Evaluating the performance of personality-based profiling in predicting physical activity Authors: Kentaro Katahira, Keisuke Takano, Takeyuki Oba, Kenta Kimura Published: 2024 ABSTRACT AbstractBackgroundProfiling or clustering individuals based on personality and other characteristics is a common statistical approach used in marketing, medicine, and social sciences. This approach improves data simplicity, supports the implementation of a data-driven decision-making process, and guides intervention strategies, such as personalized care. However, the clustering process involves loss of information owing to the discretization of continuous variables. Although any loss of information may be practically or pragmatically acceptable, the amount of information lost and its influence on predicting external outcomes have not yet been systematically investigated.MethodsWe assessed the accuracy of predicting physical activity using the clustering approach and compared it with the dimensional approach, where variables are used as continuous regressors. This analysis is based on survey data from a sample of 20,573 individuals regarding physical activity and psychological traits, including the Big-Five personality traits.ResultsA four-cluster solution, supported by the standard criterion for determining the number of clusters, achieved no more than 60–70% prediction accuracy of the dimensional approach employing the raw dimensional scale as explanatory variables.ConclusionThe cluster solution suggested by conventional statistical criteria may not be optimal when clusters are used to predict external outcomes.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40359-024-02268-6.Keywords: Physical activity, Personality, Clustering, Profiling, Prediction Background Personalization, which aims to optimize the content, channels, and modalities of an intervention or service depending on the needs and characteristics of individual recipients, has gained attention in recent years [1–3]. Within the context of personalization, various efforts have been made to incorporate psychological traits, such as personality. According to a recent review of studies on health behavior change interventions, behavioral or lifestyle data are the most commonly used information to personalize interventions (of 20/31, 65%) [3]. Among these variables, personality traits are promising factors in implementing personalization, as they are predictive of individuals’ behavior in general [4, 5] and have been shown to be informative when delivering optimal health messages (e.g., Ref [6]). Personalization of interventions requires certain rules for selecting appropriate methods and components [7] for individuals with different backgrounds and characteristics. A possible approach is to construct a predictive model using existing data that allows estimation of the effectiveness of each intervention component if delivered to a particular recipient (patient) or population [7]. For example, Cohen et al. [7] developed predictive models that aid in deciding whether to continue antidepressant medication (ADM) or transit to mindfulness-based cognitive therapy (MBCT) for preventing relapse in recurrent depression. The authors constructed models to predict the likelihood of relapse using both ADM therapy and MBCT. They demonstrated that comparing the predictions of these models may contribute to optimal treatment selection. Recent attempts have been made to build models using personality and other psychological traits as predictors of behavioral responses to interventions and treatment prognosis [8, 9]. In many cases, psychological traits, including personality, have been measured and treated as continuous variables. This is known as the trait or dimensional approach, particularly in personality research. However, when designing personalized interventions, it is expensive to develop detailed components and selection strategies for myriad combinations of multiple continuous variables. Therefore, it is effective to divide the population into a small number of distinct groups (or segments) and tailor the intervention to each group [10]. This approach is called the type or clustering approach. In the clustering-approach, statistical cluster analysis and profiling based on individual characteristics are often conducted [11–13]. However, clustering individuals sacrifice the information contained in the original continuous scales. Traditionally, dichotomization and other methods for reducing continuous values to categorical labels have been criticized for various reasons, including loss of information, reduced statistical power, underestimation of variation, and concealment of nonlinearity [14]. Regardless of these criticisms, the clustering approach has been appreciated for its simplicity and has been used to develop a knowledge base and implement personalization. For example, a clustering approach offers several benefits for messaging health campaigns [15]. By identifying groups of people who share similar psychosocial predictors, segmentation based on a clustering approach enables the crafting of targeted messages that are more effective than general messages in promoting behavioral changes [16]. Another practical advantage of the clustering approach is that predictive models no longer rely on the actual personality scores for each personality facet but rely solely on the profile, which may be obtained through different questionnaires and other data sources (interviews, meta-data, and diaries). Hence, researchers do not depend on administering the same questionnaire to each person for whom a prediction is to be made. This is critical to health sciences, marketing, and other research areas. Given these advantages and disadvantages, it is particularly important to evaluate the severity of information loss using a clustering approach. This issue has a direct impact on decision-making in practice. If