Abstract
The transition from childhood to adolescence is a crucial period for the development of healthy behaviors to be sustained later in life. With obesity a leading public health problem, the promotion of healthy behaviors has the potential to make a huge impact. The current study evaluated Stage of Change progression in a large (N=4158) computer-delivered, Transtheoretical Model-tailored intervention focusing on physical activity (PA) and fruit and vegetable consumption (FV). Markov models were used to explore stage transitions and patterns of discrete change from sixth to ninth grade. Nested model comparisons examined the consistency of these patterns across time and intervention condition. Major findings supported models in which participants were free to transition forward and backward to any of the stages, but higher probabilities were observed for remaining in the same stage or for transitioning one- or two-stages forward. Participants in the intervention group had higher probabilities of transitioning towards Maintenance, with more change occurring relative to the comparison group during transitions from grades six to eight but not for grades eight to nine.
The transition from childhood to adolescence is a crucial period for the development of healthy behaviors and is marked by substantial change in many crucial behaviors, such as diet and physical activity (Kelder, Perry, Klepp, & Lytle, 1994). The U.S. Departments of Agriculture (USDA) and Health and Human Services (HHS) recommends that maintaining energy balance (i.e. the balance of calories consumed through eating and drinking relative to the number of calories burned through physical activity) serves to reduce the risk for chronic illnesses such as cardiovascular disease, cancer, osteoporosis, and diabetes. However, research has found that physical activity (Dumith, Gigante, Domingues, & Kohl, 2011) and the consumption of fruit and vegetables decreases (Lien, Lytle, & Klepp, 2001) as adolescents transition to young adulthood. Moreover, low levels of physical activity and fitness in adolescence, particularly aerobic fitness, have been linked to higher levels of cardiovascular disease risk (Hasselstrøm, Hansen, Froberg, & Andersen, 2002). Further, decreases in fruit consumption, hours of physical education, and frequency of sports participation were associated with higher increases in standardized scores of Body Mass Index (Haerens, Vereecken, Maes, & De Bourdeaudhuij, 2010). Amidst the many changes adolescents experience during the transition to adulthood, the development of healthy behaviors signifies a modifiable mechanism to benefit overall wellness and prevent disease.
Given these observed changes, and in conjunction with the associations between certain health behaviors and chronic disease, adolescents represent an important population for intervention. Further, as being at risk for one unhealthy behavior tends to increase the odds of being at risk for another unhealthy behavior (Driskell, Dyment, Mauriello, Castle, & Sherman, 2008), the promotion of healthy behavior is an important public health issue. School-based energy balance interventions present vital opportunities for health professionals to make an impact on young people. Unfortunately, energy balance interventions in adolescents are few and have been met with mixed success (De Bourdeaudhuij et al., 2010; Ezendam, Brug, & Oenema, 2012; Lubans, Morgan, Callister, & Collins, 2009; Patrick et al., 2006; van Stralen et al., 2011).
Mauriello et al. (2010) developed a school-based computer delivered energy balance program that focused on promoting energy balance behaviors, including physical activity and fruit and vegetable consumption, in high school students. Velicer and colleagues (2013) replicated and extended these findings in middle school students. For each intervention, the guiding framework for behavior change was the Transtheoretical Model (TTM), and studies using this model have found that people vary in their general readiness to change their behavior, as well as in their attitudes and beliefs about that change, and that tailoring interventions based on these constructs leads to successful behavior change (Prochaska, 1983).
Behavior change is viewed as a temporal sequence of behavioral and cognitive modifications that individuals experience in their conscious efforts to become healthier (Martin, Velicer, & Fava, 1996; Prochaska & Velicer, 1997; Velicer et al., 2000). Within the framework of the TTM, intentional behavior change is conventionally represented as a progression through the five ordered Stages of Change: Precontemplation (PC), Contemplation (C), Preparation (P), Action (A), and Maintenance (M) (Prochaska, Redding, & Evers, 2008; Prochaska & Velicer, 1997; Prochaska, Wright, & Velicer, 2008; Velicer et al., 2000). Stage progression is a dynamic process similar to a punctuated equilibrium model such that behavior is characterized by long periods of stasis punctuated by periods of change (Reed et al., 1997). Behavioral interventions can serve to disrupt behavioral stasis and promote healthy change.
Velicer and colleagues (2013) found significant positive TTM intervention effects such that students who did not meet recommended behavioral criteria for physical activity or fruit and vegetable consumption at the beginning of the intervention were more likely to initiate energy balance behaviors, whereas students who were at criteria were less likely to relapse. As recommended by the United States Department of Health and Human Services (2001), students were considered to have met criteria if they engaged in at least 60 minutes of physical activity five days a week and have at least five servings of fruits and vegetables each day. The study developed two TTM-tailored, computer-delivered, multiple behavior interventions for implementation in New England middle schools. One intervention was designed to impact energy balance behaviors, including physical activity, fruit and vegetable consumption, and TV viewing, and the comparison intervention addressed substance use, including smoking and alcohol. Each intervention consisted of a self-directed computer-based program in which students complete a series of TTM-based assessments. After each assessment, students received stage-matched and individually tailored feedback messages based on their responses.
To better elucidate the process of behavior change, an examination of the patterns and transitions in stage change beyond traditional outcome measures (e.g. the likelihood of students at or not at criteria initiating behavior change) is warranted. This framework could inform future research on how students at different levels of readiness to change respond to intervention and help to determine specific patterns of change, including whether students in specific stages progress, regress, or stay in the same stage over the course of the intervention. Quantitative methodologies that focus on stage-sequential development are powerful methods for examining change in discrete, stage-like progressions. Markov chain approaches allow the modeling of change in a categorical variable over time and include the Manifest Markov Model, as well as the Latent Markov Model and Latent Transition Analysis (LTA), with the latter two methods incorporating measurement error (Kaplan, 2008). These longitudinal approaches model discrete change among subgroups of participants and are particularly useful for stage-sequential models in which people progress through ordered qualitative stages. These subgroups are considered to be dynamic statuses that people may move in and out of over time rather than stable classifications.
The basic Markov chain model, such as the more widely used LTA, estimates the proportion of individuals in each stage as well as the probability of either transitioning to another stage at the next time point or remaining in the same stage (Collins & Lanza, 2010; Velicer, Martin, & Collins, 1996). As such, information regarding stage membership and transitions at each time point can inform intervention efforts aimed at promoting healthy change. Three sets of parameters are estimated in LTA: 1) latent status (i.e. stage) prevalences, or delta parameters, which represent the prevalence of latent status s at time t; 2) item response probabilities, or rho parameters, which represent the probability of a given response rj,t to observed variable j, conditional on membership in a latent status s at time t; and 3) transition probabilities, or tau parameters, which represent the conditional probability of a transition to status s at time t + 1, given membership in status s at time t. For a thorough description of the LTA model, we refer interested readers to Collins and Lanza (2010) and Lanza, Patrick, and Maggs (2010). In the present study, each participant was assigned a stage by a computer algorithm at the time of measurement and given immediate, stage-matched feedback; therefore, for the purpose of this study, we treat each stage membership as a known status and do not incorporate measurement error as is done in the typical LTA. Thus, in order to focus on the patterns of transition probabilities, we fix all rho parameters such that each latent status is represented without error by a single observed variable output by the computer algorithm, and the model reduces to a manifest model (Kaplan, 2008).
