Growth Curve Analysis of the Situational Temptations Scale in a College Sample

The temptation to engage in risky drinking is culturally embedded in the social life of young adults and college students. Studies have shown that more than 2 out of 3 students report drinking in the past month (Johnston, O’Malley, & Bachman, 2003) and over 80% will drink on at least one occasion during the school year (Del Boca et al, 2004). Heavy alcohol use and frequent binge drinking by college students has received much attention over the years because many drinkers engage in risky patterns of alcohol consumption, which is at the crux of a wide range of negative college alcohol-related consequences (Curry, Southwick, & Steele, 1987; Talbott et al., 2008; Wechsler et al., 2000). The present study reports the results of a secondary data analysis that used growth curve modeling (GCM) to examine growth trajectories of situational temptations over a two year period with data from a large sample (N=1067) of college student participants in a randomized trial of the efficacy of a brief intervention for alcohol harm reduction. A mixed effects piecewise model that estimated separate linear growth effects for the treatment and post-treatment phases of the study was found to fit the data best. At baseline, male students and those with higher rates of self-reported college alcohol-related problems had significantly higher temptations scores. Growth model results showed that self-reported temptations increased rapidly over the first 6 months during the treatment phase, then gradually decreased. No differences between the treatment and control groups in growth of temptations was observed during the treatment phase (0 to 6 months), however a small significant treatment effect was found over the post-treatment phase of the study (6 to 24 months). Temptations scores were reduced over time for students who had moderate or high levels of alcohol-related problems at baseline. After adjusting for treatment and alcohol problems, gender was not related to situational temptations.

Alcohol use is culturally embedded in the social lives of many underage youth, especially those attending college. Monitoring the Future (2013), a national study that monitor trends in drug use and alcohol consumption, found approximately 1 in 5 students reported binge drinking (consuming five or more drinks in a row within the past two weeks) by senior year of high school. The research data shows the proclivity of risky drinking begins in high school and increases over the first few years of college. Studies have shown that more than 2 out of 3 college students reported drinking alcohol in the past 30 days (Johnston, O'Malley, & Bachman, 2003) and over 80% consumed alcohol at least once during the school year (Del Boca et al, 2004).
Based on the extant literature, some of the strongest predictors of alcohol misuse include having a history of alcohol-related problems (Zakletskaia, Mundt, & Fleming, 2009), male gender , involvement in collegiate sports, race and ethnicity (Zapolski et al., 2014). One important mediator of risky drinking and alcohol-related problems-what the present paper will discuss-is situational temptations Migneault, 1995;Prochaska & DiClemente, 1982).
Situational temptations for heavy drinking is an important mediator of alcoholrelated problems in the Transtheoretical model (TTM)-the TTM consists of stages of change, self-efficacy, and other variables that describe cognitive and behavioral processes of change-and situational temptations is a latent variable comprised of four primary psychosocial variables including peer pressure, social anxiety, negative affect and positive/social situations (Migneault, 1995). Other Situational temptations scales have demonstrated reliability and consistency across a wide range of health behaviors, including smoking (DiClemente et al., 1985;Velicer et al., 1990), drinking (Migneault 1995;, dieting and exercise (Rossi, et al., 2001). Herein, situational temptations will be simply stated as temptations.
As previously mentioned, underage alcohol use is deeply rooted in American culture. Hence, most youth are tempted to try alcohol before they enroll into college (Sillice et al., 2014). Although varying degrees of temptations may or may not lead to risky drinking in college, very strong temptations or urges to engage in heavy drinking can be problematic for some students who are inexperienced and/or emotionally illequipped to control his/her drinking insofar that they may experience negative consequences because of drinking. To date, there has been little research that has examined the development of temptations to drink in early college, which is why several pivotal research questions about temptations remain, such as what is the average amount of temptations students have to drink when they enroll into college, does this change over time and what are the risk factors that influences the growth of temptations? Thus, the purpose of this thesis was to fill these gaps in the literature by examining the growth trajectories of temptations in a longitudinal sample of first and second year college students. A growth curve analysis would be conducted using Mplus 7 structural equation modeling (SEM) software to build growth curve models (GCM). The dependent variable in this study is situational temptations. The primary aim of this study was to find the optimal functional form of time using repeated measures of students' total scores on the Situational Temptations scale obtained from the College based alcohol risk reduction (CBARR) study . The secondary aims were to examine the main effects of treatment on the growth of temptations and observe the changes in the growth trajectories due to gender and different levels of college alcohol-related problems.

Theoretical Framework
The theoretical framework for the current study was the Transtheoretical Model (TTM). The TTM is a model of behavior change that has integrated many key constructs from the biopsychosocial model and other psychological theories into a comprehensive theory of change that can be applied to a variety of behaviors, populations, and settings (Prochaska & DiClemente, 1982;Prochaska, DiClemente, & Norcross, 1997). The TTM conceptualizes temptations as an important latent variable and risk factor of heavy drinking and alcohol-related problems Migneault, 1995;Prochaska & DiClemente, 1982). Alternative-forms of the scale have been used to assess temptations across various maladaptive behaviors such as, diet & exercise (Rossi, et al., 2001), immoderate-heavy drinking  and smoking (Velicer et al., 1990).

