Report

Brian Lukoff Stanford University October 13, 2006 Based on a draft paper that is joint work with Eric Heggestad, Patrick Kyllonen, and Richard Roberts The decision tree method and its applications to faking Evaluating decision tree performance Three studies evaluating the method Study 1: Low-stakes noncognitive assessments Study 2: Experimental data Study 3: Real-world selection Implications and conclusions A technique from machine learning for predicting an outcome variable from (a possibly large number of) predictor variables Outcome variable can be categorical (classification tree) or continuous (regression tree) Algorithm builds the decision tree based on empirical data TRAINING SET Day Snowing? Raining? Method 1 yes yes drive 2 yes no drive 3 no yes drive 4 no yes walk 5 no no walk 6 no no walk 7 no yes drive Is it snowing? Yes No drive Is it raining? Yes drive No walk TRAINING SET Day Snowing? Raining? Method 1 yes yes drive 2 yes no drive 3 no yes drive 4 no yes walk 5 no no walk 6 no no walk 7 no yes drive Is it snowing? Yes No drive Is it raining? Yes drive Not all cases are accounted for correctly Wrong decision on Day 4 Need to choose variables predictive enough of the outcome No walk TEST SET Is it snowing? Yes No drive Is it raining? Yes drive No walk Day Snowing? Raining? Method Prediction 8 yes yes drive drive 9 no yes walk drive 10 no yes drive drive 11 yes no drive drive 12 no no walk walk 13 no no walk walk 14 yes yes drive drive Not all cases are predicted correctly Maybe the decision to drive or walk is determined by more than just the snow and rain? Ease of interpretation Simplicity of use Flexibility in variable selection Functionality to build decision trees readily available in software (e.g., the R statistical package) Outcome variable = faking status (“faking” or “honest”) Training set = an experimental data set where some participants instructed to fake Training set = a data set where some respondents are known to have faked Outcome variable = lie scale score Training set = a data set where the target lie scale was administered to some subjects So far, have used individual item responses only Other possibilities: Variance of item responses Number of item responses in the highest (or lowest category) Modal item response Decision tree method permits some sloppiness in variable selection Classification trees (dichotomous outcome case, e.g., predicting faking or not faking) Accuracy rate False positive rate Hit rate Continuous Average absolute error Correlation between actual and predicted scores Algorithm can “overfit” to the training data, so performance metrics computed on the training data not indicative of future performance Thus we will often partition the data: Training set (data used to build tree) Test set (data used to compute performance metrics) Training/test set split leaves a lot to the chance selection of the training and test set Instead, partition the data into k equal subsets Use each subset as a test set for the tree trained on the rest of the data Average the resulting performance metrics to get better estimates of performance on new data Here we will report cross-validation estimates Data sets Two sets of students (N = 431 and N = 824) that took a battery of noncognitive assessments as well as two lie scales as part of a larger study Measures Predictor variables ▪ IPIP (“Big Five” personality measure) items ▪ Social Judgment Scale items Outcomes (lie scales) ▪ Overclaiming Questionnaire ▪ Balanced Inventory of Desirable Responding Method Build regression trees to predict scores on each lie scale based on students’ item responses Varying performance, depending on the items used for prediction and the lie scale used as the outcome Correlations between actual lie scale scores and predicted scores ranged from -.02 to .49 Average prediction errors ranged from .74 to .95 SD Low-stakes setting: how much faking was there to detect? Nonexperimental data set: students with high scores on the lie scales may or may not have actually been faking Data sets An experimental data set of N = 590 students in two conditions (“honest” and “faking”) Measures Predictor variables ▪ IPIP (“Big Five” personality assessment) items Method Build decision trees to classify students as honest or faking based on their personality test item responses Decision trees correctly classified students into experimental condition with varying success Accuracy rates of 56% to 71% False positive rates of 25% to 41% Hit rates of 52% to 68% Two items on a 1-5 scale form a decision tree: Item 19: “I always get right to work” Item 107: “Do things at the last minute” (reversed) Extreme values of either one are indicative of faking Many successful trees utilized few item responses Range of tree performance Laboratory—not real-world—data Although an experimental study, still don’t know: If students in the faking condition really faked If the degree to which they faked is indicative of how people fake in an operational setting If any of the students in the honest condition faked Data set N = 264 applicants for a job Measures Predictor variables ▪ Achievement striving, assertiveness, dependability, extroversion, and stress tolerance items of the revised KeyPoint Job Fit Assessment Outcome (lie scale) ▪ Candidness scale of the revised KeyPoint Job Fit Assessment Method Build decision trees predicting the candidness (lie scale) score from the other item responses Correlations between actual and predicted candidness (lie scale) scores ranged from .26 to .58 Average prediction errors ranged from .61 to .78 SD Items are on a 1-5 scale, where 5 indicates the highest level of Achievement Striving Note that most tests are for extreme item responses Similar methodology to Study 1, but better results (e.g., stronger correlations) Difference in results likely due to the fact that motivation to fake was higher in this realworld, high-stakes setting Wide variety in decision tree quality between groups of variables (e.g., conscientiousness scale vs. openness scale) Examining trees can give insight into the structure of the assessment Some decision trees in each study used only a small number of items and achieved a moderate level of accuracy Use decision trees for real-time faking detection on computer-administered noncognitive assessments Real-time “warning” system Need to study how this changes the psychometric properties of the assessment Address whether decision trees can be effective in an operational setting—are current decision trees accurate enough to reduce faking? Comparisons of decision tree faking/honest classification with classifications from IRT mixture models Develop additional features to be used as predictor variables Explore other machine learning techniques