Original research•

Prediction of adolescent subjective well-being: A machine learning approach

Abstract

Background Subjective well-being (SWB), also known as happiness, plays an important role in evaluating both mental and physical health. Adolescents deserve specific attention because they are under a great variety of stresses and are at risk for mental disorders during adulthood.

Aim The present paper aims to predict undergraduate students’ SWB by machine learning method.

Methods Gradient Boosting Classifier which was an innovative yet validated machine learning approach was used to analyse data from 10 518 Chinese adolescents. The online survey included 298 factors such as depression and personality. Quality control procedure was used to minimise biases due to online survey reports. We applied feature selection to achieve the balance between optimal prediction and result interpretation.

Results The top 20 happiness risks and protective factors were finally brought into the predicting model. Approximately 90% individuals’ SWB can be predicted correctly, and the sensitivity and specificity were about 92% and 90%, respectively.

Conclusions This result identifies at-risk individuals according to new characteristics and established the foundation for adolescent prevention strategies.

Introduction

Happy people tend to live longer and have better physical and mental health. Adolescence is a critical period since some results suggest that positive youth development can improve long-term health.1 Furthermore, adolescent depression was a strong predictor of mental disorders during adulthood.2 For example, many investigators had reported that undergraduate students suffered from depression and are vulnerable to suicide attempt and completed suicide.3–8 In a meta-analysis, Ibrahim et al also concluded that undergraduate students were more prone to depression with high prevalence.9 Therefore, robust identification of unhappy students is critical to develop and apply specific interventions to at-risk individuals. So far, traditional approaches adopted single self-report scale such as Centre for Epidemiologic Studies Depression Scale (CES-D), Satisfaction with Life Scale (SWL) and Positive and Negative Affect Schedule (PANAS), which are not reliable since SWB was multifaceted.10 Indeed, SWB contains many dimensions such as life-satisfaction, positive emotion and negative emotion.11 For these reasons, identifying unhappy students required multivariate approaches to adequately circumscribe the multifaceted construct of SWB.

As a multivariable big-data problem,12 machine learning can provide SWB problem with solutions that would outperform classical method. As such, previous studies had applied machine learning approaches to predict SWB. For example, Bogomolov et al used machine learning to predict SWB by using real-world and online data from mobile phone.13 14 Saputri and Lee adopted the same method to predict country SWB15 and Jatupaiboon et al used electroencephalogram to train model.16 These studies showed that machine learning could predict SWB better than single scale measurements. However, these previous studies focused on adult population and their application in terms of preventive strategies towards mental health was limited. Moreover, recent learning approaches, such as ensemble methods, had shown improved classification accuracies.

Ensemble methods had been widely adopted recently because of its good performance. The general idea of ‘ensemble methods’ was essentially based on constructing a set of simple classifiers and combining them. Final decisions were given by weighted or unweighted votes from each simple classifier, which contributes to model accuracy.17 One of the most representative ensemble methods was gradient boosting algorithm. It combined a set of simple classifiers. Each of them performed on data with one distribution. Those weak classifiers generated one strong classifier which can achieve higher accuracy than other simple ones.18 Finally, their performance would be improved. Gradient boosting algorithm had many advantages. First, it was insensitive to data with non-normal distributions and outliers. Additionally, we did not have any a priori hypotheses about input variables, which should be considered by the boosting algorithm. This algorithm was also robust against the addition of irrelevant input variables due to trees’ attribute.19

Including both psychological and physiological parameters, we can take advantages of the gradient boosting algorithm to predict undergraduates’ SWB with satisfying accuracy.

Methods

Online survey design

Scores of SWL and PANAS were used to measure undergraduate SWB. Other measurement items were summarised in table 1. Scales consisted of Adolescent Self-Rating Life Events Check List, Big Five Inventory (BFI),7 Child and Adolescent Social Support Scale, CES-D Dispositional Flow Scale (DFS) General Self-Efficacy Scale (GSES) Utrecht Work Engagement Scale-Student (UWES-S) and Multidimensional-Multiattributional Causality Scale (MMCS). We also collected general information such as gender, blood type, exercise, sleep, religion, economical situations, parents’ education level and characters. Moreover, four feedback questions and time elapsing were designed for sifting reliable data.

Table 1
|
Online scale items

Data collecting

All participants came from Jining Medical University. We included all participants who signed the ethical approval document.

Students (freshmen) who were entranced in 2016 and 2017 were recruited in this study in their first year. And students (sophomores) who were entranced in 2016 also took online survey in 2017 in their second year. We recruited 10 518 survey data in total. To minimise environmental influences, students were gathered and asked to complete the online scale together.

