Has anyone published data on the effectiveness of Bayesian prediction models as an educational intervention? It seems like that would be very helpful in terms of being able to convince school districts to give them a shot.
Most of the discussions I’ve happened to see focus on http://en.wikipedia.org/wiki/Value-added_modeling , because being able to judge teachers is directly useful to school districts and lots of outsiders are interested in the topic. Since the usual approach uses multilevel model (you need to adjust for school-level effects, district-level effects, etc before you can extract a usable teacher-level effect), it’s almost Bayesian by default, and if you google ‘bayesian value-added modeling’ you’ll find a ton of material.
My experience is that school districts have a strong not-invented-here bias. For example, special education laws require research based interventions, a requirement that is generally ignored.
Maybe slightly vary the parameters to make the model “new”? Like, fit it to data from that district, and it will probably be slightly different from “other” models.
I think the hard part of refitting the model would probably just be getting access to the data—beyond that it seems like a statistician or programmer would be able to just tell a computer how to minimize some appropriate cost function.
Something like most of the marginal effort is devoted to gathering the data, which presumably doesn’t require that much expertise relative to understanding the model in the first place.
In practice, there are substantial privacy law issues, although those can be gotten around if the district is clever.
More importantly, collecting, collating, and ensuring coder reliability is expensive. What you called “marginal effort” is quite difficult for just about any large bureaucracy.
Quite a few things there. SAS’s EVAAS is generally considered the gold standard of bayesian prediction models as educational interventions; unfortunately as SAS is based in North Carolina it has yet to spread outside that particular state. Some states have similar systems being produced by similar companies.
Has anyone published data on the effectiveness of Bayesian prediction models as an educational intervention? It seems like that would be very helpful in terms of being able to convince school districts to give them a shot.
Most of the discussions I’ve happened to see focus on http://en.wikipedia.org/wiki/Value-added_modeling , because being able to judge teachers is directly useful to school districts and lots of outsiders are interested in the topic. Since the usual approach uses multilevel model (you need to adjust for school-level effects, district-level effects, etc before you can extract a usable teacher-level effect), it’s almost Bayesian by default, and if you google ‘bayesian value-added modeling’ you’ll find a ton of material.
My experience is that school districts have a strong not-invented-here bias. For example, special education laws require research based interventions, a requirement that is generally ignored.
Maybe slightly vary the parameters to make the model “new”? Like, fit it to data from that district, and it will probably be slightly different from “other” models.
But that requires effort, and school districts don’t generally want to put in effort to do things differently.
I think the hard part of refitting the model would probably just be getting access to the data—beyond that it seems like a statistician or programmer would be able to just tell a computer how to minimize some appropriate cost function.
Something like most of the marginal effort is devoted to gathering the data, which presumably doesn’t require that much expertise relative to understanding the model in the first place.
In practice, there are substantial privacy law issues, although those can be gotten around if the district is clever.
More importantly, collecting, collating, and ensuring coder reliability is expensive. What you called “marginal effort” is quite difficult for just about any large bureaucracy.
Marginal effort within the bounds of a consulting agency offering a service “tailored” to each school district.
http://www.sas.com/govedu/edu/k12/evaas/papers.html
Quite a few things there. SAS’s EVAAS is generally considered the gold standard of bayesian prediction models as educational interventions; unfortunately as SAS is based in North Carolina it has yet to spread outside that particular state. Some states have similar systems being produced by similar companies.
Particularly, if I were you I would read: http://www.sas.com/success/wf_rolesville.html