Integrating the Transtheoretical Model: A Quantitative Analysis in the Area of Sun Exposure

The constructs involved in the Transtheoretical Model (TTM) have been shown to have similar relationships to the Stages of Change across a variety of behaviors. While much work has been done investigating the way the constructs interact cross-sectionally and longitudinally , a completely integrated look at all the constructs of the TTM has not yet been successful. This study integrated all constructs of the TTM related to increase of sun protective behavior across three time points. The sample used in this study is a portion of a sample collected for three larger , multiple behavior intervention studies. Assessments were collected at baseline , 6-, and 12-month intervals. At baseline, these larger samples included 1472 people in worksites, 1816 parents, and 3875 physician patients at risk for sun exposure. Of these, 341 worksite , 4 31 parent, and 1012 physician had data at all three time points with all necessary variables. Structural equation modeling was utilized to evaluate panel designs involving seven TTM constructs at baseline, 6-, and 12-month time points. Different models were run within each pre-action Stage of Change. Due to the complexity of the model , a step approach was taken to evaluate the relationships among the constructs. The Precontemplation group showed relationships between Experiential Processes, Pros and sun protective behavior. The Contemplation and Preparation samples both showed important relationships between Cons and Confidence with behavior. There were more significant paths in the Preparation model indicating greater variance possibly due to more stage movement at follow-up time points . While not all paths found confirm expectations based on the TTM, there was strong support for the theory. Additional work needs to be done to further investigate these relationships among individual stage transitions or transition groups. A better understanding of the empirical relationships between these constructs will help further understanding of the theory and improve interventions based on the TTM Aclmowledgements This project could not have been completed without the support of many individuals. The data was collected as part of a grant funded by NCI , P01-CA50087. Thank you to all those at the Cancer Prevention Research Center who made this data set possible and available for exploration. I'd also like to thank my committee for their help and suppo1i over the last few years. Thank you Colleen Redding, Joe Fava , and Deb Reibe. Most importantly I'd like to thank my major professor , Joe Rossi who encouraged me to shoot for the moon and helped me to get there. Not only did you all serve as sounding boards for my ideas, you encouraged me to push my thinking farther and expand my perceptions of the material. Perhaps the most influential people through this process were my family. To my mom who taught me to never give up, to my dad for encouraging me to reach higher than I ever thought I could , and to my sister and brother for their love and support, thank you. Most of all, I have to thank my husband Sean. Without his endless love and support I would have never completed this journey. Without their love and encouragement I could have never reached so far to achieve my dreams .


List of Tables
While this study will examine the relationships of the TTM in the area of reducing sun exposure, these previous studies will serve as a guide for predicting relationships.
The TTM is a stage theory . As Weinstein, Rothman &Sutton (1998) describe, a stage theory must (1) have a clear categorization system for the stages, (2) have a predefined order, and (3) show that common barriers to change occur in the same stage while different barriers to change occur in different stages. These three requirements have been adequately met for the TTM and will be described below.

Defining Stage
The stage of change construct has been the most influential aspect of the TTM.
It has created what Kuhn (1970)  People are thought to proceed tlu·ough a series of stages to bring about behavior change. Stage of Change , originall y developed in the area of smoking cessation , has also been adapted to the area of sun exposure Rossi, Redding , et al., 1995;Rossi , et al, unpublished data). Individuals who are not intending to improve their sun protective behavior in the next 12 months are considered to be in Precontemplation. Those people who are intending to change in the next year, but not in the next month are categorized as Contemplators. Individuals in the Preparation stage are intending to improve sun protective behavior in the next month. People in the Action and Maintenance stages are currently engaging in sun protective behavior. Those in Action have only been using sun protection for less than 12 months while those in Maintenance have engaged in the behavior for more than 12 months. The time frame associated with sun staging is 12 months, rather than 6 as seen other staging algorithms, to attempt to control for seasonal variations.

Ordering of Stage
By definition, the stages of change fall into a temporal order.
Precontemplation is followed by contemplation, preparation, action , and maintenance, respectively. However, people do not always flow linearly through this path . Instead , people are thought to travel in a spiral pattern through the stages (Prochaska , DiClemente, & Norcross , 1992). This pattern allows people to move both forward , ' progressing, and backward, regressing , through the stages .
A latent transition analysis showed that the best fitting model of change for stages involved regression to the previous stage and one and two stage progressions (Velicer , Martin , & Collins , 1996;Evers , et al., 1998). This pattern of change is consistent with the proposed spiral of change and the ordering of the stages.

Commonalities and Differences between Stages
The stages of change are seen as the organizing construct for the TTM. The other constructs show clear patterns across the stages.
Proces ses of Change . The processes of change were originally developed by Prochaska ( 1979) as a synthesis of psychotherapy techniques. Processes of change are behavioral, cognitive, and emotional strategies that are used to bring about change .
There have been ten processes of change consistently found for a variety of behaviors (Rossi , 1992). These processes are divided into two hierarchical concepts , experiential and behavioral processes (Prochaska & DiClemente, 1983). Experiential processes include consciousness raising , dramatic relief , environmental reevaluation , self reevaluation, and social liberation. Behavioral processes include self liberation, stimulus control , counter conditioning , reinforcement management , and helping relationships.
The processes of change are expected to differ across the stages of change.
Cross-sectional data have shown that consciousness raising is highest in contemplation , while self-reevaluation is high in both contemplation and action.
Action has the highest values for stimulu s control , counter conditioning , reinforcement \ management , self-liberation , and helping relationships. Maintenance continues to have a high level for both stimulus control and counter conditioning. (Prochaska & DiClemente, 1983). Therefore, experiential processes are expected to be important in early stages, while behavioral processes are more important in later stages (Prochaska, DiClemente, & Norcross, 1992). A cross-sequential analysis confomed this expectation (Prochaska, et al., 1991 ). Experiential processes peaked in contemplation and behavioral processes peaked in action . Thus, through cross-sectional and crosssequential data, the processes of change have been shown to be related to the stages of change.
Decisional Balance. The decisional balance construct was originally developed based on Janis and Mann's ( 1977) theory of decision making. This construct was designed to measure the perceived benefits and barriers to making a behavior change. In order for change to occur, the pros must outweigh the cons of behavior change. In both cross-sectional and cross-sequential analysis, the cons of behavior change have been found to outweigh the pros of behavior change for precontemplators (Rossi, Blais, Weinstock, & Redding, 1995;Velicer et al., 1996;Velicer, et al., 1985). The pros of behavior change begin to increase, while the cons of behavior change begin to decrease in contemplation and continue their trajectories of change through the stages. This relationship of cons of change decreasing across stage , while pros of change increase across stages has been found in at least 12 different behaviors (Prochaska et al., 1994). Therefore, the decisional balance is not only a good measure showing stage differences for smoking, but it also generalizes to a variety of other health behaviors, including sun protection.

'
Confidence. Bandura (1977) originally developed self-efficacy as a measure of a person ' s confidence to do a certain behavior. Self-efficacy for sun exposure has been applied to two different situations , confidence to use sunscreen and to avoid sun exposure (Maddock et al., 1998). This situational self-efficacy for sun exposure has been found to increase across the stages of change (Rossi, et al., 1998). and an increase in self-efficacy and helping relationships.
A second longitudinal test of the model tested a priori predictions of differences on the TTM constructs based on stage movement .  ( 1996) proposed two models for how the constructs could be related . The processes of change are seen as the independent variables in the TTM (Prochaska , et al., 1991 ). Another study examined these proposed models for smoking cessation more closely . Blais (1993) examined the assumption that the processes of change are the independent variables in the model. Blais again used panel designs to test whether processes predicted pros or cons of smoking or if pros and cons of smoking predicted processes. By testing stage transitions she found that processes of change sometimes predicted the pros or cons of smoking , but pros or cons of smoking never predicted processes . Specifically, she found that experiential processes predicted change in cons of smoking [pros of behavior change] for people moving from precontemplation to contemplation. She also found that both experiential and behavioral processes predicted pros of smoking [ cons of behavior change] for the contemplation to preparation progression. Thus , the assumption that processes predict cognition and/or behavior appears to be supported. However , Blais was not able to truly test the relationship of behavior to processes and cognition due to the lack of a continuous outcome measure in the area of smoking.
These studies give some support for integrating the TTM , however, there are still serious questions on exactly how the constructs are intenelated. The original models proposed by Velicer et al. (1996) have yet to be fully tested . Blais' work shows that the processes start the process of change. But whether cognition mediates the relationship with behavior or behavior mediates the relationship with cognition is still unclear.

