Multivariate Measurement of Multiple Health Behavior Change and Its Relation to Baseline Severity

Unhealthy lifestyle behaviors have been shown to significantly increase the risk of chronic illness. Interventions changing multiple health behaviors simultaneously show promise in reducing mortality, even more than interventions focusing on individual health behaviors. Multiple health behavior change is a new field with many fundamental questions unanswered, among them how to simultaneously measure progress in multiple health-related behaviors. Previous studies have examined several potential multivariate measurement methods but none have shown consistently superior results. Furthermore various methods may best be suited to specific behaviors, study goals, or intervention contexts. This study proposed to compare several of the most commonly used measurement methods within the context of a study examining how overall baseline severity is related to a person’s ability to make positive health-related changes. This study consisted of a secondary data analysis from three randomized controlled trials for primary cancer prevention in the general population. Participants were proactively recruited based upon at risk behavior for cigarette smoking, sun exposure, or unhealthy dietary fat intake. Behaviors were examined in pairs. Participants were randomly assigned to either a treatment group which received an intervention based upon the transtheoretical model for all at risk behaviors or a control condition receiving only health behavior assessment. Health behavior change methods studied included summative indices, z-scores, standardized residuals, and progress through the stages of change. Several methods were examined to determine which method best describes the relationship between baseline severity and post-intervention outcomes. Results indicate that participants with healthier baseline behavior profiles demonstrated better post-intervention outcomes. Demographic characteristics showed comparatively smaller effects. Methods which allowed for greater detail, such as z-scores and movement through the stages of change showed greater sensitivity as evidenced by larger effect sizes. Summative indices and standardized residuals showed statistically significant results with smaller overall effects. Interventions may wish to consider tailoring based upon participant’s baseline behaviors. Future studies may wish to expand the generalizability of these methods with more diverse populations, different combinations of behaviors, and/or a different set of predictor variables.


LIST OF TABLES
. This number is expected to increase as the general population ages (Parry et al., 2011). Rates of cardiovascular disease and diabetes are also expected to climb (Mathers & Loncar, 2006).
It is well-known that certain health behaviors can decrease the risk of these diseases (Blair et al., 1996;National Research Council, 1989;USDHHS, 1991).
Despite the established association between these health behaviors and chronic illness, very few adults meet these requirements. A study by the Center for Disease Control and Prevention (2007) found that in 2005 only about 27.2% of adults ate as much as three or more servings of vegetables per day and only 32.6% ate fruit two or more times per day. In 2011, 19% of American adults smoked (CDC, 2012). In 2008, only 58% of American adults reported using sun protection methods, such as seeking shade or wearing sunscreen (CDC, 2013;NCI, 2008). Because of the strong connection between poor diet, smoking, sun exposure, and common chronic illnesses, encouraging people to engage in these health-promoting behaviors has become a major public health imperative.

Multiple Health Behavior Change Theory
There are a plethora of theories applicable to health behavior change. A review of recent research using health behavior change theory found that the most commonly employed theories were the Transtheoretical Model (TTM), Social Cognitive Theory (SCN), and Health Belief Model (HBM) (Painter et al., 2008). Until recently, most health behavior research has focused upon single behaviors. There is now evidence that the effects of healthy behaviors are synergistic, such that multiple healthy behaviors lead to greater reductions of illness and subsequently mortality (Ford et al., 2012;Loef & Walach, 2012) and that health behaviors are linked such that persons who engage in one health-promoting behavior are more likely to engage in several behaviors (Berrigan et al., 2003). Furthermore, recent research has shown that individuals who are able to make positive changes in one health-related behavior are more likely to make similar progress on a separate behavior (Paiva et al., 2012). For this reason, multiple health behavior change is gaining increased prominence as a paradigm with the potential to significantly reduce disease-promoting behaviors at a population level (Prochaska, Spring & Nigg, 2008).
Because of its novelty, multiple health behavior change (MHBC) research still has many unanswered questions. Often these questions concern fundamental conceptualizations, such as whether MHBC works via a common set of behavioral change principles that apply equally to all health behaviors, whether general health attitudes give rise to attitudes towards specific health behaviors, which in turn give rise to that behavior, or if change in one "gateway" behavior may lead to subsequent (sequential not simultaneous) change in another behavior (Noar, Chabot & Zimmerman, 2008). Other unanswered questions include what mechanisms cause multiple behavior change to have a greater impact upon behavior than intervention on a single behavior. Furthermore, it is unknown if there is a maximum number of behaviors which can be simultaneously intervened upon (Nigg, Allegrante, & Ory, 2002). While many of these issues will require empirical studies designed to investigate these questions (e.g., RCTs), existing studies may be able to shed some light on these problems.

Measurement
To determine how multiple behaviors change together, these behaviors must be measured together. Prochaska and colleagues (2008a) outline a few of the major methodological challenges of measuring multiple behavior change. These include whether to measure change in each behavior individually or to create a composite score encapsulating change in all the behaviors. Other ideas include creating an index of behaviors in which a person is now meeting recommended guidelines and is no longer at risk. Lastly, there is also the option of more holistic measures such as reductions in mortality, increased quality of life, or via some other biometric. Often, even in research when multiple behaviors are examined, they are considered simply as several single behaviors, rather than as part of an overall behavior profile.
Each behavior has its own metric, such as number of cigarettes per day for smoking and total fat intake for diet. Furthermore, individuals will be at different levels of severity for each behavior. For example, a smoker may consume a few cigarettes or a few packs of cigarettes per day; a person may never eat fruit and has not in many years or may fall only a few servings short of recommended criteria. Each individual will have different combinations of severity for different combinations of behavior. This variability of combinations will likely impact which combination of behaviors a person attempts to change. For example, they may decide to change the behavior which they perceive as the greatest health risk, the one where they are currently farthest from maintaining healthy habits, or the easy-to-achieve low-hanging fruit. This does not even begin to examine the ways in which changes in one area might have repercussions in another area, either deliberately or as a fortunate sideeffect. Additionally, there may be theoretical differences in behavior types, such as addictive vs. non-addictive behaviors and/or adoption vs. cessation behaviors (Noar & Zimmerman, 2005), which may further translate into permutations of behavior combinations, all of which must be considered when attempting to assess global severity of healthy behavior.
Some research indicates that baseline severity, or how much a person must change their behavior to meet recommended guidelines, is related to likelihood of successfully adopting healthy habits. Prior research with these data has shown that persons with relatively healthier initial behaviors are more likely to successfully change their habits, both for single behaviors such as cigarette smoking (Velicer et al., 2007;Redding et al., 2011), diet (Greene et al., 2013), sun protection (Yusofov et al., 2014) and for multiple health behaviors (Blissmer et al., 2010). This may be especially relevant in comparison to demographic characteristics, which cannot be altered and tend to have null or inconsistent effects across treatments (Blissmer et al., 2010).
A few methods of measuring MHBC have been considered. For example, recent research by Kobayashi (2012) has considered several methods of measuring MHBC in a population at risk for cigarette smoking, dietary fat intake, and poor sun protection behavior. These include number of behaviors in which a person is now meeting recommended guidelines, total progression on stages of change scores, or measures on standardized effect sizes (Kobayashi, 2012;Kobayashi et al., 2014).
Other researchers have utilized identical or similar methods including standardized residuals, optimal linear combination, and expanded intervention impact approaches (Carlson et al., 2012;Drake et al., 2013).Other research has examined multiple health behavior change by simultaneous measurement of individual behaviors which have then been analyzed separately (Blissmer et al., 2010

