Abstract
Self-determination theory (SDT) is used to predict individual differences in goal-directed behavior. A fundamental tenet of SDT is that autonomously regulated behavior is more likely to be engaged in and sustained than externally controlled behavior. Unidimensional treatment of regulation is suboptimal. The current study utilizes a multidimensional approach—polynomial regression—to evaluate the interdependent effects of autonomous and controlled regulation on physical activity. Results from three samples of healthy, younger adults demonstrate a mostly positive influence of autonomous regulation but a curvilinear effect of controlled regulation on behavior such that greater activity was associated with moderate levels of controlled regulation—an effect that cannot be identified with “unidimensional” methods. Results from Sample 3 showed that autonomous regulation was associated with greater exercise levels only when controlled regulation was moderate or high. Results suggest that controlled regulation is not wholly detrimental to behavioral promotion.
Keywords
Self-determination theory (SDT; Deci & Ryan, 2000; Ryan & Connell, 1989; Ryan & Deci, 2000) is a complex theoretical framework that has been successful in explaining a wide range of individual differences in motivational factors and behavioral engagement, including in academic (Miserandino, 1996; Patrick, Skinner, & Connell, 1993; Vallerand, Fortier, & Guay, 1997; Vansteenkiste, Lens, & Deci, 2006), environmental conservation (Osbaldiston & Sheldon, 2003; Renaud-Dubé, Taylor, Lekes, Koestner, & Guay, 2010), and health (Bartholomew, Ntoumanis, Ryan, Bosch, & Thøgersen-Ntoumani, 2011; Halvari, Halvari, Bjørnebekk, & Deci, 2010; Standage, Duda, & Ntoumanis, 2003; Wang & Biddle, 2001; Williams, McGregor, Zeldman, Freedman, & Deci, 2004; Williams, Rodin, Ryan, Grolnick, & Deci, 1998) domains. Research further suggests that SDT constructs are culturally invariant, bolstering the validity and usefulness of the framework (Grouzet et al., 2005; Rudy, Sheldon, Awong, & Tan, 2007; Vansteenkiste, Mingming, Lens, & Soenens, 2005).
Although SDT encompasses many constructs, perhaps the most widely studied is individuals’ type and degree of motivation for engaging in a behavior, or the degree to which their behavior is self-determined (Deci & Ryan, 2008; Ryan & Deci, 2000). SDT proposes multiple types of regulation of behavior: briefly, external regulation of behavior comes from motivation to engage in behavior due to external demands and pressures (e.g., because of perceived peer pressure), introjected regulation is engaging in behavior due to some internalization of these external pressures so that the individual feels guilt or shame for not meeting those external pressures (e.g., feeling guilty for letting others down), identified regulation comes from recognizing behavioral outcomes are valued by the individual, integrated regulation comes from seeing the behavior as important for the self-concept (e.g., viewing oneself as a person who engages in a particular behavior), last, internal regulation or intrinsic motivation refers to engaging in behavior because the individual enjoys it.
Researchers commonly group the regulations into autonomous (intrinsic, identified, and integrated) and controlled (external and introjected) regulations of behavior. Factor analysis has been found to support this dichotomy (Cid, Moutão, Leitão, & Alves, 2012; Vansteenkiste, Lens, De Witte, De Witte, & Deci, 2004), and use of the dichotomy is common in multiple behavioral domains (Gagné et al., 2010; Teixeira, Carraça, Markland, Silva, & Ryan, 2012). Furthermore, the dichotomy fits with theory (Ryan, Kuhl, & Deci, 1997) that argues, and evidence (Lee & Reeve, 2013; Lee, Reeve, Xue, & Xiong, 2012) that shows, that brain activity in areas associated with agency and higher order cognitive processes are associated with autonomous regulation to engage in behavior; conversely, brain areas associated with inhibition of personal agency are associated with controlled regulation. SDT developers have suggested that a treatment of motivational constructs as either controlled or autonomous is more useful than discussing intrinsic versus extrinsic motivation (Deci & Ryan, 2008; Rigby, Deci, Patrick, & Ryan, 1992).
A fundamental tenet of SDT is that autonomously regulated behavior is more likely to be engaged in frequently, to be maintained over time, and to lead to beneficial outcomes than behavior that is externally controlled (Chemolli & Gagné, 2014; Deci & Ryan, 2008; Teixeira et al., 2012). For example, the earliest SDT research on motivating children to learn in school showed that children persist at, and perform better on, activities when they are internally regulated (find tasks enjoyable), rather than when externally regulated (complete tasks for external incentives or to avoid punishment; Grolnick & Ryan, 1987; Patrick et al., 1993).
Researchers have utilized different methodologies to study this proposition regarding the relative benefits of autonomous compared with controlled regulation. The theoretical assumptions of each method determine the particular hypotheses a researcher can make. In the current article, we discuss this method–theory relationship and demonstrate that advancing analytic methods allows for better testing of standard and even more complex SDT hypotheses regarding individual differences in motivated behavior, with ramifications for behavior promotion. Specifically, we demonstrate the advantages of using an analytic method called polynomial regression for evaluating multidimensional individual differences in behavioral regulation and engagement.
Independent Effects of Regulations
Researchers have found evidence from correlational and experimental research to support the importance of autonomous regulation over controlled regulation for both initiation and maintenance of behavior using separate, or independent, variables to represent the regulations (Ng et al., 2012; Williams et al., 1998). In a recent systematic review of SDT studies in an exercise domain, a general finding was that autonomous regulation of physical activity (PA) was positively related to behavioral frequency; controlled regulation had a negative or no relationship with exercise in published samples overall (Teixeira et al., 2012). Furthermore, regarding experimental research, interventions designed to promote autonomous regulation have been found to more effectively increase PA when compared with typical control conditions (Edmunds, Ntoumanis, & Duda, 2008; Fortier, Duda, Guerin, & Teixeira, 2012); several clinical trials show support for intervention techniques that promote internalized regulation (i.e., autonomy; Ryan, Patrick, Deci, & Williams, 2008).
