Self-determination theory

SDT postulates that human motivation can be categorized into three broad taxonomies: amotivation, extrinsic motivation, and intrinsic motivation [1, 15, 16]. Amotivation pertains to individuals with the absence of intention to behave and represents unmotivated self-regulation or non-regulation [1]. Extrinsic motivation refers to an instrumentally manipulated and goal-oriented incentive to behave and depends on outcomes that are separatable from the action [1]. There are four forms of extrinsic motivation: external regulation, introjected regulation, identified regulation, and integrated regulation, representing different degrees of self-determination or autonomy [16]. External regulation is the least autonomous and most controlling form of extrinsic motivation [16]. Introjected regulation is a partially internalized form of self-regulation. Individuals with this form of regulation tend to participate in ego-oriented activities, indicating this type is more controlling rather than autonomous. Identified regulation refers to a cognitive acceptance of the values of action to achieve a preferable outcome [16]. Integrated regulation refers to the process of assimilation of identified regulation such that engaging in the behavior is completely consistent with one’s sense of self [16]. It is the most autonomous form of regulation that is extrinsically motivated. The only innately motivated type of motivation is intrinsic motivation, which refers to a natural incentive to act because of the inner satisfaction derived from that behavior and is purely self-determined [16]. In other words, when individuals are intrinsically motivated, they autonomously and freely involve themselves in an activity because of their own interest and ongoing enjoyment of doing it. Although types of motivation are conceptually distinct, they are hypothesized to be correlated and fall on the same continuum with various degrees of self-determination ordered as: amotivation (least autonomous), external regulation, introjected regulation, identified regulation, integrated regulation, and intrinsic motivation (most autonomous [16]). While the classical taxonomy of motivational types is valid and broadly used, external regulation and introjected regulation are considered as controlled motivation, whereas identified regulation and integrated regulation, in conjunction with intrinsic motivation, are classified as autonomous motivation. As a result, a revised taxonomy of motivation has been suggested, emphasizing the distinction between amotivation, controlled motivation, and autonomous motivation [17].

The development of BREQ

BREQ was initially developed by Mullan and colleagues [11] to assess individuals’ levels of exercise motivation grounded in SDT. After removing the statistically problematic amotivation subscale, a four-factor model, consisting of external regulation, introjected regulation, identified regulation, and intrinsic motivation factors demonstrated an acceptable degree of model fit [11]. Although some validity evidence was observed for BREQ, the exclusion of the amotivation dimension remained incongruent with the taxonomy of motivation hypothesized in SDT.

To make up the dimensional gap between BREQ and SDT, BREQ-2 was developed [18], featuring the inclusion of the dimension of amotivation. An excellent model fit to the data was observed (χ²(125) = 136.49, p =.23, RMSEA = 0.02, CFI = 0.95, NNFI = 0.94, SRMR = 0.05) for a group of elderly adults. Internal consistency was also reported at acceptable levels. Several researchers have since translated the BREQ-2 into different languages and have focused on the examination of various validity evidence on this scale. The original BREQ-2 has demonstrated satisfactory construct validity evidence in English [19], Greek [20], Portuguese [21], and Chinese [22, 23]. However, other studies revealed the instability of BREQ-2 when applying it to distinct cultures [24–26]. After removing item 17, recategorizing item 8 and item 18 into identified regulation and intrinsic regulation respectively, the Spanish BREQ-2 exhibited an adequate fit to the data (χ²(125) = 2.15, CFI = 0.92, TLI = 0.90, RMSEA = 0.06, SRMR = 0.05) for a group of active exercisers [26]. After removing item 17, the hypothesized hierarchical model of the Portuguese BREQ-2 showed an acceptable level of model fit (χ²(125) = 221.70, p <.001, SRMR = 0.06, NNFI = 0.90, CFI = 0.92, RMSEA = 0.04) in which external regulation and introjected regulation were regressed on controlled motivation, whereas identified and intrinsic motivation were regressed on autonomous motivation [25]. After dropping item 17 and combining identified and intrinsic motivation into autonomous motivation, the modified 3-factor Chinese BREQ-2 (external, introjected, and autonomous motivation) demonstrated an excellent model fit to the data (χ²(74) = 117.57, CFI, GFI, NFI, and TLI >0.95, RMSEA = 0.05) among a group of middle and high school students [24].