Multicultural Consideration
Many demographic factors are also associated with health-related behaviors and merit consideration. For example, Dehghan and collegues (2011) found that fruit and vegetable consumption varied based upon factors such as marital status, education, age, and gender. Differences in meeting dietary recommendations have also been found across racial groups, with non-Hispanic blacks often showing the worst outcomes (Kirkpatrick et al., 2011). Research on smoking has found significant racial differences in lifetime incidence of smoking and level of smoking severity (Trinidad et al, 2011). Incidence of sunburn has also been found to significantly differ based upon racial group (Buller et al., 2011).
In regards to MHBC, Blissmer and colleagues (2011) found that for behavioral interventions designed to promote better diet, sun protection, and smoking cessation behaviors, stage progression did not significantly differ based upon ethnicity. This is consistent with previous research which has shown that compared to the impact of variables within behaviors themselves, demographics tend to show small or nonsignificant effect sizes (Velicer et al., 2007). Nonetheless, differences in baseline severity have been found to vary across racial groups, and lack of differences in the previous studys' univariate analyses did not preclude the possibility that differences may exist when examined at the multivariate level. Therefore, ethnicity must be considered as a potentially influential covariate.

Hypotheses
Prior research has shown that improvement in more than one health-related behavior produces greater overall health improvement compared to changes in a single behavior. However, because MHBC is an emerging field, no recommended method of multivariate measurement has emerged. Therefore efficacy of several MHBC methods was considered and compared, both in terms of amount of variance accounted for and about which method would prove most sensitive or useful were made.
A person's ability to make health-related changes is inversely related to initial severity of those behaviors. This study examined how initial severity of overall health behavior was related to a person's ability to make health-related changes, using several alternative multivariate measurement methods. It was predicted that those with healthier initial profiles, as implicated by overall severity indices, will show better post-intervention outcomes.

Participants
This study consisted of a secondary analysis of integrated data from several previously collected primary studies. Data were combined from three separate randomized controlled trials. All trials examined three cancer-prevention behaviors: smoking, diet, and sun exposure. The studies utilized similar interventions, including measures, procedures, and assessment time-tables. At-risk participants were proactively recruited from the general population, rather than using clinical samples.
Study data were collected between the years 1995 -2000 and were funded by a grant from the National Cancer Institute. All study participants were at risk for at least one of the behaviors listed above.
Study 1 consisted of the parents of 9 th grade students in a northeastern state (N = 1096). Participants from Study 2 were patients from a list provided by primary care practices associated with a large health insurance company (N = 2417). Study 3 was done as part of an employee workgroup at a total of 22 worksites (N = 684). Total sample size was N = 4197. Details of sample recruitment are recorded in previous literature for Study 1 (Prochaska et al., 2004), Study 2 (Prochaska et al., 2005), and Study 3 (Velicer et al., 2004). The demographics and stage of change distribution were found to be comparable across studies (Yin et al., 2013). Participants were included in the study's main analyses if they had complete, accurate data for the dependent variables at all time-points and could be correctly classified into one of the three primary studies.. All participants were required to speak English, be over the age of 18, provide informed consent, and to be at risk for at least one health behavior.
Primary studies were approved by the appropriate Institutional Review Boards.

Intervention
Study interventions were based upon the Transtheoretical Model (TTM) (Prochaska & DiClemente, 1984). TTM is one of the most established and frequently used theories of health behavior change (Painter et al., 2008). The underlying principles of the model have been found to apply to many different health behaviors (Hall & Rossi, 2008;Prochaska, 1994;Prochaska et al., 1994). Furthermore TTMbased interventions on multiple health behaviors have been shown to lead to improvement on more than one behavior compared to controls (Prochaska et al., 2004(Prochaska et al., , 2005. Participants were randomly assigned to either a treatment or control group. The intervention consisted of a multiple behavior self-help manual based upon TTM strategies and a series of computer-generated individualized feedback reports on all behaviors found to be at risk at baseline. Participants received a five-section report for each behavior, focusing on stage of change, the pros and cons of changing, feedback on up to six processes related to change, suggestions for managing situational temptations, and strategies for taking small steps toward the next stage. Feedback also compared participant progress both to the most successful self-changers within that stage and to data from the participants' prior assessments (Redding et al., 1999;Velicer et al., 2004). Reports were mailed to participants in the intervention group at baseline, 6 months, and 12 months later. Follow-up assessments were made at 12 and months.

Measures
The health-related behaviors included diet, cigarette smoking, and sun protective behaviors. These measures, along with those designed to measure stage of change for each behavior, were used to calculate independent and dependent variables in the analyses. Number of cigarettes smoked per day was used to assess smoking severity, as it is regarded as the single best indicator of smoking severity from Fagerstrom's scale of addiction severity (Fagerstrom, Heatherton, & Kozlowski, 1990). Dietary risk was assessed by measuring total scores on healthy eating behaviors via the Dietary Behavior Questionnaire (DBQ) (Prochaska et al., 2004(Prochaska et al., , 2005Rossi et al., 1996). This scale consists of 22-items assessing food consumption over the previous month. The four subscales correspond to 1) Substitution, or replacing high-fat foods with low-fat foods, 2) Avoidance, or lessening the frequency and quantity of high-fat foods, 3) Modification, or changing cooking techniques to incorporate more low-fat foods, and 4) Fruit and Vegetables, or increasing intake of fruits and vegetables. Internal consistency for adults ranges from α = .67 to α = .84 (mean α = .75). The DBQ has been found to be sensitive to dietary change Prochaska et al., 2004).
Sun exposure was measured using the Sun Protection Behavior Scale (SPBS), a seven item scale for assessing level of sun protective behaviors during sun exposure, with higher scores reflecting more protective sun behavior (Weinstock et al., 2002).
This scale consists of two subscales, Sunscreen Use and Sun Avoidance. For adults, previous research has found good reliabilities, ranging from α = .82 for the total scale, α = .86 for sunscreen use, and α = .82 for sun avoidance. The SPBS has been found to be sensitive to the effects of interventions designed to promote sun protective behavior (Weinstock et al., 2002).
Stage of change for smoking cessation was determined via a 6-item algorithm examining baseline intentions and actions with demonstrated predictive validity (DiClemente et al., 1991). All items consisted of yes-no questions. Based upon their answers, smokers were assigned to the 1) precontemplation stage if they did not plan to quit smoking within six months, 2) contemplation stage if they planned to quit smoking within six months, 3) preparation stage if they planned to quit smoking within the next month and had made at least one attempt to stop smoking for 24 hours in the past 12 months, 4) the action stage if they had quit smoking within the previous six months, or 5) maintenance if they had successfully quit smoking for six months or longer.
Stage of change for intention to reduce risky sun behavior was assessed via a series of six questions. Stage determination followed the same format as that used for smoking, with a few exceptions. The quit attempt at the preparation stage was not included and action criteria was determined by consistently limiting time in the sun to 15 minutes or less or always using sunscreen with a minimum SPF of 15. The overall time-frame was shifted to 12 months rather than six, to account for seasonal differences in sun exposure. This method has been effectively used in prior studies (Weinstock et al., 2000(Weinstock et al., , 2002. Stage of change for dietary fat reduction was determined via a series of three questions (Greene et al., 1994). Participants answering "no" to the question, "do you consistently avoid eating high-fat foods" were assigned to precontemplation, contemplation, or preparation, based upon their intentions to change their behavior and the time-frame of this change. Participants answering "yes," were required to meet a behavioral criterion in which less than 30% of their caloric intake was from dietary fat to be classified as in the action or maintenance stage. Participants not meeting the behavioral criterion were classified as in the precontemplation, contemplation, or preparation stage of change based upon their intentions to alter their eating habits (Greene et al., 1994). This staging algorithm has demonstrated predictive validity (Greene et al., 2012) Additional measures include those utilized in the intervention to create the individualized progress reports. These consisted of the Processes of Change Inventory (Greene et al., 2013;Prochaska et al., 1988;Yusofov et al., 2014), the Situational Temptation Inventory (Velicer et al., 1990), and the Decisional Balance Inventory . All of these were administered in short form, which have been found to be highly correlated with the long-form versions, and additionally have adequate reliability.