Although the relative importance for behavior of autonomous compared with controlled regulation may be well supported in analyses that treat them as separate variables, it is possible that this separate treatment of motivation types may mask combined, interactive, or interdependent effects of the regulation types on behavior. For example, some degree of controlled regulation may be beneficial for behavioral engagement in the long term, depending on an individual’s level of autonomy. It is possible an individual with high levels of both will exercise more frequently over time, because he or she is less likely to forego exercise when time and stress increase. Indeed, lack of time and high levels of stress are two of the main reasons individuals give when asked why they do not exercise (Grubbs & Carter, 2002; Johnson, Corrigan, Dubbert, & Gramling, 1990). External pressure may be enough to get an individual to exercise, even when time pressures/stress are high, but perhaps only if that individual gets stress release (intrinsic enjoyment) from exercise (is also high in autonomy). Evaluating only the independent effects of regulations on behavior misses this complexity as does any single-dimension treatment of these regulation dimensions, including the relative autonomy index (RAI).
The RAI
A common treatment of SDT motivational variables to evaluate the proposition that autonomous regulation optimizes and controlled regulation degrades positive behavioral outcomes is use of the RAI. The RAI is a type of difference score created by subtracting individuals’ controlled regulations (external and introjected; sometimes amotivation, as well) from their autonomous regulations (identified and intrinsic; sometimes integrated, as well); each subscale of motivation is weighted to represent its relative degree of relationship to optimal self-determination with weights ranging from −3 to +3. The RAI has been used to predict behavioral outcomes (Ryan & Connell, 1989; in 23% of reviewed studies in a PA domain; Teixeira et al., 2012), and significant results are claimed as support for the proposition that behavior will be optimized as autonomous regulation is maximized and controlled regulation is minimized—that is, that interventions should promote autonomous and discourage controlled regulation. Although this intervention approach may indeed be successful, use of the RAI does not actually allow a direct test of its underlying theoretical proposition. Instead, the RAI constrains controlled regulation to have a negative relationship with behavior, in the creation of the variable as positive weights for autonomous and negative weights for controlled regulations.
The statistical problems associated with difference scores are well known: As a difference score, the RAI confounds the potentially distinct effects of different types of regulation on behavior, and it decreases statistical power for testing hypotheses. Phillips (2013) provides an explanation of how difference scores increase the risk of committing both Type I and Type II errors (finding support for a hypothesis when it is not accurate and failing to find support for a hypothesis when it is accurate, respectively). Chemolli and Gagné (2014) conclude that the RAI should, therefore, not be used to test SDT hypotheses, and the developers of the Behavioral Regulation of Exercise Questionnaire (BREQ), Behavioral Regulation of Exercise Questionnaire–Version 2 (BREQ-2), and Behavioral Regulation of Exercise Questionnaire–Version 3 (BREQ-3) no longer recommend its use (and always recommended caution in interpreting RAI results; see http://pages.bangor.ac.uk/~pes004/exercise_motivation/breq/breq.htm).
Beyond an issue of the RAI being a difference score, the theoretical proposition of a continuum of self-determination itself has been shown to be inaccurate (Chemolli & Gagné, 2014; Howard, Gagné, Morin, Wang, & Forest, in press). Chemolli and Gagné (2014) used Rasch analysis to test whether SDT constructs were on a linear continuum of self-determination/autonomous regulation as suggested by SDT, utilizing two scales of work motivation to predict work-related performance outcomes. They concluded that the SDT constructs do not fall on a continuum of self-determination or autonomy. Instead, the regulations are qualitatively different and, therefore, motivation is multidimensional; an individual can be at several different points on these regulations at one time. For example, an individual could theoretically be strongly motivated by both external and internal pressures to engage in a behavior. In fact, autonomous and controlled regulations sometimes correlate positively (e.g., Teixeira et al., 2012; Vansteenkiste, Sierens, Soenens, Luyckx, & Lens, 2009).
Relatively advanced methodology is required to evaluate the potentially complex relationships between types and degree of regulation and behavioral outcomes. Chemolli and Gagné (2014) advocate for evaluating interdependent effects of regulations using polynomials (e.g., polynomial regression), and for evaluating person-based profiles (e.g., cluster analysis) of regulation combinations and their influence on outcomes of interest.
“Person-Centered Approach”: Cluster and Profile-Based Analyses
Researchers studying motivation in an academic domain have recently utilized cluster analysis to evaluate the interdependent effects of controlled and autonomous regulation on behavioral outcomes (e.g., Boiché, Sarrazin, Grouzet, Pelletier, & Chanal, 2008; Fortunato & Goldblatt, 2006; Friederichs, Bolman, Oenema, & Lechner, 2015; Gillet, Vallerand, & Paty, 2013; Liu, Wang, Tan, Koh, & Ee, 2009; Moran, Diefendorff, Kim, & Liu, 2012). Profile-based analysis, including cluster analysis, is considered to be a “person-centered approach” to analysis, focusing on characterizing individuals by their levels of different constructs/qualities rather than taking a “variable-centered” or dimensional approach to analysis (Ratelle, Guay, & Vallerand, 2007). Ratelle et al. (2007) state that studies of SDT and academic performance that have used a dimensional approach to analysis have not been able to see whether different combinations of regulation types influence behavior differently (e.g., if individuals who enjoy learning/school and feel pressure to excel perform differently to individuals who enjoy learning/school but do not feel parental pressure to excel). They and others (Vansteenkiste et al., 2009) argue that the main benefit of cluster analysis for studying individual differences is that it allows researchers to evaluate the effect of combinations of qualitatively different motivations within person on behaviors of interest (Ratelle et al., 2007).