information loss is negligible, the clustering approach can be prioritized without risk; otherwise, effective personalization can be achieved only when an intervention is sufficiently fine-grained, that is, tailored with careful consideration of the continuum of psychological traits. To our knowledge, no previous study has systematically quantified information that is lost due to clustering based on continuous predictors. This study examines these issues in the context of predicting physical activity (PA) using psychological variables. Specifically, we investigated the extent to which personality profiling based on clustering predicts PA and clustering methods that are effective for this purpose. Insufficient PA is acknowledged as a risk factor for a number of non-communicable and chronic complications such as diabetes and cardiovascular and respiratory diseases [17, 18]. Thus, low PA is perhaps an indication of “risk of needing intervention,” therefore, predicting PA is useful for selecting the intervention target. We used data from a large online survey of approximately 20,000 people to examine psychological variables such as personality and PA [19]. Cluster profiling can also be considered for predicting PA or other outcomes [8, 20–22]. Regarding personality, we focused on the Big-Five model or five-factor model (FFM), which is a widely accepted framework for describing human personality [23]. It comprises five core traits: Neuroticism, Extraversion, Openness, Agreeableness, and Conscientiousness. We also focused on self-control [24, 25] and regulatory focus theory [26], which have been examined in the context of PA promotion [27, 28]. We examined the impact of these variables on PA prediction by incorporating two commonly used clustering methods: K-means [29] and Gaussian mixture models (GMMs) [30]. We compared the predictive accuracy based on profiling with that of the dimensional approach model, wherein each characteristic was used as a continuous variable for a predictor. Methods All methods were performed in accordance with the relevant guidelines and regulations. Data Data were obtained from a large-scale cross-sectional survey of PA and psychological factors among Japanese-speaking adults. A part of these data have been reported elsewhere [31], where details of the sampling strategy can be found. In short, these data cover N = 20,611 residents of Japan (adults aged ≥ 18 years) sampled across all prefectures in the country. The survey was distributed via an online platform, where missing responses were detected automatically and immediately. The participants could not proceed to the next page if any missing responses were detected. Thus, there were no missing values in the data. Each participant was registered in a database of potential survey panels and received an e-mail invitation to participate. Informed consent was obtained electronically from each participant prior to initiating the survey. Upon visiting the survey page, participants received study information and were asked to indicate their agreement to participate in the survey. Participation was voluntary and they could exit the survey at any time. The study was conducted using an online text-based survey that required participants to have basic literacy skills. Therefore, illiterate participants were not included in the study. The study protocol was approved by the Ethics Committee of the National Institute of Advanced Industrial Science and Technology (Approval ID: 2022 − 1279). For statistical analysis, we excluded data from (a) 30 participants who reported being less than 1.0 m tall or weighing less than 10 kg, and (b) 8 participants who reported total PA hours (sum of walking and moderate/vigorous PA) exceeding 24 h per day, as we deemed these data unreliable. The remaining data from 20,573 samples were analyzed. Measures Our survey included personality and psychological variables and a wide range of measures related to health behavior (e.g., life environment, disease history, and mobile health usage). In this study, we reported only the Big-Five personality traits, regulatory focus, and self-regulation, focusing on how these psychological variables predict physical activity. Personality The Big-Five personality traits were assessed using the Japanese version of the Ten-Item Personality Inventory (TIPI-J) [32, 33]. This scale comprises 10 items, with two items representing each of the five personality factors (extraversion, agreeableness, conscientiousness, neuroticism, and openness). Both items were reverse scored. Items were rated on a 7-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly agree). The mean of each item pair is calculated as a factor score. The Cronbach’s alpha coefficients for the TIPI-J subscales were 0.62 (Extraversion), 0.44 (Agreeableness), 0.58 (Conscientiousness), 0.58 (Neuroticism), and 0.47 (Openness). Although these values were not sufficiently high, they were comparable to those reported in previous studies using the TIPI-J. The low values are attributed to the fact that each subscale of the TIPI-J has only two items. Regulatory focus Two types of regulatory focus (promotion and prevention) were measured using the Japanese version of the Promotion/Prevention Focus Scale [26, 34]. This scale comprises 16 items, eight for each factor (promotion focus and prevention focus). Items were rated on a 7-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly agree). The sum of the scores for the eight items corresponding to each factor was used as the score for that factor. Cronbach’s alpha coefficients were 0.86 and 0.80the Promotion Prevention subscales, respectively. Self-control Self-control was measured using the Japanese version of the Brief Self-Control Scale (BSCS-J) [25, 35]. This scale has 13 items, including 9 reverse-scored items. The study respondents rated each item on a 5-point Likert scale ranging from 1 (not at all like me) to 