Wednesday, August 8, 2012
Present: Abby Crocker, Kairn Kelley (by phone), Amanda Kennedy, Rodger Kessler, Ben Littenberg, Connie van Eeghen
1. Theme: Ben: prediction is not necessarily about regression. There are many alternatives, each with strengths and weaknesses. Good to have many tools in the toolkit to choose from.
2. Presentation: Abby – “How we picked the predictive model for the NAS article.” NOTE: this is an excerpt; complete notes can be obtained from the presenter.
a. Alternative methods of predictive modeling:
i. Logistic: widely used, doesn’t have to be explained, but its limitations are not well understood, i.e. predictive power
ii. Clinical Heuristic
iii. Neural Net
iv. Bayesian Net
v. Recursive Partitioning – not well known but easy to explain
b. Board example of recursive partitioning:
i. 796 babies, of whom 253 with NAS and 531 without, the remainder missing data; total is 784 babies
ii. No babies are “grey” – they either have been treated for NAS or not
iii. Each predictor will be applied to the 784 population, e.g. sex (f/m), maternal smoking (y/n), … and evaluate the strength of each predictor.
iv. The strongest predictor (for example, maternal substance abuse) is used to determine new probabilities
1. The entire group of 784 has a 33% chance of NAS
a. 700 are positive for SA and 71% are NAS
i. The next strongest predictor (for example gestational age)
1. Older babies have a 65% of NAS
a. Smokers have a 68% of NAS
b. Non-smokers have a 15% of NAS
2. Younger babies have an 89% of NAS
b. 84 are negative for SA and 14% are NAS
i. The next strongest predictor (for example maternal age)
1. Younger moms have a 16% of NAS
2. Older moms have a 2% (no further analysis needed down this branch)
v. This model looks for the best fit and is highly dependent on how the variables are categorized
1. Cannot go below some minimum number of subjects in a branch
2. Can be forced into meaningful clinical variables, and tested that way (not by ranking the variables)
c. Summary: there are three reasons to build a model.
i. Describe the world as it is, and reduce it to something that might be important. Some of these are called exploratory studies. No single model preferred.
ii. Predict a true relationship, even if it does not describe what is true in the world completely. Try recursive partitioning.
iii. Hypothesize the relationship between a predictor and a variable, using a t test of Chi-square (if there are only 2 variables). Logistic modeling might be best.
3. Workshop Goals for 2012:
a. Journal club: identify UVM guests and articles; invite to CROW ahead of time
b. Research updates: share work-in-process
a. Aug 9, 16, 23, 30: remaining Thursdays in August will be cancelled. Everyone is encourage to attend Book Club on all August Fridays at noon. Contact Connie if you need the source document we are reviewing jointly. New schedule for fall to be set up by Doodle poll later in August.
b. Future agenda to consider:
i. Kairn – review of draft article on IRR (no Abby)
ii. Ben: budgeting exercise for grant applications
iii. Journal Club: “Methods and metrics challenges of delivery-system research,” Alexander and Hearld, March 2012 (for later in the year?). UVM authors who have published interesting design articles (Kim, Osler)
iv. Rodger: Mixed methods article; article on Behavior’s Influence on Medical Conditions (unpublished); drug company funding. Also: discuss design for PCBH clinical and cost research.
v. Amanda: presentation and interpretation of data in articles
vi. Sharon Henry: article by Cleland, Thoracic Spine Manipulation, Physical Therapy 2007
Recorder: Connie van Eeghen
Posted by Connie at 8/08/2012 09:13:00 PM