Situational Temptations Scale
The present study used the 8-item Situational Temptations Scale (see Appendix A) at each time point to measure the total strength of temptations to drink in different situations. The temptations scale was adapted by  from the original 21-item Situational Temptations Scale (Migneault, 1995).
The scale is comprised of four primary factors (Peer Pressure, Social Anxiety, Negative Affect, and Positive/Social) and the second-order factor, Situational Temptations Migneault, 1995). The subscales are used to guide individualized treatment interventions in the specific areas where the individual feels most compelled to drink (Prochaska & DiClemente, 1982). The total temptation scores was the primary outcome measure of the present study. Ruderman & McKirnan (1984) developed the Restrained Drinking Scale (RDS) to quantify college students' self-reports of their maladaptive thoughts about heavy drinking. The RDS was adapted from measures of abstemious dieting to be the first inventory of drinking restraint (Herman & Polivy, 1980;Polivy & Herman, 1985). Ruderman and McKirnan did not continue their line of research but Collins and colleagues (1989; continued to refine the scale and identified other endogenous constructs captured by the original RDS. Collins, George, and Lapp (1989) conducted a confirmatory factor analysis on the RDS and expanded drinking restraint from the single-factor structure into an multidimensional factor structure. The Temptation and Restraint Inventory (TRI; Collins et al., 1989) has three primary orthogonal factors:

Drinking Restraint and Orthogonal Temptations Factors
Govern (difficulty controlling alcohol intake), Restrict (attempts to limit drinking), and Emotion (negative affect as a reason for drinking). Collins and Lapp (1992) expanded the factor structure with two higher-order factors labeled: Concern about drinking (plans to reduce drinking/worry about controlling drinking) and Cognitive Preoccupation (thoughts about drinking). The TRI has been shown to be a robust measure of one's preoccupation to control their alcohol intake and a strong predictor of excessive drinking and negative alcohol consequences in college students (Collins & Lapp, 1992;. The RDS, TRI and Situational Temptations scales are reliable measures of college students' cognitions about restricting alcohol intake. However, there has been a long-standing debate over conceptual definitions and temptations scales (Collins & Lapp, 1992;Drummond et al., 2000;DiClemente, Prochaska & Gibertini, 1985;Marlatt & Gordon, 1985;Ruderman & McKirnan, 1984;Kavanagh et al., 2009;Schmidt et al., 2011;Sinha et al., 2011;Skinner & Aubin, 2010;Witkiewitz, 2013).

Potential Moderators of Situational Temptations
Numerous studies have shown extreme variability in drinking patterns by the time students actually reach college (Baer, 2002;Del Boca et al., 2004;Talbott et al., 2008;Wechsler & Nelson, 2008), students will increase their overall alcohol consumption within the first few weeks (Capone, et al., 2007;Del Boca, Darkes, Greenbum, & Goldman, 2004;Wechsler et al., 2000) and that weekly drinking changed considerably as a function of academic requirements and holidays (Del Boca et al., 2004). Del Boca and colleagues (2004) looked at drinking over the entire first year of college and found large individual variation in drinking rates each week. The findings from this study found alcohol consumption was consistently different for three parts of the week (i.e., lowest on Sunday-Wednesday, elevated on Thursday, highest on Friday-Saturday). Other studies have gender differences in drinking over time (Borsari & Carey, 2006;Greenbaum et al., 2005;Wechsler et al., 2002). Research suggests that gender and a history of alcohol abuse are some of the strongest predictors of alcohol misuse in the first year of college (Agostinelli, Brown, & Miller, 1995;Baer, 2002;Borsari, Murphy, & Barnett, 2007;Capone, Wood, Borsari, & Laird, 2007;Carey, Scott-Sheldon, Carey, & DeMartini, 2007;Chiauzzi, Dasmahapatra, & Black, 2013;Hingson et al., 2002;Talbott et al., 2008;Wechsler et al., 2002;Wood, Sher, & Rutledge, 2007). Moreover, it appears that men in early college with a history of alcohol-related problems may benefit from motivational interventions. To that end, the next section will focus on brief motivational interventions that have been used to lower rates of alcohol misuse and alcohol-related problems in college students.
A number of recent research trials have demonstrated lower rates of alcohol misuse and alcohol-related problems in heavy drinking college students who were given at least one brief intervention (Borsari & Mastroleo, 2013;Larimer & Cronce, 2007;Saunders et al., 2004;Wechsler et al., 2002;Wechsler & Nelson, 2008).
Motivational interviewing is one type of intervention that has garnered support in the literature with college students. As with TTM interventions, motivational interviews provide personalized feedback messages that are based on the student's self-reported drinking behaviors. The feedback reports usually consist of alcohol misuse information including a discussion about potential risk factors associated with alcohol dependence. Licensed professionals usually deliver the feedback in a very brief one-to-one session (e.g. 5-10 minutes), or in groups. The sessions are tailored to the needs of the clients who are interested in making changes to their drinking (Jouriles et al., 2010;Lojewski et al., 2010;Mastroleo, Oakley, Eaton, & Borsari, 2014).
Providing feedback in the mail is an alternative, cost effective way of introducing interventions to high-risk students. Several studies have found mailed feedback lead to significant reductions in alcohol use by heavy drinkers (Agostinelli et al., 1995;Saunders et al., 2004). For example, Agostinelli and colleagues (1995) identified 24 heavy drinkers and randomly assigned subjects to either receive feedback in the mail or to a control condition. The study found the average number of drinks consumed per week was lower in the feedback group relative to the control group after just six weeks. Similarly, Collins, Carey, and Sliwinski (2002) found participants who got his/her feedback in the mail reported consuming significantly fewer drinks per week and drank less than the control group, but the between-group differences were no longer significant after 6 months. These studies provided evidence that mailed feedback can have short term effectiveness for very heavy drinkers. However, each of these studies were limited by the relatively short-term follow-up period and small samples.
Based on previous research it is apparent that mailed feedback is effective but little research has been done to determine whether treatment outcomes may be differentiated by gender. Other research has investigated whether gender-based feedback may be as or more effective than gender-neutral treatment messages (Bishof et al., 2005;Lojewski et al., 2010). The aforementioned studies, however, did not have a mailed feedback component that delivered treatment messages differentiated by gender. Despite these shortcomings, the efficacy of direct-mail feedback warrants larger scale studies with bigger sample sizes (Larimer & Cronce, 2007) and an ample number of repeated measures and longer-term follow-ups (Cole & Maxwell, 2003;Jouriles et al., 2010;Walters and Neighbors, 2005). The present study will observe treatment interactions and gender over time.