Procedures
Preliminary Analyses. A variety of potential analytic methods have been proposed. However, as of yet, there is no established method of quantifying multiple behavior change. Therefore, several were compared. Because this research focused upon simultaneous change in multiple behaviors and the number of individuals at risk for smoking is relatively low (N = 790), only behavioral dyads were considered. The behavioral combinations examined included smoking and diet, smoking and sun protection, and sun protection and diet. Baseline levels of each behavior were operationalized by cigarettes smoked per day, overall healthy diet score, and level of sun protection behaviors.
Preliminary descriptive statistics were run and statistical assumptions including normality, linearity, and homogeneity of variance investigated. To ensure their compatibility, the three studies were compared in terms of gender, marital status, ethnicity, overall health status, and baseline scores on the three behaviors. Attrition rates were examined based upon baseline scores to determine if there were significant differences in whether or not participants completed the intervention. Differences were expected based upon previous studies (Greene et al., 2013). However these were not anticipated to adversely impact results, as the purpose of this study was to establish a new measurement method and not to determine primary outcomes.
Because these same data have been examined in prior studies focused on single behavior change, initial analyses replicated those results as a way to examine validity and reliability (Blissmer et al., 2010;Greene et al., 2013;Redding et al., 2011).
Degree of severity for each behavior was determined by calculating the difference between current baseline behavior and desired behavior for each variable.
Desired outcomes were defined as reaching the action or maintenance stage for the given behavior. For smoking, this was defined as complete cessation. To ensure consistency across behaviors, standardized scores for cigarettes were reverse scored, such that higher scores reflect healthier outcomes. For dietary behavior, this was defined as average scores on the DBQ of at least 86.20 or higher, corresponding to average total scores for those participants reaching the action or maintenance stages.
For sun protection, this was defined as average scores on the SPBS of at least 30.02, the average total score for those participants reaching the action or maintenance stage.
This severity score was then standardized by dividing it by the total sample standard deviation of each behavior, to make severity comparable across behaviors. While the mean score was taken from the subset of participants reaching action or maintenance, the pooled standard deviation from the total sample of participants with data for each behavior was used, so that all comparisons were made against the same base.
Additionally, this provided stability because only a minority of participants ever reached the desired criteria.
Separate standard deviations were calculated for each behavior at baseline and 24 months. Post-intervention outcomes scores were calculated, in the same manner as pre-intervention severity scores. A total severity index was calculated by summing individual scores and converting them into an effect size. In addition, stage of change for each behavior was determined both pre and post intervention, via the methods previously described. A composite stage of change score was determined by assigning each stage a numeric value (precontemplation = 1, contemplation = 2, preparation = 3, action = 4, maintenance = 5) and then summing stage of change scores across behaviors.
Major Analyses. The impact of severity on post-intervention outcomes was analyzed several different ways. To better compare these methods, each analysis utilized the same group of independent variables, differing only in dependent variable and statistical technique used. The first method utilized simple summative indices as the dependent variable in a series of multiple regression equations. Summation indices were calculated by simply summing number of behaviors in which the participant met recommended criteria, as defined by the action or maintenance stage post-intervention.
To rule out potential confounds, several demographic covariates were also run, including age, gender, treatment condition, baseline stage of change, and which of the three primary studies a participant was in. Age, pre-intervention severity, and postintervention severity were continuous variables. Gender, ethnic group, treatment condition, study, and baseline stage of change were run as dummy-coded categorical variables. In addition, interactions of the two independent variables, baseline severity x treatment condition were examined. Both multiple regression equations in which a normal distribution is assumed and Poisson regression equations were run. It was predicted that, consistent with previous studies (Carlson et al., 2012;Drake et al., 2013;Kobayashi, 2012) this method would show statistically significant effects but relatively small effect sizes compared to other methods (see Table 1).
Z-score methods were also used. Post intervention severity scores were analyzed as the dependent variable in multiple regression equations. The independent variable under consideration was pre-intervention (baseline) severity. To rule out potential confounds, several demographic covariates were also run, including age, gender, treatment condition, baseline stage of change, and which of the three primary studies a participant was in. Age, pre-intervention severity, and post-intervention severity were continuous variables. Gender, racial group, treatment condition, study, and baseline stage of change were run as dummy-coded categorical variables. In addition, interactions of the two independent variables, baseline severity x treatment condition were examined. Z-scores were expected to show statistically significant results and account for more of the variance than the summative index method (see Table 1).
The next set of analyses used sum of standardized residuals as a dependent variable. Sum of standardized residuals is an alternate multivariate method similar to z-scores (Carlson et al., 2012;Kobayashi, 2012;Kobayashi et al., 2014). Standardized residual were calculated via a series of multiple linear regression models with postintervention scores as dependent variables and baseline scores as independent variables. Each behavioral residual was calculated separately and then the individual standardized residual change scores were summed to create a composite. Composite standardized residual scores were run as the dependent variable in a series of multiple regression equations. For independent variables, the same covariates included in the zscore method were included in this method. Kobayashi (2014) found this method to be more sensitive to intervention effects than z-scores and it was anticipated that similar results would be found in this study, as evidenced by higher R² values. It was further anticipated that, consistent with prior research, those participants with a better preintervention profile will show better post-intervention outcomes (see Table 1).
The optimal linear combination method was examined via a series of discriminant function analysis equations. The same independent variables were included as predictors. Two dependent variables were used: the number of postintervention behaviors for which a person was still at risk or in other words had not yet progressed to the action or maintenance stage and the number of stages in which they have made progress, even if it did not reach desired criteria. This method served as an alternative to the already established multivariate measurement method of multivariate analysis of variance (MANOVAs) because previous results have found that Manovas account for very little of the variance (Kobayashi, 2012). It was anticipated that, similar to previous research, those participants who showed a better initial profile would show better post-intervention outcomes. Discriminant function analysis is not well-suited to categorical predictor variables. However dichotomous variables may theoretically be included, similar to in multiple regression equations. Therefore, the dummy-coded demographic variables were run (see Table 1).
An alternative method specific to the TTM was to measure progress through the stages of change. Stage of change progress was also measured. For this method, the five stages of change were assigned numeric values, as previously described and composite stage of change scores calculated for both pre and post-intervention severity. The resulting composites were then run in a series of multiple regression equations, similar to the procedure used for z-score examination. The same predictor covariates shall be used. Consistent with previous results, this method was anticipated to account for more of the variance than z-score or standardized residual methods (Kobayashi, 2012). It was also anticipated that those participants who show a healthier initial profile would show better post-intervention results (see Table 1).
Previous research utilizing datasets derived from these studies to investigate the same behavioral dyads found a combined meta-analytic effect size of h = 0.28 (95% CI 0.24, 0.32) for the difference in paired action rates between treatment and control groups (Yin et al., 2013). While these outcomes are substantially different from those used in this study, they represent best available estimates of possible effect sizes. Based upon this prior research and conservative estimates from Cohen's guidelines for small multivariate effects (Cohen, 1992;Rossi, 2013), estimates of required sample size to achieve power exceeding the 1 -β = .80 level were run.
Preliminary analyses showed that each behavioral dyad required an N = 485. Our subsamples were substantially greater than this, indicating adequate power. Analyses were considered statistically significant at the α = .05 level. Selection of the preferred method was determined via the R² effect size, or in other words which method accounted for the maximum amount of variance.