Hypotheses
A study where all the constructs are incorporated into an integrated model is needed to determine which model appropriately represents the change process . The ultimate goal of this study is to create a parsimonious structural model that represents the relationships among the constructs of the Transtheoretical Model. In order to answer this question more fully , stage differences will also be examined.
First , differences within each of the pre-action stages of change will be investigated. Given that the TTM is a stage model , the relationships between the constructs are expected to be different for different levels of stage. Thus stage is predicted to act as a moderator to the relationships between the constructs. For this reason , the sample will first be divided into separate samples based on the baseline stage of change and analyzed separatel y. Previous studies help form predictions about which variables may be important within each stage.
Precontemplation. The following associations related to people in Precontemplation have been discussed above: • Progression from precontemplation was associated with an increase in the pros of behavior change and an increase in self reevaluation, an experiential process of change (Prochaska, et al., 1991 ).
• The pros of behavior change are important for movement out of precontemplation .
• The pros of behavior change are expected to predict change in beh avior for early stages , however this prediction was not able to be evaluated (Velicer, et al., 1996).
• Experiential processes of change predicted change in the pros of behavior change (Blais , 1993).
Therefore , it is expected that experiential processes of change will influence the pros of beh~vior change , which will then influence behavior in this sample. There are no clear predictions based on previous research as to where self-efficacy might fit into this model. Figure 1 shows the predicted panel design for Precontemplators .
Contemplation. The following associations related to people in Contemplation have been discussed above: • Progression from Contemplation was associated with an increase in pros of behavior change and self-efficacy (Prochaska, et al., 1991).
• Cons of behavior change and self-efficacy become important in contemplation .
• Change in behavior predicts change in the cons of behavior change while pros of behavior change are expected to predict change in behavior (Velicer, et al., 1996).
• Both experiential and behavioral processes of change predict cons of behavior change for contemplation (Blais, 1993 ).
Therefore , previous research predicts that both experiential and behavioral processes will be important for predicting behavior , self-efficacy , and the pros of behavior change. Behavior is also expected to predict the cons of behavior change. The significant difference between the models for precontemplation and contemplation is the addition of the behavioral processes as independent variables and the importance of self-efficacy into the model. Figure 2 shows the predicted panel design for contemplators.
Preparat ion. The following relationships have been discussed above for the Preparation stage of change: • \_ Cons of behavior change and self-efficacy become important in later stages , such as preparation .
• Behavior predicts changes in both the pros and cons of behavior change (Velicer, et al., 1996).
• Behavioral processes of change, cons of behavior change, and self-efficacy are important for movement into action (Prochaska, et al., 1991 ).
Additionally, there are expected to be more relationships in general in this stage group than in the other two due to the transitional nature of the Preparation stage. Therefore , for people in Preparation, the behavioral processes are expected to predict behavior, which will influence both the pros and cons of behavior change. Self-efficacy will also be important in this stage, but whether it predicts or is predicted by behavior is not clear from previous Transtheoretical Model research. However, the concept of reciprocal determinism from the Social Cognitive Theory predicts that confidence both predicts and is predicted by behavior. Figure 3 shows the predicted panel design for people in preparation including only those paths specifically predicted.

Methodological Issues
While most of the predictions have support from previous research , there are a few problems in testing these models of behavior change. The pros of behavior change and some processes of change are curvilinear across stages . Since most statistical analyses assume linear relationships , relationships that are really there may not be detected due to the inadequacies of the statistical techniques . By dividing the sample into stage of change groups, it is hoped that most of the curvilinearity will be controlled.

Method Participants
The sample used in this study is a portion of a sample collected from three larger, multiple behavior intervention studies designed to increase smoking cessation, sun protective behavior, and eating a low-fat diet. Assessments were collected at baseline, 6-, and 12-month intervals. The three samples combined here were drawn from larger samples collected from a worksite, parent, and physician office populations. At baseline, these larger samples included 1474 people from worksites, 1816 parents, and 3875 physician patients at risk for sun exposure. Individuals were considered at risk if they were in either the Precontemplation, Contemplation, or Preparation stage of change. Of these , 341 worksite, 451 parent, and 1012 physician participants had data at all three time points with all necessary variables. The number of participants with complete data is low due to the design of the intervention study, which only collected Process of Change items from the treatment group. The tlu·ee samples combined are 59.8% female, 76.6% married and 94.5% white. Of this subpopulation, 32.6% (588) are in Precontemplation, 21.5% (388) are in Contemplation, and 45.9% (828) are in Preparation at baseline. Table 1 presents demographic variables broken down by the three samples. While there were significant differences between samples, these differences were expected due to the channels of recruitment.

Measures
Stage of Change. Stage of change is measured using an algorithm of several questions that assess behavioral intentions and actions for reducing sun exposure. An additional stage measure for use of SPF 15 sunscreen has also been developed. Both algorithms have been well developed across more than 10 different study populations .
Processes of Change. The processes of change are cogniti ve, emotional , and behavioral strategies utilized to bring about change. The application of the processes of change to reduction of sun exposure is under development, however , initial results show the familiar hierarchical structure representing the processes of change organized on two higher order factors , experiential and behavioral (Maddock , et al., 1998;Rossi, 1992). Due to the comple x nature of the measurement models, scores for the processes of change will act as the manifest variables making up the latent construct of the higher order experiential and behavioral processes of change for these models.
Decisional Balance. The decisional balance inventory is an 8-item scale measured on a 5-point Likert scale consisting of two subscales, the Pros and Cons of reducing sun exposure. The Pros represent positive aspects of changing behavior that can help facilitate change. The Cons represent negative aspects of changing behavior considered barriers to change. The decisional balance measure shows good internal consistency with Cronbach alphas of .78 for Pros and .74 for Cons. Decisional balance for reducing sun exposure has been found to show patterns similar to other behaviors across the stages of change (Maddock , et al., 1998;Prochaska , et al., 1994).
Confidence. Situational self-efficacy is a 7-item scale scored on a 5-point Likert , scale measuring level of confidence to practice sun protective behaviors in challenging situations. A hierarchical model structure of one general factor with two primary factors , sunscreen use and sun avoidance , best fits previous data (Maddock , et al., 1998). Both the general measure of confidence (a=.84) and the two subscales , sunscreen use (a= .88) and sun avoidance (a=.76) show good internal consistency .
Situational self-efficacy for sun exposure has also been found to have a strong relationship with the stages of change (Rossi, et al., 1998).
Sun Protection Behavior Scale. There has been no clear outcome measure yet agreed upon, nor any objective biochemical measure of ultraviolet radiation exposure that can easily be employed on a large scale intervention (Creech & Mayer, 1998).
Therefore, self-report measures have been relied upon in most studies . Consensus has been reached on those aspects of behavior that are recommended for effective reduction of UVR exposure. They include using SPF 15 sunscreen, avoiding sun exposure during peak hours, staying in the shade, and wearing sun protective clothing (Consensus Development Panel, 1991;Rossi, 1989;. The Sun . Protection Behavior Scale (SPBS) was developed to assess the level of use for these sun protective behaviors. The 9-item scale has three factors , sunscreen use, sun avoidance, and hat use. Internal consistency for these three scales has been found to be good for adult samples : sunscreen use (a=.86), sun avoidance (a=.82) , and hat use (a=.83). The SPBS has also been found to have a strong relationship to the stages of change (Maddock, et al., 1998;Rossi, et al., 1998;Weinstock , et al., 2000).

Procedure
The three samples were collected as part of larger intervention studies designed to reduce risk for smoking, diet, and sun exposure. In each sample there was another arm of intervention , either at the school, worksite, or physician office level. The treatment group analyzed in this study refers to that of the home-based expert system treatment classification , regardless of the other arms or levels of intervention.
The parent sample was collected with parents of students participating in similar high school interventions. One parent or guardian per student was recruited for the study with a 75% participation rate. The parents were randomly assigned to either home-based expert system intervent ion or no-treatment control. At the 12month follow-up , 92% of the parents were retained .
The worksite sample was implemented with employees from 24 worksites (12 treatment and 12 control). Three companies (1 treatment and 2 control sites) dropped out of the study . Independent of the worksite treatment classification, individuals were randomly assigned to either home-based expert system intervention or notreatment control condition. Of the 86% of employees screened, 78% of the eligible employees were recruited for the study. At the 12-month follow-up, 84% of the employees were retained.
The physician sample was collected from 78 physician offices. Individuals were randomly assigned to either home-based expert system treatment or no-treatment control. Of those eligible for participation , 91 % completed the baseline survey.
Approximately 80% of the sample completed the 12-month follow-up.
For all three samples, only the treatment group was asked the Processes of Change items . Therefore all analyses will be conducted on only the treatment group from each study.