In the context of determining longitudinal change in TTM-stage variables, Markov Modeling approaches are appropriate methods because they focus on discrete, rather than continuous, change. Few studies have examined TTM Stages of Change using this type of approach and this research has focused on stages of change for smoking cessation in adults using LTA (Guo, Aveyard, Fielding, & Sutton, 2009; Martin et al., 1996; Schumann, Meyer, Rumpf, Hapke, & John, 2002). Even less research has examined the effects of a behavioral intervention on stage transitions (Schumann, John, Rumpf, Hapke, & Meyer, 2006) or classification of other behaviors, like physical activity (Dishman et al., 2009) or condom use (Evers, Harlow, Redding, & LaForge, 1998). Therefore, an examination of the stage of change transitions using a Markov Modeling framework applied to novel behaviors will serve to inform and contribute to our understanding of single and multiple health behavior change.
The current study extends primary outcome findings of Velicer et al. (2013) by focusing on patterns of TTM stage change across time and intervention condition (i.e., energy balance or substance use prevention) for two behaviors (i.e. physical activity or fruit and vegetable consumption) through the examination of model based stage membership probabilities and transitions. Markov modeling was used to elucidate discrete longitudinal stage transitions in each intervention group across each behavior over the course of the study. This approach utilizes a model-comparison framework in order to promote informed and specific hypothesis testing. Thus, the primary research questions addressed are twofold: (1) What was the pattern of stage movement over time; (2) Did intervention condition affect stage transitions?
Methods
Sample
Students (N=4158) from 20 middle schools across Rhode Island participated in the study. Of the participants, 47.8% were female; with regard to ethnicity, 65.0% were white, 15.6% Hispanic, 3.8% Black, 2.4% Asian, 2.2% American Indian/Alaskan Native, 0.5% Pacific Islander, and the remaining were unknown or a combination of ethnicities. Briefly, students were randomized by school (described below), which were matched on available school-level data (e.g. percent free lunch eligible, percent English as second language, percent of students who go on to attend college, racial/ethnic composition, smoking rate, and alcohol use rate) and assigned to either the energy balance intervention or a comparison intervention (Velicer et al., 2013).
Intervention Design
As described in (Prochaska & Velicer, 1997; Velicer et al., 2000; Velicer et al., 2013), twenty schools were randomized to group using the multi-attribute utility measurement approach such that ten schools received an energy balance intervention, and ten schools received an alternate intervention (Graham, Flay, Johnson, Hansen, & Collins, 1984). The primary focus for the energy balance intervention was to increase or sustain physical activity and fruit and vegetable consumption, as well as reduce TV viewing. The primary focus for the alternate intervention was to prevent substance use acquisition for students who were not smokers or alcohol users, or to provide support and cessation information if they were. Each intervention was highly tailored based on TTM constructs (Mauriello et al., 2010); Redding et al., 1999; Velicer et al., 2013) and each intervention condition served as the comparison condition for the other, with both receiving comparable assessments and TTM-tailored intervention feedback using multimedia components for multiple behaviors. However, students in the energy balance intervention did not receive feedback on substance use behaviors, and students in the comparison intervention did not receive feedback on energy balance behaviors. Each intervention was disseminated through five 30-minute computerized TTM-tailored sessions including one in sixth grade, three in seventh grade, and one in eighth grade. A total of four assessments were completed by students in each condition early in each school year of the project (sixth, seventh, eighth, and ninth grades). More detail regarding study design and outcomes are published elsewhere (Velicer et al.,2013).
Measures
The current study focused on answering primary research questions with respect to stage transitions for two energy balance behaviors: physical activity, and fruit and vegetable consumption. The Stage of Change algorithm that was used for each behavior has been previously developed and validated (Mauriello et al., 2010). Criteria for Stage of Change were as follows: (1) precontemplation (PC; not meeting behavioral criteria and not planning to meet criteria in the next 6 months), (2) contemplation (C; not meeting behavioral criteria but planning to meet criteria in the next 6 months), (3) preparation (PR; not meeting behavioral criteria but planning to meet criteria in the next 30 days), (4) action (A; meeting behavioral criteria for less than 6 months), and (5) maintenance (M; meeting behavioral criteria for more than 6 months). Collectively, PC, C, and PR are considered “pre-action” stages, as they represent levels of readiness to change before action has been taken to modify behavior.
As mentioned previously, each computerized intervention provided students with immediate, stage-matched, and individually-tailored feedback. The results from the staging algorithm directly impacted the feedback provided at each time point and, consequently, the appropriateness and effectiveness of the intervention. As such, the present study used the algorithm-determined stage as a single indicator without error to examine intervention effects. To prevent stage misclassification, the staging algorithm used verification with specific behavioral criteria prior to asking about behavior intention. Multiple questions were used to verify the staging algorithm and achieve a reliable classification of stage instead of using multiple items to achieve reliability of latent status assignment, as is typical in LTA. Thus, the staging algorithm process was employed to achieve reliable classification rather than utilizing a latent variable, making a single indicator model preferable for these data. Specific criteria were based on recommendations from the United States Department of Health and Human Services (2001). For physical activity, this included participating in at least 60 minutes of physical activity for at least 5 days per week. For fruit and vegetable consumption, this included consuming at least five servings of fruits and vegetables each day. For example, if a participant did not report engaging in 60 minutes of physical activity at least five days a week, the participant was asked about his or her intention to engage in physical activity and would not be classified into A or M.
Markov Modeling Approach.
The traditional LTA model contains both measurement and structural components. The measurement component characterizes the discrete latent classes and the structural component determines the probabilities of status membership and stage transition. The structural component relies on autoregressive techniques to acknowledge a stochastic process in which repeated measures are linearly dependent on their own previous values. As mentioned earlier, three sets of parameters are estimated in LTA models: status prevalences (δ), transition probabilities (τ), and item-response probabilities (ρ). δ estimates represent the proportion of stage membership at time t. Statuses are mutually exclusive and exhaustive such that every individual is placed in only one stage at each time point. For a first-order model, in which each time point is related to the previous time point, τ estimates represent the probability of transitioning to a given stage at time t, conditional on stage at time t-1. As such, lower values represent a lower probability for a member of a given stage to transition (or to remain in the same stage). Hence, τ parameters reveal the underlying pattern of change and elucidate stage progression, regression, or stability. Finally, ρ estimates represent the probability of a response for a particular item, conditional on latent class membership at a specific time point. Thus, similar to factor loadings in structural equation modeling, ρ’s form the basis of latent status separation as they are used to indicate patterns of responses among discrete variables across time based on latent status membership. In LTA, high values represent high probabilities of a member of a given status endorsing a particular item. However, unlike factor loadings, ρ’s are probabilities and are scaled differently. Values close to 0 or 1 indicate latent status membership whereas values that are close to one divided by the number of response patterns indicate chance (Velicer et al., 1996).
For the current study, stage membership was calculated by a computer algorithm during the intervention to provide individually-tailored, iterative feedback to each participant at each time point. Hence, for all analyses in this study the ρ parameters were fixed to 0 or 1 to indicate each stage of change as determined by the algorithm, but we would like to emphasize that this approach can be easily adapted using LTA to include measurement error in stage classification (e.g. ρ estimation). As such, the current study focused on the estimation of stage membership probabilities (δ) and stage transition probabilities (τ), and the item-response probabilities (ρ) were fixed to 0 or 1 to indicate each stage of change as indicated by the staging algorithm that took place during the computerized intervention. The study took place over a four-year period beginning in sixth grade, with annual assessments and a follow-up assessment administered without intervention in ninth grade. Thus, three transition periods were the main focus of the study, representing the transition between sixth and seventh grade, the transition between seventh and eighth grade, and the transition between eighth and ninth grade.