RESULTS
Preliminary Analyses. Analyses were run on SPSS 19.0 and SAS 9.3.
Consistent with prior studies using the same data, the overall sample for all three studies was primarily female, married, Caucasian, and in good or very good health.
Ages ranged from 18.75 to 76 years (M = 44.34; SD = 10.50) (see Table 2 Tables 3 and 4. Descriptive statistics for the dependent variables were also calculated (see Table 5). Number of cigarettes smoked per day showed notable skew and kurtosis, unsurprising given that the majority of the sample was nonsmokers but there were several participants who smoked heavily, creating a large, positively skewed range. All other dependent variables showed signs of statistical normality, as indicated by skew and kurtosis.
Linearity was examined via simple scatterplots of baseline vs. 24 month values on the dependent variables. With the exception of a few outliers, the assumption of linearity appeared to have been upheld. Correlations among the major demographic, independent, and dependent variables were also calculated (see Table 6). While there are many other statistically significant relationships, particularly among the baseline and 24 month scores for each of the three behaviors, none of these relationships are so high that multicollinearity should be a concern (Harlow, 2014).
Comparisons across studies were made via one-way ANOVAs and chi square tests. For gender, there was a significant difference found across the three studies, χ²(2, N = 4181) = 217.037, p < .001,ϕ = .228. The worksite study contained a greater proportion of males than either the parent study or the patient, while the parent and patient studies contained a greater proportion of female participants (Table 2).
There was also a significant difference found across studies for self-reported health status, χ²(8, N = 4180) = 53.396, p < .001, ϕ = .113. Specifically, the parent study contained a greater proportion of participants who reported their health as excellent or very good and the patient study contained a greater proportion of participants who rated their health as only good or fair. The three studies contained similar proportions of participants who rated their health as poor (Table 2).
There were also significant differences found across the three studies for marital status, χ²(10, N = 4175) = 117.598, p < .001, ϕ = .168. The parent contained a greater proportion of participants who were married compared to the patient study and the worksite study. The worksite study contained a greater proportion of participants who were living with a partner compared to the parent study and the patient study. The parent study contained a lesser proportion of participants who were not married compared to the worksite study and the patient study. The parent study also contained a greater proportion of participants who were separated than the patient study and the worksite study. The patient study contained a greater proportion of participants who were widowed than the parent or worksite study. The parent study contained a greater proportion of participants who were divorced compared to the other studies (Table 2).
Significant differences were found across the studies for race/ethnicity, χ²(12, N = 4178) = 46.360, p < .001, ϕ = .105. The parent study contained a greater proportion of participants who identified as Caucasian or African-American compared to the patient and worksite study. The patient study contained a lesser proportion of Asian-American participants than the parent or worksite study. The worksite study contained a greater proportion of Hispanic participants than the other two studies (   (Table 3).
Significant differences across study were also noted for baseline sun protection stage of change, χ²(4, N = 2903) = 16.540, p < .01, ϕ = .075, p < .01. The worksite study contained a greater proportion of participants in the contemplation stage of change than the two other studies (Table 3). Overall these results indicated that there are statistically significant differences in baseline demographics and stage of change across the three studies. However this was consistent with results of prior studies (Blissmer et al., 2010) and did not suggest anything which may influence the overall results of this study.
A comparison of the dependent variables at baseline was also conducted.
Given that there were very different subsample sizes, varying based on the behavior and study, heterogeneity of variance was of especial concern (Harlow, 2014 For the whole sample, missing data was analyzed by comparing the baseline scores of those who had complete data for all three time-points with those who had missing data at either 12 or 24 months. It was found that number of cigarettes per day did not differ significantly between these two groups, F (1, 1714) = .410, η² = .00024 p = .522. There were however significant differences in baseline DBQ scores between those who had completed all three time-points and those who had not 19.73 to 20.25). These overall results indicate that those who completed the study had healthier initial profiles. It must be noted however that the absolute differences between completers and non-completers is very small, often within a single point.
95% confidence intervals and small effect sizes indicate substantial overlap between these scores. This indicates that, while the problem is missing data is noteworthy and should not be readily dismissed, its impact upon overall results may be minimal.

Single Behavior Analyses.
The reliability of single behavior analyses was done via a series of multiple Effects not reaching statistical significance were found for the worksite study (β =.0009, p = .96) and the diet contemplation stage of change (β = 0.02, p =.12). In other words, participants with better post-intervention outcomes were more likely to be older, female, Caucasian, in the treatment condition, in the preparation stage of change, not in the patient study, and have better dietary practices when the study began.
Participants were more likely to show post-intervention success if they were female, Caucasian, in the treatment condition, had progressed past the precontemplation stage of change and already used more sun protection methods pre-intervention.

