1. Introduction
With more than 1000 fatalities among construction workers every year in the United States alone [
1] and nearly
$6 billion costs in lost production, lost income, and pain and suffering [
2], the safety performance of construction workers demands improvement. Improving construction safety is challenging since workers have to execute their tasks under both physically and psychologically demanding conditions [
3,
4] to meet time and budget constraints in a project. Such working conditions can create inattentiveness [
5], anxiety [
6], or stress [
7,
8,
9], which contribute to construction workers’ unsafe behaviors [
9]. Therefore, safety managers seek practices to enhance worker focus, reduce stress, promote caution, and hone workers’ abilities to identify, acknowledge, and respond to uncertainties in the workplace, ultimately reducing human errors leading to accidents [
10].
One technique to lower stress and anxiety, enhance focus, and improve attentional performance is to implement mindfulness practices. In the last decade, an explosion of interest in mindfulness research on the concept and application of mindfulness has taken place, and mindfulness journal publications increased from fewer than 35 articles in the early 2000s to 1203 publications in 2019 (
goAMRA.org (accessed on 8 March 2021)). Originating in Buddhist philosophy and meditation practice [
11], mindfulness has been shown to be effective in treating mental and behavioral health issues [
12,
13]. Mindfulness as a trait is particularly related to attention and awareness [
14], which are essential factors in workplace safety [
15]. Because being mindful can help reduce workplace accidents and injuries, it is important to identify which individual differences, such as personality, influence the state of mindfulness of construction workers. This approach could potentially screen out accident-prone employees [
16,
17] who may need additional support or training to prevent accidents.
Considering the important role of mindfulness in curbing frequency and severity of incidents, researchers are interested in understanding the relationship between mindfulness and other common, more static human factors, such as personality [
18,
19]. Although recent research on personality shows that certain personality traits are highly related to safety and to the attentional failures that may lead to unsafe behaviors, e.g., Hasanzadeh et al. [
17], and little is known about other personal factors that may impact mindfulness. More recent studies even have suggested that the relationship between personal characteristics and safety performance may be mediated by failures in cognitive processes, such as poor selective attention or distractibility [
17]. Considering that mindfulness is particularly related to attention and awareness [
14], understanding the relationship between mindfulness and other common, more static human factors, such as personality will increase our knowledge regarding the mediating role of attention in safety performance. For instance, to what extent can a variable such as national culture makes certain employees more risk-, injury-, and accident-prone?
To address this knowledge gap, this study examines the relationship between mindfulness and variables such as personality and national culture in the context of construction safety. The results of this study offer solutions for reducing accidents in the construction industry by providing additional factors that can be used as a predictive index. Furthermore, the results may be used as inputs to design better safety training programs to enhance worker safety on job sites, which, in turn, will conceivably lead to better safety performance.
3. Materials and Methods
In order to examine the influence of the independent (i.e., personality and national culture) variables on the dependent (i.e., mindfulness) variable, this study used a regression analysis. In this section, the authors detail the study’s data collection instruments, participants characteristics, data analysis approach, and validation.
3.1. Data Collection Instruments
3.1.1. Dependent Variable
To measure mindfulness, this study used the Mindfulness Attention Awareness Scale (MAAS). Developed by Brown and Ryan [
14], and regarded as one of the most established techniques for measuring mindfulness [
86], the MAAS applies a Likert scale ranging from 1 to 6 (almost always to almost never), and participants must respond to 15-item statements in the questionnaire. The average of all statements was reported as the participant’s mindfulness score, with higher scores indicating a higher level of dispositional mindfulness and vice versa.
3.1.2. Independent Variables
To assess personality, this study used the Big Five Inventory (BFI) developed by [
50]. This 44-item questionnaire also uses a Likert scale ranging from 1 to 5 (strongly disagree to strongly agree). Like any Likert-based questionnaire, participants answer to which extent they agree or disagree with each statement. Total scores are then calculated by adding the direct and reverse score Likert value to each of the personality types, as specified by [
50].