Using a profile-based approach, Ratelle et al. (2007) found support for the multidimensional nature of motivation, with analyses resulting in three motivational profiles: high autonomous/high controlled, high autonomous/low controlled, and moderate levels of autonomous, controlled, and amotivated regulation. The high autonomous/high controlled regulation group had the best academic adjustment. Similarly, Vansteenkiste et al. (2009) found that students fit into their four hypothesized clusters of regulation corresponding to the four different combinations of high versus low autonomous regulation with high versus low controlled regulation (high/high, low/low, low/high, and high/low). They found evidence that the statistically determined clusters were differentially related to outcomes: students high in autonomous and low in controlled regulation (a “good quality motivation” cluster) demonstrated the best learning patterns. Students with “high quantity” motivation (high autonomous and high controlled motivation scores) had better outcomes (greater regulation, better performance, lower procrastination, etc.) than students with “low quantity” motivation (low autonomous and low controlled motivation scores).
Vansteenkiste et al. (2009) further list as benefits of a person-centered approach that it (a) allows researchers to determine the prevalence of individuals in “riskier groups” and (b) allows researchers to evaluate the internal validity of SDT as a multidimensional model of motivation. However, cluster analyses may limit researchers’ ability to do these things because of statistical and conceptual limitations of the analyses that stem from creating groups (profiles) from underlying dimensions that are natural continua. For example, forming groups/clusters potentially masks different relationships with outcomes among those who fall near the “middle” of the dimension(s) compared with individuals who fall closer to the cluster centers. Loss of statistical power and conceptual information with categorization of individuals on natural continua are well outlined, potential consequences of such analyses (Hunter & Schmidt, 1990).
Furthermore, regarding translation of research findings to intervention design and practice, labeling a particular combination of motivational levels a “profile” or type of individual may be misleading, if those states are relatively fluid/changeable within person. If the profiles are used to identify “high-risk” individuals, then those individuals become labeled and treated differently than individuals with another profile at the time the data were collected. At the least, research to investigate the robustness of these profiles in an academic setting and other settings is warranted before interventions are targeted and tailored for these supposed “types” of individuals.
Latent profile analysis (LPA) is another person-centered approach that could be used to analyze motivational profiles (e.g., see Graves, Cullen, Lester, Ruderman, & Gentry, 2015), and that has some advantages over cluster analysis. First, LPA allows the researcher to control for measurement error in the determination of latent variables from the manifest (measured) variables. Second, traditional cluster analyses have been criticized for being too subjective in the determination of final cluster solutions (Pastor, Barron, Miller, & Davis, 2007), and LPA avoids this by providing a more robust method for determining a final cluster solution. Third, LPA results in latent profiles to which each individual has a probability of belonging, which means that the accuracy of participant placement can be evaluated. These profiles are used to predict (are associated with) outcomes (Marsh, Lüdtke, Trautwein, & Morin, 2009). Therefore, the same potential conceptual limitations as cluster analysis exist with LPA. Furthermore, nonlinear relationships between motivations and outcomes are also not detectable by LPA.
Therefore, although profile-based analyses improve researchers’ ability to evaluate potential interdependent effects of qualitatively different motivations on outcomes, we argue that they still have limitations. Furthermore, we argue that there is no inherent reason why dimensional or variable-centered analyses preclude evaluating interdependent effects of two predictor dimensions on outcomes, for identifying (and verifying the validity of) “high-risk individuals,” or for evaluating the internal validity of SDT. Below, we present a dimensional approach called polynomial regression that offers some advantages to a person-centered analysis when studying the interdependent effects of autonomous and controlled behavioral regulation.
“Dimensional Approach”: Polynomial Regression
Instead of creating groups, or profiles, where they may not be clearly defined, polynomial regression maintains the continuous nature of two predictor variables (e.g., autonomous and controlled regulation) when evaluating their interdependent effects on a continuous outcome (e.g., academic performance, exercise frequency; see “Analysis Overview” section for details on the analysis steps). To date, no researchers have used polynomial regression to evaluate the potential interdependent effects of motivations on behavioral outcomes, but polynomial regression has been used to evaluate other structurally similar theoretical questions in health domains (Phillips, 2013; Phillips, Diefenbach, Kronish, Negron, & Horowitz, 2014) and workplace-performance domains (Edwards & Harrison, 1993). For example, Phillips et al. (2014) used polynomial regression to test the interdependent effects of medical patients’ positive and negative beliefs regarding their medications on their medication adherence. They demonstrated “nonreciprocal effects”: Individuals with indifferent attitudes (weak positive and weak negative beliefs) were more adherent than individuals with ambivalent attitudes (strong positive and strong negative beliefs). Furthermore, they demonstrated “reciprocal effects”: Individuals with accepting beliefs (strong positive and weak negative beliefs) were more adherent than those with skeptical beliefs (weak positive and strong negative beliefs) and had the best adherence overall. With regard to nonreciprocal effects of motivations on behavior, Vansteenkiste et al. (2009) found that students with “high quantity” motivation (high autonomous and high controlled motivation scores) had better outcomes (greater regulation, better performance, lower procrastination, etc.) than students with “low quantity” motivation (low autonomous and low controlled motivation scores). Polynomial regression can detect these effects without the issues associated with other methods.