\ Analysis Procedure
Structural equation modeling will be utilized to evaluate the panel designs discussed above . Structural equation modeling (Bentler, 1980) is a statistical technique that simultaneously estimates relationships between multiple constructs across time.
Due to the complexity of the model , a step approach will be taken to evaluate the relationships among the constructs (Evers and Harlow, 1997;Evers, 1998).
Step 1 involves modeling all the constructs within a single time point. This step will be done on the overall baseline sample.
Step 2 models each individual construct across the three time points.
Step 3 models evaluate longitudinal cross-lags for each pair of variable subsets.
Step 4 brings together information learned in the previous steps to pull the full model together. Steps 2 through 4 will be done within each baseline stage of change group. A final step will look at the invariance of the models based on stage movement within the stage of change groups.
Step 1 involves an exploration of the measurement structure of each of the constructs. Due to the complexity of the full models, simple, strong measures of the constructs are needed. Nunnally ( 1978) showed that a minimum of three items are needed to adequately represent a construct. Therefore, the best three items or parcels will be used to represent each measure. All TTM constructs will be evaluated in the baseline sample to confirm previously found structure. These measures will then be reduced to three manifest variables per construct. The reduced measurement structure will be confirmed in the sample of participants who have complete longitudinal data (N=l804). A final analysis in Step 1 will integrate all constructs into one model.

For
Steps 2 tlu·ough 4 the sample will be categorized by baseline stage of change. These three groups, Precontemplation (N=588), Contemplation (N=388), and Preparation (N=828), will be treated as separate samples. The following analyses will be run in each of these three samples.
Step 2 will look at each construct independently across time to investigate the nature of the relationship . Two models will be run, one including all paths, including one between baseline and 12 month, and one with that path removed. This will determine if the direct paths between baseline and 12 months are necessary in the overall models. These models will be run independently for each of the constructs within each of the three baseline stage of change samples .
Step 3 will investigate the relationships among constructs. Cross-lag models between two constructs will be run within each of the stage of change samples. These models will be run for each pairwise combination of constructs. Non-significant paths within each model will be removed and the reduced model will be analyzed again.
Chi-square difference tests and other measures of fit will determine whether specific paths are important in the overall model.
Step 4 will integrate the information found in the previous steps. Within each stage of change sample, all significant paths from previous steps will be included into one overall model including all constructs at all three time points. The overall models for each stage group will be evaluated. Several indices will be utilized to suggest model modification, including the Wald test , Lagrange Multiplier test, significance of \ proposed paths, and overall fit of the model. Careful consideration will be given prior to adding or subtracting paths in the proposed models.
These final three models will then be compared to investigate differential relationships among constructs based on baseline stage of change . There are several ways to compare models across the different stage of change populations (Maruyama, 1998). Most simply, the same model could be run on each stage of change sample and then fit indices and parameter estimates compared. Also, models within each stage of change could be modified to best fit that sample and then reduced models compared as to whether the basic processes seem to be the same. Confidence intervals could be calculated around the parameter estimates to determine if estimates in the different samples are significantly different. These two methods of comparing models are limited in that there is no direct comparison of the goodness of fit between the different samples. Also, if parameter estimates are found not to differ by confidence interval estimates, it does not necessarily mean that they are equal between the samples. A third method of comparing models between multiple samples simultaneously estimates a single theoretical model on multiple samples. This method allows for direct comparison of fit indices and parameter estimates. While the simultaneous sample analysis may yield the strongest statistical support for empirical 1, i I l estimates. Overall fit will not be able to be evaluated since the groups have different samples sizes.

