March 2006
Sun Mon Tue Wed Thu Fri Sat
      1 2 3 4
5 6 7 8 9 10 11
12 13 14 15 16 17 18
19 20 21 22 23 24 25
26 27 28 29 30 31  

Authors' Committee


Matt Blackwell (Gov)


Martin Andersen (HealthPol)
Kevin Bartz (Stats)
Deirdre Bloome (Social Policy)
John Graves (HealthPol)
Rich Nielsen (Gov)
Maya Sen (Gov)
Gary King (Gov)

Weekly Research Workshop Sponsors

Alberto Abadie, Lee Fleming, Adam Glynn, Guido Imbens, Gary King, Arthur Spirling, Jamie Robins, Don Rubin, Chris Winship

Weekly Workshop Schedule

Recent Comments

Recent Entries



SMR Blog
Brad DeLong
Cognitive Daily
Complexity & Social Networks
Developing Intelligence
The Education Wonks
Empirical Legal Studies
Free Exchange
Health Care Economist
Junk Charts
Language Log
Law & Econ Prof Blog
Machine Learning (Theory)
Marginal Revolution
Mixing Memory
Mystery Pollster
New Economist
Political Arithmetik
Political Science Methods
Pure Pedantry
Science & Law Blog
Simon Jackman
Social Science++
Statistical modeling, causal inference, and social science



Powered by
Movable Type 4.24-en

« March 16, 2006 | Main | March 21, 2006 »

20 March 2006

Making Diagnostics Mandatory

Jim Greiner

Teaching a class (see here) on the interaction between lawyers, most of whom lack quantitative training, and quantitative analysts has me thinking about the danger statistical techniques pose. As is true of those who study any branch of specialized knowledge, statisticians can abuse the trust decision makers (judges, government officials, members of the public) put in us all too easily, and often with impunity. (Of course, “we? all know that “we? would never do any such thing, even though “we? know that “everyone else? does it all the time. Gee.)

If it’s of interest (or perhaps more accurately, unless a barrage of comments tells me I’m being boring), I’ll be blogging about ways “everyone else? abuses trust, and ways “we? can try to stop it. Here’s my first suggestion: make diagnostics mandatory.

Here’s what I mean. I’ve previously blogged (see here) on the double-edged sword posed by the recent trend towards academics’ writing free software to fit models they’ve developed. One way for software-writers to lessen the danger that their models will be abused is to write diagnostics into their programming . . . and make those diagnostics hard to turn off. Suppose, for example, that some analysts are writing code to implement a new model, and the fitting process requires fancy MCMC techniques. These analysts should write MCMC convergence diagnostics into the software, and should set their defaults so that the fitting process produces these diagnostics unless it’s told not to. Perhaps, the analysts should even make it a little tough to turn off the diagnostics. That way, even if the user doesn’t look at the diagnostics, someone else (perhaps an opposing expert in a court case?) might have easier access to them.

The worry, of course, is that the output from all new software will end up looking like it came out of SAS (a package I wouldn’t wish on my worst enemy). Still, as our cognitive psychologist could probably tell us, people are incredibly lazy. Even if a user of software just has to go to a drop-down menu to look at a diagnostic, chances are he/she won’t bother.

Posted by James Greiner at 6:00 AM