Summative Indices: Multiple Regression
Preliminary results examined the distribution of participates either meeting or not meeting desired criteria for each of the three behaviors, both individually and in pairs (see Tables 7 and 8). Summative indices scores for behavioral pairs were defined as reaching the action or maintenance stage in each of the three behaviors. Summative scores could be zero, one or two behaviors (see Table 8). A small percentage of participants reached desired criteria on both behaviors post-intervention, and the majority did not reach desired criteria on either behavior.
Summative indices were first run as the dependent variable in a series of multiple regression equations in which a normal Gaussian distribution was assumed.
All participants had not reached action or maintenance criteria on either behavior preintervention. Therefore, number of pre-intervention behaviors reaching desired criteria equaled zero for all participants and could not add information on baseline severity.
This was countered by running two alternate sets of equations with plausible baseline predictor variables. The first set of equations used baseline composite stage of change, as well as the previously described set of covariates, including age, gender, ethnicity, treatment condition, patient study, and worksite study. The second set of equations ran the same covariates, plus the previously described standardized composite severity score and the standardized severity x treatment condition interaction. Because the summative index outcome variable is a count variable with a narrow range, summative indices were also analyzed as the dependent variable in a series of multiple regression equations with a Poisson distribution For smoking and sun protection behavior, there was a statistically significant overall effect, F(7, 533) = 10.76, p < .001, R² = 0.1238. Statistically significant effects were found for treatment condition, baseline composite stage of change, and ethnicity.
Effects not reaching statistical significance were found for age, gender, patient study condition, and worksite study condition (Table 10). Successful participants tended to be in the treatment condition, Caucasian, and further along on the stages of change.
The next set of analyses ran the same predictors via Poisson regression. For smoking and sun protection, there was not an overall significant effect χ² (533)  There was also an overall significant effect for smoking and diet behavior, F(7, 752) = 3.89, p < .001, R² = 0.0349. Statistically significant effects were found for treatment condition and baseline composite stage of change. Effects not reaching significance were found for age, gender, patient study condition, worksite study condition, and ethnicity (Table 12). Participants were more likely to successfully change their behavior if they were in the treatment condition and were further advanced on the stages of change.
For smoking and diet, there was not an overall significant result χ² (752) (Table 15). Participants were more likely to successfully change if they were older, female, in the treatment condition, not in the patient study, and more advanced along the stages of change at baseline.

Summative Indices: with standardized severity
The next set of equations ran the same set of covariates, as well as composite standardized baseline severity, and the standardized baseline severity x treatment condition interaction. For sun protection behavior and smoking, an overall significant effect was found F(9, 503) = 9.19, p < .001, R² = 0.1413. Significant effects were found for treatment condition and composite baseline stage of change. Effects not reaching significance were found for age, gender, patient study, worksite study, composite standardized baseline severity, baseline severity x treatment condition interaction and ethnicity (Table 16). Participants were more likely to successfully change if they were in the treatment condition and more advanced along the stages of change at baseline.
The last set of equations with summative indices used the covariates of age, gender, treatment condition, patient study, worksite study, baseline stage of change, standardized baseline severity, baseline severity x treatment interaction, and ethnicity.
For smoking and sun protection, there was not an overall significant effect χ² (503)  worksite study, baseline severity x treatment condition, and ethnicity (Table 17).
Those who changed their behavior tended to be more advanced along the stages of change, smoke less, and have better sun protection habits at baseline.
There was also an overall significant effect for smoking and diet, F(9, 713) = 7.30, p < .001, R² = 0.0843. Significant effects were found for composite baseline stage of change and composite standardized baseline severity. Effects not reaching significance were found for age, gender, treatment condition, patient study, worksite study, baseline severity x treatment condition interaction, and ethnicity (  (Table 19). Participants who successfully changed their behavior tended to smoke less and have better diets at baseline.
There was also an overall significant effect for sun protection and diet, F(9, 2771) = 56.09, p < .001, R² = 0.1541. Significant effects were found for gender, treatment condition, patient study, composite baseline stage of change, composite standardized baseline severity, and baseline severity x treatment condition interaction.
Effects not reaching significance were found for age, worksite study, and ethnicity (Table 20). Participants tended to be female, in the treatment condition, not in the patient study, be further advanced along the stages of change at baseline, have better diets and sun protection habits at baseline, especially if they were in both the treatment condition and advanced along the stages of change.
For sun protection and diet behavior there was an overall significant effect χ² (2771) = 2222.83, p = 1.0, log likelihood = -1811.60, AIC = 3819.25. Significant effects were found for gender, treatment condition, patient study, baseline stage of change, and standardized baseline severity. Effects not reaching significance were found for age, worksite study, severity x treatment interaction, and ethnicity (Table   21). Participants were more likely to succeed in the intervention if they were female, in the treatment condition, not in the patient study, be farther along the stages of change at baseline, and had better sun and diet habits at baseline.

Z-Scores
The next analytic method consisted of z-scores. Similar to summative indices, a series of multiple regression equations were run. The independent variables consisted of the same covariates, age, gender, ethnicity, treatment condition, patient study, worksite study, and composite baseline stage of change. The main predictor was the previously described standardized composite baseline severity score. Treatment condition x standardized baseline severity interaction was also run. The dependent variable was standardized composite post-intervention scores, calculated the same as pre-intervention scores.
For smoking and sun protection, there was an overall significant effect F(9, 495) = 42.51, p < .001, R² = 0.4359. Significant effects were found gender, baseline stage of change, standardized baseline severity, and ethnicity. Effects not reaching significance were found for age, patient study, worksite study, and standardized severity x treatment stage of change interaction (Table 22). Participants who successfully changed their behavior tended to be female, Caucasian, be more advanced along the stages of change and have better smoking and sun protection habits at baseline.
For smoking and diet behavior, there was an overall significant effect F(9, 696) = 53.87, p < .001, R² = 0.4106. Significant effects were found for gender, treatment condition, baseline standardized severity, and ethnicity. Effects not reaching significance were found for age, patient study, worksite study, baseline stage of change, and standardized severity x treatment condition interaction (Table 23).
Participants who successfully changed their behavior tended to be female, Caucasian, in the treatment condition, smoke less and have better diet at baseline.
For diet behavior and sun protection, there was an overall significant effect F(9, 2792) = 317.18, p < .001, R² = 0.5055. Significant effects were found for age, gender, treatment condition, patient study, baseline stage of change, and standardized baseline severity. Effects not reaching significance were found for worksite study, standardized severity x treatment condition, and ethnicity (Table 24). Successful participants tended to be older, female, in the treatment condition, not in the patient study, be further along the stages of change at baseline, and have better diet and sun protection habits at baseline.

Standardized Residuals
The next set of equations ran in a very similar fashion to z-scores. A series of multiple regression equations were run. The same covariates of age, gender, ethnicity, treatment condition, patient study, worksite study, baseline stage of change, and standardized severity x treatment condition interaction were used. The main predictor was standardized baseline severity.
The dependent variable was standardized residual scores calculated according  Table 25). Standardized residuals from individual behaviors were summed to form composite pairs (Carlston et al., 2012). Slight negative skew and notable kurtosis was found for the smoking variable (Table 25).
For smoking and sun protection, there was an overall significant effect F(9, 495) = 1.99, p < 0.05, R² = 0.0349. None of the predictors showed a statistically significant effect, although gender and ethnicity approached significance. Nonsignificant effects were found for age, treatment condition, patient study, worksite study, baseline stage of change, standardized baseline severity and standardized severity x treatment condition interaction (see Table 26). Participants tended to show better post-intervention results if they were female and Caucasian.
For smoking and diet behavior, there was an overall significant effect F(9, 696) = 2.19, p < 0.05, R² = 0.0275. Significant effects were found for gender and treatment condition. Effects not reaching significance were found for age, patient study, worksite study, baseline stage of change, standardized baseline severity and standardized severity x treatment condition interaction (see Table 27). Participants tended to succeed if they were female and in the treatment condition.
For diet and sun protection, there was an overall significant effect F(9, 2792) = 21.06, p < 0.001, R² = 0.0636. Significant effects were found for age, gender, treatment condition, patient study, and baseline stage of change. Effects not reaching significance were found for worksite study, standardized baseline severity, standardized severity x treatment condition interaction, and ethnicity (see Table 28).
Participants who succeeded tended be older, female, in the treatment condition, not in the patient study, and be more advanced along the stages of change at baseline.