Results
Step 1: Measurement Structure Step 1 examined the measurement models of each TTM construct to find the best fitting reduced model to use in later analyses . The constructs were first examined with all participants who had baseline data in the treatment group (N=3427). The simplified measurement models were then compared on a smaller sample of those participants that also have complete longitudinal data (N= 1804 ). This second analysis is to confirm that the measures are appropriate for the participants that will be included in later analyses. Each construct has slightly different sample sizes due to missing data in the full baseline sample.
Processes of Change. Three processes from each of the experiential and behavioral constructs were selected, Consciousness Raising , Dramatic Relief , and Self Reevaluation for experiential processes and Stimulus Control, Reinforcement Management , and Self Liberation for behavioral processes . These processes showed strong factor loadings in a full item model (see Appendix A). To simplify the measurement model for further analyses , scale scores were computed by summing the two items for each process . This model shows excellent fit in both the full baseline x_2 (8) Confidence. Scale scores of the two Confidence subscales caimot easily be used to represent the confidence construct since three markers are needed for stable estimates of latent constructs (Nunnally , 1978). Therefore, the three best representative items of general confidence were selected , including one item from the sunscreen use factor , and two from the avoidance factor. The factor loadings show adequate strength , although the sunscreen use item is low. However, since there are zero degrees of freedom for these models there are no fit indices .
Behavior. In addition to two subscales not being sufficient representations of a construct, as discussed above, the correlation between the two behavioral subscales is very small, indicating that one general behavior factor is inappropriate. Therefore, both subscales will be incorporated into the full model and tested independently in   Table 2 lists the correlations between the factors. The highest correlation after that between Experiential and Behavioral Processes is between Behavioral Processes and Confidence. All correlations , except that between Pros and Cons , are statistically significant. See Appendix A for more details on the process for determining the final measurement structure.
Step 2: Longitudinal Stability This step evaluated each construct independently across time (baseline, 6-, and 12-months) to investigate the nature of the longitudinal relationship. The path between the constructs at baseline and 12-months was examined to determine if this path was necessary to best represent the data. The enor for each item was also conelated between all time points. All seven constructs were run independently in the three stage samples.
Precontemplation . All seven constructs show excellent fit across the three time points. The measurement structure for all constructs is similar to that found in Step 1. Adjacent time point stability estimates range from .543 to .781. The Chisquared difference test for all constructs showed a significant (p<.01) decrease in fit when the baseline to 12-month path was removed from the model, indicating the need to retain this path in future models. These paths ranged from .160 to .288.
Contemplation. The seven constructs again show good fit across time for all seven TTM constructs with similar measurement structure as found in Step 1. Stability at adjacent time points range from .480 to .768 . The Chi-squared difference test for the Experiential Processes, Pros , Cons, Confidence, Avoidance, and Sunscreen Use constructs showed a significant decrease in fit when the baseline to 12-month path was removed from the model, indicating the need to retain this path in future models. \ This path was not necessary for model fit for the Behavioral Processes , with a non -significant Chi-squared difference test. The significant paths ranged from .194 to .283.
Preparation. The measurement models in this sample are like.those found in Step 1 and all show excellent fit across the time points. Adjacent time point stability paths range from .513 to .778. The Chi-squared difference test for all constructs showed a significant (p<.01) decrease in fit when the baseline to 12-month path was removed from the model. These paths ranged from .162 to .286.
Appendix B presents models from this step in greater detail and discusses the correlations between item error estimates . In general, the seven constructs showed consistent measurement structure, reasonable stability across the three time points, and indicate a need for estimating paths between baseline and the 12-month time point.
Step 3: Pairwise Cross-lagged Models Step 3 examined the relationships among pairs of constructs. Cross-lag models between each combination of two constructs were run within the three stage of change samples. Non-significant paths within each model were removed to produce a reduced model. Chi-square difference tests and other measures of fit were used to detennine whether specific paths were retained for future models.
Precontemplation. Three specific predictions were made for this stage sample.
First , Experiential Processes were expected to act as independent variables , specifically predicting Pros. Consistent with expectations, increases in Experiential Processes at both baseline and 6-months predict increases in the Pros of behavior change at subsequent time points. Second, the Pros of behavior change were expected to predict behavior. This relationship is less clear. High levels of Pros at baseline significantly predict higher 6-month Avoidance along with high levels of Pros at 6months predicting increases in Sunscreen Use at 12-months . However , baseline Avoidance and Sunscreen Use also predict Pros at 6-months. In contrast , the Cons of behavior change were expected to be predicted by behavior. As expected , high levels of Avoidance at baseline and 6-months significantly predict lower levels of Cons at subsequent time points. However, a complex relationship where higher levels of Sunscreen Use are predicted by higher levels of baseline Cons and lower levels of Cons at 6-months was not expected. Overall , the three predictions for the Precontemplation sample were supported in this step of the analysis. However , several additional relationships between constructs were also found. See Appendix C for a detailed report of all 21 pairwise cross-tab models for the Precontemplation sample.
Contemplation. Four sets of hypotheses were made for this stage sample.
First, Experiential Processes were expected to act as independent variab les , specifically predicting Pros and behavior. Baseline Experiential Processes did significantly predict increases in both Pros and Sunscreen Use at 6-months . However , Experiential Processes were not related to Avoidance behavior. A second hypothesis also expected Behavioral Processes to predict Pros and Behavior. These predictions were also confirmed with higher Behavioral Processes predicting higher levels of Pros and Avoidance behavior at subsequent time points . The third hypothesis expected the Pros of behavior change to influence behavior. High Pros at the 6-month time point ' significantly predict increases in both Avoidance , and Sunscreen Use. Finally, the Cons of behavior change were expected to be predicted by behavior. As expected, increases in Avoidance at both baseline and 6-months significantly predict subsequent decreases in the Cons. However , there are no significant relationships between Sunscreen Use and Cons. Thus , all predictions for the Contemplation stage were at least partially supp01ied. Additionally, several other relationships between constructs were also found. See Appendix C for a detailed report of all 21 pairwise cross-tab models for the Contemplation sample.
Preparation. Again, several hypothesized relationships were specifically predicted for this stage sample . First , Behavioral Processes were expected to be more involved in this sample, specifically, predicting behavior. While there are no significant relationships between Behavioral Processes and Sunscreen use, this anticipated relationship was found for Avoidance . Higher reported use of Behavioral Processes predicted more Avoidance behavior at all subsequent time points.
Secondly , behavior was expected to predict both the Pros and Cons of behavior change . However, there were no significant cross-lagged paths between either of the behavior subscales and Pros. While both Avoidance and Sunscreen Use predicted Cons, there were differing relationships. With Avoidance, all adjacent time point cross-lagged paths were significant, indicating that while increases in Avoidance predicted decreases in Cons , so too did increases in Cons predict decreases in Avoidance. However , only increases in baseline Sunscreen Use predicted higher levels of Cons at both follow-up time points . Finally , while no explicit hypotheses were made for Confidence , it was expected that Confidence would influence the \ behavior constructs. Higher levels of Confidence were found to be related to both higher Avoidance at all time points and higher Sunscreen Use at 6-months only.
Additionally , higher levels of Avoidance at 6-months significantly predicted higher levels of Confidence at 12-months, thus showing a somewhat reciprocal relationship.
Unlike the other two stage samples, not all hypothesized relationships were supported at this step in the Preparation sample . Additional, non-predicted relationships were also found between the constructs . A more detailed explanation of these relationships is presented in Appendix C.
The strongest regression paths are between Experiential Processes and Pros of behavior change and Confidence and Sunscreen Use at adjacent stages. The only other paths that showed a consistent relationship between baseline and 6-months and 6-and 12-months were between Sunscreen Use and Experiential Processes.
Experiential Processes were also found to significantly predict Behavioral Processes , Confidence, and Sunscreen Use but only at 6-months. Behavioral Processes at both baseline and 6-months predicted Cons at 12-months (.099, not shown), however, in opposite directions . High Pros of behavior change at 6-months significantly predict more Sunscreen Use at 12 months . Baseline Cons of behavior change predicted lower Pros at 6-months and higher Sunscreen Use at 12-months (.062, not shown) . In addition to predicting higher Avoidance, Confidence also predicted lower Cons of behavior change at 6-months . Avoidance is positively related to Confidence at 6-months and negatively related to Cons of behavior change at 12-month~. Sunscreen Use at 6-months predicts increases in Behavioral Processes as well as the relationship with Experiential Processes described above. Table 4 presents the R 2 values for all constructs at 6-and 12-months .
Confidence and previous Avoidance behavior account for 44. 7% of variance of Avoidance at 6-months. Even more variance, 60.1 %, in Avoidance behavior was accounted for at 12-months by Confidence and previous Avoidance behavior at both baseline and 6-months. Experiential Processes and previous Sunscreen Use accounted for 47.1 % of the variance in Sunscreen Use behavior at 6-months. Sunscreen Use at 12-months had 63.9% of the variance accounted for by the Pros at 6-months and Cons at baseline along with previous Sunscreen Use at both baseline and 6-months.
Contemplation. The first integrated model in this sample again included all significant regression paths discussed above in Step 3. This model shows good fit, x2(1715)=3273.01, CFI=.901, RMSEA=.048, AASR=.067. However , several paths were not significant and the Wald test indicated that they could be removed without reducing the overall fit of the model. These paths were removed until all paths remaining were significant and the Wald test did not show any additional regression paths to be removed . This final model is presented in Figure 6. This model shows good fit, x2 (1730) Almost all constructs show reasonable stability across the time points (see Table 5). Avoidance, however, is much lower than the other constructs with .285 and .286 st.ability parameters between adjacent time points. The Pros of behavior change show low stability between baseline and 6 months (.287) , but a reasonable stability between 6 and 12 months (.552). The other six constructs have a range of adjacent time stability from .532 to . 752. All baseline to 12 month paths are also significant, ranging from .159 to .285.
The strongest relationship is between Confidence and Avoidance behavior at adjacent time points, with higher Confidence predicting more Avoidance. The only other constructs showing consistent relationships across baseline to 6-months and 6-to 12-months are Cons and Confidence , however, the relationships are complex . Higher Cons at baseline predict lower Confidence at both 6-and 12-months (-. 150, not shown). In contrast, high Cons at 6 months predicted higher Confidence at 12 months.
Two other constructs strongly predicted the Pros of behavior change at 6-months , Experiential Processes and Confidence. However, these relationships were not seen between 6 and 12 months.
In addition to the above relationships, several other relationships were significant. High Experiential Processes predicted more Sunscreen Use. High Cons of behavior change at baseline significantly predicted lower Behavioral Processes and Avoidance as well as the relationships discussed above with Confidence . In addition to the relationships with Confidence discussed above, baseline Confidence also predicted higher Experiential Processes at 12-months (.097, not shown). Sunscreen Use at baseline predicted less Avoidance at 6-months. Finally, more Avoidance at 6months predicted lower Cons of behavior change at 12-months. Behavioral Processes and Pros did not significantly predict any other constructs in this sample. Preparation. The first integrated model included all significant regression paths for this sample discussed above in Step 3. This model shows good fit, The stability estimates again appear reasonable for most constructs (see Table   7). Similar to the other two samples , the reliability of the Pros of behavior change and Avoidance between adjacent time points is low. However , the other constructs all show good stability, ranging from .476 to .781 between adjacent time points and .068 and .300 between baseline and 12-months.
Several strong relationships were found in this sample. The strongest single regres~ion path is between baseline Experiential Processes and the Pros of behavior change . However , there is no similar path between 6-and 12-months , nor are Experiential Processes related to any other construct. Confidence shows a strong , consistent pattern with Avoidance behavior. Higher levels of Confidence predict more Avoidance across adjacent time points as well as between baseline and 12-months (.210 , not shown). Pros of behavior change and Behavioral Processes show a somewhat complex relationship. Higher Pros predict higher levels of Behavioral Processes at adjacent time points, however , baseline Pros predict lower levels of Behavioral Processes at 12-months (-.124, not shown).
Avoidance and Cons of behavior change are strongly related in both directions , however those relationships are both positive and negative. While higher Avoidance at baseline predicts higher Cons at 6-months, higher Avoidance at 6-months predicts lower Cons at 12-months. In contrast , higher Cons at baseline predict less Avoidance behavior while higher Cons at 6-months predict more Avoidance at 12-months .
Several additional relationships between constructs were significant. Increases in baseline Behavioral Processes predicted increases in the Pros of behavior change at 6-months . Contrary to expectations, high levels of Behavioral Processes at 6-months predicted less Avoidance. The relationships between Cons and Confidence with Avoidance are described above . Additionally, high baseline Cons of behavior change significantly predict lower Behavioral Processes and Confidence at 6-months and lower Pros at both 6-and 12-months (-.060 , not shown) . Also, higher baseline Confidence predicts higher levels of Behavioral Processes, lower Cons of behavior change , and more Sunscreen Use behavior at 6-months. Finally, the behavior constructs show additional relationships. Similar to Avoidance, more Sunscreen Use \ at baseline significantly predicts more Cons at 6-months. Also, more Avoidance at 6months predicts higher Confidence at 12-months.