Statistical Analyses
All statistical analyses were conducted in Mplus version 7 (Muthen & Muthen) and accounted for clustering within school with Robust Maximum Likelihood estimation (MLR). The full information maximum likelihood technique accounted for missing data due to attrition or missed school days when data were collected. Model fit was assessed using a number of criteria. Due to known limitations of the chi-squared distribution with large sample sizes and the large number of degrees of freedom (e.g. for an unconstrained model with five stages and four time points, the number of degrees of freedom totals 1,048,511), chi-squared goodness of fit statistics were not computed. Instead, the Akaike Information Criteria (Akaike, 1973), the Bayesian Information Criteria (Schwarz, 1978), and the Sample-size Adjusted BIC (Sclove, 1987) were primarily used to determine the best model, with lower values indicating better model fit. Because using multiple indicators of fit helps to guide the model fitting process, a Pseudo R2 measure was constructed based on McFadden’s Rho-squared (McFadden, 1974) to obtain an estimate of the percent variance explained by each model as another way of determining the best model. This pseudo R2 is calculated as 1 – (Log-LikelihoodModel/ Log-LikelihoodIntercept) in which the intercept model is a transition model where all stage movement was constrained to zero (i.e. a model with no stage transitions). In this way, the percent variance of each transition model, relative to a model with no transitions, could be examined. To date, few, if any, measures of effect size for Markov Models have been applied to determine variance accounted for in a given transition model. The use of this metric also provides an indication of the degree to which model parameters reflect an improvement in prediction from the intercept model. Thus, the application of McFadden’s Rho-squared used in this paper as an indicator of model fit and effect size is a novel development in Markov Modeling. The determination of the best fitting model was based on considering each of these criteria in conjunction with model parsimony.
Two overarching types of model comparison tests were assessed. First, intervention-specific stage movement patterns examined model parameters for each intervention condition separately for the behaviors examined. Then, intervention-effects on transition parameters examined model parameters using separate multiple-group models for the two behaviors (e.g., physical activity or fruit and vegetable consumption) with intervention as a grouping variable.
Intervention-specific stage movement patterns.
To determine the best pattern of stage movement, as well as to determine the stability of patterns over time, transition models were compared separately for each behavior within each intervention condition. A series of model comparisons were conducted to determine the best-fitting model in three steps. First, a freely estimated transition model was estimated (Model 1) with no restrictions on delta or tau parameters. Second, nested models with increasingly constrained tau parameters were compared with the free model to determine the best fitting pattern of stage movement. Model 2a restricted tau parameters to three or fewer stages forward and backward stage movement, Model 2b restricted to two or fewer forward and backward, Model 2c restricted to movement one forward and one backward. Third, to determine the stability of transition parameters, models with successive tau matrices held equivalent were compared to the freely estimated model (Model 1). Model 3a held transition parameters from grades six to seven equivalent to grades seven to eight, Model 3b held parameters from grades seven eight equivalent to grades eight to nine, and Model 3c held parameters from grades six to seven equivalent to grades eight to nine. If any two transitions were found equivalent, a follow-up model (Model 3d) added the third transition to test for equivalence across all grades.
Intervention-effects on transition parameters.
Similar to the approach above, a series of nested models were compared to determine the equivalence of model parameters across intervention condition for each behavior in three steps. First, a freely estimated multiple-group model (Model 4) was estimated. Second, sixth grade delta parameters were constrained to be equal across group (Model 5) to determine equivalence of stage distribution at the baseline assessment. Third, to identify which of the transition periods demonstrated differences in parameters estimates, each of the tau matrices for the three transitions were held equivalent across group (Model 6a for grades six to seven, Model 6b for grades seven to eight, and Model 6c for grades eight to nine).
Parsimonious model.
Finally, a parsimonious model (Model 7) was tested to integrate findings from the above approaches (i.e., intervention-specific stage movement models and intervention-effects on transition parameters). In this model, invariant transitions across time and across intervention condition were constrained to represent a final, reduced model
Results
Baseline demographics for each intervention condition are presented in Table 1. Stage membership probabilities for each time point in the free transition model with no constraints are presented in Table 2. The energy balance intervention consisted of 2,184 students and the comparison intervention consisted of 1,974 students. In sixth grade, about half of the students were in M for physical activity and about one quarter of students were in M for fruit and vegetable consumption.
Table 1.
Baseline demographics by group.
| Intervention Condition (N=2,184) | Comparison Condition (N=1,974) | Total (N=4,158) | ||||
|---|---|---|---|---|---|---|
|
|
|
|
||||
| % | N | % | N | % | N | |
|
|
|
|
||||
| Gender | ||||||
| Male | 52.2 | 1134 | 52.3 | 1028 | 52.2 | 2162 |
| Female | 47.8 | 1038 | 47.7 | 938 | 47.8 | 1976 |
| Race/Ethnicity | ||||||
| American Indian/Alaskan Native | 2.3 | 51 | 2.1 | 42 | 2.2 | 93 |
| Asian | 2.3 | 50 | 2.6 | 51 | 2.4 | 101 |
| Black, Not Hispanic | 3.1 | 67 | 4.6 | 90 | 3.8 | 157 |
| Pacific Islander | 0.4 | 9 | 0.7 | 13 | 0.5 | 22 |
| White, Not Hispanic | 64 | 1393 | 66.2 | 1303 | 65 | 2696 |
| Combination | 16.1 | 351 | 17.2 | 338 | 16.6 | 689 |
| Unknown/Not reported | 11.8 | 257 | 6.6 | 130 | 9.3 | 387 |
|
|
|
|
||||
| Mean | SD | Mean | SD | Mean | SD | |
| Age (range from 10–15) | 11.38 | 0.68 | 11.41 | 0.7 | 11.4 | 0.69 |
Note: Numbers do not always add up to the full sample size due to missing data.
Table 2.
Stage membership probabilities.
| Behavior | Time point | PC | C | PR | A | M |
|---|---|---|---|---|---|---|
| Energy Balance Intervention | ||||||
| Physical Activity | 6th grade | 0.08 | 0.13 | 0.22 | 0.09 | 0.48 |
| 7th grade | 0.05 | 0.13 | 0.21 | 0.19 | 0.42 | |
| 8th grade | 0.06 | 0.12 | 0.21 | 0.12 | 0.50 | |
| 9th grade | 0.06 | 0.14 | 0.17 | 0.19 | 0.44 | |
| Fruit and Vegetable Consumption | 6th grade | 0.12 | 0.23 | 0.34 | 0.03 | 0.27 |
| 7th grade | 0.10 | 0.21 | 0.33 | 0.16 | 0.20 | |
| 8th grade | 0.12 | 0.20 | 0.31 | 0.07 | 0.30 | |
| 9th grade | 0.14 | 0.24 | 0.30 | 0.11 | 0.20 | |
| Comparison Intervention | ||||||
| Physical Activity | 6th grade | 0.07 | 0.13 | 0.24 | 0.07 | 0.49 |
| 7th grade | 0.09 | 0.15 | 0.19 | 0.20 | 0.37 | |
| 8th grade | 0.12 | 0.12 | 0.21 | 0.18 | 0.37 | |
| 9th grade | 0.11 | 0.16 | 0.20 | 0.18 | 0.35 | |
| Fruit and Vegetable Consumption | 6th grade | 0.15 | 0.20 | 0.35 | 0.04 | 0.25 |
| 7th grade | 0.20 | 0.23 | 0.32 | 0.14 | 0.12 | |
| 8th grade | 0.24 | 0.24 | 0.29 | 0.11 | 0.11 | |
| 9th grade | 0.24 | 0.26 | 0.28 | 0.12 | 0.10 | |
Note: PC = precontemplation; C= contemplation; PR= preparation; A = action; M= maintenance.