5 (very much like me). The Cronbach’s alpha coefficient for the BSCS-J was 0.85. International physical activity questionnaire short form The Japanese version of the International Physical Activity Questionnaire Short Form (IPAQ-SF) [36, 37] was used to assess PAs. See for the validation of objective measures [38]. The IPAQ-SF assesses PA using three categories: (a) walking, (b) moderate-intensity activity, and (c) vigorous-intensity activity. For each category, participants reported the number of days and time spent on the activity (lasting for at least 10 min) over an average week (but not for the last week). The weekly time (minutes) spent on each PA category was calculated by multiplying the frequency and duration. Analysis and models Analysis flow First, we examine the clustering results based on K-means and GMM. The optimal number of clusters is examined in terms of explaining the data distribution rather than predicting the external variable using the Gap statistic [39] for K-means and the Bayesian Information Criterion (BIC) [40] for GMM. To characterize the distribution of personality scores and each cluster, we applied PCA, a method for dimensionality reduction of data, to the personality scores. Specifically, we examined if the data exhibited a cluster structure with a multimodal distribution. PCA decomposes the variance in a dataset into principal components (PCs). We visualized the structure of the clusters by color-coding them in the PC1 and PC2 score spaces. Clustering was performed using the entire dataset for analysis. Next, we compared the prediction accuracies of the dimensional and clustering approach models. To measure the prediction accuracy, data were randomly divided into training data (N = 10287) and test data (N = 10286). As a dimensional-approach model, regression models to predict total PA using each score (i.e., Big-Five personality traits, two types of regulatory focus, and self-control) as continuous predictors were considered and trained using training data. For the clustering approach, we used regression models with cluster membership (labels) as discrete predictors of PA. The predictive performances of these “clustering-approach” models were then compared with those of the dimensional-approach models. We systematically explored how the number of clusters included in the clustering approach model influenced the predictive performance and expected the performance to improve with the increase in number of clusters. Specifically, we focused on the predictive power of the clustering approach using the number of clusters selected in the established metric relative to the dimensional-approach model. We explored the number of clusters required to achieve a predictive performance comparable to that of the dimensional-approach model. Variables used for clustering and predictors For both the clustering- and dimensional-approach models, we considered five sets of variables by combining the Big-Five scores (TIPI-J) and the regulatory focus/self-control variables. The first set included only the Big-Five scores, whereas the second included regulatory focus (promotion and prevention focus scores) to the first set. The third set added a self-control score to the first set, whereas the fourth set included all the scores. The fifth set of variables did not include the Big-Five scores and comprised only the regulatory focus and self-control scores. The models in which Variable Sets 1–5 were included in the multiple linear regression analysis as explanatory variables were designated as Models 1–5, respectively. The variables were standardized such that the training data had a mean of 0 and a standard deviation of 1. This transformation was then applied to the test data, resulting in the transformed test scores deviating slightly from a mean of zero and a standard deviation of one. Outcome The primary outcome predicted by the models was total PA level (min/week), as assessed using the IPAQ-SF. Owing to the skewed distribution of total PA, it was log-transformed and used as a dependent variable in the main analysis. As an additional analysis, we also conducted logistic regression to predict a binary outcome indicating if the study participants satisfied the PA level of > 23 METs-h/week (the results are shown in the Supplementary Information A). Japan’s Ministry of Health, Welfare, and Labour established this 23 MET-h/week criterion in 2006. Predicting if individuals do not meet this condition is useful because candidates who fail to reach this level may require interventions. In our overall dataset, 8,915 participants met the criterion of 23 METs-h/week or more for PA, while 11,658 did not. Dimensional-approach models The model formulas for the dimensional-approach models are as follows: Model 1: PA ~ 1 + E + A + C + N + O. Model 2: PA ~ 1 + E + A + C + N + O + Promotion + Prevention. Model 3: PA ~ 1 + E + A + C + N + O + Self-control. Model 4: PA ~ 1 + E + A + C + N + O + Promotion + Prevention + Self-control. Model 5: PA ~ Promotion + Prevention + Self-Control. where E = Extraversion, A = Agreeableness, C = Conscientiousness, N = Neuroticism, O = Openness. In the above model formulas, the variable on the left side of the “~” represents the outcome variable, while those on the right side represent the predictor variables and “+1” indicates the intercept. Note that while including covariates such as age and gender in the model increased the prediction accuracy of PA (as shown in Supporting Information C), these covariates were excluded to ensure that the R² value is zero when personality traits or other psychological characteristics are not included, making the R² a ratio scale with zero as a baseline, thus enabling easier interpretation. Clustering-approach models The clustering methods used were K-means and GMMs. K-means clustering was performed using the k_means() function from the scikit-learn library for Python, using