Examining Intervention Effects Over Time
Cross-sectional research is the most common approach used to examine proximal variables of interest and identify average developmental trends based on the results of an intervention. Cross-sectional research, however, provides limited information about the dynamic changes in alcohol use. It is likely that drinking will vary over the first few years of college, hence longitudinal data are collected over important developmental periods and these research designs are well equipped to evaluate causal relationships (Cole & Maxwell., 2003;Duncan et al., 2006;Hox & Stroel, 2005;Singer & Willet, 2003). Longitudinal randomized control trial (RCT) studies are considered the gold standard for research (Larimer & Cronce, 2007). RCTs reduce bias by assigning subjects to one of the different treatments under study. The secondary data for this study comes from a longitudinal RCT.

The Current Study
To date, little scientific research has been conducted to assess the growth trajectories of situational temptations to drink in college. Moreover, no studies have observed factors that can potentially moderate the growth of temptations in a sample of underage heavy drinkers. This study will fill the current gap in research by utilizing GCM to model the growth of temptations in a large sample of college students.
The present study utilized a longitudinal dataset that contains a large sample of college students (N=1067) that were followed for over two years of early college. The secondary data comes from a proactive study on college student drinking. The college based alcohol risk reduction (CBARR) collected data from freshmen and sophomore drinkers matriculating in the fall of 2001 . The CBARR's randomized control trial (RCT) design provided an excellent opportunity to observe behavior change for up to two years of early college and equally important, the interventions' influence on risky drinking. RCT studies are superior to cross-sectional studies insofar that repeated measures on several key variables in large clusters of students can be analyzed with sophisticated longitudinal models. Participants were randomized into treatment or control conditions based on his/her gender and readiness to change heavy drinking at baseline, which were two prognostic factors that were related to the primary outcome measures of the CBARR trial. The eligibility criteria to participate in the CBARR included having consumed at least one drink in the previous year, no prior history of treatment for substance abuse, and a score below16 on the AUDIT-alcohol use dependence identification test (Babor & Higgins, 2001). All of the survey data was collected by a telephone research center .
Many of the participants were assessed up to six times over 24 months. Three brief individually tailored computer generated feedback reports were mailed to participants in the treatment group (N=534), after the baseline, 3 and 6 month telephone survey assessments. The individualized feedback reports were based on the students' responses to the survey assessments on up to 17 different constructs, including situational temptations, decisional balance, stage of change for heavy episodic ("binge") drinking, measures of alcohol use, blood alcohol content during heavy drinking episodes, normative feedback of peer drinking and beliefs.

Research Methods
Growth curve analysis can analyze repeated measurements to obtain a growth curve that represents the dynamic rate of change in human behavior across time.
Growth curve modeling (GCM) is widely used in the social sciences to estimate the growth of factors over a period of time (Bullock, Harlow, & Mulaik, 1994;Cole & Maxwell, 2003;Del Boca et al., 2004;Heck & Thomas, 2015;Little, 2013;van Geert, et al., 2012). Two common analytic methods used to construct GCMs are structural equation models (SEM) and multi-level models (MLM). The SEM approach combines the best features of path analysis, regression and factor analysis to model withinperson and between-person change over time or situations (Hox & Storl, 2005;Kline, 2006;Preacher et al., 2008). In contrast, the MLM (also called hierarchical linear models) model for change, includes a hierarchy of nested effects in the model to disentangle the influences of within-person and between-person change simultaneously (Singer & Willet, 2003). Both statistical models provide equivalent mathematical results. Additionally, GCM is more flexible than traditional ad hoc categorization procedures like ANOVA and MANOVA because ANOVA and MANOVA can only include complete case data and these antiquated methods are ill equipped to handle unequally spaced time points (Bryk & Raudenbush, 1992;Duncan, Duncan & Stryker, 2006). Furthermore, GCMs can also include time-varying variables to study different phenomena that affect the individual growth trajectories in the model (Duncan, Duncan & Stryker, 2006). Hence, GCM was the optimal statistical method for the specific research aims of this study, which will be discussed next.

Research Aims
The primary aim of this study sought out to evaluate the functional form of time to bridge the current gap in the literature. The best unconditional GCM would provide robust estimates of the inter-individual variability in the sample and estimate the systematic changes in temptations over time via the slope. Given the centrality of the growth processes and SEM approach, the statistical models were expressed using parameters that represented specific quantities and subgroups of interest (e.g. intercepts, slopes, variances, covariates, etc.). The present study will increase our understanding of the temporal development of temptations to drink and will add to the current literature on alcohol interventions for college students.

