The Institute for Social Research and the James Morgan Fund for New Directions in the Analysis of Complex Interactions will provide up to two $15,000 awards to stimulate new and innovative uses of SEARCH Program senior graduate students and post-doctoral scholars at any accredited university in the U.S. to explore and expand our understanding of complex interactions among explanatory variables in describing behavior.
We know there are complex combinations operating and the availability of large, rich databases provides an opportunity to use a different approach. The ISR program SEARCH was created for analysis of survey data with samples of 1,000 or more, and a clear dependent variable—a behavior or situation to be explained or predicted. And it was designed to look for non-linearities and non-additives (interaction effects) assumed away by multiple regression. When the dependent “variable” is a dichotomy, or a set of categories, one can call the result a decision tree, and change a measure of improvement to a chi square.
SEARCH splits the data sequentially on the best binary split of the best variable for reducing the unexplained variance, or in the case of a categorical dependent variable, for increasing the chi square. With ranks, it tries the first against the rest, then the first two, etc. With non-ranked categories like race or region, it tries each group against all the rest, keeping the best of the best. The reason this works is that a very few splits on any predictor will exhaust its explanatory power. (Kalton, 1967, in JDS article) There are stopping rules to avoid overdoing it.
The idea was first promoted by an article in the Journal of the American Statistical Association (June, 1963) and used extensively in a book Productive Americans, 1966. There are already many published articles using multiple regression, and explorations whether SEARCH would produce useful new interpretations is the challenge.
SEARCH has been widely used in marketing, but is not the same as cluster analysis, data mining, or neural networks, which often do not specify what is to be explained or predicted, and some tend to start with a complete multilevel details and make risky decisions on how to combine. The ISR SEARCH program’s first splits are based on substantial frequencies and hence are safe. As a guide for decision-making or diagnosis, SEARCH can help decide how far down the branching diagram to continue. (For a more extensive overview, see this PDF document.)
The most revealing analysis would take data already analyzed with standard linear additive models and see whether SEARCH provided new and useful findings.
A fuller history and explanation can be found in Morgan, J. N. (2005). History and Potential of Binary Segmentation for Exploratory Data Analysis (PDF). Journal of Data Science, 3, 123-136.
For an explanation of the underlying algebra, see this PDF document.
For a more detailed discussion about SEARCH, and a step-by-step description of how to access to the program, go to http://bit.ly/SEARCH-ISR.
For instructions on how to run SEARCH in Stata, go to http://bit.ly/SEARCH-Stata.
Form and Scope of the Award:
The Morgan Fund will make up two awards of $15,000 each for stipends that may be used to cover salary and/or other expenses that can be suggested by the applicant and funded at the discretion of the Awards Committee. Funds can also be used to purchase of consulting time with relevant staff at the University of Michigan Institute for Social Research.
$10,000 will be award at the time of selection and $5,000 will be awarded when the project is completed and deliverables are submitted.
Applicants are encouraged to use datasets developed at the Institute for Social Research (e.g. American National Election Study, Panel Study on Income Dynamics, Health and Retirement Study, Surveys of Consumer Attitudes, Monitoring the Future, etc.) or data archived at the Inter-university Consortium for Political and Social Research, but are not limited to these.
These are the criteria for eligibility:
- Applicants must be admitted to PhD candidacy in a graduate program or post-doctoral scholars at an accredited university in the U.S.
- Special preference will be given to students using the award to advance their dissertation research
Applications should be no more than three pages with description of the proposed activity addressing the following:
- A description of the analysis and a rationale
- Statement about the motivation and value of the proposed project
- Examples of relevant literature that would be improved by the analysis
- Sufficient detail about the relevant hypotheses and design
- Description of the kinds of data that will be used
- Statement of deliverables (e.g. dissertation chapter, conference paper, etc.)
- A recommendation letter from your advisor
- An up-to-date transcript
- An up-to-date Curriculum Vitae
A progress report will be required on August 1, five months after the award is made.
December 2 → Announcement
January 15 → Applications are due
February 28 → Decisions are announced
March 14 → Initial payments made
August 1 → Report and deliverables submitted by recipients
September 15 → Second payment made
Applications are accepted in both electronic and hardcopy format. Send electronic applications to Jennifer Puckett at firstname.lastname@example.org. Hardcopy applications should be sent to Jennifer Puckett, PO Box 1248 Ann Arbor, MI 48106-1248 (postmarked by January 15, 2014). Note, letters of recommendation may be sent directly by the recommender in either electronic or hard copy.
For questions about the application and application process, please contact Patrick Shields at email@example.com or 734.764.8369.
For help in correcting set-ups or advice on combining a data management program with SEARCH are available from Peter Solenberger, who can be reached at firstname.lastname@example.org.
SEARCH developer Jim Morgan is available to comment on proposals. He can be reached at email@example.com.