. The paper, entitled “The Connected Science Model for Innovation – the DARPA Role”, summarizes many of the same thoughts as the slides above but is far more comprehensive and easier to follow.

In that paper, Bonvillian uses DARPA, “the primary inheritor of the WWII connected science model embodied in Los Alamos and MIT’s Rad Lab,” as “a tool to explore the deep interaction between the US military leadership and technology leadership.” He argues for a third innovation factor beyond the two identified by Solow and Romer, namely, the way that the two factors – R&D and talent – interact with each other. He also points out that “the US, following the war, shifted to a highly-decentralized model… [which] was predominantly a basic-science focused model… and left what later became known as a “valley of death” between research and development stages.” DARPA, however, emerges as one major exception to this model.

Bonvillian studies in some details three groups of innovators:

  1. Thomas Edison’s “Invention Factory” at Menlo Park in New Jersey.
  2. Albert Loomis and MIT’s Rad Lab.
  3. The Transistor Team at Bell Labs.

He concludes that “a common rule set seems to characterize successful innovation at the personal and face-to-face level,” including “ensuring a highly-collaborative team or group of great talent, a non-hierarchical, flat and democratic structure where all can contribute, a cross-disciplinary talent mix, including experimental and theoretical skills sets…, organization around a challenge model, using a connected science model able to move breakthroughs across fundamental, applied, development and prototype stages, cooperative and collaborative leaders able to promote intense, high morale and direct access to top decision makers able to implement the group’s findings.”

The second half of the paper focuses on DARPA as a “unique model combining institutional connectedness and great groups.” For instance, we learn how J.C.R. Licklider “laid the foundations for two of the 20th century’s technology revolutions, personal computing and the internet,” as “an office director at DARPA.” Bonvillian provides a list of DARPA’ s key advantages, from “small and flexible” to “project-based assignments organized around a challenge model”, but also goes over some of the challenges the agency faces in the 21st century, in particular as its role in the war against terror has pushed it away from the hybrid model that had bridged university and industry efforts “through a process that envisioned revolutionary new capabilities, identified barriers to their realization, focused the best minds in the field on new approaches to overcome those barriers and fostered rapid commercialization and DoD adoption.”

The paper, well-researched and well-argumented, is a must-read for anyone who cares about science and technology policy or the changing landscape of innovation.

Accountable Care Organizations

The Affordable Care Act has led to the creation of a new type of entity, the Accountable Care Organization, defined by the Centers for Medicare and Medicaid Services as a “legal entity… composed of certified Medicare providers or suppliers” who “work together to manage and coordinate care for a defined population of Medicare fee-for-service (FFS) beneficiaries.” The main incentive for participation is that “ACOs that meet specified quality performance standards are eligible to receive payments for shared savings if they can reduce spending growth below target amounts.”

In contrast with pure FFS models, which are volume-based payment models, shared-savings programs are supposed to introduce some level of financial accountability. In a recent Health Affairs paper entitled “The Design and Application of Shared Savings Programs: Lessons from Early Adopters” (September 2012), Weissman et al review various shared-savings programs and discuss in depth the case of the Massachusetts Patient-Centered Medical Home Initiative.

The authors note that, in contrast with capitation (a fixed payment per person), shared savings do not necessarily expose providers to downside financial risk, which occurs when the cost to the provider is greater than the payment by the insurer. Therefore, they can be viewed as “stepping stones on the way to more fully accountable care”.

They group the real-life shared-savings arrangements they studied, based on an analysis of twenty-seven organizations, according to three categories:

1. savings calculations including:
  • benchmark or control group selection,
  • definition of gross expenditures considered,
  • definition of outliers excluded from the calculations (expenses above a cut-off point that varies between $50,000 and $500,000 for the organizations considered),
  • use of risk adjustment (to account for the underlying health of participating populations and make sure providers would not be discouraged from serving sicker-than-average patients),
  • choice of patients included in the calculations, usually based on primary care provider.
2. savings distribution (how to split the savings between insurers and providers), with a focus on reducing the risk of overpayment for the insurers:
  • use of a risk threshold (“the savings percentage that providers must achieve to become eligible for savings distributions”, typically 2% to 5%),
  • use of a shared savings starting point (“the point above which savings are shared”), which may or may not be equal to the risk threshold,
  • possible distribution cap or maximum payout,
  • sharing proportion kept by insurers, i.e., payers, typically 20% to 80%,
  • savings allocation between the practices, for instance according to the number of patients in the practice.
3. performance measurement and support to transition into full ACOs: definition of eligibility criteria for practices to receive savings, criteria that serve both as “gate” and as “ladder”:
  • “Practices must pass through the gate to be eligible to share in any net savings”,
  • “A ladder can be used to vary the magnitude of distributions according to the level of savings and the level of quality performance.”

