Predicting Academic Difficulty in Veterinary Medicine: A Case-Control Study

Abstract

A veterinary education is costly and time-consuming, so it is in the best interests of applicants and colleges alike that admissions processes be fair and select applicants who are likely to succeed. We employed a case-control study to explore whether any of 28 admissions variables used by a veterinary college located in the Midwest region of the United States predicted which students would encounter academic difficulty in the veterinary curriculum. Participants were selected from the veterinary classes admitted between 2008 and 2017. We defined academic difficulty cases ( n = 55) as any students dismissed from the program or placed on academic probation. Controls ( n = 220) were selected at random from the same classes, excluding the academic difficulty cases, students with honor code violations, or any who had exited the program early for any reason other than participation in a concurrent program. Admissions variables included: gender, citizenship, underrepresented status, state of residence, age, interview scores, GPA (science), GRE scores, undergraduate credits, participation in honors courses, community college credits, repeats/withdrawals of required undergraduate courses, course load, and admissions committee review criteria including work experience, animal/vet experience, references, essays, leadership, personal development, special circumstances, and overall committee score. Zero-order correlations for academic difficulty were significant for underrepresented status, age, GPA (science), verbal and quantitative GRE scores, repeats/withdrawals, and references. When combined in logistic regression, only science GPA, verbal GRE, and references significantly and independently predicted struggler status.

Using Machine Learning in Veterinary Medical Education: An Introduction for Veterinary Medicine Educators

“…During model specification, the authors calculated zero-order correlations [13], which means that the authors pre-determined potential correlations between two independent variables and only selected variables with zero-order correlations to include when building a single logistic regression model. The data included in the model adhered to specific assumptions including (1) the relationship between the logit (also known as log-odds) of the outcome and each of the included continuous independent variables were all linear, (2) there were no highly influential outlier data points, (3) there were no highly correlated independent variables (i.e., absence of multicollinearity), (4) the observations were independent of each other, and (5) the sample size was sufficient which is typically considered at least 10 observations of the least frequent outcome for each independent variable [14].…”