Physical Activity
Intervention-specific stage movement patterns.
See Table 3 for results of stage movement pattern models specific to each intervention condition for physical activity stages. In the energy balance intervention, as indicated by McFadden’s Rho-squared, Model 1 accounted for the most variance (35%) relative to a stable model with no transitions. Indices of relative model fit (e.g. lower AIC, BIC, and SBIC) revealed that the free stage movement model (Model 1) was favored over more restricted movement models (Models 2a-c). Models with transition matrices held successively equivalent (Models 3a-d) demonstrated that, compared to the free transition model (Model 1), parameter estimates in each of the three transitions tended to be different across grades six through nine. It is noted that Model 3a has slightly lower BIC and SBIC values than Model 1, but accounts for the same amount variance (McFadden’s Rho-squared = 35%), and thus will be taken into consideration during final model fitting of the parsimonious model.
Table 3.
Model fit statistics for stage movement patterns within intervention condition for Physical Activity.
| Stage Movement Pattern | # FP | −LL | AIC | ΔAIC | BIC | ΔBIC | SBIC | ΔSBIC | Pseudo R2 | ||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Energy Balance Intervention | −13826.79 | ||||||||||
|
| |||||||||||
| Model 1 | Free movement | 89 | −8997.00 | 18171.99 | -- | 18678.19 | -- | 18395.42 | -- | 0.35 | |
| Model 2 | Stage movement | a: 3F, 3B** | 77 | −9410.96 | 18975.92 | 803.92 | 19413.86 | 735.67 | 19169.22 | 773.80 | 0.32 |
| b: 2F, 2B | 66 | −10117.77 | 20367.54 | 2195.54 | 20742.91 | 2064.73 | 20533.22 | 2137.80 | 0.27 | ||
| c: 1F, 1B | 49 | −11337.69 | 22773.37 | 4601.38 | 23052.06 | 4373.87 | 22896.38 | 4500.96 | 0.18 | ||
| Model 3 | Stage stability | a: τ grades 6–7 vs 7–8 | 69 | −9041.88 | 18221.75 | 49.76 | 18614.19 | −63.99 | 18394.97 | −0.45 | 0.35 |
| b: τ grades 6–7 vs 8–9 | 69 | −9139.61 | 18417.22 | 245.23 | 18809.66 | 131.48 | 18590.44 | 195.02 | 0.34 | ||
| c: τ grades 7–8 vs 8–9 | 69 | −9123.29 | 18384.57 | 212.58 | 18777.01 | 98.83 | 18557.79 | 162.37 | 0.34 | ||
| d: τ grades 6–9 | 49 | −9192.00 | 18481.99 | 260.24 | 18760.68 | 146.49 | 18605.00 | 210.03 | 0.34 | ||
|
| |||||||||||
| Comparison Intervention | −12668.84 | ||||||||||
|
| |||||||||||
| Model 1 | Free movement | 89 | −7877.22 | 15932.45 | -- | 16429.58 | -- | 16146.83 | -- | 0.38 | |
| Model 2 | Stage movement | a: 3F, 3B | 73 | −8253.01 | 16652.02 | 719.57 | 17059.78 | 630.20 | 16827.86 | 681.03 | 0.35 |
| b: 2F, 2B | 64 | −9035.75 | 18199.49 | 2267.05 | 18556.98 | 2127.40 | 18353.65 | 2206.83 | 0.29 | ||
| c: 1F, 1B | 49 | −10303.68 | 20705.37 | 4772.92 | 20979.07 | 4549.49 | 20823.40 | 4676.57 | 0.19 | ||
| Model 3 | Stage stability | a: τ grades 6–7 vs 7–8 | 69 | −7891.96 | 15921.92 | −10.52 | 16307.34 | −122.24 | 16088.13 | −58.70 | 0.38 |
| b: τ grades 6–7 vs 8–9 | 69 | −7888.33 | 15914.66 | −17.79 | 16300.08 | −129.50 | 16080.86 | −65.96 | 0.38 | ||
| c: τ grades 7–8 vs 8–9 | 69 | −7888.64 | 15915.28 | −17.17 | 16300.69 | −128.89 | 16081.48 | −65.35 | 0.38 | ||
| d: τ grades 6–9 | 49 | −7902.28 | 15902.56 | −19.37 | 16176.26 | −131.08 | 16020.58 | −67.54 | 0.38 | ||
Note: #FP = number of free parameters; −LL = log likelihood; AIC = Akaike information criterion; BIC = Bayesian information criterion; SBIC = sample size adjusted BIC;
= the best loglikelihood was not replicated after multiple iterations of start values; all Δ models compared to Model 1 of respective condition (except Model 3d, which is compared to Model 3c); Pseudo R2 (McFadden’s Rho-squared) is provided relative to a stable model with all transitions fixed to 0; Stage movment models (Models 2a-c)indicate the maximum number of stages allowed for forward (F) or backward (B) movement; Stage stability models (Models 3a-d) indicate which transitions (τ) are held equivalent; bolded, italic text indicates the lowest value of model fit; differences in number of free parameters between intervention conditions due to constraining transitions from PC/C/PR to M to zero in the comparison condition for each transition.
For the comparison intervention, indices of relative model fit revealed that a free transition model (Model 1) was favored and had the most improvement in fit (McFadden’s Rho-squared = 38%) over more restricted transition models (Models 2a-c). Unlike the energy balance intervention, however, models with successive transition matrices held equivalent across all grades (Models 3a-d) revealed lower fit compared to the free transition model. Model 3d, which held parameters across each transition equal, demonstrated the best fit and was the most parsimonious model. Thus the transition matrices were considered to be consistent across grades in the comparison condition but not the intervention condition.
Intervention effects on transition parameters.
Table 4 presents fit indices for multiple-group models testing equivalence of parameters across intervention condition. Relative fit indices for Model 5 revealed that baseline delta parameters could be held equivalent without negatively impacting model fit, signifying that baseline stage distribution was equivalent across intervention groups. For Models 6a-c, relative model fit indicated that holding transition parameters equivalent across intervention groups from grades six to seven, as well as for grades seven to eight, resulted in poorer model fit but not for transitions from grades eight to nine. This suggests that intervention differences between conditions from grades six to eight disappeared after the conclusion of the intervention in eighth grade, where they became equivalent to the comparison condition.
Table 4.
Model fit results for stage movement across intervention condition for Physical Activity.
| Stage Movement Pattern | # FP | −LL | AIC | ΔAIC | BIC | ΔBIC | SBIC | ΔSBIC | Pseudo R2 | ||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Model 4 | Free movement | 154 | −19746.21 | 39800.43 | -- | 40775.42 | -- | 40286.07 | -- | 0.33 | |
| Model 5 | Baseline δ equal across intervention | 150 | −19749.35 | 39798.71 | −1.72 | 40748.37 | −27.04 | 40271.74 | −14.33 | 0.33 | |
| Model 6 | τ equal across intervention | a: grades 6–7 | 138 | −19840.42 | 39956.83 | 158.13 | 40830.53 | 82.15 | 40392.02 | 120.28 | 0.32 |
| b: grades 7–8 | 138 | −19816.05 | 39908.1 | −48.73 | 40781.79 | −48.73 | 40343.29 | −48.73 | 0.33 | ||
| c: grades 8–9 | 138 | −19757.35 | 39790.69 | −117.41 | 40664.38 | −117.41 | 40225.88 | −117.41 | 0.33 | ||
| Model 7 | Parsimonious | 87 | −19813.29 | 39800.58 | 0.16 | 40351.39 | −424.03 | 40074.94 | −211.13 | 0.33 | |
Note: #FP = number of free parameters; −LL = log likelihood; AIC = Akaike information criterion; BIC = Bayesian information criterion; SBIC = sample size adjusted BIC; all Δ models compared to Model 4 of respective behavior; Pseudo R2 (McFadden’s Rho-squared) is provided relative to a stable model with all transitions fixed to 0; bolded, italic text indicates the lowest value of model fit.