Hypotheses
Consistent with the literature, it was hypothesized that treatment would reduce the average rate of growth in temptations over time (DiClemente et al., 1985;Witkiewitz, 2013). The study by DiClemente and colleagues (1985) observed situational temptations in former smokers for four years and found high temptations scores took upwards of six months to reduce to moderate levels after receiving TTM interventions (DiClemente et al., 1986;Velicer et al., 1990). More recently, the study by Witkiewitz (2013) looked at the change in temptations and drinking outcomes in a national sample of alcoholics in outpatient treatment and found cravings to drink decreased over 3 years. It is important to note that a major limitation of that study was the single-item cravings measure was defined as temptations to drink-that measure was also used to construct the GCM. The current study, observed the influence of TTM treatment interventions on the average rate of growth in temptations to drink for two years.
It was hypothesized that high levels of college alcohol-related problems would be significantly associated with higher situational temptations at baseline and over time. Finally, consistent with a large body of literature, it was hypothesized that gender differences would be found at baseline and in the growth of situational temptations over time (Bishof et al., 2005;Lojewski et al., 2010).
In summary, the growth of temptations to drink will be evaluated with GCM, which will provide some additional insight into the changes in temptations as a function of time.

Participants
The sample (N=1067) ranged from 17 to 44 years old (M=18.82, SD=1.81) and 88% of the sample was white (n=939) and 56% were female (n=596). Participants were randomly assigned by stage of change and gender to one of two experimental group conditions: (1) treatment (n= 534) and (2) assessment-matched control (n=533).
The experimental groups were found to be comparable at baseline on gender and all measures of alcohol use and alcohol-related problems (Laforge, et al., 2001). All of the participants were given identical assessments at baseline and were scheduled for post-baseline assessments at 3, 6, 12, 18 and 24 months.

Measures
Situational Temptations Scale. The eight-item Situational Temptations scale  is the primary outcome measure. The 8-item scale was adapted by  from the 21-item scale (Migneault, 1995).
Respondents were asked to rate "how tempted you are to drink a lot" (e.g. "When I am offered a drink") with a reduced scale that consists of eight items that were judged on 5-point Likert scales (1=Not at all tempted and 5= Extremely tempted). The total score is the summed from each subscale. Total scores may range from 8 to 40, with higher scores representing stronger temptations to drink. The Cronbach's alpha-coefficient (α) reliability estimate was .89 for the total sample.
instrument that measures self-reported frequencies of personal and social problems commonly associated with college student alcohol use. Study participants were asked questions such as, "as a result of drinking, how often have you felt sad, blue or depressed" and students rated how often they experienced problems as a result of drinking alcoholic beverages within the past six months on a five-point Likert scale ( 1= Never/Almost never and 5= Very often). The CAPS score at baseline was used and coefficient alpha was .83 in the sample.

Categorical College Alcohol Problems Scale (CCAPS). The CAPS variable
was dummy-coded and total problem scores were reclassified as CCAPS to aid the descriptive comparisons of alcohol-related problems. The distribution of the sum scores for CAPS was non-normal and zero-inflated at baseline. The scores ranged from 0 to 21 (M=3.9, SD=3.7). CCAPS is a categorical variable with four mutually exclusive and all inclusive levels created from the CAPS score. These categories represent approximate quartiles of the sample distribution of the CAPS variable Students with a CAPS score of 0 reported "no alcohol problems" in the previous three months, and are used as the reference group for CCAPS in the regression models.
Students who reported any alcohol related problems were compared to the reference to optimize the variance in the models as well aid in the descriptive comparisons of college problems in the GCM analyses. The CCAPS variable was entered into the statistical models with three indicator or dummy variables (coded 0 or 1) for the "Low", "Medium," and "High" categories. Therefore, the model estimates shown in the statistical model in Tables 4 and 6 represent differences in situational temptations for each level of CCAPS compared to the reference level; "None." Treat. Feedback reports were four to nine pages long and offered either primary or secondary prevention messages, depending on the individual's alcohol risk status. The intervention group received three feedback reports in the mail starting after the first assessment at baseline. Feedback reports were generated based on students' responses on 17 different constructs, such as frequency of high-risk alcohol use; motivational readiness; decisional balance measures of alcohol expectancies; situational temptations; cognitive barriers to change; and behavioral, experiential, and cognitive processes that have been found to be related to behavioral change (Saunders et al., 2004). The control group did not receive any feedback. Treat was dummy coded (0=Control and 1=Treatment) and is the moderator that represents the main effect of treatment in the conditional GCM analyses.
Female. Gender were dummy coded (0=Male and 1=Female). Female is the moderator that represents the main effect for gender in the conditional GCM analyses.

Analytic Plan
The ensuing models would be developed in two steps: The first step is a factor analysis that maps the temptations repeated measures variables onto the latent variables (i.e. intercepts, slopes). This step involved comparing a series of increasingly complex "unconditional" GCM models to identify the best fitting functional form of

Statistical Approach
The models were constructed using Mplus 7.11 latent variable software. Six  (Duncan, Duncan & Stryker, 2006). All of the GCMs were estimated with full information maximum likelihood estimation and the unstructured (default) covariance-variance matrix. There were 33 response-patterns where missingness in the data occurred and 33% attrition rate at the final wave of data collection. Therefore, missing data was handled using full information maximum likelihood (FIML) and the standard errors for the parameter estimates were computed using the observed information matrix (Kenward & Molenberghs, 1998). Any observations with missing data on the covariates were deleted from the model.

Modeling Time
To estimate the initial status and slope growth factors in MPLUS, time was represented by fixed time scores entered directly into the model software. Time since baseline at each survey assessment (0, 3, 6, 12, 18 and 24 months post-baseline) was rescaled to represent time since baseline in years (0, .25, .5, 1, 1.5, and 2.0) to facilitate the convergence of the polynomial growth curve models. Thus, the fixed time score for the baseline assessment was entered as 0, at 3 month assessment the fixed time score was 0.25, at 6 months as 0.5 and so forth.