Discriminant Function Analysis: Summative Indices
The next set of analyses consisted of a series of discriminant function analyses (DFA). Previous studies have found that multivariate analysis of variance (MANOVA) may be used as a multiple health behavior change measurement method.
However it accounts for relatively little variance (Kobayashi, 2012) and may not be well-suited towards variables with low correlation between the dependent variables.
Therefore DFA, which is mathematically equivalent to MANOVA, was suggested as an alternative (Kobayashi, 2012).
DFA may also not be ideally suited to all MHBC studies. The dependent variables must be categorical and independent variables are usually continuous (Harlow, 2014). DFA was run twice, once with summative indices, or the number of behaviors reaching desired criteria post-intervention (zero, one, or two), and once using composite post-intervention stage of change, which allowed for measuring progress not reaching desired criteria. The same predictor variables were run, including age, gender, ethnicity, treatment condition, patient study, worksite study, baseline stage of change, standardized baseline severity, and standardized severity x treatment condition interaction. Categorical variables were dummy-coded, as described in previous analyses. Because discriminant loadings do not have significance tests, loadings exceeding |0.3| were considered meaningful (Harlow, 2014). It was recognized that due to unequal group sizes among the outcome variable and the categorical nature of several predictor variables, DFA may not be ideally suited to this type of analysis.
The first set of analyses used summative indices as the dependent variable.
For smoking and sun protection, there remained an overall significant effect and a medium effect size (see Table 29). Of the two linear combinations, only the first was statistically significant F(18, 1004) = p < .001, eigenvalue = 0.1657, canonical correlation = .3770. The discriminant loadings were meaningful for treatment condition, baseline stage of change, and standardized severity x treatment condition interaction (see Table 30). Classification error rates came to .4771, indicating a correct classification of 52.29%.
For smoking and diet summative indices, there was an overall significant effect and a small to medium effect size (Cohen, 1992) (see Table 31). Of the two linear combinations only the first was statistically significant, F(18, 1424) = 4.14, p < .001, eigenvalue = 0.0939, canonical correlation = .2929. The discriminant loadings were meaningful for treatment condition and standardized baseline severity (see Table 32).
Classification error rates came to .5120, indicating a correct classification of 48.80%.
For sun protection and diet summative indices, there was also an overall significant effect including a medium effect size (see Table 33).  Table 34). Error classification rates came to .4834, indicating a correct classification rate of 51.66%.

Discriminant Function Analysis: Stage of Change
The next set of analyses used post-intervention stage of change as the dependent variable. Stage of change was defined as 1 = precontemplation, 2 = contemplation, 3 = preparation, 4 = action, and 5 = maintenance. Final postintervention stage of change for each behavior was summed for form a composite score, providing more detailed information than post-intervention summation scores.
The same set of predictor variables was used.
For smoking and sun protection, there was an overall significant effect including a large effect size (see Table 35). Of the eight linear combinations only the first was statistically significant F(72, 3024.6) = 3.33, p < .001, eigenvalue = 0.3796, canonical correlation = .52457. Of the discriminant loadings, treatment condition, baseline stage of change, and standardized baseline severity (see Table 36). Error classification rate came to .6803. Correct classification rate came to 31.97%.
For diet behavior and smoking, there was an overall significant effect including a large effect size (see Table 37). Of the eight linear combinations, the first was statistically significant F(72, 4302) = 3.03, p < .001, eigenvalue = 0.2306, canonical correlation = .4328. Meaningful discriminant loadings were found for baseline stage of change and standardized baseline severity (see Table 38). The error classification rate came to .7567, indicating a correct classification rate of 24.33%.
For diet behavior and sun protection, there was an overall significant effect including a large effect size (see Table 39). Of the eight linear combinations, the first F(72, 16820) = 14.29, p < .001, eigenvalue = 0.3672, canonical correlation = .51822 and second F(56, 14895) = 2.40, p < .001, eigenvalue = 0.0284, canonical correlation = .1661 were statistically significant. For the first combination, meaningful discriminant loadings were found for treatment condition, baseline stage of change, and standardized baseline severity (see Table 40). The error classification rate came to .7491, indicating a successful classification rate of 25.09%.

Post-Intervention Stage of Change
The final set of analyses proceeded similarly to summative indices. A series of multiple regression equations were run. The dependent predictor consisted of composite post-intervention stage of change, also known as composite baseline stage of change. Post-intervention stage of change composites could range from two to ten with a mean score of 4.4 or 4.60. Skew and kurtosis were within acceptable range (see Tables 41 and 42). This allowed for greater detail, including accounting for intervention progress which did reach desired criteria. The independent predictors included the same covariates of age, gender, ethnicity, treatment condition, patient study, worksite study, and composite baseline stage of change. Equations were run once with only these covariates, using composite baseline stage of change as a measure of baseline severity, calculated identically to post-intervention severity.
Equations were next run these predictors and inclusion of standardized baseline severity and standardized severity x treatment condition interaction, to ensure consistency with prior analyses.
For smoking and sun protection there was an overall significant effect F(7, 533) = 24.61, p < .001, R² = .2443. Significant effects were found for treatment condition and baseline stage of change. Effects not reaching significance were found for age, gender, patient condition, worksite condition, and ethnicity (see Table 43).
Successful participants tended to be in the treatment condition and more advanced along the stages of change at baseline.
There was also an overall significant effect for smoking and diet, F(7, 752) = 14.58, p < .001, R² = .1195. Specifically, there were significant main effects for treatment condition and composite baseline stage of change. Effects not reaching significance were found for age, gender, patient study, worksite study, and ethnicity (see Table 44). Participants who succeeded in the intervention tended to be in the treatment condition and be more advanced along the stages of change at baseline.
There was also an overall significant effect for diet and sun protection, F(7, 2773) = 88.33, p < .001, R² = .1823. Significant effects were found for age, gender, treatment condition, patient study and baseline stage of change. Effects not reaching significance were found for worksite study and ethnicity (see Table 45). Participants who successfully changed their behavior tended to be older, female, in the treatment condition, not in the patient study, and be more advanced along the stages of change at baseline.

