Regression Models in the Social Sciences

Regression Models in the Social Sciences

Chris Snijders

Utrecht University

Department of Sociology

Heidelberglaan 1

3584 CS Utrecht

The Netherlands

tel (31) 30 253 4140

fax (31) 30 253 4405

email C.Snijders@fsw.ruu.nl

Problem Description:

Social scientists often test their theoretical models on the basis of survey-data. Most of the time, some dependent variable is being approximated with a set of independent variables using regression analysis. To say something about the validity of the theoretical model, it makes sense to compare its results to the results from a neural network on the same data. As a result, contrary to most applications of neural networks, social scientists are happy if neural networks do not outperform their theoretical models: this simply implies the theoretical model does a relatively good job. In addition, the neural networks can be used to "scan" the often quite messy survey data for unexpected new relations.

Position of user

Currently, I am a postdoc at Dept. of Sociology, Utrecht University and Interuniversity Center for Social Science Theory and Methodology (ICS), Utrecht, The Netherlands.

How NNs were applied

As a user of the Educator version, I mostly use the standard MLPs. The data I use are gathered through surveys (say, about 1500 respondents). On the average, I try to approximate one or two dependent variables with about fifteen independent variables. Usually, this is a matter of minutes. A larger part of the time is devoted to finding the "right" architecture for the MLPs. I hope and think Neurosolutions for Excel™ can take over some of this burden.

Results that were obtained

Regression analysis in social sciences typically reach an R2 of about 0.3. Neural networks perform a little (R2=0.32) to a lot (R2=0.5) better. When neural networks perform a lot better than standard regression analysis, this is an indication that the theoretical model is a poor approximator of reality and needs some additional work. "Solving" the problem as in character recognition is, of course, out of the question.

How were the results applied

Using neural networks has lead to two different results. First, it provides a benchmark against which to compare the results from standard regression. If the neural network cannot (or hardly) outperform my theoretical model, I am convinced that the model is good. Usually, I keep such results with neural networks to myself. Social scientists who appreciate the power of neural networks are scarce. In addition, I make use of neural networks to explore the data and thereby as a source of inspiration for new theoretical models. This implies that most of the direct benefits of neural networks are not visible to most of the academic world. It will take some time to let the social sciences appreciate neural computation.

Download this Application Summary in Word97 format.
If you don't have Word97....Download the Word Viewer.