To compute the national culture aspect of participants, this study utilized a questionnaire adapted from [
54]. National culture has four dimensions:
power distance (five items),
individualism vs.
collectivism (six items),
uncertainty avoidance (five items),
masculinity vs.
femininity (four items). National culture was also computed based on a Likert scale like personality. The scale ranges from 1 to 5 (strongly disagree to strongly agree) for national culture.
3.2. Participants
A total of 156 participants (30 construction workers and 126 students) aged 18–62 years (mean = 26.81, standard deviation = 9.55) were recruited to provide the data; however, one data point was removed from the student sample because it was incomplete. Therefore, a total of 155 participants (30 construction workers and 125 students) were considered for analysis. The student respondents were from the department of civil engineering at George Mason University. Since previous studies have shown that experience plays an important role in determining safety performance of people involved in construction activities [
87,
88], student sample data were divided into two groups: students with experience and novice students. The students with experience sample were between 21 and 40 years old (mean = 24.42, standard deviation = 4.46), novice student sample were between 18 and 40 years old (mean = 23.44, standard deviation = 4.33), and the construction workers’ sample were between the ages of 19 and 62 years old (mean = 39.69, standard deviation = 13.89). Participants were recruited through on-campus fliers, posting an invitation flyer at construction sites, and stopping by construction companies’ main offices. All participants provided written informed consent, and workers were given
$15 gift cards, whereas students received classroom credit points as compensation after finishing the questionnaires. All procedures were approved by the Institutional Review Board (IRB) of George Mason University.
3.3. Data Analysis Approach
In order to analyze data that contains multiple independent variables, generalized regression approaches are appropriate to select which variables have the most significant effect on the response variable. Before conducting any regression analysis, the assumptions for regression should be tested. Therefore, the research team first tested the distribution of the data and the existence of potential outliers. The assumptions for regression were checked for each dataset separately (i.e., students with experience, novice students, and workers). Plotting the distribution of the mindfulness score for construction workers, it was found that the data were normally distribution with no outliers (Shapiro–Wilk: p-value 0.73 > 0.05). For novice student data, one data point was an outlier (same data point as the outlier) and removed from data in the analysis. By removing the outlier, the distribution of the dependent variable changed from non-normal (Shapiro–Wilk: p-value 0.01 < 0.05) to normal (Shapiro–Wilk: p-value 0.07 > 0.05). The mindfulness score of students with experience was normally distributed with no outliers (Shapiro–Wilk: p-value 0.09 > 0.05).
Then, to confirm that the relationship between the independent variables and the response is linear, the research team evaluated the scatter plot of each independent variable with the response variable. There was no pattern to show violation of linearity, such as curvilinear or cubic, on the dataset. To address the assumption regarding the variance of the residuals of the response variable, the authors plotted the residuals versus predicted values and observed no pattern for the datasets. All observations proved to be independent of each other, and finally, the distribution of the residuals of the dependent variable were normally distributed for students with experience (Shapiro–Wilk p = 0.97) as well as construction workers (Shapiro–Wilk p = 0.17). However, the distribution of the residuals of the dependent variable were not normally distributed for novice students (Shapiro–Wilk p = 0.003). Plotting the distribution of the residual for novice students revealed that one data point was an outlier and removed from the analysis. After removing this data point, the distribution of the residuals of the dependent variable changed from non-normal (Shapiro–Wilk: p-value 0.003 < 0.05) to normal (Shapiro–Wilk: p-value 0.117 > 0.05).
In any regression approach, the first starting point to do any type of regression is linear regression. Simple linear regression, also known as ordinary least squares (OLS), attempts to minimize the sum of error squared. Even though this regression method is key to understanding the nature of the regression model, it usually oversimplifies the analysis and thus important variables may not be selected as they should be. The other disadvantage of linear regression is that it is prone to multicollinearity, which means that if there is high correlation between the predictors, removal of one or more variables from the analysis may be necessary, which is problematic because those variables may be important predictors for the response variable at hand. Hence, a more refined type of approach is needed to better select variables while also simplifying the model by retaining the significant variables and removing the variables that do not contribute to the prediction of the response variable.