Post-Intervention Stage of Change with Standardized Severity
The next set of analyses included the same set of covariates in addition to standardized baseline severity and standardized severity x treatment condition interaction. For smoking and sun protection, there was an overall significant effect, F(9, 503) = 20.70, p < .001, R² = .2703. Significant effects were found for baseline stage of change and standardized baseline severity. Effects not reaching significance were found for age, gender, treatment condition, patient study, worksite study, standardized severity x treatment condition interaction, and ethnicity (see Table 46).
Participants who changed their behavior tended to be more advanced along the stages of change at baseline, smoke less, and have better sun protection habits at baseline.
For smoking and diet behavior, there was an overall significant effect F(9, 713) = 15.02, p < .001, R² = .1594. Significant main effects were found for baseline stage of change and standardized baseline severity. Effects not reaching statistical significance were found for age, gender, treatment condition, patient study, worksite study, standardized severity x treatment condition interaction, and ethnicity (see Table   47). Participants who changed their behavior tended to be more advanced along the stages of change, smoke less, and have better diets before intervention.
Lastly, for sun protection and diet there was an overall significant effect, F(9, 2771) = 106.61, p < .001, R² = .2572. Significant main effects gender, treatment condition, patient study, baseline stage of change, standardized baseline severity, standardized severity x treatment condition, and ethnicity. Effects not reaching significance were found for age and worksite study (see Table 48). Participants who successfully changed their behavior tended to be female, an ethnic minority, in the treatment condition, not in the patient study, be more advanced along the stages of change at baseline, have better diet or sun protection habits at baseline. This was especially so if participants were both in the treatment condition and had better habits at baseline. A summary of effect size measures for each analytic strategy is presented in Table 49 DISCUSSION

Summary of Analyses
The main purpose of this study was to investigate which analytic method produced the best measure of effect, as evidenced by the most inclusive effect size. All analytic methods included the same set of independent variables. compound further if behavioral triplets were used rather than pairs. Therefore DFA, while showing promise, is recommended to be used with caution.
The last set of analyses utilized post-intervention stage of change as the dependent variable. This analysis essentially used the post-intervention equivalent of the pre-intervention composite stage of change. This method also performed well, better than standardized residuals and summative indices although not as well as zscores. Multiple regression was used and although post-intervention stage of change is technically a count variable, the variable was found to mimic a normal curve sufficiently that it could be treated as a continuous variable. Use of regression allowed for use of R² effect size as well as continuous and dummy-coded predictor variables, making it a good overall method. However, post-intervention stage of change is inherently dependent upon the transtheoretical model. While this model is broad in scope and has been found to apply to many behaviors (Hall & Rossi, 2008), there are many other interventions which may wish to utilize MHBC measurements. In those cases, an alternative method, such as z-scores might be necessary.
A few trends were noted among the independent variables also. These analyses utilized the same set of covariates, age, gender, treatment condition, primary study and baseline stage of change, differing only on the presence or absence of standardized baseline severity and standardized baseline severity x treatment condition interaction. Standardized baseline severity added a large amount of variance, almost invariably leading to a large multivariate effect size (Cohen, 1992). Furthermore, the interaction term of baseline severity and treatment condition was rarely statistically significant. In addition, the other measure of pre-intervention severity, composite baseline severity, was also often statistically significant even amongst smaller samples and remained significant with the addition of standardized baseline severity. Severity appears to be a stronger predictor of post-intervention success than any of the covariates, including treatment condition.
That is not to say that the covariates did not provide intriguing information.
Treatment condition was frequently statistically significant if the sample size was large or only the demographic covariates were used. Fortunate, as practitioners would hope their intervention would meaningfully impact behavior. Just as notable, treatment condition was often not statistically significant if standardized baseline severity was included as an independent variable. Treatment condition was a stronger predictor of treatment success than the demographic variables. This may be partly due to this variable being a dummy-coded comparison with other continuous variables such as standardized severity, being more suited to multiple regression comparison. Taken at face value, this further highlights that post-intervention severity is more strongly influenced by pre-intervention severity and stage of change compared to treatment condition. In other words how prepared a person is to make a healthy lifestyle change is more determinant of their success than whether or not they receive an intervention.
As has often been shown in the literature and clinical practice, interventions given to those who are unready or in an earlier stage of change in which they are unprepared for meaningful action will have little effect.
Curiously the interaction effect of standardized severity and treatment condition was also rarely statistically significant, even with a large sample size. One would intuitively suspect that persons who are ready for behavior change might most strongly respond to an intervention designed to help them change their behavior. This was not the case. Perhaps once a person has decided on their own to make lifestyle changes, they will seek out resources that will help them make changes on their own, regardless of what interventions are available. It is further possible that those who were in the control condition but were already close to desired criteria, sought out ways to improve their behavior without prompting from the researcher. The very act of being in a health behavior study might have provided sufficient motivation for an already motivated subgroup.
The statistical significance of age and gender tended to vary based upon sample size, indicating that while these may have some effects, they tended to be less important than other variables. Usually when the demographic covariates were statistically significant, participants showed better post-intervention outcomes if they were older, female, and Caucasian. However for each of these trends there was at least one analysis in which better outcomes were found if the participants were younger, male or from an ethnic minority. Overall this highlights that when designing interventions for certain behavioral combinations, those interventions should be tailored to certain subsamples which may respond differently.
Overall, the independent variables were consistent with prior studies.
Statistical significance tended to vary based upon sample size, with variables being much more likely to be statistically significant in the larger subsample of sun protection and diet behavior. In smaller samples, the demographic covariates tended to show non-significant or small effects, consistent with prior studies showing that demographic effects were inconsistent or small (Blissmer et al., 2010). This was especially so in comparison to more malleable concerns such as baseline stage of change and baseline severity.