the reticulate package in R. The cluster size for the K-means clustering was evaluated using the gap statistic [39]. The gap statistic was computed using the clusGap function in the cluster packages for R. Similarly, the GMM was performed using the GaussianMixture function in scikit-learn. The initial values for the K-means and GMM were determined by k-means + + algorithm [41]. The Bayesian information criterion (BIC) was used to evaluate the number of clusters [40]. A smaller BIC indicates a more probable model. For both K-means and GMM, optimization was executed starting from 30 random initializations, and the best solution was selected. The number of clusters for K varied from 1 to 30. However, for the GMM, as the number of clusters increases, the likelihood tends to become extremely high owing to the assignment of components to a small number of samples, and this can make the performance of BIC unstable and less reliable. Thus, the BIC was evaluated for cluster numbers of up to 20. In addition, from the perspective of the computational cost, the Gap statistic was evaluated for cluster numbers of up to 20. Based on the patterns observed for up to 20 clusters, the likelihood that an optimal solution would include a cluster size of > 20 was considered low. For the clustering-based approach, cluster labels were transformed into dummy-coded dichotomous variables and used as predictors in the regression model. The model formula is: PA ~ 1 + cluster Here, “cluster” is a factor variable that was converted into dummy coded variables in R. Metrics for prediction performance Prediction accuracy was evaluated using the explained variance (R²) for continuous PA. For binary outcomes indicating adherence to the recommended PA level, prediction accuracy was assessed using the area under the Curve (AUC) for the receiver operating characteristic (ROC) curve. These prediction accuracy metrics were computed using the test data rather than the training data. To obtain 95% confidence intervals for these metrics, 1,000 bootstrap samples were generated from both the training and test data. Clustering and regression model fitting was performed on each bootstrap sample and the predictive performance of the regression model was evaluated using the bootstrap sample from the test data. Results Demographics and descriptive statistics Table 1 present the demographic information and descriptive statistics of the scores for each scale. This table also includes the profiles of the four clusters generated by K-means clustering using the Big-Five scores (Variable Set 1) for the overall samples (N = 20,573). The respondents’ mean age was 52.65 (SD = 17.77), ranging from 19 to 94 years. Table 2 presents the Pearson’s correlation coefficients among the nine variables, comprising the five Big-Five scores, two regulatory focus scores, a self-control score, and total PA (raw value rather than a log-transformed one). Weak correlations with coefficients between r = .1 and r = .2 were observed for all the variables, except for total PA and prevention focus. Table 1Demographic information, descriptive statistics of personality scores and profiles of the four clusters generated by K-means using the Big-Five scoresK-means (Big-Five only, K = 4)CharacteristicOverall, N = 20,573Cluster 1, N = 3,445Cluster 2, N = 5,817Cluster 3, N = 6,553Cluster 4, N = 4,758Age52.66 (17.77)43.91 (15.12)55.01 (17.54)49.02 (17.38)61.11 (15.98)Women10,701 (52%)1,943 (56%)3,216 (55%)2,977 (45%)2,565 (54%)BMI22.14 (3.70)22.31 (4.43)21.78 (3.43)22.38 (3.75)22.12 (3.32)Extraversion3.73 (1.33)2.62 (1.03)2.90 (0.95)4.21 (0.96)4.91 (1.05)Agreeableness5.01 (1.04)4.40 (1.08)5.63 (0.65)4.25 (0.77)5.73 (0.70)Conscientiousness4.15 (1.22)2.70 (0.93)4.43 (0.95)3.90 (0.86)5.20 (0.94)Neuroticism3.99 (1.23)5.38 (0.94)3.87 (0.99)4.22 (0.84)2.82 (0.95)Openness3.80 (1.15)2.75 (0.95)3.26 (0.90)4.17 (0.83)4.69 (1.01)Promotion focus27.88 (6.66)24.19 (6.87)27.10 (6.20)27.94 (5.69)31.43 (6.56)Prevention focus27.97 (6.32)31.66 (6.69)28.32 (5.95)27.89 (5.51)25.00 (6.06)Self-control43.32 (7.85)36.04 (7.22)45.14 (6.52)41.27 (6.12)49.20 (6.51)Total PA (in min/week)270 (60, 600)180 (0, 420)240 (60, 540)240 (30, 600)420 (160, 750)Note Mean (SD); n (%); For PA, Median (interquartile range, IQR) of the raw values are shown; BMI = Body Mass Index Demographic information, descriptive statistics of personality scores and profiles of the four clusters generated by K-means using the Big-Five scores Note Mean (SD); n (%); For PA, Median (interquartile range, IQR) of the raw values are shown; BMI = Body Mass Index Table 2Correlations between variables (Pearson’s correlation coefficients, TotalN = 20,573)EAC N OPromotion focusPrevention focusSelf-controlEA0.03C0.260.33N− 0.29− 0.32− 0.43O0.360.080.27− 0.28Promotion0.250.170.30− 0.210.40Prevention− 0.26− 0.11− 0.200.45− 0.150.24Self-control0.210.390.63− 0.470.200.22− 0.29Total PA0.140.150.13− 0.110.110.16− 0.020.14Note E = Extraversion, A = Agreeableness, C = Conscientiousness, N = Neuroticism, O = Openness, PA = Physical Activity (raw value) Correlations between variables (Pearson’s correlation coefficients, TotalN = 20,573) Note E = Extraversion, A = Agreeableness, C = Conscientiousness, N = Neuroticism, O = Openness, PA = Physical Activity (raw value) Number of clusters First, we evaluated the number of clusters in an unsupervised manner (exclusive of the outcome variable to be predicted) for the clustering methods, including K-means and GMM. The Gap statistic was calculated to determine the optimal number of clusters (K) for K-means clustering (Fig. 1). Primarily, we considered Variable Set 1 (Big-Five personality only) and Variable Set 4 (full variable set), which yielded the best predictive performance in the dimensional approach, as shown later. A higher Gap statistic value indicates better clustering. For Variable Set 1, K = 4, and for Variable Set 4, K = 5 were deemed optimal, as