Applied to SDT hypotheses, polynomial regression accounts for the intensity and quality of each regulation and evaluates the interdependent effects of the two regulation dimensions on behavior/outcomes without the potential problems associated with artificial grouping of individuals from natural continua. For example, a researcher can evaluate the differing effects on behavior of having high levels of autonomous and high levels of controlled regulation compared with having high levels of one but not the other regulation or low levels of both regulations, and can still see whether there are differential effects on outcomes of having moderate levels of one or both the regulations. Although these relationships are testable using profile-based analyses, statistical power to test the hypotheses is greater when predictor variables are left continuous.
Furthermore, polynomial regression allows researchers to test even more complex hypotheses than profile-based analyses are able to evaluate—namely, any relationships that involve nonlinear relationships between variables, including interactive effects between predictors and/or curvilinear effects of one or both predictors on the outcome. Theoretically, the effect of quality of motivation on an outcome may depend on its intensity as well as the intensity of another regulation. Polynomial regression tests for interaction and curvilinear effects simultaneously, thereby overcoming limitations of even other regression analyses for testing moderation/interaction (e.g., the X × Y term may appear significant when it is tested incrementally to X and Y separately, but the term driving the false interaction may indeed be X2 or Y2; see Edwards, 2001, 2008). For example, polynomial regression could detect an effect where students’ autonomous regulation more strongly predicts academic performance as levels of controlled regulation decreases (and could even detect whether this increase in the effect of autonomous regulation was exponential vs. linear). An example of a possible curvilinear effect that polynomial regression could detect is the case where students’ autonomous regulation is most beneficial for school attendance at moderately high levels but slightly less beneficial at extreme high and moderately to extreme low levels. Although cluster-based analyses could detect some interactions, the loss of statistical information from forming categories of individuals makes detecting/interpreting interactions less likely, and complex interactive effects (that can occur simultaneously with curvilinear effects) are not detectable with profile analyses.
Study Overview
In the current study, the potential interdependent effects of autonomous and controlled regulation on PA are evaluated in three separate samples, collected between 2012 and 2016. The first sample is a relatively large college-student sample, offering the benefit of a larger sample size for testing polynomial regression (larger sample sizes provide more reliable regression estimates; Kelley & Maxwell, 2003). Sample 2 was collected in the same population as Sample 1 (from a medium-sized, private university in a large urban area of the United States) to cross validate the results from Sample 1 using a prospective study design and an objective measure of PA (via accelerometers). Sample 3 was also collected to verify results of Samples 1 and 2, with a relatively large sample of a similar population to Samples 1 and 2 but from a different geographic area (large, public university in a small town of the United States) and with prospectively measured, self-reported exercise frequency as the outcome. Levels of autonomous and controlled regulation were measured in all samples and used to predict respective measures of PA. PA is a target behavior for healthy adults to prevent chronic illness, and for those who already have chronic illness to manage the illness, and for optimizing quality of life/subjective well-being (Centers for Disease Control and Prevention [CDC], 2004). Recommendations for PA made to adults for preventing or managing chronic illness include engaging in moderate to vigorous PA for at least 30 min, most if not all days of the week (U.S. Department of Health and Human Services, 2010).
Polynomial regression is used to analyze the data, as described in the “Analysis Overview” section, and to evaluate the following, general hypothesized relationships:
Hypotheses
Although we have no a priori hypotheses regarding potential interactive or curvilinear effects of the motivational dimensions on PA, these types of effects are possible with polynomial regression and will be evaluated.
Method
Sample 1 Participants and Procedure
Participants in Sample 1 were from an initial recruited sample of 553 college students, from an anonymous subject pool in the psychology department at an urban, private university in the United States. All measures were assessed concurrently in an online questionnaire that was piloted to take approximately 40 to 50 min to complete. Participants were compensated with partial course credit upon completion of the online questionnaire. Thirty-seven individuals were excluded from analyses due to being on a nationally competitive athletic team (meaning that their PA is not “leisure time” activity, but for a sport and dictated by a coach/trainer); 25 were excluded for failing to accurately respond to two random responding checks in the surveys (e.g., “Please answer ‘mostly untrue’ for this question”). Data from an additional 15 students were excluded because of incomplete data; pairwise deletion was used to handle missing data, because the data may not have been missing at random. Therefore, analysis was conducted with a final sample of 476 students: 71% identified as female, 31% identified as minority race or ethnicity, and the average age was 19.40 years (SD = 2.04 years).
Sample 2 Participants and Procedure
Participants in Sample 2 were from an initial recruited sample of 123 healthy adults from the same university as Sample 1. Inclusion criteria were 18 years or older, no participation in the previous survey study (not in Sample 1), and willingness to engage in at least two exercise sessions per week for the 4 weeks of the study. Participants were recruited through the psychology department research subject pool and university-wide emails. Data analysis was conducted with a final sample of 106 participants; excluded were 11 participants who were on nationally competitive athletic teams at the time of the study and six additional individuals who failed two random responding checks. Approximately 75% were female, 25% identified as minority race or ethnicity, and the average age of participants was 25.40 years (SD = 11.60 years).
Baseline survey measures, including measures of SDT constructs, self-reported PA, and demographic characteristics, were evaluated in computer-administered surveys. Participants then received accelerometers (Fitbit Zips; Fitbit.com) and instructions on using the Fitbits for the subsequent 4 weeks. Daily emails reminded participants to wear their Fitbits. At the end of 4 weeks, participants returned the Fitbits, completed a computer-administered survey, and were compensated with partial course credit and US$20 cash (individuals in the subject pool) or with US$40 cash (individuals not in the subject pool).