Models containing statistical Interactions.
Inclusion of variables in the statistical models compared in this study followed the hierarchy principal for modeling statistical interaction (Cox and Donnelly, 2011).
The hierarchy principle of interaction modeling requires that a statistical model with one or more statistically significant interactions terms must also contain of the all lower order terms and main effects for each of the variables involved in the higher order interaction, regardless of whether the lower order terms are statistically significant. For example, a model with a statistically significant term representing the cubic polynomial function of time, also includes terms for the quadratic and linear functions of time, even if the lower order terms do not appear to be statistically significant. This principal is needed in most instances for valid interpretation of the interaction effects.

Mixed-Models
The fixed components (aka. fixed effects) of a mixed-model (0, 1, 2 ….n) accounts for the deviation from the mean of the sample distribution of scores (i.e. inter-individual variability). The random components (random effects) of the mixed model (b0i, b1i, b2i) accounts for the deviations or dispersion of an individual's scores (i.e. intra-individual variability) over time (Duncan et al., 2006;Little, 2013). The fixed effect estimate for the intercept (0) parameter is the grand mean for the sample at baseline. The mean change over time is the fixed slope (1, 2 ….n). The random effect estimates the variation due to individual differences surrounding the fixed effects (i.e. "residual variances and covariances" in GCM output in MPLUS) and are represented in the mixed model by b0i, b1i. Finally, the unexplained variation or error term (eij,) in the mixed model represents the residual error that was left over after all of the variation due to the fixed and random effects have been accounted for.

No Change Model
The null model estimated the intercept without including the random effect in the model. The next model included the random effect term to account for individual differences around the intercept. The random intercept model is expressed with the following function: Yij= (β0 + b0i) + εij (1) The function, Yij= represents temptations scores at for individual i at observation time j, β0 is the intercept, b0i is the random effect for the intercept for individual i and εij is the unexplained residual error at time j. This model assumes no change over time, hence the intercept only model is sometimes referred to as the "no change model."

Growth curve Models
Next the second, to model the biphasic rate of change during the treatment phase (baseline to 6 months) and the average change that occurred immediately following the treatment phase (6 months to 24 months), respectively. Finally, constructing these different models in seriatim concomitantly tested the hypothesis that the average growth trajectory of temptations was nonlinear over the length of the study.

Model Selection
Model fit refers to the ability of that particular model to reproduce the data (i.e., the variance-covariance matrix). Constructing the unconditional GCM with the best functional form of time was an iterative procedure, in which theory led to a model and a number of goodness of fit statistics could be used to evaluate the fit of the models (Bentler, 1990;Cheung & Rensvold, 2002;Kline, 2006;Preacher, Wichman, smaller model was better, then the larger model was selected as a reasonable (perhaps better) representation of temptations. Two or more alternative models (including the null) were compared in terms of model fit. The selection of the final model was aided by several model fit statistics (Hox and Stoel, 2005) and the chi-square difference test.
The Comparative Fit Index (CFI) is a statistic used to indicate the relative improvement in the fit of the measurement model compared to a statistical baseline model. A value of .90 indicates good fit and estimates ≥.95 indicates excellent fit (Bentler, 1990;Kline, 2006). The root mean square error of approximation (RMSEA) is a parsimony-adjusted, absolute model fit index that accounts for the degrees of freedom in the model and sample size. Estimates below .10 indicated acceptable fit, <.05 was good fit and a value of zero indicated the best fit (Kline, 2006). The standard root mean squared residual (SRMR) is an estimate of the standardized differences between the observed and the predicted covariances. SRMR residuals should be close to zero for a very close fit, but estimates below .08 indicate acceptable model fit (Kline, 2006). These statistics (e.g. CFI, RMSEA, and SRMR) are used to assess the extent to which the covariances predicted by the model correspond to the observed covariances in the data. Models that met the minimum acceptable model fit using the criteria mentioned previously, were required before the structural parameters (i.e. causal paths) of the models were interpreted.
In addition, information-based indices such as Akaike's information criterion E(Yij) = (β0 + b0i) + (β1+ bnj)(tn) + β2(Treati) + β3 (Femalei) + β4(CCAPSi) + β5(tn*Treati) + β6(tn* Femalei) + β7(tn * CCAPSi) + β8(Femalei * Treati) + β9(Femalei * CCAPSi) + β10(CCAPSi * Treati) + β11(tn * femalei * CCAPSi) + β12(tn * femalei * Treati) + β13(tn * CCAPSi*Treati) + εij The final conditional model shown in Table 6 in the next chapter contains only statistically significant model terms. It is the best fitting model resulting from the comparison of fit statistics for the conditional models first examined, and then examined in the full multivariable model. The full multivariable model piecewise linear mixed GCM described in equation (3) is an example of the full hierarchical interaction model that was to be evaluated to identify the best fitting final model of the growth in situational temptations. The values of the fixed independent variables (Treat, Female and CCAPS level) were determined at baseline and therefore do not vary over time. Fixed effect terms were included to test hypotheses concerning potential main and moderation effects of each of the independent variables; treat (3), female (4) and baseline alcohol problems, CCAPS (5). Specifically, the multivariable piecewise model included terms for the fixed and random intercept, β0, and b0i, respectively. Time (tn) is represented in the piecewise multivariable model by fixed and random effect terms for each of the two linear time segments, β1(t1, b1i) covering the intervention period from baseline to 6 months, and β1(t2, b2i) covering the post-intervention period from 6 months to 24 months post-baseline. Following the hierarchy principal of interaction modeling, the initial multivariable model includes all of the fixed and random effects terms for time as well model terms for all three way higher-order interactions and subordinate lower order interactions and main effect terms for all three hypothesized independent variables. Tests of whether situational temptation growth was moderated over time in the piecewise model therefore requires two interaction terms for each potential moderator (moderator* t1) and (moderator* t2), corresponding to the test of whether growth was moderated during the either intervention phase (0 to 6 months), or during the post intervention phase (6 months to 18 months), respectively. Two-way interaction terms of interest related to the study examine hypotheses concerning whether growth in situational temptations differed over time conditional on the value of each of the potential moderators (b5, b6 and b7).