Data processing

Figure 1 shows the steps of our data processing. First, we dropped samples with one score feedback questions, such as ‘this survey is meaningless’, ‘harry when answering questions’, ‘hard to understand this questionnaire’ and ‘answers do not reflect the truth’. In addition, only answering times within 99% CI were included. Second, the data size was reduced to 10 272 in total. Then, dummy features (table 1) were encoded to one-hot codes and binary classifying features were encoded to 0/1 codes, which were the proper format for machine learning. Third, standardisation was adopted to eliminate different orders’ problem. After principal component analysis of SWL and PANAS, the foremost two components were calculated as the ‘pca’ score. Whole data were divided into two datasets according to freshmen and sophomore. Data 1 contained all freshmen’s information (N=6 886), and data 2 contained all sophomores’ information (N=3 386). The top 30% samples with ‘pca’ were labelled as 1 (data 1: N=2062; data 2: N=1016), and the bottom 30% were labelled as 0 (data 1: N=2063; data 2: N=1016). The remaining 40% of data were excluded in this study. After random shuffle, each dataset was divided into the training set and testing set with a conventional ratio 7 (data 1: N=2887; data 2: N=1422) to 3 (data 1: N=1238; data 2: N=610).

Figure 1
Request permissionrequest permission icon
Figure 1

Flowchart of the study.

Feature selection

Each item of every scale was input as a feature, we included 298 features in total. To avoid overfitting and facilitate practical application, fewer features would be better. We assessed all 298 features simultaneously with elastic net regularisation, which can avoid correlated factors overfitting. This method would remove uninformative features and assign low weight to correlated ones. We selected the best 20 features from data 1 and data 2 separately. Further model construction would only consider these 20 features. This number was selected to balance prediction accuracy and usability in practical analysis. We also assessed models with 10 or 30 features.

Machine learning algorithm

This research applied computational language python and Gradient Boosting Classifier (GBC) in scikit-learn (sklearn, http://scikit-learn.org/stable/index.html), a machine learning module in python, to build our predicting model. It was the model parameters which constructed one specific model. As an ensemble method, GBC combined multiple weak classifiers to produce more accurate prediction significantly; it was much better than any base classifiers. Initially, one specific classifier fits the data. The next new classifier would re-weight parameters to the direction of gradient descend, which would minimise loss of function. Finally, a model with the minimum error on training data (70% dataset) was generated. Model performance was evaluated by confusion matrix, receiver operating characteristic curve (ROC) and area under curve (AUC). Considering different life patterns between freshmen and sophomores, two models on data 1 (GBC1) and data 2 (GBC2) were built separately.

Tuning parameters

To achieve better predicting accuracy, tuning hyperparameters of model was necessary. For example, model accuracy with different settings of two main important hyperparameters n_estimators and learning_rate has a typical pattern. With many important hyperparameters (eg, max_depth, min_samples_split, min_samples_leaf, max_features) and trade-off problems, specific algorithm makes this process efficient. A machine learning algorithm named GridSearchCV was applied to process optimal hyperparameters searching in GBC.

Results

Label construction

Subjective well-being included three dimensions: life satisfaction, positive affect and negative affect.20 21 We took the SWL score and PANAS two subscores (positive and negative affect scores) as SWB measurement. After principal component analysis, we took foremost two components which can explain freshmen 82.5% (86.0%, sophomores) variance as the ‘pca’ score for further label tagging. The ‘pca’ score was negatively related with happiness, and it was calculated by formula listed below. The top 30% points were 1.87 (2.69, sophomores) and bottom 30% points were −2.48 (−2.47, sophomores). Observations with a score higher than 1.87 (2.69, sophomores) were labelled as unhappy while individuals with a score lower than −2.48 (−2.47, sophomores) were considered as happy.

Freshmen: ‘pca’=0.536×component 1+0.289×component 2

Sophomores: ‘pca’=0.509×component 1+0.352×component 2

Model performance

After tuning parameters, GBC1 with 0.06 learning_rate enabled the model to achieve best performance. GBC2 with default parameters was the best. Figure 2 shows models’ ROCs. GBC had tiny advantage on predicting sophomores’ SWB. Table 2 shows various model performance measurements on different sets of feature numbers. Models with 10, 20 and 30 predictors had no significant difference.

Figure 2
Request permissionrequest permission icon
Figure 2

ROAUC.AUC, area under curve; ROC, receiver operating characteristic curve.

Table 2
|
Model evaluation

Feature importance

All 20 selected features’ relative importance for predicting undergraduate SWB are shown in figure 3. Risk and protective factors with huge difference patterns can be observed between freshmen and sophomores. The top three predictors of freshmen were “I felt fearful” (CES-D), ‘get nervous easily’ (BFI), “I was bothered by things that usually don’t bother me” (CES-D). The top three predictors of sophomores were questions of CES-D: “I was happy”, “I felt fearful”, “I felt that I could not shake off the blues even with help from my family or friends”.