Discussion
The Process Due to the complexity of the proposed integrated models , a step approach was adopted to simplify model development. By breaking down the model into smaller steps, information was gained in each step that allowed progression towards the final integrated models. Along the way, several important observations were made.

In
Step 1 we found that the majority of the measurement structures of the TTM constructs were very strong. However, more work may be needed on the behavior scale due to the small correlation found between the two subscales of Avoidance and Sunscreen Use. Due to such a small correlation, both subscales were included in later models rather than a single indicator of behavior. Similarly, the Confidence scale, when limited to only three items, was weaker than other constructs due to the underlying two subscale structure. The one item from the Sunscreen scale had a much lower factor loading than the two items from the Avoidance subscale. Part of the limitation of this factor structure is an artifact of using only three indicators for the general Confidence factor in the reduced model. When examining the full item level structure presented in Appendix A, the Confidence measure shows good fit when several indicators on both subscales are utilized. These difficulties with both the Confidence and behavior measures could be lessened by looking at the two behaviors independently. The Confidence measure could also be reduced to either the avoidance or sunscreen use subscales . \ The measurement model including all baseline measures shows interesting correlations between factors. The h_ ighest correlation after that between Experiential and Behavioral Processes was between Behavioral Proce sses and Confidence. This strong correlation may be due to the Behavioral Processes being driven by the self liberation scale score . Self liberation measures a person ' s commitment to making a behavior change. This construct appears similar to that of a person ' s Confidence to make a behavior change. Therefore, this high correlation demonstrates the similarity between these two constructs of the TTM that are seldom examined together.
Step 2 investigated the stability of the TTM constructs over the three time points . All measures were reasonably stable when examined independently in all three samples. However, when pairwise cross-lag models and the final integrated models were examined, a few constructs showed lower stability. The Pros , for example , showed much lower stability in the integrated model than when examined independentl y in Step 2 . This lack of stabilit y when compared with other constructs may explain why the Pros did not act as a good predictor for other constructs in the integrated models.
Step 3 was reall y just a stepping stone to reduce the number of paths included in the integrated models. Several constructs showed significant relationships within this step that did not hold up in the integrated models . These relationships could be true associations , but simply be too weak to hold up in the larger more complex models . Alternativel y, these paths could simpl y be artifacts , with two constructs showin g a relationship that is explained better by the relationship with a third variable that 01: lY enters in the inte grated -models .

The Models
Precontemplation. The final integrated model for the Precontemplation sample at least partially supported all hypothesized relationships. Experiential Processes strongly predicted the Pros of behavior change at subsequent time points.
The Pros predicted Sunscreen Use while Avoidance behavior significantly predicted the Cons between 6-and 12-months. In addition to these hypothesized paths, Confidence also strongly predicted Avoidance behavior at subsequent time points.
While not specifically predicted this and other paths are consistent with model expectations .
However, a few paths were significant but in the opposite direction than expected. Sunscreen Use predicted Pros in addition to being predicted by it, as was the expected direction. Similarly, while the Cons were predicted by Avoidance behavior as expected, contrary to predictions baseline levels of Cons also significantly predicted Avoidance. Also unexpectedly , Sunscreen Use predicted Experiential and Behavioral Processes.
While there is not a clear picture of relationships for the Precontemplation sample, all the hypothesized relationships were at least partially supported. Several constructs showed unexpected relationships , especially relationships in the opposite direction. These may indicate a need to rethink assumptions that have been made about the order of changes in the constructs. A person's behavior may have an impact on their thoughts about changing behavior. It is important to understand that a person's past behavior influences future behavior as well as other more cognitive constructs (Sutton , 1994 ). However , in order to understand the process of behavior change, discovering which variables influence behavior may help us to better recognize how to bring about change in behavior. Those constructs that predicted increases in sun protective behaviors in this sample were Experiential Processes, Pros of behavior change, and Confidence . Experiential Processes also predicted both the Pros and Confidence. These relationships also represent the strongest paths in the integrated model. Therefore, for people in the Precontemplation stage, interventions designed to increase these constructs may have the most impact on increasing sun protective behavior.

Contemplation . The final integrated model for this sample only partially supported some of the hypothesized relationships. Experiential and Behavioral
Processes were expected to predict the Pros of behavior change, Confidence , and behavior. The Experiential Processes were found to significantly predict Pros and Sunscreen use. However, they did not predict Confidence, nor did the Behavioral Processes predict any other constructs. The Pros of behavior change were expected to predict behavior, however, there were no significant relationships between these two constructs. The Cons of behavior change were hypothesized to be predicted by behavior. While this was partially supported between 6-and 12-months , Cons significantly predicted Avoidance behavior between baseline and 6-months.
Several additional relationships were also found. Similar to the Precontemplation model, Confidence was a strong predictor of Avoidance behavior.
Confidence also strongly predicted increases in the Pros of behavior change . In ' addition to the predicted relationship between behavior and Cons discussed above , Cons acted as a predictor of Behavioral Processes and Confidence, with lower Cons indicating more process use and higher Confidence. Overall , Confidence and the Cons of behavior change showed the most predictive relationships with the other constructs. Both also directly predicted Avoidance behavior as well as other TTM constructs. Therefore , it appears that these two constructs may be the most important for intervention during this stage.
Preparation. The final integrated model for the Preparation sample at least pai1ially supported three of the four hypothesized relationships. As expected the Cons of behavior change were strongly predicted by both measures of behavior. However, the directions of those relationships were somewhat surprising . Higher levels of sun protective behavior at baseline predicted higher Cons. This is contrary to the prediction that small behavioral steps will help to reduce the Cons of behavior change.
However, the expected negative relationship was found between Avoidance behavior at 6-months and Cons at 12-months. These relationships may indicate that small steps , like using sunscreen may enhance the Cons , however , once sufficient behavior chai1ge has been made, the Cons appear to be reduced.
Expectations about Confidence being important in this stage were also confirmed. Confidence predicted Behavioral Processes, Cons of behavior change , and both measures of behavior. Two of the strongest relationships in the model are between Confidence and Avoidance. Another hypothesis that was supported in the final model was that of Behavioral Processes predicting behavior. However , again this was in the opposite direction with more process use predicting less behavior.
The one specific hypothesis not supported was that of behavior predicting the Pros of behavior change . Experiential and Behavioral Processes along with Cons were the only constructs to significantly predict Pros. As in the previous two samples, the relationship between Experiential Processes and Pros is strong between baseline and 6-months. However , as expected, Experiential Processes showed no other significant relationships. Overall , Confidence was the best predictor of behavior in this sample.
Additionally , there were more relationships between constructs in this sample than in the previous two. This may be a sign of greater variance in this sample due to the changing nature of the Preparation stage of change.
Similarities Across Stage Models. The strongest paths in these integrated models are consistent across the three stage samples. Experiential Processes were only expected to be important in Precontemplation and Contemplation . However, the strong relationship between Experiential Processes and Pros was found in all three samples. Additionally , Confidence was found to significantly predict Avoidance behavior in all samples. There were no specific predictions for Confidence although it was expected to become impo1iant in the later stages. Other than these relationships, there were no other comparable paths across all three stages.
In addition to these similarities across all three samples, there were several corresponding paths between pairs of stage groups. Both the Precontemplation and Contemplation stages showed relationships between Experiential Processes and Sunscreen Use and the lack predictive paths from Behavioral Processes or Pros to any other construct. Experiential Processes were expected to be imporiant in both of these stages. However , Pros were expected to predict behavior in both stages and Behavioral Processes were expected to become important in the Contemplation sample.
The model from Preparation also showed similarities to the model from the Contemplation group. Cons and Avoidance behavior were related in both of these stage samples. As predicted, Avoidance predicted the Cons. However , there was a reciprocal relationship with Cons also predicting Avoidance . These relationships were in the expected direction with lower Cons predicting more Avoidance, similarly, more • Avoidance behavior predicting lower Cons. The Cons of behavior change also significantly predicted Confidence in both these stage samples. Most of the paths were in an expected direction with lower Cons predicting higher Confidence.
However, in the Confidence group, lower Cons at 6-months predicted lower Confidence at 12-months. Another reciprocal relationship was found in the Preparation group with Confidence also predicting Cons.