Sample 3 Participants and Procedure
Participants in Sample 3 were from an initial recruited sample of 436 college students, from a subject pool in the psychology department at a large, public university in a small town in the United States. Baseline survey measures, including measures of SDT constructs and demographic characteristics were evaluated in the laboratory, via computer-administered surveys. Participants reported their frequency of exercise sessions in the preceding week at follow-up, which occurred after a minimum of 2 weeks (and maximum of 4 weeks) postbaseline. Participants were compensated with partial course credit; some participants used Fitbits as part of a substudy within the sample and were further compensated with US$20 cash. No participants failed the random response checks added to the survey (“Please answer ‘mostly true’ to this item”). Seventeen participants did not complete the follow-up survey, and 10 participants had missing data from the baseline and/or follow-up survey on the analyzed variables. Pairwise deletion was used for these cases, because it was not clear whether the data were missing at random. This yielded a final sample of 409 participants: 71% identified as female, 20% identified as minority race or ethnicity, and the average age was 19.13 years (SD = 2.07 years).
Measures
Behavioral regulation
Participants’ motivations to engage in exercise (exercise regulation types) were evaluated using the BREQ-2 (Markland & Tobin, 2004). A BREQ-3 that includes integrated regulation was developed after the start of data collection for Sample 1, so we used BREQ-2 for all samples for consistency. Example items for each subscale are the following, each having response options from 1 = not true for me to 5 = very true for me: External Regulation, “I exercise because other people say I should” and “I feel under pressure from my friends/family to exercise” (four items; α = .86, .78, and .78 for Samples 1, 2, and 3, respectively); Introjected Regulation, “I feel guilty when I don’t exercise” and “I feel ashamed when I miss an exercise session” (three items; α = .85, .77, and .87 for Samples 1, 2, and 3, respectively); Identified Regulation, “I value the benefits of exercise” and “I think it’s important to make the effort to exercise regularly” (four items; α = .82, .73, and .80 for Samples 1, 2, and 3, respectively); and Intrinsic Regulation, “I exercise because it’s fun” and “I enjoy my exercise sessions” (four items; α = .93, .91, and .90 for Samples 1, 2, and 3, respectively).
As in the literature, controlled regulation is calculated as the average of individuals’ external and introjected regulation, and autonomous regulation is calculated as the average of individuals’ identified and intrinsic regulation (Ryan & Deci, 2000; Vansteenkiste et al., 2009). Amotivated regulation was not included as a dimension, because it can be argued that lack of autonomous and lack of controlled regulation is equivalent to being amotivated (and, therefore, inclusion of all three of these regulations would add redundancy; Vansteenkiste et al., 2009).
Self-reported PA
We measured participants’ moderate to vigorous PA in Samples 1 and 3 via self-reports. In Sample 1, participants reported their PA in a modified version of the International PA Questionnaire (IPAQ; Ainsworth et al., 2006; Booth, 2000), which assessed their days per week and minutes per day of moderate and of vigorous PA. An average score for minutes per day of moderate to vigorous PA was calculated to represent the variable. In Sample 3, participants reported their frequency of days in the past week of engaging in exercise sessions (equivalent to 30+ min of moderate- to vigorous-intensity activity over the course of a day, in at least 10-min bouts), assessed 2 weeks to 1 month postbaseline.
Objective PA
In Sample 2, data were collected using Fitbits. Fitbits (Fitbit.com) provided participants’ average steps per day for the month of the study. We used software to collate and analyze the data from Fitbit.com; the software is available for research purposes from Fitabase.com. Participants who wore their Fitbits (verified by checking whether any activity was logged each day, on Fitbit.com via Fitabase.com) at least 23 out of the 30 days (more than 75%) of the study were included in the analysis; all participants who were given a Fitbit met this requirement.
Analysis Overview
The data were first evaluated for univariate and multivariate outliers, because outliers can greatly influence regression coefficients and clustering procedures (see Tabachnick & Fidell, 2007). The factor structure of the regulation items was assessed in each sample using exploratory factor analysis with maximum likelihood extraction and oblique rotation (direct oblimin). Parallel analysis was used to determine the number of factors to extract with eigenvalues above the level expected by chance (not just greater than 1; see O’Connor, 2000). Parallel analysis indicated two factors fit the data for all samples, and these factors corresponded to autonomous (intrinsic and identified) and controlled (external and introjected) regulations. Cross loadings existed for two of the introjected regulation items, but the stronger loadings were with the factor that included the external regulation items. Even with cross loadings, the factors did not strongly correlate with each other, r(466) = .20 in Sample 1, r(105) = −.03 in Sample 2, and r(408) = .18 in Sample 3. Variables representing autonomous and controlled regulation were, therefore, calculated as theoretically appropriate and as in existing SDT research. In all samples, polynomial regression and response surface analysis were used to evaluate the interdependent effects of autonomous and controlled regulation on PA, as described in the next sections.
The exploratory polynomial regression analyses 1 were conducted using the following steps (Edwards, 2002; Phillips, 2013): First, the two predictors (e.g., autonomous and controlled regulation) were mean centered to improve interpretation of the results and to reduce multicollinearity of the predictor terms with their higher order terms entered in subsequent steps of the regression. Second, higher order terms (e.g., quadratic terms include the squared terms for each regulation and the interaction term, that is, product of the two regulation variables) were created so that polynomial models of increasing order can be entered into the regression, predicting PA, in separate steps (first step is the linear model, with the two predictors, X and Y; second step is quadratic, with X2, XY, and Y2 terms; third step is cubic, with X3, X2Y, XY2, and Y3 terms, etc.). The highest order polynomial that explains significant incremental variance in the outcome was kept as the final model and graphed to aid interpretation of the results. Parameters were statistically evaluated to verify specific relationships of interest in the final graph. Resources for graphing and determining statistical parameters for the final polynomial surface are available from Jeffrey R. Edwards’s personal website, from online supplemental material to Shanock, Baran, Gentry, Pattinson, and Heggestad (2010), and in online supplemental material from Phillips (2013). Statistical packages, such as STATA, can also be used to graph the surface of the final polynomial model.