Although BREQ-2 addressed continuum of behavior regulation from amotivation to intrinsic motivation and demonstrated evidence of validity in various settings, there remained a conceptual discrepancy between SDT and BREQ-2 in terms of the dimensions of self-regulation. SDT divides extrinsic motivation into four types: external regulation, introjected regulation, identified regulation, and integrated regulation [16]. In BREQ-2, the integrated dimension of behavior regulation was excluded due to its definitional similarity to intrinsic motivation [18]. While both represent a sense of volition, the integrated and intrinsic motivations differ in that the former is still influenced by external forces, whereas the latter is influenced solely by internal pleasure [16].

These theoretical discrepancies considered, a new BREQ instrument in English was developed which included the integrated regulation subscale but omitted the amotivation subscale [27]. The construct validity evidence of the five-factor model was confirmed by model fit indices indicating a good model fit (χ²(142) = 253.82, p <.01, CFI = 0.93; IFI = 0.93, RMSEA = 0.09). The new integrated regulation subscale demonstrated satisfactory concurrent and predictive validity evidence. González et al. [14] expanded on Wilson’s [27] and Markland and Tobin’s work [18] by developing Spanish BREQ-3, featuring the inclusion of both amotivation and integrated regulation subscales. The new Spanish BREQ-3 showed satisfactory validity and reliability evidence among adult exercisers, and measurement invariance was stablished across gender and age [14]. By using a similar manner, Cid et al. [13] developed a Portuguese version of BREQ-3, following the removal of one item from each subscale, the Portuguese BREQ-3 demonstrated satisfactory construct validity evidence (χ²(120) = 331.86, p <.001, SRMR = 0.06, NNFI = 0.91, CFI = 0.93, RMSEA = 0.06) with evidence of factorial invariance across gender. The English BREQ-3 is also available online for download [28]. To our best knowledge, no previous study has examined the validity evidence for English BREQ-3.

Zhong and Wang [9] developed a Chinese version of BREQ-3 by merging the integrated regulation scale [29] into BREQ-2, which demonstrated an acceptable model fit (χ²(194)= 565.46, p <.001, CFI = 0.90; TLI = 0.88; RMSEA = 0.06). However, the development of integrated regulation Subscale of the English version of BREQ-3 was adapted from the work of Pelletier [30] rather than Mclachlan and colleagues [29], resulting in an integrated regulation subscale in the Chinese version of the BREQ-3 that is not aligned with the English version. Specifically, the English version contains items: “I exercise because it is consistent with my life goals; I consider exercise part of my identity; I consider exercise a fundamental part of who I am; and I consider exercise consistent with my values”. The Chinese version includes items: “it is consistent with my values, goals and aims in life; it is essential to my identity and sense of self; it is genuinely part of me; and doing exercise and being myself are inseparable”. While the first three questions are fairly similar in meaning, the last question is phrased noticeably different. This discrepancy may lead to misunderstanding and reduced ability to make comparisons. Sperber et al. [31] argued that failing to maintain the meaning of the original items might result in conclusions that appear culturally distinct but are actually the result of instrument inequivalence. Hence, using congruent integrated regulation subscales for both English and Chinese BREQ-3 is necessary to ensure cross-cultural equivalence of BREQ-3.

Limitations of confirmatory factor analysis

Previous researchers have predominantly relied on confirmatory factor analysis (CFA) to examine the theoretically assumed dimensionality of BREQ-3 [13, 27]. Although this statistical approach has been widely used in the examination of the construct validity evidence for psychological instrument in various contexts, it has several limitations. CFA implicitly assumes that items loading on one latent factor have no effect on other factors (cross-loadings) and thus coerce them to 0. This assumption may lead to a biased parameter estimation [32]. In addition, the exclusion of cross-loading estimation in a CFA model may result in an overestimation of the correlation between latent factors as the only way to represent the association between indicators and other factors is to inflate the factor correlations [33]. To overcome these limitations, exploratory structural equation modeling (ESEM) was developed to model data based on CFA with advantages to estimate cross-loadings between indicators and latent factors, thus providing a more realistic parameter estimation [34]. Another improvement on the traditional CFA model is the use of bi-factor modeling approach. Bi-factor CFA (BCFA) assumes that the covariances among all indicators can be explained by a general latent factor and the remaining covariances can be further explained by specific latent factors [35]. A prominent advantage of BCFA is the independence of latent factors. In other words, all of the latent factors estimated from a BCFA model are assumed to be uncorrelated, thus making results more interpretable [36]. ESEM has also been integrated with bi-factor modeling into a broader statistical framework called bi-factor exploratory structural equation modeling (BESEM [37]), which comprises the advantages of both BCFA and ESEM. To align with SDT, the present study used two global factors to model controlled (external regulation and introjected regulation) and autonomous motivation (identified regulation, integrated regulation, and intrinsic motivation) and used six specific factors to model amotivation, external regulation, introjected regulation, identified regulation, integrated regulation, and intrinsic motivation. A previous study found redundant information shared by specific motivation factors in the Chinese BREQ-2 [24], therefore a bi-factor approach will provide evidence to determine if specific motivational types provide additional information in explaining the common variance shared among items beyond the common variance explained by the two global factors. It is recommended to compare different structural equation models as it is preferred to select a more parsimonious factor architecture rather than a more complex one when assessed models exhibited parallel model fit [33]. All four models are presented in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Graphical illustration of four alternative structural equation models.

