Family Experiences in Adolescence Predict Young Adult Educational Attainment: A Machine Learning Approach

• In Event: 3-195 - Leveraging Machine Learning Methods for Research on Child Development in the Big Data Era

Integrative Statement

Predicting future outcomes is a key mission of developmental science (Yarkoni & Westfall, 2017). Developmental scientists have recently recognized the potential added value of using machine learning (ML) algorithms, which, facilitated by cross-validation, can train predictive models based on observed data and test predictive performance on new observations (Whelan et al., 2014). To train predictive models, existing research focused on explaining observed developmental outcomes provide prior knowledge that can potentially enhance performance of predictions (Craft et al., 2017). In the case of education achievement, both theoretical work and empirical findings from explanatory studies suggest that adolescents’ family experiences are related to later educational attainment (Gordon & Cui, 2012). Data driven, ML algorithms can add to this knowledge by identifying the specific and complex combinations of family experience that are most predictive of future educational attainment. In particular, ML methods can complement explanatory studies by identifying the relative contribution of a massive number of potential predictors without assumptions of linearity or exclusivity – feature importance – thereby capturing the complex developmental mechanisms and numerous factors that influence developmental outcomes (Brick et al., 2017). Accordingly, this study addressed two aims: (1) to use a constellation of family experiences in adolescence that have been identified as significant explanatory variables to train predictive models for young adult educational attainment; (2) to identify the relatively important predictive family factors, including indices of family structure, resources and interactions, in predicting young adult educational attainment.

We used public data from the National Longitudinal Study of Adolescent to Adult Health (Add Health). Based on 41 publications that examined relations between family experiences and educational attainment using Add Health data, we identified 26 family-related variables from Wave I (adolescent mean age = 15.96, SD = 1.79) that captured family structure (e.g., single-parenthood, family size), family resources (e.g., family income, parent education) and family interactions (e.g., adolescent-parent closeness) and that might predict educational outcomes at Wave IV (young adult mean age = 28.83, SD = 1.79), specifically, college entrance (66% yes vs. 34% no) and college completion (33% yes vs. 67% no). A Random Forest Classifier, an ensemble method based on decision tree classifiers that allow for nonlinearities and complex high-order interactions among predictors (Breiman, 2001), was used to build predictive models for the two binary outcomes. Viable data from 4,830 cases (all cases with less than 25% missing data and with multiple imputation to handle the remaining missingness; Brick et al., 2017) were randomly sliced into a training set (80%) and a test set (20%). Model tuning was done with ten-fold cross-validation.

Results indicate that educational attainment can be predicted relatively accurately, with predictive accuracy obtained from the test set of 72% for college entrance and 77% for college completion. Family income was identified as the most important predictor for both outcomes, followed by mothers’ and fathers’ education, parent age, and adolescent-father closeness. Complementing and adding to existing propositions about the importance of adolescents’ family experiences for later educational attainment, the ML results highlight the important predictive role of family resources.

Authors

• Xiaoran Sun, The Pennsylvania State University
  Presenting Author

• Nilam Ram, Penn State University
  Non-Presenting Author