3.4. Penalization Methods
Traditional regression methods such as stepwise, forward, and backward selection suffer from high variability and low prediction accuracy, especially when there is a correlation between variables or multiple predictors [
89,
90]. In response to these shortcomings, using penalized regression methods have gained traction among researchers due to their higher prediction accuracy and computational efficiency [
91]. Using penalized estimates in a regression model, the user accepts some bias in order to reduce variance. Similar to ordinary least squares (OLS) estimation, penalized regression methods estimate the regression coefficients of the predictors by minimizing the sum of squares of the residuals; however, in contrast to OLS methods, the penalized regression places a constraint or penalty, e.g., [
92] on the size of the regression coefficients, which causes the coefficient estimates to be biased. The introduction of the penalty improves the prediction capability of the model by decreasing the variance of the coefficient estimates.
Generalized regressions with no penalties are based on the least square estimation method, which is an unpenalized fit and provides no simplification (no variable selection) and no shrinkage of parameters. Alternatively, penalized regression selects variables by minimizing the sum of the squared residuals while also adding a penalty proportional to the size of the regression coefficients. If the size of the penalty on a specific parameter is large enough, it causes the regression coefficient to shrink towards zero. Hence, some variables will be removed from the analysis, which will simplify the final model by selecting fewer variables. In the same token, shrinkage is done by continuously shrinking the regression coefficients by introducing some degree of bias, e.g., [
92,
93,
94] in the coefficient estimates. More often, the introduction of bias tends to reduce variance, resulting in a model with a better prediction performance.
In this study, we utilized a type of generalized penalized regression called elastic net regression, which can select variables by introducing a penalty in the regression model. In many statistical models, the typical technique behind the penalized regression analysis is using an estimation method called maximum likelihood. This estimation method delivers the best fit based on the observed data. By applying a penalized likelihood instead, better prediction on the response variable can be achieved.
The purpose of introducing a penalty is to achieve two main purposes. The first purpose is to allow the model to perform variable selection by removing unimportant predictors, and the second purpose is to apply shrinkage of estimation parameters. By optimizing the penalized likelihood, the regression model is simplified (fewer predictors), overfitting (weak prediction performance) can be avoided, and issues that arise from multicollinearity (high correlation between predictors) can be resolved [
95]. In other words, by applying penalized estimates in regression, some degree of bias is accepted in order to reduce variance. The penalization regression methods and corresponding penalties are shown in
Table 2.
In ridge regression, the coefficients on the predictors are shrunk by imposing a penalty (i.e., βj2)—also written as L2 penalty—such that the ridge coefficients minimize a penalized sum of residual squares. The disadvantage of the ridge regression is that it shrinks the coefficients to non-zero values to prevent overfitting and keeps all the variables. Hence, this approach is not a viable option to reduce variable selection.
As with ridge regression, Lasso regression has a shrinkage approach but with a subtle difference: The L2-penalty is replaced by L1-penalty (i.e., |βj|). This method shrinks the less important variable coefficients to zero and therefore can remove the variables that are deemed not significant predictors for the response variable. Even though Lasso both shrinks and selects by removing variables, studies have shown that Lasso tends to yield a model that is more parsimonious (less complex) than the elastic net approach. The other shortcoming of this method is that in the case of collinearity, the Lasso model selects the variable with the strongest correlation with the response variable and drops the other variables from the model.
In this study, we selected to use the elastic net approach for our data analysis. This method was proposed by [
92] and utilizes an algorithm called LARS-EN, which was adapted from LARS for Lasso [
96]. This method also uses both ridge and Lasso regression penalties and combines the techniques from the two methods by learning from their limitations to improve on the regularization of the model. Elastic net regression can be written as follows:
where
= (x
1, x
2, …. x
p) are input variables,
is the alpha parameter, λ complexity parameter and y is the response variable. As can be seen from the above equation, when alpha is zero, the regression equation becomes ridge regression, and when alpha is one, the equation becomes Lasso.