Limitations
The primary limitations of these analyses lie in the sample. First, three large samples were combined together to provide sufficient sample size for the stage group analyses. Each of these three samples represents a different segment of the general population. They were not equivalent on demographic characteristics at baseline . If demographic variables have any influence on the constructs examined in these analyses , combining these samples may mask such differences. However , it was hoped that by combining the three samples , the total sample would better represent the general population.
A second, and more important limitation of the sample is that all participants who were included in this study were from the treatment group . The Processes of Change were only collected on those individuals in the treatment conditions. This may have both theoretical and statistical implications. By only exploring the relationships between constructs on those who have undergone an intervention, natural change within the population is not modeled. DiClemente and Prochaska (1982) found that the TTM could be used to explain natural change as well as change initiated by a clinical intervention, however, processes were utilized differentially by these two groups. Self-quitters rated processes such as feedback, stimulus control and social management less important than those who participated in formal cessation programs (DiClemente & Prochaska, 1982). Therefore, relationships found among the treatment group may not be similar to that found among individuals who did not complete an intervention. Individuals in the control group may report less process use than those in the treatment group. Since the Processes were not collected from the control groups in these samples, that comparison cannot be empirically tested. Additionally, there may be statistical limitations to only examining the treatment group. The intervention may be creating more variance than would normally be found between these variables as people progress through the stages. This higher level of variance may create inflated correlations and relationships between the TTM constructs. Again, without the ability to compare between a treatment and control group, the relationships found must be carefully interpreted.
A final limitation related to the sample is that of missing data. Due to the need for responses to so many variables, a large proportion of the sample could not be utilized . Since we only included individuals from the treatment group at risk for sun exposure, the analyses in this study are only using 52. 7% of those collected at baseline. Therefore, almost half the possible sample was not available for at least one of the follow-up time points. This missing data needs to be further analyzed in order to determine if it is missing at random or missing in a more systematic way that may impact the results of these models. While this sample may have its limitations , the complex model structures used in these analyses requires a large amount of data, both in sample size and number of variables. The number of existing longitudinal data sets that include all TTM construct variables on a sufficient number of individuals for the above analyses are few . As in most secondary data analyses, we must make the best of the data we have and look to future studies to collect data better suited to these types of analyses.
In addition to sample limitations, there may be issues with the theoretical basis for making our hypotheses. While most of the previous research has begun in the smoking cessation area, this study focused on the area of reducing sun exposure . A continuous behavioral outcome measure was needed to test how well the TTM constructs predict actual behavior. The Sun Protection Behavior Scale provides such a measure. There is no similar continuous measure in the area of smoking cessation .
While there have been several proposed methods of measuring smoking outcome (Velicer et al., 1992, Velicer et al., 1996 there has yet to be clear consensus on the best measure. Previous studies have found that the relationships of the TTM · generalize well to a variety of behavioral areas (Prochaska et al., 1994;Rossi & Redding, 2001 ). The hypotheses for this study were developed based on previous research from the smoking cessation literature. Some of the discrepancies between \ expected and actual relationships found in this study may be due either to differences between the content area of sun protective behavior and smoking cessation or basic mis-specifications of the hypothesized relationships. This confound makes clear interpretation of the results difficult. It is hoped that the findings in the area of sun exposure could be generalized to other areas. However, future studies are needed to verify the generalizability of the relationships between TTM constructs in other areas .

Future Directions
Perhaps a better test of the stage aspect of the TTM would be to test individual stage transitions rather than simply differentiating on baseline stage of change. While the breakdown used in this study was an attempt to create more homogenous subgroups that would follow similar patterns of change, that may not have happened .
Of those who started in the Precontemplation stage, only 57.3% were still in However , due to the complexity of the models found in this study , a fmiher break down of the samples may not be feasible. The size of the sample may not allow for individual stage transitions to be tested . For example, even the largest transition between Preparation and Action has only 139 people . This would not be sufficient to create a stable matrix between the 21 manifest items at three time points, or 63 manifest variables. A more basic breakdown could be examined between those individuals who progressed from their baseline stage compared to those who remained stable. However , first a system for determining stage transition groupings would be needed. Some individuals either progressed or regressed a stage at the 6-month interval , only to return to their baseline stage at 12-months. These individuals could be considered to be either stable, if looking between baseline and 12-months , or progressing or regressing , if looking between baseline and 6-months . Some determination would have to be made as to who would be included in the stable subgroups and who would be classified in the progressing subgroups. Norman et al. (1998) developed a classification system based on dynamic typology clustering that could be used for partitioning stage transitions. The stage samples could be further divided based on this classification system. However, the sample might still be too small to fully test all subgroup samples.
To simplify the models so that they may be tested on these smaller samples , a more basic model testing method could be used rather than including all constructs included in this study. One possibilit y would involve removing entire constructs from \ individual models. Those constructs not found to be statistically impo1iant in the stage samples in this study could be removed in further stage transition models. For example, the Cons of behavior change showed little involvement in the Precontemplation , stage. While this construct was more important in later stages , it could be removed from analyses involving stage transitions from the Precontemplation stage. Alternatively, rather than removing a single construct, a set of analyses similar to Step 3 could be compared between the stable and progressing subgroups. This may at least create a glimpse of where differences lie between those who progress through the stages and those who remain stable . A third possibility could be examined where select variables are modeled at specific time points, rather than having all constructs at all time points . Instead of using the full cross-lagged models analyzed in this study, more theoretically driven models could be proposed for different stage transitions.
Those constructs that are seen as only predictors, such as the Experiential and Behavioral Processes, would only be included from the baseline time point where as those seen as outcomes, such as behavior would only be included at the follow-up time points. This type of simplified model was run on a sample of smokers in Precontemplation (Evans et al., 1999). Strong relationships were found as expected with Processes predicting Decisional Balance that then predicted behavior. However , by looking at the constructs at only one time point each, non-expected directional relationships could not be tested. Although these simplified models may not provide as stringent a test of the relationships between TTM constructs , they may allow models to converge using the smaller sample sizes involved with the stage transition sub-sa(nples.
The next step, after more work has refined the models in the area of sun protective behavior , is to examine the relationships between these variables in other health behaviors. Ideally, this type ofresearch would return to the foundation of the TTM , to smoking cessation. However , as discussed above, there are concerns about what behavioral criteria could be used in this area. Additional areas, such as exercise , diet, condom use, or alcohol use where the TTM has been applied have continuous outcome measures that would make them more suitable for these types of analyses.
The analyses described in this study, along with the further analyses described above could be applied to these other health behaviors to test the generalizability of the relationships between the TTM constructs . The only limitations on which behaviors could be analyzed are those having large enough samples with all TTM constructs and inclusion of a continuous outcome measure. Once these data sets exist, secondary data analysis to investigate the underlying relationships between the TTM constructs could be done.

Implications for Intervention
There was no clear determination with these analyses as to whether cognition predicted behavior or behavior predicted cognition. Paths in both directions were found in all stage models. If we had determined that cognition clearly influenced behavior, interventions should be focused around changing a person's pros and cons before teaching them how to modify their behavior. This is primarily the state of the current interventions based on the TTM. However, if we had found that behavior changes have an impact on cognition, there may be a remarkable change to interventions. Instead of focusing on a person ' s pros and cons, more emphasis should \ be given to making small changes in behavior. Small steps, such as reducing the amount of time spent in the sun should be encouraged to increase confidence and reduce the cons of behavior change. Because the direction of these relationships was not clearly defined by these analyses , both strategies should be used in building interventions to increase behavior change.