cm = controlled motivation; am = autonomous motivation; amoti = amotivation; exter = external regulation; intro = introjected regulation; ident = identified regulation; integ = integrated regulation; intri = intrinsic motivation; CFA = confirmatory factor analysis; ESEM = exploratory structural equation modeling; BCFA = bi-factor confirmatory factor analysis; BESEM = bi-factor exploratory structural equation modeling.

https://doi.org/10.1371/journal.pone.0265004.g001

In light of the aforementioned limitations, this study aimed to investigate the psychometric properties of the Chinese BREQ-3. Specifically, this study aimed to: (a) translate the English BREQ-3 into Chinese; (b) examine the best representation of the factor configuration of Chinese BREQ-3 using CFA, BCFA, ESEM, and BESEM models; (c) examine measurement invariance for the best-fitted model across gender, and (d) test the concurrent validity evidence and reliability for the best-fitted model by correlating types of motivation with theoretically related psychological constructs, such as positive and negative affect [38] and basic psychological needs [20].

Participants

Undergraduate college students (N = 825) from several public universities in the mainland of China participated in this study. Student athletes were excluded to ensure the external validity of this study. The participants’ average age was 19.72 (±1.25). Respondents (n = 3, 0.36%) who did not report gender information were excluded from this study. Table 1 shows the background information for the participants.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Background information for the participants (N = 822).

https://doi.org/10.1371/journal.pone.0265004.t001

BREQ-3 translation to Chinese

A combined translation approach was used to translate the English version of BREQ-3 into simplified Chinese [40]. Three bilingual students, who have been studying in the United States for two to three years independently translated the instruments. The bilingual translators included two doctoral students majoring in Physical Education and an undergraduate student majoring in Sport Management. Each individually translated instrument was scrutinized by the other two bilingual translators. Any discrepancies between translated instruments were discussed in a meeting among the translators. This procedure continued until three translators agreed with each other and constructed a final version of translated instrument. The translated instrument was then back-translated by a bilingual Chinese professor who has been living in U.S. for over 15 years. The original and back-translated versions of the instrument were assessed by two monolingual English-speaking professors in sport and exercise psychology. If any differences were identified, the monolingual professors provided three forward-translators detailed explanations of the difference between two instruments. Following discussion, the three translators modified the wording of the problematic items and the back-translator translated the adapted items into English again. The process continued until the monolingual reviewers considered the two English versions were identical. The final product of this translation process, the Chinese BREQ-3, was used in this study.

Procedures

The study was approved by the Institutional Review Board (IRB) of the first author’s institution. Convenience sampling approach was used to recruit participants in this study. After obtaining permission to conduct this study, the researchers contacted professors in China willing to distribute the surveys through personal networks. After obtaining permission, the students taught by those professors were invited to participate in this study.

An implied consent form, Chinese translated version of BREQ-3, I-PANAS-SF, and BPNES were presented to the potential participants using Wen Juan Xing (TongdaoLiepin, Beijing), an online questionnaire management software, and the questionnaires were virtually sent to students via Wechat (Tencent, Shenzhen), an instant mobile messaging app. Students were asked to voluntarily fill out the questionnaires either at the end of a class or at their preferred time and location, depending on the professor’s preference. Completing the questionnaires took approximately 15 minutes. The finished questionnaires were automatically stored online in a password protected account and were only accessible to the primary researcher of this study. Data were downloaded to the primary researcher’s laptop encrypted with passwords for analysis purpose.

Data analysis

R (Version 4.0.2, R Core Team, www.r-project.org) was used for data preparation and Mplus [41] was used to analyze data. Descriptive statistics for the items of BREQ-3 was computed. The Shapiro-Wilk test and the Mardia estimate of multivariate Kurtosis were used to assess univariate and multivariate normality of the item scores respectively. Items with z scores higher than 3.29 or less than -3.29 were considered outliers [42]. Robust maximum likelihood estimator (MLR) was utilized to examine the fitting of the factor structure of Chinese BREQ-3 as it provided robust fit indices with the occurrence of non-normality and outliers and was suitable for Likert scale with five or more response categories [43].