Limitations and Future Directions
There are other variables which researchers may wish to examine in future studies. It was determined that baseline severity and the calculated change indices were so highly correlated that they could not be included in the same analyses.
Therefore amount of change was excluded as a possible independent variable. Other variables which this study did not examine include level of effort. This could be measured via the transtheoretical model's processes of change, decisional balance, self-efficacy, and temptation constructs, which might show significant effects between level of effort, post-intervention success and level of change.
The previous methods examined MHBC at the individual level. Another method, the expanded intervention impact score, is based upon an intervention's total effect upon the general population (Drake et al., 2013) via the intervention impact formula of intervention impact = efficacy times participation (I = E x P) (Velicer & Prochaska, 1999). The formula could be expanded to include multiple behaviors (I = Σ number of behaviors (n) (En x Pn) (Prochaska, Velicer et al., 2008b), where P is the proportion of individuals who are at risk for each behavior, and E is the estimated efficacy of intervention for each behavior, defined as the percentage of participants meeting recommended guidelines at follow-up. However, because this is a populationbased measure, it cannot be used to examine the impact of baseline severity upon individual outcomes and was unsuitable for this study. Future studies interested in MHBC at the population level may wish to utilize this method.
There are a few further limitations which future research may wish to consider.
Only three cancer-prevention behaviors were considered and those behaviors were examined in pairs. There are other behaviors worth consideration, such as exercise, compliance with prescribed medication regimes, and responsible alcohol intake.
Behaviors may also be modified in triplets or even with four or more behaviors. It is currently unknown if there is a maximum number of behaviors which may be simultaneously modified. The increase both in potentially modifiable behaviors and behavioral combinations leads to the question of which behaviors are best changed together. Past studies of behavioral pairs have found that change organized around an intuitive theme, such as healthy energy balance with diet and exercise, showed greater effect sizes than those behaviors which are less obviously linked, such as smoking and sun protection (Yin et al., 2013). There may also be behaviors which do not lend themselves as well to MHBC. Behavioral pairs involving smoking for example have been found to show comparatively smaller effect sizes (Yin et al., 2013). These results are in conflict with the findings of this study which showed that the two behavioral pairs involving smoking have similar effect sizes to sun and diet. Broad results cannot be drawn from such a narrow comparison however and it is worth exploring which behavioral combinations product the greatest overall behavior change.
Behaviors may also be aimed at reducing the risks of other illnesses, such as diabetes or heart disease. These behaviors have growing prevalence rates, similar to cancer and tend to have risk factors similar to cancer. Interventions with the aim of promoting general as well as illness-specific health have the potential to greatly improve public health.
Furthermore this study was comprised of primary prevention data from persons who are not ill. Research has shown that cancer survivors, far from having more careful health behavior, show comparably high rates of risk behaviors such as smoking, diet and risky sun protection behavior. Indeed younger cancer survivors are more likely to smoke than non-cancer controls (Coups & Ostroff, 2005). Yet health behavior change during treatment or to prevent remission is now considered a key part of cancer treatment (Demark-Wahnefried, 2005;Pinto & Trunzo, 2005). Studies are currently underway to determine optimal ways to encourage healthy behavior change in cancer survivors. Encouraging multiple healthy behaviors amongst this population certainly has a place and great potential for improved public health.
Sample Limitations. Additionally, despite its large size, this sample was relatively homogenous. The lack of racial/ethnic diversity is of particular concern, as previous studies have shown that differences in health-related behaviors tend to vary along ethnic lines (Buller et al., 2011;Trinidad et al., 2011). Ideally, ethnicity would be included in the study as a moderating variable so that its independent influence and effect size could be considered. However a few practical limitations precluded this.
The sample was overwhelmingly Caucasian. The subsample sizes were so unbalanced that any independent effects of ethnicity could not be found. Indeed, ethnicity was rarely found to be statistically significant, even when the sample was so strongly powered that other normally non-significant effects, such as patient or worksite study, were statistically significant. Furthermore because ethnicity groups were so uneven, at best, ethnicity could only be dichotomized as Caucasian vs. ethnic minority. This meant that minority groups such as African-American, Asian-American, Native American, Hispanic, and multiracial were all lumped together even though there is certain to be considerable differences in health concerns and health behavior across these groups. Because the purpose of this study was to evaluate methodologies rather than to evaluate an intervention or describe health-related behaviors, these limitations did not negatively impact the overall findings. However future studies with different aims should take these difficulties into consideration by selecting a more diverse sample.
There were a few other limits in generalizability. Females outnumbered males in the study. Most participants were middle-aged. The samples were recruited from the northeastern United States. Results may not generalize to people from different geographic regions. Senior citizens and children have different health concerns than middle-aged adults, which would be reflected in different health behaviors.
Additionally the majority of participants reported good health. In 2000 45% of the American public had a chronic health condition with 21% having multiple chronic conditions. These numbers are projected to grow over the next several decades (Anderson & Horvath, 2004). Therefore this relatively healthy sample may be less than representative of the general public.
Female participants were usually shown to have better habits than males. On the surface this might reflect a greater health awareness or concern among females. It might also reflect the behaviors. Women tend to more concerned with sun protection (Weinstock et al., 2000) and diets compared with men (Dehghan, Akhtar-Danesh, & Merchant, 2011). Although clinical samples have generally shown men to have higher success rates with quitting smoking, other studies using the general population disagree and say neither gender is more likely to quit smoking (Jarvis et al., 2013). In other words, for two of the three behaviors, women tend to perform better than men and the third behavior shows comparable gender rates. Men may be more inclined towards other behaviors such as exercise adoption (Loprinzi & Cardinal, 2012).

Methodological Limitations.
Because the primary purpose of this study was to establish a multivariate methodology for health behavior change research, several methodological limits must be addressed. To begin with, only three behaviors were studied and those behaviors studied in pairs. This was deliberately done so that pairs, the most basic unit of multiple health behavior change, might first be examined. There is as yet no theoretical reason why pairs should behave differently than three or more behaviors. However this has not yet been established.
There were also differences in how desired criteria were determined. Smoking cessation had a clearly defined public health guideline which participants strove towards, namely zero cigarettes per day. Dietary behavior was measured via the DBQ and sun protection via the SPBS. Desired criteria scores were determined by examining the average scores of persons meeting either action or maintenance stage of change post-intervention. Therefore the exact 'desired criteria' scores for this study will differ from cut-off scores for different studies.
The variability of public health guidelines must also be considered. At the time of data collection, excessive dietary fat intake was regarded as a risk factor for chronic illness. Since then research has shifted from quantity to quality. Certain types of fats are currently shown to have protective effects against illnesses such as cancer (Schwab et al., 2014;Zheng et al., 2013 Smoking has an advantage over diet and sun protection in that there is an unambiguous public health goal, complete cessation. There is also an easy to measure behavior, cigarettes per day. A disadvantage is that this method relies upon a single item measure. Single item measures are widely reported as less reliable than scales (Wanous & Hudy, 2001) although some studies indicate they perform comparably (Ginns & Barrie, 2004). Smoking behavior showed notable skew and kurtosis. Most people, even those who smoke and have no immediate plans to quit, are aware of the health risks (Hammond et al., 2006). Heavy smokers are rare and the rate of heavy smoking is decreasing (Jamal et al., 2014). This leads to a positive skew, with most participants smoking few cigarettes and a few heavy smokers. Such data violates the statistical assumptions inherent in most analyses, including multiple regression and discriminant function analysis. Transformations or the removal of outliers can often be used to correct the problem (Kobayashi, 2012;Osbourne & Water, 2002 (Harlow, 2014). Because regression has the option of dummy-coding as a method of handling categorical predictor variables and has multiple methods for handling slight violations of statistical assumptions, is much more ubiquitous, and easy to use, multiple regression may be a preferred method over discriminant function analysis.
This study used complete case analysis, excluding those participants who did not provide complete data or could not be classified into one of the primary studies.
Comparisons of the study indicated that baseline differences between participants who completed the intervention and those who dropped out or provided incomplete data tended to be minimal or nonexistent. Because the study's main purpose was methodological rather than descriptive, these differences did not detract from the overall results. However future research may wish to utilize more advanced missing data techniques such as multiple imputation. This may also strengthen the case for multiple regression as a favored MHBC method, there are quite a few multiple imputation methods specialized for regression (Graham, 2012)  Despite these limitations, this study had several valuable strengths. Three pairs of behaviors, including two adoption behaviors and one cessation behavior were examined with consistent results. The study successfully replicated much of the previous research, while also discovering several areas ripe for future research. Crossmethods comparisons were made with the same set of predictor variables, such that any differences in effect size might be definitively attributed to how the dependent variable was calculated. multiple health behavior change in a primary cancer prevention intervention.
Standardized measures of baseline severity were examined as a primary predictor.
Overall results found that multiple health behavior change methods which allow for greater detail, such as z-scores and movement through the stages of change, account for a greater amount of variance than simpler methods such as summative indices.
Standardized residuals do not appear to be well-suited towards research in which standardized severity is a primary predictor variable. Given the ease of use and ubiquity of multiple regression, this method may be preferred over the more esoteric discriminant function analysis.