The TTM as a Theory
A theory is a general statement positing relationships that are expected to explain some aspect of the world around us. Theories are used to examine why something happens. A good theory should clearly explain the intenelationships between constructs, be based on empirical data , and be able to be falsified (Blalock, 1969). The TTM was developed and continues to be developed based on empirical data. The relationships between variables are specified and have initially been examined, primarily using cross-sectional data . The above analysis attempted to further specify these relationships longitudinally.
A theory can also be evaluated based on certain criteria: falisifiability, reliance on facts, clear and reasonable assumptions, clear and operationalized constructs, clear prediction of relationships , parsimony , and generalizability (Chafetz , 1978). The TTM as discussed above , has a strong reliance on empirical data . The assumption that there are stages of change has been carefully examined above and found to be appropriate. Researchers using the TTM are careful to define their constructs.
Intensive measure development has been undertaken for each construct. There are already clear relationships between stages and the other constructs. This study demonstrated different patterns of relationships between variables among the stages of \ change. While not creating a completely clear picture , the above research does add to the understanding of relationships between constructs . While finding parsimonious patterns of relationships was a goal of the study , simple relationships were not found .
However , this may simply reflect the fact that behavior change is not an easily explainable process. Behavior change is not easy, otherwise anyone who simply wanted to change would with little difficulty. Past research has shown how difficult change is, even with strong intentions. While parsimony is admirable , it may not be realistic in the area of health behavior change. The final criteria for evaluating a theory is generalizability. The model has been applied to a variety of areas within the health behavior arena and in a multitude of other areas. While the simple application of the model to other behaviors without evaluation of the relationships in the new behavior has been criticized (Joseph, Breslin & Skinner, 1999), these applications have shown remarkable similarity to the relationships found in smoking cessation .
Interventions based on the TTM in other health related areas have been found to be effective in producing behavior change Rossi , Redding, et al., 1995;Weinstock et al., 2000) . Studies similar to this one replicated in other behavior areas may help further enhance those interventions.
The overall strength of the TTM comes from its reliance on explaining relationships found in empirical data and the ability of the model to generalize to a variety of behaviors . While there is still work to do to further understand the underlying question of why and how people change, the TTM provides at least a rough framework for that analysis.  RMSEA=.065 (see Figure Al). To simplify for further analyses, the three best processes from each of the Experiential and Behavioral constructs were selected, Consciousness Raising, Dramatic Relief, and Self Reevaluation for Experiential processes and Stimulus Control, Reinforcement Management , and Self Liberation for Behavioral processes. To simplify the measurement model for further analyses, scale scores were computed by summing the two items for each process. Figure A2 presents

Confidence
The proposed model with one higher -order factor for the two confidence subscales showed good fit in the baseline sample, x2c8)=167.89, CFI=.982, RMSEA=.077, AASR=.021 (see Figure AS). Scale scores of the two factors cannot easily be used to represent the confidence construct since three markers are needed for stable estimates oflatent constructs (Nunnally, 1978). The three best representative items were selected for a general confidence factor, including one item from the sunscreen use factor, and two from the avoidance factor. Figure A6 presents this reduced item model in both the full baseline sample and longitudinal baseline sample .
There are zero degrees of freedom for these models , therefore there are no indices of fit. The factor loadings show adequate strength, although the sunscreen use item is low.

Behavior
The proposed seven item , two-factor structure was examined initially to simplify the structure to six items . The model presented in Figure   Due to the small correlation between the factors , a single higher order measure of sun protective behavior could not be formed by items or parcels from these two subscales .
Therefore, both scales will be incorporated into the full model and tested independently in future steps.  Figures B2 and B4), respectively. Figure B3 shows that two of the three items representing the Pros show non-significant adjacent correlated errors. The Avoidance model has one item with all non-significant correlated errors along with one item with the baseline to 6-month errors not significantly correlated (see Figure B6).

Contemplation
The seven constructs again show good fit across time for all seven TTM constructs. Table B2 presents fit indices for the models with and without the baseline to 12.:month stability paths. The Chi-squared difference test for the Experiential Processes, Pros, Cons, Confidence, Avoidance, and Sunscreen Use constructs showed a significant decrease in fit when this path was removed from the model , indicating the need to retain this path in future models. This path was not significant for the \ Behavioral Processes . Figures B8 to B 14 present the models for the seven constructs that include this baseline to 12-month path. Correlated error paths on items across the time points were also estimated. Experiential Processes ( Figure B8), Behavioral Processes ( Figure B9), Cons ( Figure Bl 1), and Sunscreen Use ( Figure B14) show all correlated error paths are significant. Both the Pros and Confidence scales have two items each with non-significant error correlations (see Figures B 10 and B 12, respectively). Only three of the nine correlated error terms were significant for the Avoidance factor (see Figure B13).
Preparation Table B3  Avoidance model has one item with all non-significant correlated errors along with one item with the baseline to 6-month errors not significantly correlated (see Figure   B20).  Step 3 examined the relationships among pairs of constructs. Cross-lag models between each combination of two constructs were run within the three stage of change samples. Non-significant paths within each model were removed to produce a reduced model. Chi-square difference tests and other measures of fit were used to determine whether specific paths were retained for future models.
Precontemplation. increases in Experiential Processes at both baseline and 6-months predict increases in the Pros of behavior change at subsequent time points (see Figure C2). Several additional relationships were found , with higher levels of baseline Experiential Processes significantly predicting higher Behavioral Processes ( Figure C 1 ), Confidence ( Figure C4), Sunscreen Use ( Figure C6), and Avoidance at 6-months , along with lower Avoidance levels at 12-months ( Figure CS). Additionally , high levels of Experiential Processes at the 6-month time point significantly predicts higher Sunscreen Use, and Avoidance at 12-rnonths along with lower levels of Cons ( Figure   C3). Sunscreen Use is the only construct found to significantly predict future levels of Experiential Processes ( Figure C6).
Behavioral Processes, while seen as independent variables in the overall model, were not expected to be highly involved for the Precontemplation sample.
However , several significant relationships were found. Overall, high levels of Behavioral Processes predict higher Pros ( Figure C7) and Avoidance ( Figure C 10 Figure C8). Behavioral Processes at 6-months also predicted increased Sunscreen Use at 12-months ( Figure Cl 1). While expected to be act as independent variables, the Behavioral Processes were significantly predicted by Pros, Confidence (Figure 9), and Sunscreen Use.
Pros were predicted to act as a mediator between Experiential Processes and behavior, with Experiential Processes predicting Pros, which would then predict behavior. As discussed above, Experiential Processes were found to predict increases There were no specific hypotheses related to the Confidence construct.
Despite this, several relationships were significant. Overall, high levels of baseline Confidence were found to predict high levels of 6-month Behavioral Processes , Pros , Avoidance ( Figure C19), Sunscreen Use and 12-month Sunscreen Use ( Figure C20).
Baseline Confidence is also negatively related to 6-month Cons. Additionally , high levels of Confidence at 6-months predict high levels of both Avoidance and Sunscreen Use and low levels of Cons at 12-months. High 6-month Confidence levels were predicted by high Experiential Processes , Avoidance, and Sunscreen Use and lower levels of Cons. Despite the lack of hypothesized relationships, Confidence appears to prediq and be predicted by several TTM constructs.
The behavior subscales were analyzed separately with each TTM construct.
Overall , higher levels of baseline Avoidance behavior predict higher levels of Pros and Confidence and lower levels of Cons at 6-months . Six month Avoidance only significantly predicts decreases in 12-month Cons . Additionally , high levels of baseline Sunscreen Use predict increases in Experiential Processes , Pros, Confidence, and Avoidance ( Figure C21) at 6-months. Finally , 6-month Sunscreen Use predicts increases in both Experiential and Behavioral Processes at 6-months.
Contemplation. Non-significant paths were again removed from each combination of pairwise cross-lagged constructs. Tables C2a and C2b present the fit indices and Chi-squared difference for all pairwise models . Almost all reduced models in this sample show non-significant Chi-squared difference tests , indicating that those parameters that were removed are not necessary for the fit of the model. The reduced model for Confidence shows a slightly significant (p<.05) Chi-squared difference. However, all the paths removed from this model were all non-significant.
Experiential Processes were expected to act as independent variables , specifically predicting Pros and behavior. Baseline Experiential Processes did significantly predict increases in both Pros ( Figure C23) and Sunscreen Use ( Figure   C27) at 6-months. Experiential Processes did not significantly predict any other constructs (see Figures C22, C24, and C26). However, high 12-month Experiential Processes were significantly predicted by high 6-month Pros and Sunscreen use along with high Baseline Confidence ( Figure C25).
, Behavioral Processes were also expected to predict Pros and behavior , acting as independent variables. These predictions were confirmed with higher Behavioral Processes predicting higher levels of Pros ( Figure C28) and Avoidance behavior ( Figure C3 l) at subsequent time points . In addition , 6-month Behavioral Processes were predicted by Cons ( Figure C29), with higher Cons predicting increased levels of Behavioral Processes. There were no significant relationships between Behavioral Processes and either Confidence ( Figure C30) or Sunscreen Use ( Figure C32).
The Pros of behavior change were hypothesized to be influenced by both Experiential and Behavioral Processes as well as to influence behavior. As discussed above , Experiential and Behavioral Processes were found to be significant predictors of Pros. In addition, increases in 6-month Pros were predicted by lower Cons ( Figure   C33) and higher Confidence ( Figure C34). High baseline Pros only significantly predict lower Cons at 6-months. However, high Pros at the 6-month time point significantly predict increases in Experiential Processes, Avoidance ( Figure C35), and Sunscreen Use ( Figure C36). Thus, all predictions for Pros were at least partially supported in this sample.
The Cons of behavior change were expected to be predicted by behavior.
Increases in Avoidance at both baseline and 6-months significantly predict subsequent decreases in Cons ( Figure C38). There are no significant relationships between Sunscreen Use and Cons ( Figure C39). As discussed above, high baseline Pros also predicted lower Cons of behavior change. In addition, high baseline Cons significantly predict lower Behavioral Processes , Pros, and Avoidance at 6-months along with all follow -up time points of Confidence ( Figure C37). Six month Cons are also m;gatively related to Confidence at 12-months . While more constructs were significantly related to Cons than expected, all significant relationships were in an anticipated direction.
There were no hypotheses made for Confidence , however several empirical relationships were found . High baseline Confidence levels signific antly predict higher Pros at 6-months and Experiential Processes at 12-months . Additionally , high Confidence predicts increases in Avoidance behavior at subsequent time points ( Figure C40). Conversel y, higher baseline Sunscreen Use behavior predicts lower levels of Confidence at 6-months ( Figure C41 ). Cons and Confidence show a complex relationship. High levels of Cons at baseline predict lower Confidence at both follow-up time points . However , high levels of Cons at 6-months significantly predict higher levels of Confidence at 12-months.
The behavior constructs were predicted to influence the Cons of behavior change. Behavior was also expected to be influenced by Experiential and Behavioral Processes along with the Pros of behavior change.
For Avoidance, these predictions were at least partially supported . Avoidance negatively predicts Cons at subsequent time points. Also as predicted , high Behavioral Processes predict higher Avoidance behavior at subsequent time points .

