Making Diagnostics Mandatory 

 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.