they had the highest values and were also higher than the mean + standard error (SE) of the number of clusters K -1 [39]. For GMM, the BIC indicates that the optimal number of clusters is between 5 and 12, as these values minimize the BIC. Evaluation of number of clusters based on a statistical metric. (A) Evaluation of K-means by Gap statistic. A higher value indicates better clustering. Error bars indicate standard errors. (B) Evaluation of number of clusters by BIC. A lower value indicates better clustering Cluster profile Figure 2A and B show the cluster centers for each cluster of K-means, where the number of clusters was four (Table 1; sample sizes for each cluster are reported in Supporting Information D). Clusters were ordered by increasing the mean values of extraversion. Panel A shows the results for the cluster solution with only the Big-Five (Variable Set 1), and Panel B shows the results for the cluster solution with full variable sets (Variable Set 4). The solid line represents the results using the entire dataset, while the squares (‘□’) and triangles (‘△’) denote the results using only the training data and test data, respectively. The close alignment of the cluster centers between the test and training data indicates that the results obtained from K-means are stable. Regarding the Big-Five scores, Panel B, with the addition of promotion focus, prevention focus, and self-control, did not change significantly from Panel A. Cluster 1 has higher Neuroticism and higher prevention focus, corresponding to the “overcontrol” type in the Asendorpf-Robins-Caspi (ARC) typology [20, 21]; Cluster 4 is the opposite, with lower neuroticism and lower prevention focus, corresponding to the “resilient” type in the ARC typology. Cluster 3 can be described as the reserved type, characterized by high agreeableness and conscientiousness. Cluster 2 is a mirror image of Cluster 3. Figure 2C and D show the cluster centers for the GMM with four clusters. In contrast to K-means, results using the test data, training data, and the entire dataset differ significantly, indicating that the GMM lacks stability. Therefore, these clusters cannot be considered interpretable profiles for the data. Cluster centers for the four cluster solutions. (A) Cluster solution with K-means using only Big-Five (Variable Set 1). (B) Cluster solution with K-means with full variable sets (Variable Set 4). Error bars indicate standard error of the mean. (C) Cluster solution with GMM using only Big-Five (Variable Set 1). (D) Cluster solution with GMM with full variable sets (Variable Set 4). The solid lines represent clustering of results using the entire data, while symbols indicate the results of clustering using training data and test data separately Figure 3A and B show the distribution of clusters obtained by plotting the first and second principal components of the PCA using the five Big-Five variables employed in the clustering process. The four cluster solutions obtained from the Big-Five alone are represented by distinct colors. The variance explained by these two principal components was 64%. In the K-means clustering (Fig. 3A), the data were divided into three segments along the first principal component, and the middle segment was further divided in the direction of the second principal component. This is a typical tendency when K-means clustering is applied to data with a unimodal distribution, where the cluster structure is not distinct. In contrast, the GMM does not correspond directly with the principal components (Fig. 3B). Distribution of clusters on the principal component scores obtained by PCA for Variable Set 1 (Big-Five only). (A) K-means, (B) GMM, (C) Principal component loadings, and (D) Distribution of PC scores. PC = Principal Component The first principal component had positive loadings on Extraversion, Agreeableness, Conscientiousness, and Openness and negative loadings on Neuroticism (Fig. 4C). The second principal component had a positive loading on Agreeableness. Cluster 1 in K-means was the group with a low PC 1 score (i.e., high neuroticism and low others), whereas Cluster 4 was the opposite. Cluster 2 had a high PC 2 score (i.e., high agreeableness and low extraversion and openness), and Cluster 3 had the opposite. Evaluation of predictive accuracy for a different number of clusters. R2 for total PA of test data when using Variable Set 1 (A) and Variable Set 4 (B) are plotted. Error bars indicate 95% confidence interval obtained by the bootstrap samples (both test data and training data). The solid horizontal line indicates the R2 values of the dimensional approach model presented in Table 1 (dashed lines represent 95% CI). The colored lines represent the R2 values of the clustering-approach model (blue for K-means and orange for GMM). Error bars indicate 95% CI Based on the distribution of the PC scores, no clear cluster boundaries with decreasing densities were observed (Fig. 4D), indicating that these clusters may not represent typical “types” within the population. The clusters obtained by K-means may be better interpreted as “segments” divided along the principal component axes. Comparing prediction performance between clustering-approach and dimensional-approach First, we examined the extent to which a dimensional approach model, specifically, a linear regression model, can explain total PA based on the test data. In this model, we included scores from the Big-Five personality and other psychological characteristics, such as regulatory focus and self-control, as the explanatory variables. These variables were directly included in the model as continuous variables without any categorization. Table 3 shows the goodness-of-fit metrics (R², AIC, and BIC) for Models 1–5, using the five predictor sets. Model 4 with all the variables had the highest R2 values for both training and test data (0.054, 0.051, respectively). Thus, Model 4 