Experiential Processes do not show significant predictive relationships with
Avoidance . While there were no specific predictions between behavior and Confidence , Confidence was a strong predictor of Avoidance at adjacent times.
Additionally , baseline Cons and Sunscreen Use ( Figure C42) were negatively related to Avoidance at 6-months.
Only two of the expected relationships were partially supported for the Sunscreen Use behavior subscale . High baseline Experiential Processes predicted higher levels of Sunscreen Use at 6-months . However , high levels of Sunscreen Use at 6-months significantly predicted higher Experiential Processes at 12-months. The only relationship between Pros and Sunscreen Use was that higher Pros at 6 months influenced higher Sunscreen Use at 12-months. Sunscreen Use did not show any predictive relationship with the Cons of behavior change , nor was it predicted by Behavioral Processes. While not predicted , high levels of Sunscreen Use at baseline predicted lower levels of Confidence and Avoidance behavior.
Preparation. As in the previous two samples, non-significant paths were removed from each combination of cross-lagged constructs. All reduced models show non-significant Chi-squared difference tests , indicating that the parameters that were removed are not necessary for the fit of the model. These values and fit indices are presented in Tables C3a and C3b.
Experiential Processes were not expected to play a large role in this sample.
Accordingly, only Pros and Cons showed significant relationships with Experiential Processes . Higher Experiential Processes at baseline were strongly related to higher Pros at 6-months ( Figure C44 ). The relationship with Cons is a bit more complex , where lower Cons at 12-months were significantly predicted by lower Experiential Processes at baseline and higher Experiential Processes at 6-months ( Figure C45). All other constructs did not show significant relationships with Experiential Processes (see  Figure C53), this anticipated relationship was found for Avoidance. Higher reported use of Behavioral Processes predicted more Avoidance behavior at all subsequent time points ( Figure C52). However, Behavioral Processes at 6-months were also predicted by baseline Avoidance. In addition to this relationship with behavior, Behavioral Processes were also found to relate to several of the other TTM constructs. The only cross-lagged path not significant between Behavioral Processes and Pros was that from Behavioral Processes at 6-months to 12-month Pros ( Figure   C49). High endorsement of Behavioral Processes was significantly related to lower Cons reported at later time points ( Figure CSO). Additionally, high baseline Cons significantly predicted less endorsement of Behavioral Processes at 6-months. Finally, higher Confidence at baseline predicted higher endorsement of Behavioral Processes at 6-months, which showed a positive predictive relationship with Confidence at 12months ( Figure C51).
Behavior was expected to predict the Pros of behavior change. However , there were no significant cross-lagged paths between either of the behavior subscales and Pros (see Figures C56 and C57). Despite this lack of expected relationships , all other TTM constructs did show predictive relationships with the Pros. As discussed above, Pros both predicted and were predicted by Experiential and Behavioral Processes of Change. Additionally, low levels of Cons at baseline predicted higher levels of Pros at both 6-;-and 12-month time points ( Figure C54). Finally , high levels of Confidence at baseline were also related to higher levels of Pros at 6-months (Figme C55).
Similar to the Pros , the Cons of behavior change were expected to be predicted by the behavior scales. While both Avoidance and Sunscreen Use predicted Cons, there were differing relationships. With Avoidance, all adjacent time point crosslagged paths were significant, indicating that while increases in Avoidance predicted decreases in Cons, so to did increases in Cons predict decreases in Avoidance ( Figure   C59). However, only increases in baseline Sunscreen Use predicted higher levels of Cons at both follow-up time points ( Figure C60). Confidence also was negatively related to Cons at subsequent time points ( Figure C58). Additionally, high levels of Cons at baseline predicted lower levels of Confidence at 6-months. As discussed above , Cons were also predicted by Experiential and Behavioral Processes and also predicted Behavioral Processes and Pros.
While no explicit hypotheses were made for Confidence , it was expected that they would influence the behavior constructs. Higher levels of Confidence were found to be related to both higher Avoidance at all time points ( Figure C61) and higher Sunscreen Use at 6-months only ( Figure C62). Additionally, higher levels of Avoidance at 6-months significantly predicted higher levels of Confidence at 12months. As discussed above Confidence also predicted Behavioral Processes , Pros , and Cons , and was predicted by Cons of behavior change.
Behavior was expected to be predicted by Behavioral Processes and to predict both the Pros and Cons of behavior Change. Behavioral Processes did significantly prediq later levels of Avoidance behavior , however, they did not predict future Sunscreen Use . Additionally , both Avoidance and Sunscreen Use predicted the Cons of behavior change , as discussed above , however ; neither showed any significant predictive relationships with the Pros of behavior change. In addition to these relationships, both behavior subscales were also predicted by Confidence , as discussed above. When examined together , baseline Avoidance significantly predicted 6-month Sunscreen Use while 6-month Sunscreen Use predicted 12-month Avoidance ( Figure   C63). n.s.    .675 .509 .273 Figure C12 .  .225 Figure C26.

Processes of Change
The following feelings, thoughts, and experiences can affect the summer sun exposure habits of some people. We would like to know how often you may have had similar feelings, thoughts, and experiences during the past summer using the following scale: 1= Never, 2 = Rarely, 3 = Sometimes, 4 = Often, 5 = Always Consciousness Raising 2. I look for information about the risks of getting too much sun. 17. I think about what I've seen on TV or in magazines about the health risks of sun exposure. Dramatic Relief 3. It bothers me when I see someone whose skin has been damaged by the sun. 18. I get upset when I see someone aged by too much tanning. Self Reevaluation 13. I think about the damage to my appearance that will result from too much sun. 28. I worry that too much sun will make my skin look bad .

Stimulus Control
8. I avoid social situations where I could get too much sun. 9. I cover up when I know I will be out in the sun for awhile. Reinforcement Management 10. I reward myself when I avoid the sun. 25. I am rewarded by others for reducing my sun exposure. Self Liberation 12. I make commitments to reduce my sun exposure. 27. I tell myself that if! try hard enough I can avoid the risk from sun exposure.