Support for Hypothesis 1 (that PA will be greater as autonomous and controlled regulation both increase) would be found if the slope of the surface (i.e., change in PA) along the line of congruence (i.e., when autonomous and controlled regulation are equal to each other in magnitude and valence) is significantly greater than 0. Support for Hypothesis 2 (that PA should be greater as autonomous regulation increases and controlled regulation decreases) would be found if the slope (i.e., change in PA) along the line of incongruence (i.e., when autonomous and controlled regulation are maximally different from each other) is significantly greater than 0. As is used to interpret multiple regression results (e.g., interactions), the regression coefficients for the final polynomial model can be used to graph and interpret any curvilinear or interactive effects. These statistical tests of the slopes can be accomplished using resources freely available on the Internet: In supplemental material, Phillips (2013) provides SPSS syntax for calculating and statistically evaluating these and other parameters of interest for “congruence relationships.” We refer the interested reader to Edwards (2002) and Phillips (2013) for more detailed information on polynomial regression, in general.
Results
Sample 1
No univariate outliers were detected. The linear model of the exploratory polynomial regression was significant, R2 change = .15, F(2, 473) = 40.53, p < .001, and the quadratic model predicted significant incremental variance to the linear model, R2 change = .02, F(3, 470) = 3.06, p = .03. The cubic model did not predict significant incremental variance in the outcome to the quadratic model, R2 change = .004, F(4, 466) = 0.59, p = .67. Figure 1a is a graph of the final model, and Table 1 reports all regression results for the linear and quadratic models for all three samples. In Sample 1, three individual terms were significant predictors of exercise minutes per day: the linear terms of autonomous regulation (B = 14.83, β = .40, t = 7.73, p < .001; 95% confidence interval [CI] = [11.06, 18.60]) and controlled regulation (B = 4.48, β = .11, t = 2.07, p = .04; 95% CI = [0.23, 8.73]), and the squared term for controlled regulation (curvilinear effect: controlled regulation on PA; B = −4.33, β = −.13, t = −2.29, p = .02; 95% CI = [−8.04, −0.62]).

Figures 1a, 1b, and 1c show the final, best fitting polynomial models for Samples 1, 2, and 3, respectively.
Exploratory Polynomial Regression Results for Three Samples.
Note. Regression coefficients for the linear and quadratic models are presented. The final models graphed for all samples are quadratic (see figures). The outcome for Sample 1 is concurrently assessed, self-reported moderate- to vigorous-intensity exercise minutes per day; for Sample 2 is accelerometer-measured average steps per day; for Sample 3 is prospectively assessed, self-reported exercise days in the previous week.
p < .07. *p < .05. **p < .01. ***p < .001.
Regarding the hypotheses, although they are most appropriate for a linear final model, the slopes of the surface (i.e., changes in self-reported PA) were in support of both hypotheses: The slope along the line of congruence, representing whether PA is different when both autonomous and controlled regulation are high versus both low (Hypothesis 1), was significant, slope = 16.07, t(475) = 6.42, p < .001; 95% CI = [12.04, 20.11]. The slope of the surface along the line of incongruence, representing whether PA is different when autonomous regulation is high and controlled regulation is low (a “high quality motivational profile”) versus when controlled regulation is high and autonomous regulation is low (i.e., “low quality motivational profile”) was significant, slope = 10.27, t(475) = 3.73, p < .001; 95% CI = [4.19, 16.35]. Because the slope was positive, this provides support for Hypothesis 2. However, because the slopes of a surface are calculated at a single point (in this case, when both predictors are equal to 0, which is the average level of the regulations in the sample, because the predictors are mean centered), any nonlinear change in the outcome with changes in the predictors, as occurs in a quadratic model, limits the usefulness of these slopes in specifying how PA changes (other than relative overall level) with changes in regulations. Because the quadratic model was the best fitting model, there are more complex relationships between the variables that are not captured by slopes of the surface at one point. For Sample 1, the significant quadratic term equal to the square of controlled regulation indicates that there is a curvilinear relationship between controlled regulation and PA; furthermore, because the interaction term was not significant, we can conclude that this curvilinear effect is the same at all levels of autonomous regulation. In total, PA was greatest when autonomous regulation was high but controlled regulation was at a moderate (near average) level (Figure 1a).
Sample 2
No univariate outliers were detected. The linear model of the polynomial regression was significant, R2 change = .12, F(2, 103) = 6.77, p = .002, and the quadratic model predicted marginally significant incremental variance to the linear model, R2 change = .06, F(3, 100) = 2.41, p = .07. Because Sample 2 was meant as a replication sample for the results of Sample 1, which was a quadratic model, we graphed the quadratic model from Sample 2 (Figure 1b) and evaluated the significance of individual terms in the model. As in Sample 1, the squared term for controlled regulation (the curvilinear effect of controlled regulation on PA) was significant (B = −968.18, β = −.21, t = −2.28, p < .03; 95% CI = [−1,811.70, −124.65]), and the coefficient for autonomous regulation was significant (B = 1,146.71, β = .36, t = 3.52, p = .001; 95% CI = [500.16, 1793.26]).