The general mechanics of the elastic net is executed in two steps. First, the algorithm finds the ridge regression coefficient and on the second step uses a Lasso-sort of shrinkage of the coefficients. To eliminate the limitations found in Lasso, the elastic net includes a quadratic section of the penalty, which increases variable selection. It is worth noting that the quadratic section of the penalty (i.e.,
β
j2) when used in isolation (α = 0), becomes ridge regression. The other advantage of elastic net regression is grouping. If there is a very high correlation among independent variables, then this method performs well in incorporating variables into the model that aids in better prediction accuracy, unlike the Lasso approach, which tends to select only one variable from the highly correlated independent variables. In addition, simulation studies done on real-world data have shown that the elastic net approach often performs better than Lasso [
92].
3.5. Validation
Cross-validation technique utilizes different samples of data to increase the overall accuracy of the predictive model [
97]. In this study, k-fold cross-validation is used as it is the most widely used method for estimating prediction error [
94]. The choice of k is usually 5 or 10, but there is no hard and fast rule [
98]. For a relatively small dataset (as the case here), k = 10 is chosen and therefore used in the place of k, yielding the 10-fold cross-validation. In 10-fold cross-validation, training data are randomly broken into 10 groups or folds of approximately equal sizes. The first part or fold is used for the validation set and the rest of the data are fit for the remaining folds. This is repeated 10 times with a different part used for error estimation.
4. Results
The descriptive statistics of variables studied are presented in
Table 3. On average, workers have higher mindfulness, power distance, and conscientiousness scores. Average scores for extraversion and openness are almost the same for students with experience, novice students, and workers data sets. Looking at the standard deviation (SD) columns, both the student samples have a much higher neurotic SD than workers’ sample. The workers’ SD is higher for conscientiousness as compared to the students with experience and novice students’ sample.
The correlation matrix between all variables considered in the analysis for the students with experience data (ES), novice students (NS), and for the construction workers’ data (W) appears in
Table 4. There were no issues with multicollinearity (i.e., correlation greater than 0.7) between independent variables (see
Table 4). In addition, the Cronbach alpha of personality, national culture and mindfulness are computed, and all the items are greater than 0.7 (see
Table 5). To test whether the common method bias exists in the collected data, the research team also conducted Harman’s single factor test. Since the cumulative percent of variance was 20.85% (less than 50%), the impact of common method bias was not a substantial threat.
As described above, the research team applied the adaptive elastic net method of estimation to select the independent variables that are significant predictors of mindfulness score. The adaptive estimation utilizes a modified version of the L
1 penalty via weights generated from the maximum likelihood estimates of the predictors to improve the overall fit of the model. K-fold (10-fold) cross-validation compared the training set with the testing set of the data. As the alpha value can be a value within zero and one (α ∈ (0, 1)), the model can move between ridge (α = 0) and Lasso regression (α =1). To select an α-alpha value to be used in the elastic net regression, a set of values (0.10, 0.25, 0.5, 0.75, and 0.90) were compared using a standard error value and the best alpha value was selected for fine tuning.
Table 6 shows the different alpha values and the corresponding errors. Alpha values of 0.5, 0.25, and 0.75 were chosen to be used in the elastic regression with the least error for students with experience, novice students, and workers datasets, respectively.
The results of the adaptive elastic net regression with 10-fold cross-validation appear in
Table 7. As one can see, the variance inflation factor (VIF) values are less than five, assuring no collinearity between independent variables. On the estimate column, the elastic net regression model equates to zero for some of the independent variables. This means the model has removed the variables from the analysis by shrinking the coefficient all the way to zero and selected fewer variables that explain the response variable. The negative sign of the estimates depicts an opposite relation between the predictors and the response variable. Higher
agreeableness scores positively correlate with mindfulness among novice students, and higher
neuroticism scores negatively correlate with mindfulness for both student samples. For workers, the significant predictors are
uncertainty avoidance and
conscientiousness, which all show significant positive correlation with mindfulness score.
