Using neural networks to predict MBA student success

College Student Journal, March, 2004 by Bijayananda Naik, Srinivasan Ragothaman

Predicting MBA student performance for admission decisions is crucial for educational institutions. This paper evaluates the ability of three different models--neural networks, logit, and probit to predict MBA student performance in graduate programs. The neural network technique was used to classify applicants into successful and marginal student pools based on undergraduate GPA, GMAT scores, undergraduate major, age and other relevant data. The results of this study show that the neural network model performs as well as the statistical models and is a useful tool in predicting MBA student performance. Several limitations of this study are discussed.

Introduction

Master of Business Administration (MBA) program admission directors are carefully reexamining the admission criteria they use to admit student into their graduate programs. This is a response to the bleak enrollment prospects facing business schools across the country. Educational institutions consider a variety of factors when making admission decisions. Some of the evaluation criteria normally used include the following: overall undergraduate grade point average (GPA), junior/senior GPA, undergraduate major and institution, Graduate Management Admissions Test (GMAT) score, references, goals statement, a personal interview and others.

Traditionally, academic researchers have developed several statistical models (e.g., discriminant analysis, multiple regression, stepwise regression) to predict an applicant's success in the MBA program (Pharr et. al, 1993; Wright and Palmer, 1997). However these statistical models usually assume multivariate normality and homoscedastic variances. When these assumptions are violated in real world data structures, the predictive ability of regression models is diminished (Hardgrave et. al., 1994). Another problem with the usage of regression models found in earlier research is a rather skewed distribution of graduate GPA (Abedi, 1991). Moreover, since statistical models use only objective information, potentially relevant subjective information is often disregarded. Another problem with multiple regression and stepwise regression studies is the low value of R2. For example, Pharr et al. (1993) reported an R2 of only 0.18 percent.

Because of the aforementioned limitations of standard statistical models, neural network and rule induction techniques have become increasingly popular for decision-making (Ragothaman and Naik, 1994; Weiss and Kulikowski, 1991). The objective of this paper is to evaluate the ability of three different models, namely, logistic regression, probit analysis and neural networks, to predict MBA student performance.

Prior Research

Many of the prior studies involving graduate student performance have used linear regression models to estimate student academic performance based on their preadmission record of achievement. Wright and Palmer (1994) have identified a few potential deficiencies in studies that have relied on linear regression. Gayle and Jones (1973), and Baird (1975) found a significant positive relationship between Graduate Records Examination (GRE) scores and graduate grade point average (GGPA) in graduate students. Deckro and Woundenberg (1977) studied nine variables as possible predictors of academic success among Kent State MBA students. Ahmadi et. al. (1997) used bivariate regression models to examine relationships between graduate GPA and a number of factors and found that undergraduate GPA and GMAT were significant variables in predicting academic success.

Paolillo (1982) employed step-wise regression in his study and found that the applicant's junior and senior undergraduate grade point average was the first variable to enter into the equation. Schwan (1988) found GGPA to be significantly correlated with GMAT score, undergraduate grade point average, and junior/senior grade point average among Murray State University MBA students. Wright and Palmer (1997) used analysis of variance to examine GMAT scores among groups of graduate students and concluded that the use of total GMAT scores in the admission process may be misleading and more attention may be paid to specific verbal and quantitative components.

Adams and Hancock (2000) used analysis of variance and correlation analysis to show that amount of work experience prior to enrollment is a predictor of success and showed better association with success than GMAT and undergraduate GPA. Braunstein (2002) used correlation and regression techniques to determine the variables most closely related to academic success of recent graduates of an MBA program. They found undergraduate GPA and GMAT score to have the strongest positive correlation with the graduate GPA. Other statistically significant variables included gender, type of undergraduate degree, and years of work experience.

Wilson and Hardgrave (1995) compared the use of regression techniques with classification techniques to predict graduate student success in an MBA program. They concluded that nonparametric procedures, such as neural networks, performs at least as well as statistical methods and are worthy of further investigation. Arnold and Chakravarty (1996) used linear discriminant analysis and neural networks to classify applicants into the executive MBA program at a private university. They reported that the neural network model appeared to perform better in capturing the complex interactions among the predictor variables in their research.