Hypothesis Testing
The conditional GCMs evaluated the effects of the main effects and moderators of the growth in situational temptations over the two year period. It was hypothesized that the growth of temptations will be significantly lower in the treatment group.

7(tn*CCAPS).
Although, in this study no specific hypotheses were proposed concerning the potential for the moderation of the treatment effect by either gender or level of baseline alcohol problems, the approach to modeling used in this study can provide answers to these potentially important questions. Since, all three-way interactions were evaluated the model can provide new insight into whether effect of treatment on situational temptations differed by gender (b12), or by the level of alcohol problems reported during the three months prior to baseline (b13). Specifically whether significant gender moderation in the treatment effect was observed during the intervention phase (0 to 6 months.) is addressed by the test of the term β12(t1 * female * treat). Whether there is evidence for gender moderation of the treatment effect during the post-intervention phase (6 to 24 months) of the study can be determined by the test of the β12(t1 * femalei * treati) model term. Similarly, research questions concerning moderation of the treatment effect by the levels of baseline alcohol problems can be tested with the β13(t1 * ccaps * treat) interaction term.

Descriptives
The summary statistics for the primary dependent variable at baseline, situational temptations, revealed that the variable was normally distributed (M= 18.4,  Table 1. It shows that there was substantial dependency among the repeated observations, which was addressed in the longitudinal analyses. Figure 1 displayed the boxplots over time. These findings suggest that the analysis of situational temptations data over time can be performed employing statistical methods for normally distributed outcome data. The CAPS variable was dummy-coded and total problem scores were reclassified as CCAPS to aid the descriptive comparisons of problems. Students who reported no alcohol problems at baseline were used as the reference group. Table 2 shows the sample characteristics and descriptive statistics for each level of CCAPS.

Sample Characteristics
The reference group (CCAPS0= "No problems") contained approximately 23% of the sample and reported the lowest temptations scores (M= 13.9,SD= 4.9). The high problems group also contained approximately 23% of the sample. Figure 2 shows the box-plot for baseline temptations scores were linearly associated with the CCAPS variable.

Unconditional Models
Full information maximum likelihood estimation was used to estimate the growth curve models in Mplus 7. There were 33 response-patterns where missingness in the data occurred and 33% attrition rate at the final wave of data collection. The following section describes the results from the baseline model and the proceeding steps of the unconditional and conditional GCMs. The models were constructed with and without the random effect term included in the models. The model fit statistics for the fixed effects and mixed-models results and the LLRT are shown in Table 3.

Baseline (intercept only) unconditional model.
The baseline model estimated just the intercept in the model, which is the average score of situational temptations at baseline. The model was restricted to just this parameter, therefore the random effect term in the model was estimated to be zero.

Unconditional Quadratic Model
The LLRT test indicated that the fixed quadratic slope model was a statistically better model than the random linear model (see table 3  Although the model fit statistics for the fixed-effects and mixed-effects models were both well suited for these data, these results indicated that the best model of temptations needed to account for random variation in change in the rate of change over time. These results are consistent with the literature that suggests that that the growth trajectories of student's temptations to drink would be nonlinear and declining over time. However, in order to test for differences in the growth rates during and after treatment, the piecewise model was examined next.

Unconditional Piecewise Model
Based on the GCM findings from the previous models, the fixed piecewise model included the random effects on the intercept and the first slope. The ΔAIC (-124.40), ΔBIC (-110.43) and ΔCFI (.03) and provided further evidence that the piecewise model was statistically the best unconditional model. Therefore, the random piecewise model was selected as the best unconditional GCM of situational temptations. Figure 5 shows the unconditional Piecewise model with two line segments that were postulated to represent different intervals of time. Based on the recommendations of Wold (1974), the location of the spline knot that connected the two line segments was at (or near) the inflection point of the change curve. The first slope denoted the intervention period from baseline (0 months) to 6 months. The random Piecewise model's unstandardized regression coefficient estimates for the linear slope during intervention period was positive and statistically significant from baseline to SE=.277,p< .001) (data not shown). The second slope denoted the postintervention period between 12, 18, & 24-months. The regression coefficients for the second slope was negative and was marginally statistically significant (p<.06) between 6 months to 24 months (b=-.145, SE=.08) (data not shown).

Conditional GCM Results
As shown above, the piecewise model with terms for estimating the fixed and random effects for initial status and the slopes for the intervention (t1) and postintervention (t2) time periods was found to be best fitting unconditional model of the growth of situational temptations for these data. The second step of this analysis compared a series of models to evaluate the contribution of conditional variables to the piecewise growth model. Table 5 Table 4. Statistically significant predictors and moderators of situational temptation from Table 4 were then entered simultaneously into a multivariable GCM and the statistically significant predictors and moderators were retained in the final model, shown in Table 6.

Treatment
As shown in Table 4, treatment was not statistically significantly related to temptations at baseline (b= .216, SE= .336, p= .52) or during the treatment phase (b= .085, SE= .549 p= .877), although during the 6 to 24 month post-treatment period, a small but statistically significant reduction in situational temptations was observed for participants in the treatment group (b= -.317, SE= .159 p< .05). These results suggest that the average score of temptations grew over the first six months, regardless of which condition students were placed in. However, there was a small significant difference found in the second slope, indicating that the treatment led to lower temptations scores, on average, relative to the control group.

