A Neural Network Approach To Identifying Adolescent Adjustment - Statistical Data Included

Adolescence, Spring, 2001 by Jyotsna Nair, Satish S. Nair, Javad H. Kashani, John C. Reid, Venkatesh G. Rao

ABSTRACT

This study examined the relationship between the quality of adjustment in adolescents and a set of psychiatric diagnoses, personality traits, parental bonding, and social support variables. One hundred fifty adolescents were administered the Millon Adolescent Personality Inventory, the Parental Bonding Questionnaire, the Social Support Questionnaire, and the Diagnostic Interview for Children and Adolescents. A neural network approach was then utilized, and it was found that several of the variables (e.g., Major Depressive Disorder, Conduct Disorder, and Societal Conformity) had a significant role in classifying adolescents into three groups: maladjusted, nominally adjusted, and well-adjusted.

Few studies have identified risk and protective factors in both dysfunctional and well-adjusted adolescents. Kashani et al. (1987) found that well-adjusted youths had goad self-concepts, caring parents, and satisfactory social support systems. Using the same sample, the present study sought to determine whether a neural network approach would offer additional information. A neural network is a nonlinear regression model that can predict outputs (or effect variables) using several inputs (or cause variables) and quantify complex relationships between such cause-and-effect variables (McCord-Nelson & Illingworth, 1991). Neural networks have been applied in several areas, including mental health (Cohen & Servan-Schreiber, 1992; Kashani et al., 1996; Nair et al., 1996). Here, a neural network model was created to ascertain how various measures would relate to adolescents characterized as well-adjusted, nominally adjusted (some dysfunctions), and maladjusted (serious dysfunctions).

METHOD

Data Collection

One hundred fifty youths between 14 and 16 years of age were drawn from a sample of 1,700 midwestern public school students. There were equal numbers of boys and girls. Ninety-five percent were Caucasian, and the rest were Asian or African American. Other characteristics are described in the study by Kashani et al. (1987).

Diagnoses were made based on data collected from adolescents and their parents using the Diagnostic Interview for Children and Adolescents (DICA; Herjanic et al., 1975; Herjanic & Reich, 1982). The DICA diagnoses used in this study were Oppositional Defiant Disorder, Conduct Disorder, Anxiety, and Major Depressive Disorder (Depression). The Millon Adolescent Personality Inventory (MAPI; Millon et al., 1972) was used to obtain information about adolescents' personalities. The Millon scales used were Cooperative, Forceful, Sensitive, Personal Esteem, Social Tolerance, Family Rapport, Impulse Control, and Societal Conformity. Additional inputs to the neural network were the Parental Care and the Parental Overprotection scales of the Parental Bonding Instrument (PBI; Parker et al., 1979), and the Satisfaction Rating and Total Number of People scales of the Social Support Questionnaire (SSQ; Sarason et al., 1983). Gender was also included in the neural network model, bringing the number of inputs to 17 (see Figur e 1).

The 150 adolescents were divided into three groups, based on clinical interviews conducted by expert psychiatrists (Kashani et al., 1987), and these were used as the model outputs. Group 1 consisted of troubled adolescents with psychiatric disorders. These maladjusted adolescents had one or more DSM-III diagnoses, experienced impaired functioning, and were in need of treatment. Group 2 consisted of nominally adjusted adolescents. They were not free of symptoms, but were not in need of treatment. Group 3 consisted of well-adjusted adolescents. They were free of psychiatric syndromes or symptoms. Seven adolescents were dropped from the analyses due to missing data.

Neural Network Modeling

A multilayered back-propagation neural network was used (17 inputs from each of the 143 adolescents, comprising the input patterns, and the three binary outputs). The network was exposed to the data, and the parameters (weights and biases) were adjusted to minimize error, using a back-propagation training algorithm. The procedure for analyzing the data involved three stages: (1) The inputs (personality, DICA diagnoses, parental bonding, social support, and gender variables) were mapped to the three outputs (maladjusted adolescents, nominally adjusted adolescents, and well-adjusted adolescents) using a neural network. (2) Each input was varied, one at a time, across its minimum to maximum range to determine how the increase in that input affected the output (i.e., whether group membership would change). This perturbation process (also called contribution analysis) was used to identify the variables most related to troubled adolescents and those most related to well-adjusted adolescents (i.e., variables associ ated with risk and protection). (3) Statistical analyses were conducted to confirm the relative effect of the input variables on the outputs.

Validation

The ability of the neural network model (a 17 X 20 X 5 X 3 structure was used) to perform the classification was examined by setting aside 20% of the patterns (or observations) as validation (or testing) data. In this cross-validation approach, the training involved repeatedly exposing the network to the remaining 80% of the patterns (training data) for several epochs, where an epoch is one complete cycle through the network for all cases. (Data were normalized before training.) Simultaneously, the prediction errors in the testing data were monitored. Typically, the training errors (in this instance, for the 80% set) drop consistently while the testing errors (for the 20% set) drop and then increase with continued training. The optimal number of training epochs is achieved when the training and testing errors are both acceptable. After the optimal number of training epochs is determined, the data are pooled (i.e., the training and testing data are combined) and the network is trained for this number of epoch s using the combined set. A network trained in this manner is considered generalizable, in the sense that it can be used to make predictions.