Figure 3
Request permissionrequest permission icon
Figure 3

Top 20 features for predicting undergraduates’ SWBCES-D: items of CES-D; sleep: self-reported sleep quantity; DFS_loss: one of nine dimensions of DFS, loss of self-consciousness; BFI: items of BFI; MMCS: items of MMCS; ASLEC: items of ASLEC; DFS_clear: one of nine dimensions of DFI, clear goals. ASLEC, Adolescent Self-Rating Life Events Check List; BFI, Big Five Inventory; CES-D, Centre for Epidemiologic Studies Depression Scale; DFS, Dispositional Flow Scale; MMCS, Multidimensional-Multiattributional Causality Scale; SWB, subjective well-being.

Discussion

Main findings

The present paper constructed a machine learning model for predicting undergraduate students’ SWB with an accuracy of about 91%. Meanwhile, important predictors for SWB were displayed and analysed. Personality and depressive symptoms affect both freshmen’s and sophomores’ SWB most. In addition, this paper presented a machine learning method, GBC, to predict undergraduate SWB. Two models were built on freshmen data (GBC1) and sophomore data (GBC2). The prediction accuracy achieved 90.47% (GBC1) and 90.98% (GBC2). Different SWB patterns between freshmen and sophomore were explored according to different important predictors. As far as we know, this work might be the first one adopting machine learning method to predict adolescent happiness in the Chinese population. Furthermore, two possible 20-item questionnaires for interpreting were generated (table 3). Rather than evaluate happiness with redundant factors, these 20 self-reported questions may diagnose SWB more efficiently. Students with this simple self-test can monitor their mental health, and psychological consultation teachers could identify at-risk individuals easily by evaluating the scores of each question. For example, a depressive student with a relatively low sleep situation may receive sleep therapy.

Table 3
|
Twenty features of predicting undergraduate SWB

Freshmen and sophomores shared 11 predictors. Most of those items measuring depressive symptoms were also important predictors. For both freshmen and sophomores, half predictors came from CES-D. Depression and happiness are ‘mirror images’, and the relationship between depression and SWB had already been reported.22 23 Reasonably, happier students were less depressed. Since depression had been proved to be related to genomic background,24 the relationship between SWB and depression suggested the possibility of ‘biological happiness’ in addition to ‘sociological happiness’. A quantified SWB level could be possible in the future. Moreover, a Genome-wide Association Study (GWAS) had reported that parts of ‘SWB’ SNPs were significantly associated with depression symptoms.25 Both SWB and depression may share some of the same genetic factors. Clinically, well-being therapy (WBT) was a psychotherapeutic strategy. It aimed to increase patients’ mental health, and guide themselves to a state of positive emotion by emphasising on self-observations. WBT has been proved as a successful way in easing depression.26 Those results indicated a close relationship between happiness and depression, which may contribute to antidepression therapy in the future.

The importance of personality questions was about the same as depressive items. Many studies had reported the strong relationship between personality and happiness.20 27 In this study, consistent with previous studies, agreeableness, openness, conscientiousness and extraversion were positively correlated with happiness while neuroticism was negatively related to happiness. Since people with agreeableness, openness, conscientiousness and extraversion were more likely to get involved in positive social network, these traits would contribute to positive enjoyment or life satisfaction as well. The happiness level of people with these personalities can increase at the same time. On the contrary, people with neuroticism tended to suffer more misfortunate feelings, which resulted in less joy or pleasures.28

Two dimensions of DFS: loss of self-consciousness and clear goals can affect undergraduates’ SWB slightly. DFS was designed as the measurement of flow. ‘Flow’ was first put forward as an intrinsically optimal state that resulted from intense engagement with daily activities.29 In other words, people who were facing certain challengeable activities with matched skills would generate positive emotions during acting. Being intrinsically motivated, everyone can gain happiness because of a sense of euphoria and satisfaction.30 We suppose college students can achieve SWB if they can be guided to experience more flow states.

Limitations

First, all participants came from the same college, which may not represent the undergraduate population perfectly. In the future, we would try to seek for global cooperation to understand happiness predictors generally. Furthermore, no validation studies had been conducted on our 20-item questionnaires, which may possibly be used to measure undergraduate SWB. The reliability and effectiveness needed to be further studied.

Implications

This work provides a new evaluating approach on SWB, and contributes to understand SWB predictors. The 20-item questionnaires may inspire further simple self-reported SWB measurement, which could benefit individualised happiness detection.

Naixin Zhang obtained a bachelor degree from the department of life science, Nankai University, Tianjin, China in 2017. She is now working on the master program in the Bio-X Institutes, Department of Life Science and Technology, Shanghai Jiaotong University, Shanghai, China. Her research interest include susceptibility genes of mental disorders, such as schizophrenia and major disorder depression, pharmacogenomics, such as different effectiveness of anti-depressant venlafaxine on patients with single nueleotide polymorphism, and subjective well-being of adolescent by machine learning method.

author bio image
Article metrics
Altmetricopen-url
Dimensionsopen-url