Female
Female was evaluated first as the only predictor in the GCM then again in the models with the other factors. In the results from single predictor unadjusted model shown in Table 4, females at baseline had significantly lower situational temptations score than males (b= -0.85, SE= .337, p< .01), but gender was not a moderator of growth in situational temptations during either the intervention or post-intervention periods. However, when controlling for the effects of Treat and/or CCAPS in the multivariable conditional models, Female was not significant (Table 6). In addition, the LLRT indicated the final model was statistically better with Female trimmed from the model. Since temptations did not differ by treatment group at baseline this finding suggests that the observed baseline gender difference may be due to the gender differences the rate of alcohol related problems during the three months prior to baseline. The findings indicate that the growth trajectories of temptations would not be altered or change across time based on small differences between men and women's average scores at baseline.

College alcohol-related problems
In contrast, the level of baseline alcohol problems as measured by CCAPS was significantly related to situational temptations at baseline and over time in both unadjusted and adjusted models (Tables 4 and 6). Students who experienced a score of 1 or greater had significantly higher baseline situational temptation scores. After controlling for the effects of treatment and gender, those in the Low (b= 3.626, SE= .385 p< .000), Moderate (b= 6.173, SE= .417 p< .000) and High (b=8.546, SE= .415 p< .000) levels of CCAPS had significantly higher temptations to drink at baseline, compared to the no problems reference group (CCAPS=0) ( Table 6). Additionally, as shown in Table 6, the Moderate (CCAPS2) and High (CCAPS3) groups decreased their temptations during the intervention phase (b= -1.66, SE= .810, p< .05) and the post-intervention phase (b= -1.55 SE= .815, p= .06), respectively). The results indicated students with moderate and high levels of alcohol-related problems at baseline situational temptations had greater reductions in temptations relative to their peers who had no alcohol related problems at baseline. Table 6 shows the statistical significant fixed effect parameter estimates for the final model. The full multivariable model that was initially evaluated included the growth parameters, all of the predictors and the two and three-way interaction terms.

Final GCM Model
None of the three-way interactions in the saturated model was significant, and were not reported in any of the tables. Taken together, the final model results indicated that the average change in temptations was different during each phase of the intervention.
Finally, moderate and high levels of college alcohol-related problems were significantly associated with higher temptations scores.

Discussion
The present study investigated the growth of situational temptations over two years of early college and examined the extent to which treatment, gender and alcoholproblems influenced the growth of temptations over time. The quadratic and piecewise growth trajectories supported the hypothesis that the growth was nonlinear over time, however, the Piecewise model was the best statistical GCM. An important finding from the conditional models was that treatment was associated with the growth of temptations, which is in line with other research that found small reductions in temptations to drink in samples of alcoholics (Witkiewitz, 2013). These findings are also consistent with other research that found situational temptations to smoke decreased over several years following TTM interventions (DiClemente, Prochaska, & Gibertini, 1985;Prochaska, Velicer, Guadagnoli, Rossi, & DiClemente, 1991). The differences that were found between students that received and those who did not receive any feedback were not very large. In fact, the absolute difference between the groups' average scores was less than a 1 point. Nonetheless, the finding that the TTM feedback reports led to a small but statistically significant reduction in temptations is a promising result because of what this could entail for clinicians that provide feedback to their clients.
This study found that students who reported moderate alcohol problem levels p< .05) or high levels of problems (p=.07) lowered their temptations during the intervention period. One plausible explanation for the reduction in that the reduction in temptations may be due to regression to the mean, which refers to the statistical phenomenon that if a variable is at the extreme at the first measurement it will be closer to the average at a subsequent measurements. That is, the finding that temptations scores were reduced for students in the highest two CCAPS groups at the 6 month time point, may simply be due to random statistical variation. Another explanation is that students who had high temptations at the beginning could have reduced their temptations based on their experiences with drinking during the first few weeks of school. Further, students who came to college with many alcohol related problems may have reduced their temptations to drink because of alcohol-related negative consequences experienced during the school year and had an immediate impact on how tempted they were to drink by themselves or with their peers. The latter explanation is partially supported by the evidence from this study. The moderate and high CCAPS groups had significantly decreased growth in temptations scores during the first six months of the study (b=-1.16, se=.81 and b=-1.06, se=.24, respectively), and smaller but non-significant reductions in their temptations scores during the post-intervention period (b=1.55, se=.82 and b=-0.41, se=.24, respectively).
Contrary to my expectations, gender was not related to the growth of temptations nor was it supported in the final GCM. Gender differences in drinking are attributed to rates of alcohol absorption, elimination and impairment. Some studies report gender differences related to treatment Carey et al., 2007), however, the results from this study find no evidence that gender was a moderator of the treatment effect on situational temptations; the test of the gender x treatment x time interaction was not statistically significant. Therefore, I concluded that treatment was equally effective for men and women. These results are consistent with studies that that have found no gender differences in the effectiveness of brief gender tailored alcohol interventions (Bischof, Rumpf, Meyer, Hapke, & John, 2005;Lojewski et al., 2010).
Although the growth of temptations was significantly mitigated by personalized feedback after a period of initial growth, the clinical significance of these findings may not be important. With that said, consider that the average growth of temptations was less than a full point and consider that the average increase in temptations was also less than a full point in the control group. What this may suggest is that temptations to drink takes a long time to change and that for the most part, students' temptations to drink will increase slightly and gradually level-off. Therefore, college administrations should consider investigating other important cognitive factors that are susceptible to change like Self-Efficacy, Resistance or Barriers to drinking to consider the distal effect of these variables on risky drinking.