 The authors then describe an initiative in Massachusetts, called the Patient-Centered Medical Home Initiative, which helps understand more concretely the issues involved in making Accountable Care Organizations become a reality. A key element was to develop simulations “to estimate the probability of a practice’s generating savings attributable to random fluctuations in patient costs” and it was subsequently decided to “weight the savings proportion by the probability that it was achieved by actual practice innovations rather than chance.”

For instance, if the shared savings proportion is set to 40% and a 1% shared savings is achieved, which simulations determined had a 73% chance of occurring through real improvements rather than chance, the provider will receive 40x73=29% of the savings.

Weissman et al suggest the following design principles for shared savings arrangements:

  • Pilot programs “should generally give greater weight to ensuring that the incentives result in outcomes that are mutually beneficial instead of focusing on the narrower goal of risk protection for the payer [insurer].”
  • Eligibility criteria – when savings will be shared (“gate”) and to which extent (“ladder”) – help build acceptance for payers and providers and thus should be given particular attention.
  •  “Real” savings, i.e., which savings will be recognized as real, should be agreed upon by payers and providers.
  • Selecting the appropriate organizational units for savings measurement and savings distributions is also particularly important.
  • Risk adjustment in calculation of savings “is needed both to ensure fairness across providers and to minimize incentives to avoid high-risk patients.”
  • Rare, costly events such as transplants, trauma care and cancer care should be excluded.

This article, very clear and thoroughly researched, will appeal to a wide range of researchers interested in healthcare payment models.

Blog Updates

This year will see a few changes at my blog. First, as you may have noticed, I've shortened the name to "Engineered" instead of "Thoughts on..." Second, and more importantly, I've decided to adjust the focus of the blog to match more closely my current research. For the foreseeable future, a new post on healthcare finance will be published every Monday, describing a concept or summarizing a paper that I find particularly interesting. My hope is that the blog will become a resource for operations researchers or members of the general public interested in this area, in addition to a medium for disseminating my own research results throughout the semester and beyond. There will also be a post on science and technology policy, innovation or a related topic (advanced manufacturing, STEM education, etc) many Thursdays.

Decision-making under uncertainty using robust optimization

Decision-making in most applications involves dealing with uncertainty, from random stock returns in finance to random demand in inventory management to random arrival and service times in a service center. This leads to the following two questions:

  1. how should this randomness be modelled?
  2. how should system performance be evaluated, and ultimately optimized? 

When assessing possible answers, it is fair to ask what makes a decision-making methodology more appealing than another. Intuitively, an “ideal” approach exhibits strengths in both dimensions outlined above, in the sense that (1) it allows for a modelling of randomness that does not require many additional assumptions, especially in settings where the validity of such assumptions is hard to check in practice, and (2) it allows for efficient optimization procedures for a wide range of problem instances, which in these days of “big data” most likely includes instances of very large scale. In other words, an “ideal” approach should build directly upon the information at hand – specifically, historical observations of the randomness – without forcing this information into rigid constructs that are difficult to validate even with the benefit of hindsight, and while preserving computational tractability to remain of practical relevance for the decision maker.

Interestingly, the method of choice to model uncertainty has long scored poorly on not just one but both criteria above. Indeed, probability theory assigns likelihoods to events that, when defined as a function of several sources of uncertainty, may occur in many different ways (think for instance of the sum of ten random variables taking a specific value, and all the possible combinations of those ten random variables that achieve the desired value as their sum). To assign a probability to this complex event, it is then necessary to resort to an advanced mathematical technique called convolution, made even more complex when random variables are continuous – and this is just to compute a probability, not optimize it. Additional hurdles include the fact that the manager observes realizations of the random variables but never the probability distributions themselves, and that such distributions are difficult to estimate accurately in many practical applications.

So why have probabilities become the standard tool to model uncertainty? A possible answer is that they were developed to describe random events in the realm of pure mathematics and have since been used beyond their intended purpose to guide business managers in their optimization attempts. This might have happened because academicians and practitioners alike lacked other tools to describe uncertainty in a satisfactory manner, or because the systems under early consideration remained simple enough to fit easily within the probabilistic framework. But as the size of the problems considered by today’s managers increases and today’s business environments are defined more and more by fast-changing market conditions, it seems particular urgent to bring forward a viable alternative to probabilistic decision-making that exhibits the attributes outlined above. This is precisely what researchers in a field called “robust optimization” have done. (Disclosure: robust optimization is also one of my key research areas.)

By now robust optimization has been a hot topic in operations research for almost two decades, and I won’t go over its distinguished contributions to fields as varied as portfolio management, logistics and shortest path problems, among many others, in this post – for that purpose, the interested reader is referred to one of the several overview papers available on the Internet.