Parsimonious model.
A final, parsimonious model (Model 7) was developed to integrate the findings based on the intervention-specific stage movement patterns (Models 1–3) and the intervention effects on transition patterns (Models 4–6). See Figure 1 for a summary of equivalence findings from stage movement pattern and transition pattern models estimated across time and intervention condition. In the parsimonious model, transition matrices within the energy balance intervention were allowed to vary from transitions from grades six to nine and transition matrices within the comparison intervention were constrained to equivalence across grades six to nine. The transition parameters across intervention condition for grades eight to nine were constrained to equivalence, while the transition parameters across intervention condition for grades six to eight were allowed to vary. Indices of relative model fit for the parsimonious model indicated that there was a very slight increase in AIC but a large a drop in the BIC and SBIC compared to the free, multiple-group transition model (Model 4). The parsimonious model maintained a McFadden’s Rho-squares of 33% and thus, we interpreted this model to be the best fitting model. We briefly note that because Model 3a in the intervention condition also demonstrated adequate comparable model fit to Model 1, we tested an alternate parsimonious model that held intervention condition transition probabilities for grades 6 to 8 equal; however, McFadden’s Rho-squared was lower (32%) and relative model fit indices were not improved so this model was not further considered.
Figure 1.
Summary of findings for transition matrices across grade transitions and intervention condition for physical activity and fruit and vegetable consumption.
Note: transition (tau) matrices were equivalent (≈) or non-equivalent (≠) across time or intervention condition. These findings informed the parsimonious model for each behavior.
Figure 2 presents a visual model of the transition parameter estimates for stage movement in the intervention condition for Model 7. In this depiction, the intervention parameter estimates are shown with each stage transition such that probabilities of remaining in a stage are presented inside circles, and probabilities of transitioning to another stage are presented along solid (for forward transitions) or dashed (for backward transitions) lines. The smaller, italic text represents accompanying parameter estimate represents the difference between the intervention and comparison conditions. For example, in Figure 2A, the probability of staying in PC for Physical Activity during the transition from grade six to seven is .24 and this value is .20 lower than the parameter estimate in the comparison condition, indicating that students are less likely to stay in PC at this time point if the receive the intervention than if they received the comparison intervention.
Figure 2.
Physical Activity stage movement from grades six to nine for intervention condition.
Note: PC = precontemplation; C = contemplation; PR = creparation; A = action; M = maintenance; solid arrows indicate forward stage movement, dashed arrows indicate backwards stage movement; values inside circles represent staying in that stage; smaller values in italics reflect the differences from comparison intervention parameter estimates (note that the transition from 8th-9th grade were constrained to equal)..
An examination of the transition estimates reveals that in the comparison condition, no participants transitioned from pre-action stages (PC, C, and PR) to M during any of the transitions. Participants in the intervention condition did, however, transition from pre-action stages to M, but only during transitions from sixth to eighth grade. The probability of transitioning into M from any of the other stages was consistently higher across all grades in the intervention condition. Backwards transitions were generally lower in the intervention condition, though the differences from the comparison intervention were fairly small. Probabilities with corresponding standard errors for all transitions in Model 7 are presented in Supplementary Tables 1-3.
Fruit and Vegetable Consumption
Intervention-specific stage movement patterns.
See Table 5 for model results of stage movement pattern models specific to each intervention condition. For the energy balance intervention, indices of relative model fit reveal that the free stage movement model (Model 1) was favored over more restricted movement models (Models 2a-c). As indicated by McFadden’s Rho-squared, Model 1 accounted for the most improvement in fit (35%) relative to a model with no transitions, compared with Models 2a-c. Next, models with successive transition matrices held equivalent (Models 3a-d) demonstrated that parameter estimates in each of the three transitions were significantly different when compared with the freely estimated model (Model 1). Overall, Model 1 explained the most variance and had the lowest values for model fit compared to Models 2–3, indicating that a free transition model fits best.
Table 5.
Model fit statistics for stage movement patterns within intervention condition for Fruit and Vegetable Consumption.
| Stage Movement Pattern | # FP | −LL | AIC | ΔAIC | BIC | ΔBIC | SBIC | ΔSBIC | Pseudo R2 | ||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Energy Balance Intervention | |||||||||||
|
| |||||||||||
| Model 1 | Free movement | 89 | −9866.72 | 19911.43 | -- | 20417.62 | -- | 20134.86 | -- | 0.35 | |
| Model 2 | Stage movement | a: 3F, 3B | 77 | −10297.53 | 20749.06 | 837.63 | 21187.00 | 769.38 | 20942.36 | 807.50 | 0.33 |
| b: 2F, 2B | 66 | −11034.97 | 22201.94 | 2290.51 | 22577.32 | 2159.70 | 22367.63 | 2232.77 | 0.28 | ||
| c: 1F, 1B | 49 | −12304.66 | 24707.33 | 4795.89 | 24986.01 | 4568.39 | 24830.33 | 4695.48 | 0.19 | ||
| Model 3 | Stage stability | a: τ grades 6–7 vs 7–8 | 69 | −9949.36 | 20036.72 | 125.28 | 20429.16 | 11.53 | 20209.93 | 75.08 | 0.35 |
| b: τ grades 6–7 vs 8–9 | 69 | −10008.02 | 20154.04 | 242.61 | 20546.48 | 128.86 | 20327.26 | 192.40 | 0.34 | ||
| c: τ grades 7–8 vs 8–9 | 69 | −9993.36 | 20124.72 | 213.29 | 20517.16 | 99.53 | 20297.93 | 163.08 | 0.35 | ||
| d: τ grades 6–9 | 49 | −10087.59 | 20273.18 | 236.46 | 20551.86 | 122.71 | 20396.18 | 186.25 | 0.34 | ||
|
| |||||||||||
| Comparison Intervention | −14246.07 | ||||||||||
|
| |||||||||||
| Model 1 | Free movement | 89 | −8839.94 | 17857.89 | -- | 18355.02 | -- | 18072.26 | -- | 0.38 | |
| Model 2 | Stage movement | a: 3F, 3B** | 73 | −9273.69 | 18693.37 | 835.49 | 19101.13 | 746.11 | 18869.21 | 796.95 | 0.35 |
| b: 2F, 2B | 64 | −10162.78 | 20453.56 | 2595.67 | 20811.05 | 2456.03 | 20607.72 | 2535.45 | 0.29 | ||
| c: 1F, 1B | 49 | −11467.91 | 23033.82 | 5175.93 | 23307.52 | 4952.50 | 23151.85 | 5079.58 | 0.20 | ||
| Model 3 | Stage stability | a: τ grades 6–7 vs 7–8 | 69 | −8854.29 | 17846.57 | −11.31 | 18231.99 | −123.03 | 18012.78 | −59.49 | 0.38 |
| b: τ grades 6–7 vs 8–9 | 69 | −8856.88 | 17851.76 | −6.13 | 18237.18 | −117.84 | 18017.96 | −54.30 | 0.38 | ||
| c: τ grades 7–8 vs 8–9 | 69 | −8846.32 | 17830.64 | −27.25 | 18216.06 | −138.96 | 17996.84 | −75.42 | 0.38 | ||
| d: τ grades 6–9 | 49 | −8865.05 | 17828.09 | −18.48 | 18101.80 | −130.20 | 17946.12 | −66.66 | 0.38 | ||
Note: #FP = number of free parameters; −LL = log likelihood; AIC = Akaike information criterion; BIC = Bayesian information criterion; SBIC = sample size adjusted BIC;
= the best loglikelihood was not replicated after multiple iterations of start values; all Δ models compared to Model 1 of respective condition (except Model 3d, which is compared to Model 3c); Pseudo R2 (McFadden’s Rho-squared) is provided relative to a stable model with all transitions fixed to 0; Stage movment models (Models 2a-c)indicate the maximum number of stages allowed for forward (F) or backward (B) movement; Stage stability models (Models 3a-d) indicate which transitions (τ) are held equivalent; bolded, italic text indicates the lowest value of model fit; differences in number of free parameters between intervention conditions due to constraining transitions from PC/C/PR to M to zero in the comparison condition for each transition.