Model Limitations
It is important to note that the line plot in Figure 4 indicated that the average growth trajectory for the entire sample peaked at or near the end of the first year of college. The growth then leveled off and began to decrease over the following year.
This model indicated that the increase in temptations over the first year would change and the trajectory of growth would appear to flatten out over time. In contrast, the Piecewise model in Figure 5 assumed the growth changed right at the six month time point. These models forced the average rate of change to fit the curve, however, the proclivity of both types of polynomials it that they alter the verticality of the slope(s) of change. In other words, these models may suggest students' temptations will increase and that personalized feedback has a supplementary effect 18 months after treatment, but these models make assumptions about the growth that may not be replicated in other samples.

Study Limitations
The results of this study should be considered in light of several limitations.
The first limitation was that the entire sample was used to construct the unconditional GCMs. Although the large sample size was a strength of the current study, the sample data should have been partitioned into complementary subsets. Then cross-validating techniques on several smaller datasets would be able to compare the statistical models to the results obtained with the full sample. By skipping this step, in theory, could have increased the probability of having committed a type 1 error. Type 1 errors can lead to the improper rejection of the null hypothesis or the competing models.
The second limitation was that this study considered only a handful of moderator variables in this study. This is an important limitation insofar that all of the predictors in the models were either categorical or was polyotomized (CCAPS) into an ordinal-level variable based on the distribution of the CAPS quasi-continuous scale.
Moreover, none of the covariates that were considered in this study measured alcohol consumption. More specifically, number of binge episodes and peak number of drinks should have been considered since these variables are strongly related to alcoholrelated problems (Capone et al., 2007;Kavanagh, et al., 2009;Migneault, 1995;Nelson et al., 2009) and situational temptations is a mediator of risky drinking (Migneault, 1995;. A third limitation of this study was the lack of close and/or equally spaced waves of data. An optimal design would have more assessments and data collection waves closer together and over a longer period of time. The additional time points would allow other equally plausible models to be evaluated as alternative explanations of the growth curve (Bullock, Harlow & Mulaik, 1994;Little, 2013;Singer & Willet, 2003). For instance, a cubic function could have been another plausible model that could have been examined.
Finally, another limitation was that no formal test of covariate 'missingness' was performed. Without a thorough evaluation of the problematic cases, transcription errors, omitted predictors, potential outliers and high-leverage cases could have biased the findings. While the FIML procedure was used in Mplus to provide some protection against covariate dependent missing data bias (Graham, 2003) it is possible that the type of missingness that occurred was missing completely at random (MCAR), missing at random (MAR) or covariate-dependent dropout (CDD), and could have biased the results.
One final limitation of this study is that regression diagnostics were not conducted for the statistical models. This means that the statistical assumptions of the normal linear mixed model used for the unconditional and conditional models were not systematically examined, so we do not know whether these assumptions were satisfactorily met, or whether these models were correctly specified, or if the parameter estimates and fit indices may be biased due to the influence of extreme cases. However, preliminary exploration of the univariate and bivariate characteristics of the variables and preliminary examination of the linear regression fixed estimates did not indicate any obvious distributional problems with the data. Nevertheless, without a thorough evaluation the linear mixed model assumptions and problematic cases for possible transcription errors and high-leverage cases, it is not known if, or to what extent the results presented in this thesis are biased in some way.

Future Directions
Future studies ought to consider using time-varying predictors in the GCM and include drinking outcomes in the model. For starters, including a baseline predictor such as the total number of drinks per week would explain more of the total variation at baseline and possibly the slopes as well. Static variables such as race and ethnicity or Greek life status are also important subgroups to consider in future studies.
Finally, alcohol use is both a psychological and physiological condition but most researchers rely on latent psychological factors to define alcohol misuse and some use technology to construct mappings of neurological structures in the brain and surveys. However, the problem with studying temptations is that the term has moderate overlap in different perspectives, which leads to some neurological studies that consider "cravings" to drink as temptations. Neurological researchers have hypothesized a biological explanation for cravings to drink. Chemical reactions in the brain stimulate neural pathways that cause behavior. Researchers have studied neural pathways neural pathways in the brain that are related to temptations and cravings to drink alcohol with functional magnetic resonance imaging (fMRI) machines. Recent fMRI studies have shown that substance cues and situational stressors can stimulate certain structures in the brain that lead to bothersome thoughts about drinking (Schmidt et al., 2011;Sinha et al., 2009).
The term cravings is more descriptive of alcohol dependence (Drummond et al., 2000;Witkiewitz, 2013)

College Alcohol Problems Scale
Study participants rated how often they experienced the following problems as a result of drinking alcoholic beverages within the six months on a 5-point Likert scale, with 1= Never/almost never to 5= Very often.
1. As a result of drinking, how often have you felt sad, blue or depressed?
2. As a result of drinking, how often have you engaged in unplanned sexual activity?
3. As a result of drinking, how often have you felt nervousness or irritability?
4. As a result of drinking, how often have you driven under the influence?
5. As a result of drinking, how often have you felt bad about yourself?
6. As a result of drinking, how often did you NOT use protection when engaging in sex?
7. As a result of drinking, how often did you have problems with appetite or sleeping?
8. As a result of drinking, how often were you involved in illegal activities associated with drug use?
(Odd numbered items capture the dimension of alcohol related 'personal' problems and the even numbers capture the dimension of alcohol related 'social' problems.)