Because the results conceptually replicate the final model in Sample 1, we take this to be a replication of Sample 1’s findings, and the marginal significance a result of smaller sample size rather than due to chance. It is, however, possible that the objectively measured outcome resulted in a weaker relationship between variables, because self-reported and objective measures of activity correlate only moderately. If the linear model was to be taken as the final model, no graph of the surface (plane) would be necessary, because only autonomous regulation was significantly related to Fitbit activity (B = 1,065.42, β = .33, t = 3.61, p < .001; 95% CI = [479.79, 1,651.04]), not controlled regulation (B = −292.36, β = −.08, t = −0.86, p = .40; 95% CI = [−970.68, 385.95]). However, the significant curvilinear term in the quadratic model explains why the linear term for controlled regulation is nonsignificant, supporting the decision to interpret the quadratic rather than the linear model for Sample 2: Similar to when one variable acts as a moderator of the relationship between another variable and an outcome, the controlled regulation variable is acting as a moderator of its own relationship to PA—that is, the relationship between controlled regulation and PA depends on the level of controlled regulation.
Regarding the hypotheses, as in Sample 1, the slopes of the surface at a single coordinate of predictors (when both are at their average level, or 0) should not be interpreted as the most meaningful effect of the combined regulations on PA, because the quadratic model is taken as a final model. However, the slopes of the surface (i.e., changes in Fitbit steps per day) were in support of Hypothesis 2 but not in support of Hypothesis 1. The slope of the surface along the line of congruence (test of Hypothesis 1) was marginally significant, slope = 842.42, t(105) = 1.80, p = .08; 95% CI = [−105.67, 1,651.78], and the slope of the surface along the line of incongruence (test of Hypothesis 2) was significant, slope = 1,451.00, t(105) = 3.10, p = .002; 95% CI = [444.55, 2,271.00]. However, as in Sample 1, the curvilinear term for controlled regulation, in the quadratic model, was significant, highlighting a more complex relationship than these slopes can capture. As can be seen in Figure 1b, PA was maximized as autonomous regulation increased but when controlled regulation was at a moderate (vs. high or low) level. Visual comparison of Figures 1a and 1b (and the fact that the linear term for controlled regulation had a significant, positive relationship to PA in Sample 1 but no significant relationship to PA in Sample 2) highlights a slightly different effect of high levels of controlled regulation: In Sample 1, higher levels of controlled regulation had less of a negative effect (compared with moderate levels) than is the case in Sample 2.
Sample 3
No outliers were detected. The quadratic model was the final model for the outcome; the linear model explained 9% of the variance in self-reported number of exercise days at follow-up, F(2, 406) = 19.53, p < .001; the quadratic model explained 2% incremental variance in the outcome, F(3, 403) = 2.78, p = .04; the cubic model did not predict significant incremental variance, F(4, 399) = 0.42, p = .79. Figure 1c shows the graph of the final model, and Table 1 reports the regression results in full. In Sample 3, the linear term of autonomous regulation (B = 0.68, β = .31, t = 6.03, p < .001; 95% CI = [0.46, 0.91]) and the interaction term (autonomous × controlled regulation; B = 0.46, β = .15, t = 2.41, p = .02; 95% CI = [0.08, 0.83]) were significant, and the squared term for controlled regulation (B = −0.32, β = −.10, t = −1.91, p = .06; 95% CI = [−0.66, 0.01]) was marginally significant.
Regarding the hypotheses, the slopes of the surface (i.e., changes in self-reported exercise days) were in support of both hypotheses: The slope along the line of congruence, representing whether exercise is different when both autonomous and controlled regulation are high versus both low (Hypothesis 1), was significant, slope = 0.64, t(408) = 4.52, p < .001; 95% CI = [0.37, 0.91]. The slope of the surface along the line of incongruence, representing whether PA is different when autonomous regulation is high and controlled regulation is low (a “high quality motivational profile”) versus when controlled regulation is high and autonomous regulation is low (i.e., “low quality motivational profile”) was significant, slope = 0.73, t(408) = 3.49, p = .001; 95% CI = [0.32, 1.14]. Because the slope was positive, this provides support for Hypothesis 2. Because the quadratic model was the best fitting model, there are more complex relationships between the variables that are not captured by slopes of the surface at one point. The quadratic terms representing the interaction between autonomous and controlled regulation and the squared controlled regulation indicate that exercise days per week are influenced by specific, relative levels of both motivation types: As seen in Figure 1c, participants who exercised the most had high levels of autonomous and controlled regulation, and at low levels of controlled regulation, even high autonomous regulation was associated with few exercise days in the week. The curvilinear effect of controlled regulation is similar to that seen in both Samples 1 and 2, but the interaction effect between autonomous and controlled regulation in the current sample differs from the previous samples (for which high autonomous regulation was associated with greater PA, for all levels of controlled regulation).
Discussion
In the current article, we reviewed methods used by researchers to test a major proposition of SDT, that behavior should be optimized when an individual is more autonomous and less controlled in his or her regulation of that behavior. Commonly used methods such as independent treatment of regulations as predictors and the RAI are limited in evaluating this proposition; both these methods preclude evaluation of more complex, interdependent effects of autonomous and controlled regulation on behavior (e.g., when both are high vs. low). Furthermore, due to statistical problems with difference scores, the RAI should not be used. More recently, cluster analysis has been used to study such interdependent effects in an academic domain. We have argued that many of the benefits of cluster analysis for testing SDT hypotheses are shared by a “dimensional approach” polynomial regression, and that polynomial regression may have advantages over cluster or other person-centered analyses, which can lose statistical power and/or conceptual information when individuals are fit into groups, or profiles, although LPA avoids the statistical problems associated with forming artificial categories as in cluster analysis.
