This chapter briefly describes the development of the Career Interest Profiles, an application of Holland's (1997) RIASEC work environments, for the 1,122 careers in the database. Each career profile consists of six numerical scores in invariant order (R-I-A-S-E-C) indicating how descriptive and characteristic the occupation is for each of the work environments. These profiles were instrumental in the development of the CareerMatch™ system, allowing users to compare their own Interest Inventory scores to that of each career in the database and to determine which careers best “fit” their interest profile.

Two time-honored occupational classification methods – empirical (discriminant analysis) and judgment (expert ratings) – were used in the development of the Career Interest Profiles. The reliability, validity, and economy of the methods were compared. On all criteria examined, the judgment method did at least as well as or better than the empirical method. Discriminant analysis was very efficient in classifying an occupation into one RIASEC category (assigning a primary code), but provided little or no information for the other five categories. In contrast, the judgment method generated a reasonable distribution of values across the work environments.

With the empirical approach, numerical RIASEC profiles were developed for all the careers in the database using discriminant functions – a type of classification analysis used to produce probabilities of group membership. In this case, the discriminant analysis would yield probabilities of group membership for each of the RIASEC categories.

A stepwise discriminant variable selection procedure was carried out on 67 predictor variables to identify a smaller predictor set. (Predictor variables included data on worker functions, general educational development, aptitudes, temperaments, GOE codes, physical demands, and environmental conditions.) The analysis was carried out using two independent development samples – one from a study conducted by Holland in 1982 and the other from a present-day study of 150 careers. The stepwise selection procedure considered the relative contribution of each of the 67 variables with respect to overall discriminatory power of the model (as measured by Wilks’ Lambda). After analysis, seven predictor variables were chosen for the final model, based on their discriminatory power as well as performance of overall error rates and average squared canonical correlations. The final discriminant model was then run on all the careers in the database using the 7 identified predictor variables.

In general, this approach was very successful in identifying primary group membership, but there were concerns about the reliability of non-primary group probabilities. This statistical method is generally applied to the identification of single group membership, rather than probabilities across multiple categories. Therefore, a complementary approach would be needed to assess and verify numerical profiles across the all six RIASEC categories, including both primary and non-primary groups.

In order to overcome the inherent limitations of the empirical approach, an additional approach, called the judgment method, was used to generate numerical profiles. For this method, three expert judges rated occupations on how descriptive and characteristic they were of each of the six RIASEC work environments. Using a seven-point scale, the judges assessed the RIASEC categories for each career in the database. Profiles were then developed from the mean scores of the three judges and submitted for review by a secondary panel.

In general, the ratings from trained judges proved both effective and reliable. In particular, it proved more effective than the empirical method in deriving numerical profiles for the five non-primary RIASEC categories. Statistically, the posterior probability levels compared to the discriminant method were much less extreme among the six RIASEC categories, providing reasonable proportions for all six categories. Cross-classification tables and values for Cohen’s Kappa were also examined to assess the degree of agreement between the two methods. Cohen’s Kappa is a measure of agreement between method pairs that involve unordered categories, with higher values indicating a higher level of agreement, and values above .70 considered acceptable. The two methods described here had a Kappa of .72.