In the CFA model, items were specified to regress on a priori factor with no cross loading permitted. All factors were allowed to correlate with each other. In the BCFA model, items were specified to regress on two general factors (except for amotivation) as well as on their specific factors with no cross loading permitted. All factors were not allowed to correlate with each other. In the ESEM model, each item was allowed to regress on every factor with oblique target rotation used to make cross-loadings as close to zero as possible. All factors were allowed to correlate with each other. In the BESEM model, each item was allowed to regress on every factor with oblique target rotation used to make cross-loadings as close to zero as possible. All factors were not allowed to correlate with each other.

The following goodness of fit indices were analyzed for each of the four models: chi-square (χ²), comparative fit index (CFI [44]), Tucker-Lewis index (TLI [45]), and root mean square error of approximation (RMSEA [46]). Akaike information criterion (AIC [47]), Bayesian information criterion (BIC [48]), and Akaike’s Bayesian information criterion (ABIC [49]) were used to assess whether model fit outperformed model complexity. The cut-off values for the indication of a good model fit for those indexes are: p >.05 for χ², CFI >0.95, TLI >0.95, RMSEA <0.06 [50], and p-close for RMSEA >.05 [51]. No rule of thumb was used to determine the adequacy of AIC, BIC, and ABIC values, but lower values generally suggest a better trade-off between model fit and model complexity [52]. A decrease in CFI less than 0.01 or an increase in RMSEA less than 0.015 between assessed models was used to test model difference [53, 54]. A multi-group model, depending which model has the best fit, was conducted to test the measurement invariance between men and women. The steps of invariance testing followed the suggestions from Putnick and Bornstein [52]. First, we fitted the best model separately to the men and women groups. We then proceeded to test for configural (no equality constraints), metric (constrain factor loadings), scalar invariance (constrain factor loadings and intercepts), and strict invariance (constrain factor loadings, intercepts, and residual variances) between groups. A decrease in CFI less than 0.01 or an increase in RMSEA less than 0.015 from less-constrained model to the more-constrained model was used as evidence of gender invariance [54]. Next, concurrent evidence of validity was assessed by regressing the scores obtained from the best model onto the scores from I-PANAS-SF and BPNES [52].

The composite reliability omega (ω_c) and hierarchical omega (ω_h) coefficients were calculated to assess the model-based reliability [55]. Composite omega is the ratio of the estimated true variance of the score of a measure to its total variance, thereby indicating internal consistency [56]. A composite omega value greater than 0.80 indicates that the internal consistency is satisfactory [57]. A hierarchical omega index measures the proportion of total variance that can be traced to individual differences in the general factor [56]. If the hierarchical omega is being applied to specific factors in a bi-factor model, it measures the proportion of total variance that can be traced to individual differences in the specific factors. Hierarchical omega value greater than 0.50 is indictive of satisfactory factor reliability [58]. If the hierarchical omega value is larger than 0.80, the total score should be considered unidimensional [59]. The explained common variance (ECV) was used to determine the degree to which the common variance among a set of items is explained by a general factor [60]. ECV greater than 0.85 should be regarded as essentially unidimensional. Item explained common variance (I-ECV) was used to assess the item common variance that is attributable to a general factor [60]. I-ECV greater than 0.85 indicates an item essentially reflects the general factor [60]. To determine the extent to which item correlations inform the general factor, the percentage of uncontaminated correlations (PUC) was computed. The greater the PUC, the more saturated the correlation matrix is with information useful for estimating the parameters for the general factor, and the less probable it is that the parameter estimations in a unidimensional model would be biased. When the PUC is less than 0.80, ECV is greater than 0.60, and the hierarchical omega is greater than 0.70, the existence of some multidimensionality is not severe enough to reject the instrument’s interpretation as largely unidimensional [58].

Data screening

Prior to analysis, data screening was conducted. No missing values were found. Outliers were identified within item 8 (19), item 14 (20), item 20 (20), and the amotivation subscale (12). We decided not to delete the outliers because they resulted from the ceiling and flooring effect of the responses rather than unexplainable errors [61]. The Shapiro-Wilk test indicated that none of the item scores met univariate normality (all ps <.05). The Mardia estimate of multivariate Kurtosis indicated deviation of the item scores from multivariate normality (p <.05). Potential biases of the non-normality and outliers were compensated by using robust estimation procedure [61].