Although the hypotheses in the current study were related to linear effects of different combinations of type and degree of regulations on PA, the results from all three samples indicated a more complex set of relationships between these variables; namely, in Samples 1 and 2, an overall positive effect of autonomous regulation on PA level but a curvilinear (or conditional) effect of controlled regulation on PA, at all levels of autonomous regulation, such that moderate levels of controlled regulation resulted in greater PA than either high or low extreme of controlled regulation. This curvilinear relationship between controlled regulation and exercise frequency was also observed in Sample 3, but an additional interaction effect was found such that the effect of autonomous regulation on exercise frequency was not uniformly positive—at the lowest levels of controlled regulation, exercise frequency was no different between those with high and low autonomous regulation. The quadratic model highlights these important effects that would not be realized if the terms were used only as linear terms or as difference scores, as they have been to-date in the literature. Indeed, the many studies that report zero relationship between controlled regulation and behavior may be missing a curvilinear effect of controlled regulation on the behavior. It is possible a reanalysis of existing data with polynomial regression may reveal similar results.
The current results provide evidence that controlled regulation is not universally detrimental to behavioral engagement. For example, if an individual exercises only because of enjoyment of the activity, he or she may be more likely to give up exercising when he or she perceives a lack of time to exercise. If the individual enjoys exercise but also exercises due to a partner’s expectations or feeling bad when he or she does not exercise, then he or she may be more likely to exercise in the face of time constraints. The curvilinear relationship of controlled regulation to exercise frequency indicates that too much or too little of these more extrinsic motivations is detrimental to engagement in regular exercise. Controlled regulation may, therefore, not be something researchers should necessarily aim to minimize in behavior-change interventions. Theoretically, nonlinear relations of interest (such as the curvilinear and interaction relationships in the current results) indicate that the influence of different qualities of motivation on behavioral engagement depends on their absolute and relative (to other motivation types’) quantity. The current study demonstrated a curvilinear effect of controlled regulation on exercise engagement (frequency)—it is possible that this effect would not exist for well-being or other type of outcome, however. For example, although individuals with some controlled regulation exercised more than individuals with very low/no controlled regulation (for a given level of autonomous regulation), they may not have reaped as many positive mental benefits from their exercise.
The degree of variation in responses on autonomous and controlled regulations indicates that there may not be clusters of individuals on these regulations with respect to exercise. It may be that there are more clearly defined and discernable groups in academic domains (Ratelle et al., 2007; Vansteenkiste et al., 2009).
Regarding arguments that a “person-centered” approach to data analysis is preferable to a dimensional approach for determining individual difference factors in behavior, the current study has demonstrated that these arguments are not always supported. Both approaches can evaluate complex, conditional, or interactive effects of different regulation types on outcomes; but dimensional approaches have the additional advantage that they optimize statistical power and maintain conceptual distinctions between individuals who may characterize a group to more or less of an extent. Furthermore, with regard to basic research required for optimal intervention design and targeting, whereas person-centered analyses may identify high-risk individuals, dimensional analyses such as polynomial regression can do the same but without assuming that these motivational profiles are necessarily stable and definable enough to warrant giving some individuals and not others an intervention. Dimensional or variable-centered analysis does not necessarily overlook the individual and can be a valuable and even preferable approach to evaluating individual differences in motivational “profiles” via a person-centered analysis.
Although polynomial regression holds some advantages over cluster analyses in the case of two predictor dimensions, as addressed in the current study, profile-based analyses may have an advantage in the case of more than two predictor dimensions, in terms of interpretability of results (i.e., condensing complex relationships into identifiable trends that can be statistically evaluated). Because SDT posits at least five potentially distinguishable types of behavioral regulation, it is possible that the two overarching dimensions (autonomous and controlled regulation, as commonly used in the literature) should not be used in every case for characterizing the data. Polynomial regression can have more than two predictors, but methods for interpreting and statistically evaluating the resultant multidimensional relationships do not exist, to the authors’ knowledge. If the motivation items do not factor into autonomous and controlled regulation, the researcher could instead use polynomial regression to evaluate the interdependent effects of intrinsic and extrinsic regulation or of intrinsic/autonomous and introjected regulation, for example, where appropriate given a priori hypotheses. Another advantage of profile-based methods is that multivariate analyses, such as MANOVA, are easily conducted with determined profiles of individuals, but polynomial regression is conducted on a single outcome at a time. Although it is mathematically possible to use polynomial regression for multiple outcome variables, guides to researchers on how to conduct such analyses do not exist.
The current study focused on individuals’ type and degree of behavioral regulation, even though SDT encompasses other relevant constructs for behavior prediction. Sheldon, Ryan, Deci, and Kasser (2004) have shown that goal contents and motives have independent effects on psychological well-being. Therefore, there may be independent and even interdependent effects of these elements on individuals’ exercise behaviors and outcomes, but polynomial regression and cluster analysis would be limited in their ability to evaluate these hypotheses. Both types of analyses are also limited in evaluating change within person over time in the interdependent effects of behavioral regulations on outcomes of interest. Research has shown that absolute and relative levels of different regulations change over time (Mata et al., 2009; Rodgers, Hall, Duncan, Pearson, & Milne, 2010), and so it is possible that the interdependent effects of regulations change over time (e.g., with behavioral experience). Future research could utilize polynomial regression to evaluate differences in the interdependent effects of autonomous and controlled regulation on behavior for those in an initiation versus maintenance stage or could be used to evaluate individuals before and after receiving a behavioral intervention.
Although the current results do not necessarily suggest techniques be developed to promote controlled regulation of behavior, they do suggest that focusing on autonomous regulation alone may not be the optimal strategy or outcome for interventions to develop regular exercise (and perhaps other behaviors, such as studying). Future research could empirically determine whether the optimal technique is to maximize autonomous regulation and to minimize controlled regulation versus promoting, to some degree, moderate levels of controlled regulation, for different behavioral domains (academic, exercise, environment conservation, etc.).
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
Supplemental Material
Supplementary material is available online with this article.
Notes
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