explained 5.4% and 5.1% of the variation in PA in the training and test data, respectively. Model 4 had the lowest AIC and BIC values, indicating that it is the best model. Table 3Multiple regressions predicting total physical activity (N = 10287 for training data, N = 10286 for test data)Variable SetR2-trainR2-testAICBICModel 1Big-Five0.0450.04547569.1147619.78Model 2Big-Five + Focus0.0560.05047458.0647523.21Model 3Big-Five + Self-control0.0470.04547555.5147613.42Model 4 Big-Five + Focus + Self-control 0.057 0.051 47445.1 47517.5 Model 5Focus + Self-control0.0420.03347598.347634.49Note R2-train = R-squared for training data; R2-test = R-squared for test data; AIC = Akaike’s Information Criterion, BIC = Bayesian Information Criterion. The results for Model 4, determined as the best model, are highlighted in bold font Multiple regressions predicting total physical activity (N = 10287 for training data, N = 10286 for test data) Note R2-train = R-squared for training data; R2-test = R-squared for test data; AIC = Akaike’s Information Criterion, BIC = Bayesian Information Criterion. The results for Model 4, determined as the best model, are highlighted in bold font Table 4 presents the estimation results for the dimensional models. In Model 1, which includes only the Big-Five scores as predictors, the four Big-Five variables (Extraversion, Agreeableness, Conscientiousness, and Openness) showed a positive impact on total PA (s > 0.1), except for Neuroticism ( = -0.01, 95%CI = [-0.06, 0.05]). These results indicate that despite the correlations among the Big-Five variables (Table 2), the four Big-Five variables contributed independently toward explaining PA. In Model 4, promotion focus, and self-control had a positive effect ( = 0.29, 95%CI = [0.23,0.35]; = 0.13, 95%CI = [0.06,0.19], respectively), whereas the impact of conscientiousness was eliminated ( = 0.01; 95%CI = [-0.05,0.08]). Table 4Estimates of regression models (total N = 10287)Model IDPredictorBetaCItp-valueModel 1E0.26[0.21,0.31]9.67< 0.001A0.31[0.26,0.36]11.88< 0.001C0.12[0.07,0.18]4.33< 0.001N-0.01[-0.06,0.05]-0.330.74O0.11[0.06,0.16]4.19< 0.001Model 2E0.23[0.18,0.28]8.36< 0.001A0.28[0.23,0.33]10.7< 0.001C0.07[0.02,0.13]2.570.01N-0.01[-0.07,0.05]-0.250.8O0.02[-0.03,0.08]0.790.43Promotion focus0.3[0.24,0.35]9.76< 0.001Prevention focus-0.01[-0.07,0.05]-0.330.74Model 3E0.26[0.2,0.31]9.53< 0.001A0.29[0.24,0.34]10.88< 0.001C0.06[-0.01,0.12]1.80.07N0.02[-0.04,0.07]0.580.56O0.11[0.06,0.17]4.23< 0.001Self-control0.13[0.07,0.19]3.95< 0.001Model 4E0.23[0.17,0.28]8.34< 0.001A0.26[0.21,0.31]9.75< 0.001C0.01[-0.05,0.08]0.360.72N0.01[-0.05,0.07]0.350.73O0.03[-0.03,0.08]0.90.37Promotion focus0.29[0.23,0.35]9.47< 0.001Prevention focus0.01[-0.05,0.07]0.210.83Self-control0.13[0.06,0.19]3.87< 0.001Note E = Extraversion, A = Agreeableness, C = Conscientiousness, N = Neuroticism, O = Openness, PA = Physical activity Estimates of regression models (total N = 10287) Note E = Extraversion, A = Agreeableness, C = Conscientiousness, N = Neuroticism, O = Openness, PA = Physical activity Figure 4 shows the R2 values of the clustering approach model for predicting total PA (for test data), where dummy-coded cluster labels obtained with Variable Sets 1 (Panel A) and 4 (Panel B) were used as the predictors. The solid blue lines indicate the K-means results, and the dashed orange lines denote GMM results. The solid horizontal black line indicates the R2 values of the dimensional approach model presented in Table 1 (dashed lines represent 95% CI). For K-means using only the Big-Five (Variable Set 1, Fig. 4A), R2 was approximately 0.03 for K = 4, which is approximately 70% of the dimensional-approach model. When the number of clusters exceeds nine, R2 exceeds 0.04 and approaches 0.045, which is the R2 of the dimensional-approach model. In Variable Set 4 (Panel B), which has the full set of variables, the R2 values are slightly higher than those of Variable Set 1; however, the general trends remain consistent. When the number of clusters was K = 4, R² was approximately 60% of that in the-dimensional model. Comparing the R2 values of the predictive models using K-means and GMM cluster labels, K-means outperformed the GMM in most cluster numbers (Fig. 4). This is consistent with the results of a previous study comparing K-means and GMM methods on different datasets [42]. Ref [42] argues that when the predicted variable is a linear function of the explanatory variables, K-means provides a segmentation that leads to the best prediction—consistent with our results. Supporting Information A shows the results for predictive performance for calculating a binary outcome indicating if the participants satisfied the physical activity criterion of 23 METs-h/week. For this case, the prediction accuracy of K-means (area under the curve [AUC] for test data) was approximately 70–80% of that achieved by the dimensional approach when the number of clusters was four (Fig. S1). No further increase was observed in the prediction accuracy when the number of clusters exceeded seven or eight. While we have focused on the total PA as the primary outcome variable, the PA measured using the IPAQ-PA consisted of walking, moderate activity, and vigorous activity. In Supporting Information B, we report the predictive accuracy of separate PA activities. Discussion Clustering of individuals based on personality and other characteristics is useful in marketing, business, and basic scientific research, such as medicine and psychology. Although grouping individuals results in loss of information, the extent to which this occurs has not been sufficiently examined. In this study, we evaluated the prediction of PA through clustering based on personality and other psychometric characteristics by comparing it with a dimensional-approach model (i.e., linear regression). Several studies in psychology and medicine have indicated that loss of information