Goodness of fit

Next, we evaluated the goodness of fit among the four competing models (Table 2). Given the large sample size, none of the models fit well based on the χ² statistics (p <.001). According to other goodness of fit indices, the CFA model fit the data poorly. The ESEM model showed almost excellent fit. When examining the bi-factor models, the BCFA model showed poor fit and the BESEM model demonstrated excellent fit. The decrease in AIC, BIC, and ABIC values indicate that the increase in model fit outperformed the increase in complexity from ESEM to BESEM. While initial inspection of the goodness of fit indices indicated strong evidence for selecting the BESEM model over other alternatives, further examination of the parameter estimates will assist in determining the optimal structure of the Chinese BREQ-3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Goodness of fit of alternative structural equation models.

https://doi.org/10.1371/journal.pone.0265004.t002

Parameter estimates

Table 3 shows the standardized factor loadings for the CFA and ESEM models. Each item was significantly loaded on their priori-specified factors in the CFA model. In the ESEM model, item 13, 19 under the identified regulation subscale, item 3 under the intrinsic motivation subscale, and all items under the integrated regulation subscale were not significantly loaded on their intended factors. Cross-loadings of those items suggested that they were associated with adjacent or even distant factors.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Standardized factor loadings for confirmatory factor analysis and exploratory structural equation models.

https://doi.org/10.1371/journal.pone.0265004.t003

Table 4 shows the standardized factor loadings for the BCFA and BESEM models. The factor loadings for the amotivation items were significant and high in both BCFA and BESEM model. While all the items from the external regulation and introjected regulation subscales significantly loaded on the controlled motivation factor in both the BCFA and BESEM models, all of these items also significantly loaded on their specific factors (except for item 18 in the BCFA model). While all the items from the identified regulation, integrated regulation, and intrinsic motivation subscales significantly loaded on the autonomous motivation factor, most of them did not load on their specific factors or had significant but relatively low factor loadings. The findings from the goodness of fit statistics and the parameter estimates indicate that the BESEM model should be accepted for further analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Standardized factor loadings for bi-factor confirmatory factor analysis and bi-factor exploratory structural equation models.

https://doi.org/10.1371/journal.pone.0265004.t004

Factor correlations

Factor correlations for the CFA and ESEM models were calculated and are reported in Table 5. Overall, the factor correlations in the ESEM model are substantially lower than those in the CFA model. For both model types, factors generally correlated most closely with the most proximal factor. More distal factors demonstrated non-significant or negative relationships.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Factor correlations for the confirmatory factor analysis and exploratory structural equation models.

https://doi.org/10.1371/journal.pone.0265004.t005

Measurement invariance

Table 6 shows the measurement invariance test across men and women. CFI, TLI, and RMSEA values represented an excellent fit of the BESEM model for both men and women groups. Comparing the configural, metric, scalar, and strict invariance tests, there was no substantial decrease in the CFI values or increase in RMSEA values.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Measurement invariance testing between men and women.

https://doi.org/10.1371/journal.pone.0265004.t006

Concurrent evidence of validity

Table 7 demonstrates the concurrent evidence of validity for the Chinese BREQ-3 by regressing the global and subscale scores to their theoretically related constructs. When including the amotivation, controlled motivation, and autonomous motivation as the only predictors, amotivation and controlled motivation positively predicted competence, PA, and NA. In addition, controlled motivation positively predicted autonomy and relatedness. Autonomous motivation positively predicted autonomy, competence, relatedness, PA, and negatively predicted NA. This model explained a substantial portion of the variance for autonomy (55%), competence (55%), relatedness (37%), PA (50%), and NA (35%). Adding specific types of motivation into the model resulted in a slight improvement in the explained variances (ΔR² = 0%, 2%, 1%, 0% 2% for autonomy, competence, relatedness, PA, and NA respectively).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Standardized coefficients and explained variances for types of motivation regressed on covariates.

https://doi.org/10.1371/journal.pone.0265004.t007

Reliability

Table 4 displays the composite omega(ω_c), hierarchical omega (ω_h), ECV, I_ECV, and PUC for different types of motivation based on the BESEM model. All motivation factors demonstrated satisfactory composite omega (ω_c >0.80). While the controlled motivation, autonomous motivation demonstrated satisfactory hierarchical omega (ω_h >0.50), the external regulation, introjected regulation, identified regulation, integrated regulation, and intrinsic motivation subscales had poor hierarchical omega (ω_h <0.50). The ECV for the controlled motivation was less than 0.85 and was greater than 0.85 for the autonomous motivation. The I_ECV values for all the items under the controlled motivation were less than 0.85 (except for item 18), whereas the I_ECV values for autonomous motivation items were above or close to 0.85 (except for item 13). The PUC was relatively low for the controlled motivation and high for the autonomous motivation.