For the comparison intervention, a free transition model (Model 1) was favored over more restricted transition models (Models 2a-c) with 38% improvement in fit, relative to a model with no transitions. Unlike the energy balance intervention, however, models with successive transition matrices held equivalent (Models 3a-d) revealed better relative fit compared to the free transition model and each had McFadden’s Rho-squared of 38%. Model 3d, which held parameters across each transition equal, demonstrated the best relative fit and was the most parsimonious model. Thus the transition matrices can be considered to be consistent across time in the comparison condition but not the intervention condition.
Intervention effects on transition parameters.
Refer to Table 6 for a presentation of fit indices from multiple-group models testing equivalence of parameters across intervention conditions. For Fruit and Vegetable Consumption, relative fit indices for Model 5 revealed constraining baseline delta parameters to equal across intervention group a drop in BIC and SBIC, but not the AIC, and McFadden’s Rho-squared remained the same. Since two of the fit criteria agreed and the variance accounted for did not change, we interpreted this to support that baseline stage distribution were relatively equivalent across intervention groups. For models 6a-c, relative model fit indicated that holding transition parameters equivalent across intervention groups from grades six to seven, as well as for grades seven to eight, resulted in poorer model fit but not for transitions from grades eight to nine. Consistent with findings from the physical activity analyses, this suggests that intervention differences from grades six to eight disappeared after the conclusion of the intervention in eighth grade, where the became equivalent to the comparison condition.
Table 6.
Model fit results for stage movement across intervention condition for Fruit and Vegetable Consumption.
| Stage Movement Pattern | # FP | −LL | AIC | ΔAIC | BIC | ΔBIC | SBIC | ΔSBIC | Pseudo R2 | ||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Model 4 | Free movement | 154 | −21578.75 | 43465.49 | -- | 44440.48 | -- | 43951.14 | -- | 0.33 | |
| Model 5 | Baseline δ equal across intervention | 150 | −21585.45 | 43470.91 | 5.41 | 44420.57 | −19.91 | 43943.94 | −7.2 | 0.33 | |
| Model 6 | τ equal across intervention | a: grades 6–7 | 138 | −21675.75 | 43627.51 | 156.6 | 44501.2 | 80.63 | 44062.7 | 118.76 | 0.33 |
| b: grades 7–8 | 138 | −21654.48 | 43584.97 | −42.54 | 44458.66 | −42.54 | 44020.15 | −42.54 | 0.33 | ||
| c: grades 8–9 | 138 | −21585.3 | 43446.6 | −138.37 | 44320.29 | −138.37 | 43881.79 | −138.37 | 0.33 | ||
| Model 7 | Parsimonious | 87 | −21645.49 | 43464.98 | 18.38 | 44015.79 | −304.51 | 43739.34 | −142.45 | 0.33 | |
Note: #FP = number of free parameters; −LL = log likelihood; AIC = Akaike information criterion; BIC = Bayesian information criterion; SBIC = sample size adjusted BIC; all Δ models compared to Model 4 of respective behavior; Pseudo R2 (McFadden’s Rho-squared) is provided relative to a stable model with all transitions fixed to 0; bolded, italic text indicates the lowest value of model fit.
Parsimonious model.
A final, parsimonious model (Model 7) integrated findings from the intervention-specific stage movement patterns (Models 1–3) and the intervention effects on transition patterns (Models 4–6). In this model, transition matrices within the energy balance intervention were allowed to vary across grades six through nine and transition matrices within the comparison intervention were constrained to equivalence across grades six through nine. The transition parameters across intervention condition for grades eight to nine were constrained to equivalence, while the transition parameters across intervention condition for grades six to eight were allowed to vary. In this model, indices of relative model fit revealed a slight increase in the AIC but a large drop in the BIC and SBIC compared to the free, multiple-group transition (Model 4). The parsimonious model maintained a McFadden’s Rho-squared of 33% and thus, we interpreted findings to indicate that Model 7 is the best fitting model.
Parameter estimates for the intervention condition for Model 7 are displayed in Figure 3, with differences from the comparison intervention displayed in smaller, italic text. An examination of these estimates reveals that in the comparison condition, no participants transitioned from pre-action stages (PC, C, and PR) to M during any of the transitions. Participants in the intervention condition did, however, transition from pre-action stages to M, but only during transitions from sixth to eighth grade during the intervention phase of the study. The probability of transitioning into M from any of the other stages was consistently higher across all grades in the intervention condition. Backwards transitions were generally lower in the intervention condition, though the differences from the comparison intervention were fairly small. Probabilities with corresponding standard errors for all transitions are presented in Supplementary Tables 1-3.
Figure 3.
Fruit and Vegetable stage movement from grades six to nine for intervention condition.
Note: PC = precontemplation; C = pontemplation; PR = preparation; A = action; M = aaintenance; solid arrows indicate forward stage movement, dashed arrows indicate backwards stage movement; values inside circles represent staying in that stage; smaller values in italics reflect differences from comparison intervention parameter estimates (note that the transition from 8th-9th grade were constrained to equal).
Discussion
This study represents a secondary data analysis focused on highlighting the mechanisms of behavior change. It is the first to systematically examine intervention effects on stage transitions in two important energy balance behaviors, physical activity and eating behavior. Major findings demonstrated that for both behaviors, a freely estimated transition model was favored when compared to models containing stage movement to one, two-, or three-stages forward or backward. Given the support for the free movement model, and considering that measurements were taken at yearly intervals, it is evident that a considerable amount of forward and backward stage movement occurred between intervention assessments. This is consistent with previous research that also favored models with higher numbers of stage transitions for smoking behavior from measures taken at yearly intervals (Schumann et al., 2006). Previous research using data taken at shorter intervals (i.e. six month) favored more restricted models with one- or two- stage movement patterns (Martin et al., 1996; Schumann et al., 2002). This is not surprising, considering that some stage movement patterns are constrained by the time requirements needed for an individual to be characterized in the A or M stages (i.e. the individual must maintain specific behavior at recommended criteria for less than or greater than six months, respectively). As such, a critical consideration for any longitudinal study is to ensure that the window of measurement is adequately capturing the process of change for the behavior being studied.
Another key finding from this study highlighted the intervention effects by demonstrating that the longitudinal pattern of transitions from grades six to nine in the comparison intervention (i.e., substance use prevention) was relatively stable across grade level, while the pattern of transitions in the intervention condition (i.e., energy balance) demonstrated a dynamic and changing transition pattern. This finding was observed for both physical activity and fruit and vegetable consumption behaviors, further outlining the success of the intervention in promoting behavior change. In addition, a successful behavioral intervention should promote positive change (e.g., towards M) and prevent negative change (e.g., movement towards PC). In the current study, the intervention condition for both physical activity and fruit and vegetable consumption demonstrated mostly positive differences in the forward moving arrows (see values in Figures 2 and 3 with solid lines on the upper arrows, differences from the comparison condition are presented in italic font) and a general trend of negative differences in transition probabilities for backwards stage movement in the intervention condition relative to the comparison condition.