due to dichotomization or categorization [14, 43, 44]. However, most of these studies focused on loss of statistical power or introduction of bias, and to our knowledge, no studies have yet quantitatively discussed the decrease in prediction accuracy due to the discretization of predictors using actual data. In addition, we are not aware of any studies that examined the impact of the number of clusters or clustering methods on predictive accuracies. Indeed, this study has addressed these issues. The main finding of this study is that for a commonly used number of personality clusters (profiles: approximately three to five), the accuracy of clustering-based predictions is significantly lower than that of dimensional-approach models. To achieve a comparable prediction accuracy of the dimensional-approach model, the number of clusters should be more than 8. This implies that, even if the number of clusters is optimized within the unsupervised learning framework, it does not necessarily result in the best clustering method in terms of prediction. Determination of the number of clusters depends on the objective. Clustering can be used to understand and interpret the profile of each cluster, which in turn can be used to design interventions. Therefore, the tradeoff between predictive accuracy and interpretability should be considered. Individuals who base their decisions on clusters should be aware that they sacrifice the predictive power or rich information inherent to continuous variables. It is challenging to tailor an intervention while preserving the continuum of personality traits and other individual differences, including the use of personality scores (without clustering), to predict the preference for or suitability of health messages, for which a predictive model has been developed as a knowledge base [6]. Intervention components do not have to be delivered in a deterministic manner, and a predictive model can be used to adjust the probability or mixture rate of multiple intervention components. In this study, the dimensional approach models with personality traits as predictor variables explained a maximum of 5% of the variation in PA, suggesting inadequate predictive performance of the model. However, this poor performance was not surprising given that personality scores are known to have weak correlations with PA (r = .1; i.e., R2 = 0.01) or even lower [45]. The prediction performance could have improved if health-related or PA-specific trait variables were included in the analysis, as the trait measures assessed only general tendencies [46]. Indeed, a model using other PA-specific traits (e.g., process of change) in our dataset explained 19% of the variance in PA [31]. Future research can consider examining the extent of information loss when clustering is based on variables with high predictive power. In this study, the distribution of trait scores did not show a clear cluster structure. Many clustering methods, such as K-means and GMM, yield clustering results that segment the feature (trait score) space, even for unimodally distributed features with no clear cluster structure [42, 47]. Moreover, nonnormality in the distribution leads to the selection of solutions with multiple clusters, even in a unimodal case [47]. It is conceivable that this was also the case in the data obtained in this study. Although it may still yield a useful segmentation for prediction, the user should be aware that the boundaries of that segmentation are not naturally derived from the distribution of features (i.e., variables) but are determined by the nature of the clustering methods. As a future work, it remains necessary to examine the extent of information loss resulting from profiling in data where clear cluster structures exist for self-reported personality and psychological variables, although we are not aware of such data. While applying clustering methods to data where distinct clusters are not apparent and there is a “diffuse cluster structure,” the K-means method tends to divide the data uniformly. By contrast, mixture models and other methods that assume generative processes are heavily influenced by the detailed properties of the clustering method. As demonstrated in this study using the GMM, these models can exhibit instability. Therefore, results obtained using such methods should be interpreted with caution. From a research-practice perspective, in cases where no clear cluster structure exists, individuals near the boundaries of the segments are more susceptible to slight variations caused by differences in clustering methods, initial values, or sample variations, and this could lead to suboptimal interventions. In such cases, it may be beneficial to consider more detailed data so that experts can manually determine the intervention. As such, when personalizing interventions based on clustering, it is essential to carefully consider the results based on original data distribution instead of relying solely on the profile from the clustering results. This study has several limitations. The first is the use of a self-reported measure of PA, which may have been biased because of traits that may be related to personality. Moreover, it may be necessary to conduct a similar investigation using an objective measure of PA, such as a pedometer (cf. Ref [48] that provides evidence highlighting the differential relationship between the Big-Five personality traits and PA, based on the method of measurement). Second, the unsupervised clustering methods used in the clustering approach model were limited to K-means and GMM. This study recommends the use of K-means for PA prediction, as it has a better prediction accuracy than the GMM. However, we did not discuss other commonly used methods, such as hierarchical cluster analysis or latent profile analysis. We do not expect these methods to provide better methods in terms of prediction than