Tuesday, September 21, 2010
The September 2010 issue of Harvard Business Review has an insightful piece, “The Judgment Deficit,” by Amar Bhide. It explores the role of statistical models in the centralization of decision making and how this played a role in the recent downfall of th

Professor Bhide notes that managers recognize the need for balance between centralized command-and-control and individual initiative and judgment.  But he observes that in recent times “a new form of centralized control has taken root – one that is the work not of old-fashioned autocrats, committees, or rule books but of statistical models and algorithms.” 

These mechanistic technologies “have value under certain circumstances,” he observes, but can be misused or overused. Bhide cites Friedrich Hayek’s 1945 article: “The Use of Knowledge in Society,” who says “It is a dispute as to whether planning is to be done centrally. . . or is to be divided among many individuals . . . .Which of these systems is likely to be more efficient depends mainly on the question under which of theme ew can expec that fuller use will be made of the existing knowledge.”  Well, with data analytics and statistics, that now becomes more possible.  But given the experience of the financial systems meltdown, does it really work?

Centralization?  Centralization in decision making is needed in some cases.   But “how to centralize – whether through case-by-case judgment, a rule book, or a computer model – is as difficult a question as how much [to centralize].”

“Technologically advanced societies couldn’t function without some centralized control.” For example, government regulation is needed in finance, food, and worker safety arenas. And, economies of scale in areas such as rail transport have been key to development.  In addition, specialization and coordination of labor have moved from the production line to the development and delivery of services.

But, “The information technology revolution has shifted the balance between judgment and rules, giving a strong economic and psychological boost to judgment-free decision making.” . . However, “. . .the inability to cope with context-specific information makes centralized organizations inflexible.”

Decentralization?  Bhide expands on when decentralization and individual judgment are appropriate.  Information in emergencies, like Hurricane Katrina, cannot be quickly and accurately conveyed to a central manager.  On-the-ground managers need the discretion and flexibility to figure out a solution in the moment, given the specific context of the situation.  For example, is the priority drinking water, shelter, or security?

But even beyond emergency situations, centralized control tends to not work because it cannot adapt to the needs of innovation.  Innovations tend to be one-of-a-kind developments, notes Bhide, and centralized systems tend to force uniformity, not diversity.

Creating a dynamic, yet cohesive, culture in large organizations is hard.  This is the challenge of major corporations such as IBM or GM – how do you create an ecosystem of companies, manufacturers, and software developers to innovate and produce results?  When does centralization for economies of scale and common standards help vs. hurt?  The same challenge faces government.

Bhide observes that effective decentralization has pre-conditions.  It demands mechanisms to coordinate independent initiatives.  The key conditions are the use of dialog and relationships:  “. . . dynamic societies and organizations rely on dialogue and relationships to a greater degree than do top-down systems, in which a few tell the many what to do.”

“Established relationships complement dialogue in sharing information and facilitating coordination.  Doing business repeatedly with the same parties reduces ambiguities and misunderstandings. . . “  [an aside:  in the government world, this is interpreted as “no bid” contracts and cozy relationships].

“Relationships can also smooth the way for making adjustments when things go wrong. . . . Without a relationship in place, a buyer is less likely to take the circumstance of the situation into account and more likely to default to a prescribed response.”

These elements of decentralization, however, are typically not amenable to statistical computer models.

A balancing act?  What is needed is a balancing act in the use of centralized controls (especially computer models), says Bhide:  “The challenge is to keep control by human authority – or computer models -- within judicious limits. . . .

[The use of] Computerized controls work best with inanimate products or processes that can be physically shielded so that variations in conditions (such as the temperature or humidity inside a plant or product casing) can be minimized and when feedback from measured outcomes can be continuously used to adjust or improve the decision-making algorithms.

 . .  Computers also shine when . . .the number of possible outcomes is vast. . . but they conform to well-specified rules [e.g., chess or air traffic routing]. . . Conversely, human judgment is favored when shielding is difficult, outcomes are ambiguous, and the possibilities are open-ended.”