Finally, a multiple-group comparison of transition parameters across intervention condition revealed that intervention effects declined after eighth grade and became comparable to comparison group transition, which coincided both with the conclusion of the intervention and with the transition from middle to high school. The underlying reasons for the diminished intervention effect by ninth grade are complex and may be confounded, as these changes may be due in part due to the cessation of the intervention in eighth grade but may also be due to changes in lifestyle, environment, and behavior associated with the transition from middle school to high school. For example, students who played athletics in middle school may not play in high school, thereby resulting in reduced physical activity. Similarly, the transition to high school may also produce diet changes, as students in high school have more independence and diverse options for lunch and snacks throughout the day.
A major limitation to this work was the lack of a true no-treatment control group that would provide an indication of normative or natural change. This limitation was also described in the primary outcome study, which highlights the choice in a two-treatment control comparison trial as more cost effective, maximized school participation, and met curriculum demands of participating schools (Velicer et al., 2013). In addition, future studies may consider using multiple measures to determine stage of change in order to incorporate measurement error and represent truly “latent” statuses. As discussed previously, the design of the current study precluded use of multiple indicators due to the nature of the built-in stage classification checks within the staging algorithm. However, a follow-up study that examined the stage classification using LCA could integrate the validity of the staging algorithm into the model. Finally, currently available fit indices, such as the AIC and BIC, have been criticized for their assumptions and reliability in real world data (Dziak, Coffman, Lanza, & Li, 2012) and tend to favor more complex models. Due to the nature of LTA, the models tested in this paper all contained extremely large numbers of estimated parameters. It is largely unknown how robust these indices are with complex LTA models. Moreover, the use of a pseudo R-squared value, such as McFadden’s rho-squared, has not been previously used with LTA and should be investigated further in future studies.
Findings from this study can serve as a springboard for the design of future interventions for physical activity and eating behavior interventions, or extend this method to other behaviors. Future research should focus on maintaining intervention effects after the intervention ends, especially during difficult life transition periods. Interventions should also focus on increasing the probabilities of staying in or transitioning into A or M and decreasing the probability of staying in or transitioning into PC. This is especially important for the transition from grades eight to nine, where no participants in either condition transitioned to M. Future studies may be designed to determine whether there are specific challenges for students during this transition and whether the effects of the intervention diminish after it concludes. Further, given the tendency for high probabilities for staying in PC, but low probabilities for transitioning into PC, interventions may strive to identify individuals who consistently remain in pre-action stages to tailor the intervention to their specific needs. For example, someone who is not ready to change and remains that way consistently may have different needs than an individual with a more unstable stage membership pattern. Similarly, a person who is unstable in A or M may need more feedback or help than a person who is consistently in M. In this way, interventions may be developed to be sensitive not only to whether a person is ready to change at a given time, but also to whether they tend to remain that way or fluctuate in their readiness. Interventionists may also be interested in examining the effects of covariates such as gender, socioeconomic status, and race/ethnicity, due to known differences in physical activity and diet behaviors that exist for many demographic subgroups.
The analytic approach presented in this work represents a step away from reliance on fitting of smooth, continuous, and linear trajectories used in many conventional methods by using a person-centered approach to place focus on categorization of discrete patterns of change. As is the case with stage-based behavior change, the true nature of some behaviors is best characterized by nonlinear movements or punctuated equilibrium. LTA-type approaches can provide critical information that permits researchers to improve future interventions by localizing change events in stage-based models and identifying different pathways to change. The method employed herein builds upon and extends previous work by Cleveland, Lanza, Bray, Turrisi, and Mallett (2013), which examines intervention effects as covariates, by utilizing a multiple-group approach to examine differences in movement patterns across intervention conditions. It also extends other LTA approaches that model multiple-group intervention effects at follow-up and subsequently test for differential responses across classes separately in a separate generalized linear model (Mackesy-Amiti et al., 2013). The model-based approach employed in this work allows for simultaneous estimation of all probabilities, placing clearer focus on the process of change by highlighting where the change occurs (e.g., high transition probabilities may indicate a period with lots of growth). Finally, the calculation of an index such as McFadden’s Rho-squared offers an indication of how much improvement in fit a proposed model offers, relative to a more basic, null-like model that can be determined by the research. As such, the model-comparison approach used here highlights the formulations of hypotheses for specific model parameters within and across multiple groups, facilitating a comprehensive way to test and compare transition models that would be useful when studying patterns of longitudinal behavior change.
Supplementary Material
Acknowledgments:
The ideas and opinions expressed herein are those of the authors alone, and endorsement by the authors’ institution or the NIH is not intended and should not be inferred.
Funding:
This paper was partially supported by Grants DA020112 from NIDA, T32MH019927 from NIMH, and Grant G20RR030883 from NIH.
Role of the Funders/Sponsors:
None of the funders or sponsors of this research had any role in the design and conduct of the study; collection, management, analysis, and interpretation of data; preparation, review, or approval of the manuscript; or decision to submit the manuscript for publication.
Footnotes
Article Information
Conflict of Interest Disclosures: Each author signed a form for disclosure of potential conflicts of interest. No authors reported any financial or other conflicts of interest in relation to the work described.
Ethical Principles: The authors affirm having followed professional ethical guidelines in preparing this work. These guidelines include obtaining informed consent from human participants, maintaining ethical treatment and respect for the rights of human or animal participants, and ensuring the privacy of participants and their data, such as ensuring that individual participants cannot be identified in reported results or from publicly available original or archival data.