The Wald test (also called the Wald Chi-Squared Test) is a test that examines whether explanatory variables in a model are significant. For variable estimates that are non-zero, the Wald test is calculated for significance, and the corresponding p-value is computed. The lower 95% and upper 95% show confidence interval (CI) ranges at 0.05 significance. CIs less than zero signify a negative relationship of predictors with the response variable, and positive relationships appear in CI ranges greater than zero. The corresponding p-values for Wald statistic for the estimates of each variable are used to select the predicator variables if the p-values are found to be significant (i.e., p <0.05).
The variables included in the equation are significant predictors with negative coefficients, showing negative direction with the response variable (mindfulness) and vice versa with positive coefficients. The prediction expression for the dependent variable mindfulness for students with experience, novice students, and workers can be written as Equations (2) to (4), respectively:
Our data show that for construction workers, conscientiousness personality trait, and uncertainty avoidance of national culture dimensions all positively correlate with mindfulness. Uncertainty avoidance of the culture dimension (i.e., the largest estimator) is associated with minimizing taking risks, and individual that have high uncertainty avoidance scores are likely to give importance to have instructions that are detailed and closely follow instructions and procedures for standardized work. This signifies that by avoiding uncertainty, workers are mindful of not taking short cuts that undermine their safety. The higher conscientiousness signifies higher mindfulness score because mindfulness is associated with higher awareness of their surroundings. Therefore, these variables are important indicators of mindfulness among workers, and lower measure of these variables can imply lower mindfulness and consequently lower safety performance.
Within the student data, agreeableness positively affects mindfulness for novice students, while neuroticism was found to be negatively associated with mindfulness for both students with experience and novice students. A higher degree of agreeableness shows that individuals can go along with the people around them and are less combative in nature, which is related to non-judgmental attitude of mindfulness. Neuroticism trait of personality is associated with anxiety and stress, which negatively impact mindfulness of individuals. The results confirm the negative association of this personality trait with mindfulness, and higher value of neuroticism scores can be used as a predictor of lower mindfulness. With respect to cultural dimensions, for students with experience, higher individualism was associated with lower mindfulness. This could be because students with experience are much younger and possibly more individualistic than workers; however, in construction sites, workers become more risk averse due to the existence of hazards and care more about their co-worker’s safety.
6. Conclusions
High-risk organizations that operate in complex, high-hazard domains for extended periods of time have been using mindfulness to increase productivity and record better safety outcomes [
44]. A growing number of studies are showing the benefits of using mindfulness to enhance safety and increased work performance in the workplace and suggesting that incorporating mindfulness can reduce incident occurrence. Unfortunately, limited was known regarding the extent to which personal characteristics of workers might impact their mindfulness state.
This paper examines ways to predict mindfulness of individual using relatively static factors—namely, by measuring personality traits and national culture. Using elastic net regression, the results show that certain personality traits and national cultural dimensions are associated with individuals’ dispositional mindfulness and the extent of influence of these variables can vary according to the levels of working experience in the construction industry. By detecting workers with lower mindfulness, the results of this study will enable safety managers to develop a more mindful workforce at construction sites by providing targeted training programs for workers with lower mindfulness or assigning those less mindful workers to activities that do not require higher levels of attention.
There are some limitations related to this study that are worth mentioning. First, the sample size of the study is limited to participants in the Northern Virginia region. Although the participants in the study came from a diverse background (workers and students consisted of Hispanic, black, and white), more data should be collected from diverse population of workers across the United States to generalize the results. Second, as limited number of variables are considered in the model, future studies should be conducted to investigate the role of other personal characteristics on mindfulness of construction workers. Third, potential sources of common method bias such as item characteristics, item context, or measurement context may affect the results of the findings [
117], even though the variables passed the Harman’s single factor test (20.85% < 50%). Despite these limitations, this study contributes to the body of knowledge and have the potential to reduce cost of safety programs by enabling safety managers to provide personalized interventions for their workers according to their personal characteristics.