For today’s post I will focus on a specific aspect of robust optimization that is currently gaining much traction, specifically, the use of robust optimization as a “one-stop” decision-making methodology that (a) builds directly upon the data at hand, (b) remains tractable, intuitive and insightful in a wide range of settings where probabilistic models become intractable and (c) leads to the axioms of probability theory as a consequence of the modelling framework.

While (a) and (b) have been standard arguments in favor of robust optimization for at least a decade, the connection with the world of probabilities provided in (c) is a novel development that should appeal to many decision makers. The resulting framework is at the core of a paper by Chaithanya Bandi and Dimitris Bertsimas – my former dissertation adviser – published last summer in Mathematical Programming.

They suggest modelling sources of uncertainty not as random variables with probability distributions but as parameters belonging to uncertainty sets that are consistent with the limit laws of probabilities (which bound the deviation of a sum of independently and identically distributed random variables from their mean, using their standard deviation and the square root of the number of random variables). Other asymptotic laws can be incorporated to the uncertainty set in specific circumstances.

The objective of the problem then becomes to optimize the worst-case value of some criterion, instead of its expected value as would have been the case in stochastic optimization.

Bandi and Bertsimas discuss:

  • Using historical data and the central limit theorem,
  • Modeling correlation information,
  • How to use stable laws to construct uncertainty sets for heavy-tailed distributions that have infinite variance,
  • Incorporating distributional information using “typical sets”, first introduced in the context of information theory, which exhibit the properties that (i) the probability of the typical set is nearly 1 and (ii) all elements of the typical set are nearly equiprobable, and are defined in the Bandi and Bertsimas paper in terms of uncertainty sets with examples drawn from common distributions.

The rest of the paper shows the results obtained by applying this framework to three applications.

Performance analysis of queueing networks. (Some of the work described in this section is joint work between Bandi, Bertsimas and Nataly Youssef - download their paper here.) The authors introduce the concept of a robust queue where arrival and service processes are modelled by uncertainty sets instead of assigning probability distributions and derive formulas on the worst-case waiting time for the n-th customer in the queue. They connect their results with those obtained in traditional queueing theory and derive results for systems with heavy-tailed behavior such as Internet traffic that are not believed to have been previously available.

Further, they analyze single-class queueing networks using their framework and present computational results where they compare their approach, dubbed the Robust Queueing Network Analyzer (RQNA), with results obtained using simulation and the Queueing Network Analyzer (QNA) developed in traditional queueing theory. Their observation is that RQNA’s results are often significantly closer to simulated values than QNA’s.

Optimal mechanism design for multi-item auctions. (Full paper.) In this problem, an auctioneer is interested in selling multiple items to multiple buyers with private valuations for the items. The auctioneer’s goal in the robust optimization approach is to maximize the worst-case revenue, with his beliefs on buyers’ valuations modelled through uncertainty sets. His decisions are the allocation of items and the payment rules, which should satisfy the following properties:

  • Individual Rationality: Bidders do not derive negative utility by participating in the auction, assuming truthful bidding, i.e., bidding their true valuation of an item.
  • Budget Feasibility: Each buyer is charged within his budget constraints.
  • Incentive Compatibility: the total utility of the i-th buyer under truthful bidding is greater than the total utility that Buyer i derives by bidding any other bid vector.

Bandi and Bertsimas provide a robust optimization mechanism (ROM) that solves the problem and consists of (1) an algorithm to compute the worst-case revenue before a bid vector is realized, and (2) an algorithm to compute allocations and payments afterward, which also uses the worst-case revenue as input.

In addition, they investigate the special case where the buyers do not have any budget constraints, for which they compare the resulting algorithms in their model with a mechanism called Myerson auction. They argue that their method in that setting exhibits stronger robustness properties when the distribution or the standard deviation of the valuations is misspecified.

Pricing multi-dimensional options. In this finance problem, Bandi and Bertsimas (along with their collaborator Si Chen) propose to model the underlying price dynamics with uncertainty sets and then apply robust optimization rather than dynamic programming to solve the pricing problem. They illustrate their approach in the context of European call options and reformulate the problem as a linear problem.

The approach has the flexibility to incorporate transaction costs and liquidity effects. It also captures a phenomenon known in finance as the implied volatility smile, which can be explained in the context of robust optimization using risk aversion arguments. Bandi and Bertsimas further give examples of the accuracy of their method relative to observed market prices. You can download that paper here.

The authors’ central argument is that “modelling stochastic phenomena with probability theory is a choice” and that “given the computational difficulties in high dimensions, we feel we should consider alternative, computationally tractable approaches in high dimensions.” They provide compelling evidence that robust optimization is well-suited for that purpose.