References
- U.S. Department of Health and Human Services. (2001). The Surgeon General’s call to action to prevent and decrease overweight and obesity. U.S. Department of Health and Human Services, Public Health Service, Office of the Surgeon General, Rockville, MD. Available from: U.S. GPO, Washington. [PubMed] [Google Scholar]
- Akaike H (1973). Information theory and an extension of the maximum likelihood principle. In Petrov BN and Csaki F (Eds.), Second international symposium on information theory (pp. 267 281). Budapest: Academiai Kiado. [Google Scholar]
- Cleveland MJ, Lanza ST, Ray AE, Turrisi R, & Mallett KM (2012). Transitions in first-year college student drinking behaviors: Does drinking latent class membership moderate the effects of parent- and peer-based intervention components? Psychology of Addictive Behaviors, 26, 440–450. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Collins LM, & Lanza ST (2010). Latent class and latent transition analysis: With applications in the social, behavioral, and health sciences (Vol. 718): John Wiley & Sons. [Google Scholar]
- De Bourdeaudhuij I, Maes L, De Henauw S, De Vriendt T, Moreno LA, Kersting M, . . . Sjöstrom M (2010). Evaluation of a computer-tailored physical activity intervention in adolescents in six european countries: The activ-o-meter in the helena intervention study. Journal of Adolescent Health, 46(5), 458–466. [DOI] [PubMed] [Google Scholar]
- Dishman RK, Thom NJ, Rooks CR, Motl RW, Horwath C, & Nigg CR (2009). Failure of post-action stages of the transtheoretical model to predict change in regular physical activity: A multiethnic cohort study. Annals of Behavioral Medicine, 37(3), 280–293. [DOI] [PubMed] [Google Scholar]
- Driskell M-M, Dyment S, Mauriello L, Castle P, & Sherman K (2008). Relationships among multiple behaviors for childhood and adolescent obesity prevention. Preventive Medicine, 46(3), 209–215. [DOI] [PubMed] [Google Scholar]
- Dumith SC, Gigante DP, Domingues MR, & Kohl HW (2011). Physical activity change during adolescence: A systematic review and a pooled analysis. International Journal of Epidemiology, 40(3), 685–698. [DOI] [PubMed] [Google Scholar]
- Dziak JJ, Coffman DL, Lanza ST, & Li R (2012). Sensitivity and specificity of information criteria. The Methodology Center and Department of Statistics, Penn State, The Pennsylvania State University
- Evers KE, Harlow LL, Redding CA, & LaForge RG (1998). Longitudinal changes in stages of change for condom use in women. American Journal of Health Promotion, 13(1), 19–24. [DOI] [PubMed] [Google Scholar]
- Ezendam NP, Brug J, & Oenema A (2012). Evaluation of the web-based computer-tailored fataintphat intervention to promote energy balance among adolescents: Results from a school cluster randomized trial. Archives of Pediatrics & Adolescent Medicine, 166(3), 248–255. [DOI] [PubMed] [Google Scholar]
- Graham JW, Flay BR, Johnson CA, Hansen WB, & Collins LM (1984). Group comparability a multiattribute utility measurement approach to the use of random assignment with small numbers of aggregated units. Evaluation Review, 8(2), 247–260. [Google Scholar]
- Guo B, Aveyard P, Fielding A, & Sutton S (2009). Using latent class and latent transition analysis to examine the transtheoretical model staging algorithm and sequential stage transition in adolescent smoking. Substance Use & Misuse, 44(14), 2028–2042. [DOI] [PubMed] [Google Scholar]
- Haerens L, Vereecken C, Maes L, & De Bourdeaudhuij I (2010). Relationship of physical activity and dietary habits with body mass index in the transition from childhood to adolescence: A 4-year longitudinal study. Public Health Nutrition, 13(10), 1722. [DOI] [PubMed] [Google Scholar]
- Hasselstrøm H, Hansen S, Froberg K, & Andersen LB (2002). Physical fitness and physical activity during adolescence as predictors of cardiovascular disease risk in young adulthood. Danish youth and sports study. An eight-year follow-up study. International Journal of Sports Medicine, 23(S1), 27–31. [DOI] [PubMed] [Google Scholar]
- Kaplan D (2008). An overview of markov chain methods for the study of stage-sequential developmental processes. Developmental Psychology, 44(2), 457–467. doi: 10.1037/0012-1649.44.2.457 [DOI] [PubMed] [Google Scholar]
- Kelder SH, Perry CL, Klepp K-I, & Lytle LL (1994). Longitudinal tracking of adolescent smoking, physical activity, and food choice behaviors. American Journal of Public Health, 84(7), 1121–1126. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Lanza ST, Patrick ME, & Maggs JL (2010). Latent transition analysis: Benefits of a latent variable approach to modeling transitions in substance use. Journal of Drug Issues, 40(1), 93–120. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Lien N, Lytle LA, & Klepp KI (2001). Stability in consumption of fruit, vegetables, and sugary foods in a cohort from age 14 to age 21. Preventive Medicine, 33(3), 217–226. [DOI] [PubMed] [Google Scholar]
- Lubans DR, Morgan PJ, Callister R, & Collins CE (2009). Effects of integrating pedometers, parental materials, and e-mail support within an extracurricular school sport intervention. Journal of Adolescent Health, 44(2), 176–183. [DOI] [PubMed] [Google Scholar]
- Mackesy-Amiti ME, Finnegan L, Ouellet LJ, Golub ET, Hagan H, Hudson SM, … & Garfein RS (2013). Peer-education intervention to reduce injection risk behaviors benefits high-risk young injection drug users: a latent transition analysis of the CIDUS 3/DUIT study. AIDS and Behavior, 17(6), 2075–2083. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Martin RA, Velicer WF, & Fava JL (1996). Latent transition analysis to the stages of change for smoking cessation. Addictive Behaviors, 21(1), 67–80. [DOI] [PubMed] [Google Scholar]
- Mauriello LM, Ciavatta MMH, Paiva AL, Sherman KJ, Castle PH, Johnson JL, & Prochaska JM (2010). Results of a multi-media multiple behavior obesity prevention program for adolescents. Preventive Medicine, 51(6), 451–456. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Muthen LK, & Muthen BO (1998-2015). Mplus User’s Guide Seventh Edition. Muthen & Muthen: Los Angeles, CA. [Google Scholar]
- Patrick K, Calfas KJ, Norman GJ, Zabinski MF, Sallis JF, Rupp J, . . . Cella J (2006). Randomized controlled trial of a primary care and home-based intervention for physical activity and nutrition behaviors: Pace+ for adolescents. Archives of Pediatrics & Adolescent Medicine, 160(2), 128–136. [DOI] [PubMed] [Google Scholar]
- Prochaska JO (1983). Stages and processes of self-change of smoking: Toward an integrative model of change. Journal of Consulting and Clinical Psychology, 51(3), 390–395. [DOI] [PubMed] [Google Scholar]
- Prochaska JO, Redding CA, & Evers KE (2008). The transtheoretical model and stages of change. In Glanz K, Rimer BK, & Viswanath K (Eds.), Health behavior and health education: Theory, research, and practice (4 ed., pp. 170–122). San Fransico, CA: Jossey-Bass Inc. [Google Scholar]
- Prochaska JO, & Velicer WF (1997). The transtheoretical model of health behavior change. American Journal of Health Promotion, 12(1), 38–48. [DOI] [PubMed] [Google Scholar]
- Prochaska JO, Wright JA, & Velicer WF (2008). Evaluating theories of health behavior change: A hierarchy of criteria applied to the transtheoretical model. Applied Psychology, 57(4), 561–588. [Google Scholar]
- Schumann A, John U, Rumpf H-J, Hapke U, & Meyer C (2006). Changes in the “stages of change” as outcome measures of a smoking cessation intervention: A randomized controlled trial. Preventive Medicine, 43(2), 101–106. [DOI] [PubMed] [Google Scholar]
- Schumann A, Meyer C, Rumpf H-J, Hapke U, & John U (2002). Naturalistic changes in the readiness to quit tobacco smoking in a german general population sample. Preventive Medicine, 35(4), 326–333. [DOI] [PubMed] [Google Scholar]
- Schwarz G (1978). Estimating the dimension of a model. The Annals of Statistics, 6(2), 461–464. [Google Scholar]
- Sclove SL (1987). Application of model-selection criteria to some problems in multivariate analysis. Psychometrika, 52(3), 333–343. [Google Scholar]
- van Stralen MM, Yildirim M, te Velde SJ, Brug J, van Mechelen W, & Chinapaw MJ (2011). What works in school-based energy balance behaviour interventions and what does not? a systematic review of mediating mechanisms. International Journal of obesity, 35(10), 1251–1265. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Velicer WF, Martin RA, & Collins LM (1996). Latent transition analysis for longitudinal data. Addiction, 91(12s1), 197–210. [PubMed] [Google Scholar]
- Velicer WF, Prochaska JO, Fava JL, Rossi JS, Redding CA, Laforge RG, & Robbins ML (2000). Using the transtheoretical model for population-based approaches to health promotion and disease prevention. Homeostatis in Health and Disease, 40(5), 174–195. [Google Scholar]
- Velicer WF, Redding CA, Paiva AL, Mauriello LM, Blissmer B, Oatley K, . . . Prochaska JO (2013). Multiple behavior interventions to prevent substance abuse and increase energy balance behaviors in middle school students. Translational Behavioral Medicine, 3(1), 82–93. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.