K-means; however, discussing these methods will be a future research topic. Finally, this study focused on the problem of predicting PA. In many cases, predicting responsiveness to intervention rather than predicting PA or other outcomes is required for practical applications. Predicting responses to an intervention is a more complex problem, and further research is required to determine the generalizability of the findings to a broader context. Nevertheless, the study framework and findings may be useful for addressing this challenge. Recommendations for researchers Based on these results, we discuss how researchers can use clustering to personalize interventions and make predictions. First, as discussed in this study, when the target outcome variable is available in the data, it is advisable to check the prediction accuracy when determining the number of clusters and clustering methods. Although there are several metrics for determining the number of clusters in an unsupervised manner, such as Gap statistic and BIC, the number of clusters determined by these metrics is not always optimal for prediction purposes, as indicated by our results. Using test data, which are separated form training data used for clustering, to assess the predictive accuracy is useful for mitigating the impact of overfitting. Additionally, it is also beneficial to examine data distribution using a dimensionality reduction technique, such as PCA, to determine if clear cluster boundaries exist. If clear boundaries are observed, groups on opposite sides of the boundaries may have qualitatively distinctive characteristics and treating them as distinct groups may be effective (e.g., in such cases, distinct mechanisms may exist, such as genetic mutations or distinct environmental factors). In such cases, a clustering method with a mixture distribution model, such as GMM, may be appropriate because it can effectively capture the cluster structure of the data. Conversely, when variables are considered to follow a continuous, unimodal distribution, we recommend K-means as the first choice of clustering method because it often provides stable results and a good empirical performance, as shown in this study. There is a trade-off between prediction accuracy and interpretability: the more clusters there are, the better the prediction, but the more difficult it is to interpret the profile. Researchers and practitioners should determine the number of clusters, considering how many clusters they are prepared to accept and how much loss of predictive accuracy they are willing to accept. In this study, linear and logistic regression models, with only the main effect of each score as a predictor, were used as benchmarks for prediction accuracy. However, in practice, there may be cases wherein there is an interaction between the predictor variables. For example, cases have been reported in which a combination of two or more personality factors can alter the prediction. Specifically, neuroticism is thought to be associated with negative health outcomes, but when combined with high conscientiousness, it predicts a healthy behavior [49]. In such cases, the clustering approach may perform better than dimensional models using only the main effects. If such results are obtained or assumed a priori, it may be useful to use machine learning methods that can also manage interactions and nonlinearities between variables, such as the support vector regression and random forests, as a benchmark for comparison. Conclusions This study examined the performance of profiling based on clustering methods in the context of explaining physical activity through personality. The results indicated that profiles based on the number of clusters selected by conventional statistical criteria had approximately 60–70% of the predictive performance compared to linear regression using continuous predictors. These findings underscore the need for caution when profiling using unsupervised statistical clustering methods. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Abbreviations Extraversion Agreeableness Conscientiousness Neuroticism Openness Openness Physical Activity Body Mass Index International Physical Activity Questionnaire Moderate to Vigorous Physical Activity Gaussian Mixture Model Principal Component Analysis Author contributions All authors conceived and designed the study. K.Katahira, T.O., and K.T. performed the survey. K.Katahira and T.O. analyzed the data and interpreted the results. K.Katahira wrote the manuscript, and K.T. revised the manuscript. All authors reviewed the manuscript. Funding This work was partially supported by JSPS KAKENHI Grant Number: JP20H00625 (to K.Katahira). Data availability Data cannot be shared publicly because consent for data sharing was not obtained from each participant. Data is available from the corresponding author (k.katahira@aist.go.jp) for researchers who meet the criteria for access to confidential data, which is granted by the ethics committee of the National Institute of Advanced Industrial Science and Technology. Declarations Ethics approval and consent to participate The study protocol was approved by the Ethics Committee of the National Institute of Advanced Industrial Science and Technology (Approval ID: 2022 − 1279). All the participants provided informed consent. The study was conducted using an online text-based survey that required participants to have basic literacy skills. Therefore, illiterate participants were not included in the study. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement Data cannot be shared publicly because consent for data sharing was not obtained from each participant. Data is available from the corresponding author (k.katahira@aist.go.jp) for researchers who meet the criteria for access to confidential data, which is granted by the ethics committee of the National Institute of Advanced Industrial Science and Technology.