Sunday, October 07, 2012

Baumol's cost disease: medicine, education and post-AI disruption

William Baumol was born in 1922. In 2012, 90 years later, he's listed as first author on a new bookThe Cost Disease: Why Computers Get Cheaper and Health Care Doesn't.

Damn. It's one thing to win the brain lottery, but winning the longevity lottery is really piling on. Even if all he did is read the page drafts he's doing pretty well.

That's not the most irritating thing about Baumol though. The most irritating thing is that I keep forgetting about his fundamental insight, one that I first blogged about 8 years ago...

... The disparity between rapid productivity growth in mechanized sectors and slow productivity growth in human-service jobs produces Baumol's disease—named after the economist William J. Baumol. According to Baumol, in a technological economy falling prices for manufactured goods and automated services eventually increase the relative cost of labor-intensive services such as nursing and teaching. Baumol has predicted that the share of gross domestic product spent on health care will rise from 11.6 percent in 1990 to 35 percent in 2040, while the share spent on education will rise from 6.7 percent to 29 percent.

The shifting of relative costs need not in itself be a problem. If Americans in 2050 or 2100 pay far more (as a percentage of their spending) for health care and education than they did in 1900, they may still be better off—if they pay correspondingly less for other goods and services. The problem is that as the relative cost of services like education and health care rises, more and more Americans will find themselves in service-sector jobs that, unlike the professions, have historically been low-wage...

Today Education and Health Care are famously afflicted by Baumol's disease. Law used to be, but then full-text search decimated legal employment (and yet, legal costs have not fallen ....).

Baumol argues that even if these professions remain labor intensive, and even if health care comes therefore to claim 50% of our GDP, that we'll be able to afford it nonetheless.

His argument is persuasive, but is that likely to happen? College education today is experiencing widespread disruption including iTunes Ucoursera (Caltech, University of Toronto and many more), edX (MIT, Harvard, Berkeley), California open-source eTexts, Stanford Online, Khan Academy and numerous for-profit ventures. Education is deep in whitewater times.

Health care, particularly medical care, isn't changing as quickly. The fundamental tasks of sorting out what's going on with a particular patient, and how best to manage that problem in their personal context, and then how to manage the patient's psyche and health -- those haven't changed much [1] over the past century. 

We're accumulating more health care data though -- for better and for worse [3]. "Analytics" is the "hot" area in health care IT now, including running Google/Facebook style algorithms against large clinical and financial data sets [2].

That doesn't necessarily sound disruptive, unless you know that the techniques used in extracting meaning from large data sets are the same technologies that power our post-AI world. (Yeah, I used the forbidden acronym.) If you know that, then you know "Analytics" can be thought of as the current pseudonym for "Medical AI". Whether it's disruptive or not remains to be seen, but I suspect that we'll get to health care cost disruption well before health care hits 50% of a much larger future GDP.

 [1] It's interesting to read articles written in the 1970s during the early days of diagnostic lab testing. They imagined patients walking into a series of lab test queues staffed with low wage workers, then emerging with a set of diagnoses and plans. Similar plans arose during the last period of genomic enthusiasm. They will come again ... 
[2] The base stats is generally pretty simple stuff, if only because more complex algorithms don't scale well to terabyte data sets. The trick is that simple stats on large data sets enabled by cheap computation can produce surprisingly useful answers. This is best described in the terrific Halevy, Norvig and Pereira paper: The Unreasonable Effectiveness of Data.
[3] In 1996 I was part of a theater-style presentation called "Dark Visions: 1996-2010" that included a fanciful and intentionally dramatic timeline of dystopic data sharing. By 2005 India was the world center of clinical AI, and by 2006 elite health care providers had moved to more private paper records. Maybe we were a bit hasty :-).

See also:

No comments: