Validation Study of a Behavioral Based Instrument for Dietary Fat Consumption

This study aimed to validate the dietar y behavior que stionnaire (DBQ) developed by a group of researchers from several different fields including nutrition , nursing and psycholog y (Rossi et al , 2008) adapted from the work of Kristal and colleagues (Kristal et al, 1990). The new instrument was designed to mea sure the behavior of dietar y fat intake and involved 22 items with 4 theoretical constructs: Substitute , Moderate fat intake , Modify cooking , and Increase healthful foods . This study utilized multiple psychometric technique s including PCA , CF A, factorial invariance comparison (multi-group CFA) , MANO VA and ANOVA using data from Rhode Island parents of high school student s. The new data correctly replicated the original 4 factor models. The resu lts from facto rial invariance comparison using Cheung and Renswo ld's method suggested that the factor structure , loading pattern as well as covariance structure are invariant between pre -action stage group and post-action stage group and between females and males. All 4 theoretical constructs had adequate Cronbach's alphas (range: 0.62 0.77) . MANOVA results suggested that the DBQ was correctly discriminated by TTM stages of change. Psychometric findings sugge sted that the new measurement had goo d reliability and validity , and is ready for the use in future fat intake research, especially in Tran stheo retical model based behavior al research. ACKNOWLEDGMENTS I would like to thank my major professor Dr. Joseph Rossi for his exceptional support and incredible guidance throughout my years at the Psychology department. I would also like to thank Dr. Greene and Dr. Redding for serving on my graduate committee. Many thanks to Dr. Susan Rossi for being a chair of the committee. Finally , I would like to thank Dr. Redding for her generous permission to access her Parents Study data.


LIST OF FIGURES
consumption of Increase healthful foods has the potential to prevent certain diseases.
Unfortunately , most Americans consume excess dietary fat and consume too few Increase healthful foods (Rossi et al., 2008. Encouraging people to reduce dietary fat and increase healthful foods consumption will reduce the risk of many serious diseases. To assess and monitor dietary interventions, a standard behavioral assessment tool is needed. The assessment tool must be reliable, comprehensive, and functional. The result of this study is expected to yield a psychometrically sound and validated dietary instrument. The purpose of this study is to validate one of the asse ssment tools , the dietary behavior measure (Rossi et al, 2008) from a psychometric perspective.

Dietary Behavior Questionaire (DBQJ
The dietary behavior questionnaire (OBQ) was developed by a group of researchers from severa l different fields including nutrition , nursi ng and psychology (Rossi et al, 2008) based on the work of Kristal and colleague s (Kristal et al, 1990).
The most meaningful difference of these measures compared with most other dietary measurements was that they focused on behaviorally based fat consumption rather than frequency or amount of fat consumed. Food frequency questionnaire s have been widely used in dietary research. These mea sures assess how often several foods and nutrients are consumed (Block , 1982;Medlin , 1988). These measures have had high functionality (Willett, 1994), but numerous of researchers have criticized their validity and accuracy (Birkett & Boulet , 1995;Briefel et al, 1992). Kri stal and colleague s developed a first generation behaviorally based fat consumption measure utilizing an anthropological theory (Jerome, 1976). They also found that measurement utilizing behavior was more sensitive to dietary change than a food frequency questionnaire (Kristal et al, 1994) . However, several studies have failed to replicate the factor structure of their instrument (Birkett & Boulet, 1995;Greene et al, 1994). Rossi and colleagues developed a new version of a behavioral based fat consumption measurement (Rossi et al, 2008).
In the creation of this new measure, the researchers developed several criteria: the measurement should be short, inexpensive to conduct, appropriate for use in survey and intervention research, low in subject response burden, suitable for telephone interview or self-administered mail format, and acceptable for regular readministrations and behaviorally based to allow for individual feedback on specific behaviors. They employed a sequential method for scale construction that utilized both qualitative and quantitative analyses to make their measurement more psychometrically powerful (Redding et al, 2006). Literature review, conducting of focus groups, item generation , and expert review were included in the initial qualitative analyses. In quantitative analyses, they utilized item analysis, principal components analysis (PCA), measurement modeling, and validation on subsequent samples. Many of the initial items were adapted from existing versions of the Food Habits Questionnaire (FHQ) (Kristal et al, 1990). At the base point , they included 61 potential items from the FHQ and utilized a 5-point Likert scale (I = never to 5 = almost always). In this process , the study collected 434 respondents by mailing surveys to randomly assigned Rhode Island residents. PCA was performed to ascertain the measures structure and shorten the questionnaire. Based on the PCA results , they reduced the instrument to 22 items from the initial 61 questions.  (Rossi et al., 2008).

The Transtheoretical Model (TTM)
This new version of behavioral based fat consumption measurement involved the Transtheoretica l Mode l (TTM) (DiClemente & Prochaska, 1982;Procha ska, DiClemente & Norcross, 1992). The TTM is a usefu l framework for understanding behavioral change. The model postulates constructs including ten processes of change (Prochaska, Vel icer, DiClemente , & Fava, 1988), five different stages of change (Di Clemente, et al., 1991 ), deci sional balance  and selfefficacy and temptation (Ve licer , DiC!emente , Rossi, & Prochaska , 1990). The TTM was or iginally applied to smoking behav ior, and the applicatio n expanded across other health behaviors such as alcohol use, sunscreen use, condom use, exercise and dietary fat intake (Gree ne et al., 1994;Vallis et al., 2003;Prochaska et al., 1994). In the conceptual dimensions of the TTM, "the stages of change" construct is the core. "The stages of change" refers to the tempora l nature of a behaviora l change . The construct contains five distinct stages: precontemplation (PC) , contemplation (C) , preparation (P), action (A) and maintenance (M) (Prochaska, DiClemente , & Norcross, 1992). A person in the precontemplation stage would have no intention of changing his current behavior. A person in the contemplation stage would be thinking about changing his current behavior within the next six months. In the preparation stage , a person would think about changing his current behavior within the next 30 days. A person in the action stage is engaging in the target behavior but has been doing so for less than 6 months. Finally, a person in the maintenance stage has been actively engaged in the target behavior for over six months. The target behavior in this study is reducing the amount of dietary fat. The processes of change contain ten overt and covert strategies that individuals utilize in order to modify, adopt, or eliminate a target behavior. The I 0 processes of change can be divided into two categories: experiential and behavioral.
The experiential category is defined as processes that promote change through the use of emotional and/or cognitive strategies. The behavioral category is defined as processes that promote change through the use of specific strategies and actions.
Decisional balance refers to an individual's balance of pros (advantages) and cons (disadvantages) for adopting a specific behavior. Self-efficacy refers to an individual ' s confidence in their ability to perform the desired behavior. Temptation refers to how tempted an individual is to not engage in the desired behavior in a variety of situations. The basic application strategy of the TIM to a target behavior involves assessing the stage of change, then based on that assessment information, building and promoting the use of individual tailored processes of change aimed at moving one through the stages. Both cross-sectional and longitudinal studies have shown evidence to support the rationale of the TTM (DiClemente et al., 1991). The DBQ was built as a reliable measure, especially useful for TIM based interventions of fat reduction and is also applicable for other non -TTM based studies as well.

Interventions
Currently, a number of clinic-based options for the treatment of obesity are available. TTM tailored interventions have been adapted for this particular behavior as well as intervening on multiple behaviors , such as healthy eating, exercise and managing emotional distress simultaneously (Johnson et al., 2008). Intervening on multiple healthy behaviors simultaneously is one of the most effective ways to promote healthy life style (U.S. Department of Health and Human Services, 2000).
The multi -intervention strategy is an effective way to reduce health care costs (Edington,200 I). A number of studies support an advantage in using TTM tailored interventions (Prochaska et al., 2004(Prochaska et al., , 2005. Jones et al. (2003), for example, compared usual care with TTM tailored treatment in participants with diabetes and showed a significant advantage in using TTM tai lored treatment. They recruited 1,029 individuals with type I and type2 diabetes. All individuals were in one of three preaction (i.e . precontemplation, contemplation and preparation) stages for one of three behaviors; self-monitoring of blood glucose (SMBG) , healthy eating and smoking. were also evident in Hemoglobin AIC measurements, an index of diabetes severity

Hypotheses
This study was designed to investigate the validity of the dietary behavior questionnaire (DBQ) (Rossi et al, 2008). The most macro level hypothesis was that the DBQ should be a reliable and consistent measure across different target populations, with the new data set replicating the same factor structure as the original study showed (Rossi et al, 2008) . Figure 1  1 a. The new data should represent that the first factor (i.e. Substitute) is measured by question items 1 to 5 (see Table I and Figure I). This hypothesis demonstrates that question items 1 to 5 are reliably measuring substituting high fat foods with lower fat alternatives .
The new data should represent that the second factor (i.e. Moderate fat intake) is measured by question items 6 to IO (see Table I and Figure   1 ). This hypothesis demonstrates that question items 6 to 10 are reliably measuring eating high fat foods less often and in smaller portions.
le. The new data should represent that the third factor (i.e. Modify cooking) is measured by question items 11 to 15 (see Tab le 1 and Figure 1 ). This hypothesis demonstrates that question items 11 to 15 are reliably measuring reducing dietary fat in food preparation. l d . The new data sho uld repre sent that the fourth factor (i.e. Increase healthful foods) is measured by question items 16 to 22 (see Table l and Figure 1 ). This hypothesis demonstrates that question items 16 to 22 are reliably measuring eating more fruits, vegetables, and grains.
2. The final correlated 4 factor model should provide an adequate fit to the data; CFI>.90 and RMSEA <.08. This hypothesis demonstrates that the DBQ model shou ld fit well w ith the new data.
3. The final corre lated 4 factor model also should have a potential to have one higher order factor.
4. The TIM stages should represent a similar factor structure (configura l invariance) and similar fit parameter matrices (i.e. metric invariance). The hypothes is explains that the DBQ's stability of across different stages.
5. Male data and female data shou ld represent similar factor structure (configura l invariance) and similar fit parameter matrices (i.e. metric invariance). The hypothesis explains the DBQ 's stability across gender.
6. The new data shou ld show significant mean differences across the TTM 's stages of change , and should have adequately high effect sizes. The hypothesis explains that the DBQ should have "known groups va lidity" (Redd ing et al., 2006).

Participants
This stud y is a seco ndary data analysis of a data set w hich was collected in the "Parent Study 2002" (Redding et al., 2002). The Paren t Study was app roved by the Institutional Review Board for Human Subjects and recruited 2547 target parents who provid ed informed consent. After the participants agreed to participate the study , they we re randomized into one of three intervention group s: I). Benchmark expe rt system (ES) (diet, sun, smoking); 2). New Behaviors ES (exe rcise, stress , alcohol); and 3).
Enhanced ES (diet, sun, smoking), using URN randomization based on gender (ma le; female). Since the aim of the "Parents Study 2002" targeted sets of multiple beh av iors, the participants in different groups were given diff erent beh av ioral surveys. This stud y on ly investigated participants surve yed about dietar y behavior (i.e. participants who completed the diet behavi or questionnair e) . In the Parent Study , the part icipating parent s' spouse or significant other we re also recruit ed into the study. Their group was not randomized, but linked to the group of target parent. There was a stron g correlation between the pare nts and their partner s. Thus , we conducted the analyses after excluding partner s from the data . After excludin g the part ners and participants w ith drop-out and missing data , 1366 participants remained. negat ive item s, for higher fat food consuming habit s and low healthful food consuming habits, and appear alternatel y. The number of positiv e and negative items per scale is not equal , but contain s enough items to control response biases related to acquiescence . All the DBQ items utili ze a five-point Likert scale (5=Almost always, 4=Ofte n, 3=Sometimes. 2=Rarely, 1 =Never) which has been preferred by several researchers (e.g., Reddin g et al., 2006 ). Thi s study excluded one item "How often did you reduce the amount of butter , margarine, or oil in a recipe to cut down on fat?", and focused on 22 item s in the mea sure. Reasons for this exclusion were 1) The original mea sure developed by Ro ssi and colleagues (Prochaska et al., 2004;Rossi et al., 2008 ) did not contain thi s item, 2) Any theoreti cal supp ort for adding this item could not be found , 3) There was no theoretical information about which factor (sub co nstruct) reasonably high coefficient Alpha. These psychometric results have supported that it is reasonable to mea sure these set of constructs and supported that some part of background assumptions for conducting SEM (Joreskog, 1967(Joreskog, , 1969 procedures.

Analyses
To assess the validity of the dietary behavior questionnaire (DBQ)'s psychometric properties, this study conducted seve ral psychometric procedures including multivariate analysis of variance (MAN O VA), principal components analysis (PCA), confirmatory factor analysis (CF A), and invariance nested model comparisons. Some of these psychometric procedures are included in the structural equation modelin g (SEM) framework. In this study, AMOS6 was used for SEM procedures. Other psychometric procedures were conducted, such as calculating descriptive statistics, PCA, and MANOVA using SPSS14 and SAS 9.2.
As an initial step of the analysis , the overall samp le was divided into two subgroups to conduct a "cross -validation" approach . These two subgroups were randomly selected from the overall sample. This procedure was done using SPSS 14.
In this procedure , descriptive statistics of demographic data were checked for each half of the data and are presented in Table 3. The cross-validation approach was adapted to only PCA and reliabi lity analysis (i.e . Cronbach's alpha). In general, PCA is conducted on the first subgroup as an exploratory group and CF A is conducted on the second subgroup as a validation group in a measure development study. However , since the goal of this study was validating an instrument and did not intend modifying the scale, the PCA and the reliability ana lysis were conducted on the first and the second subgroups to see the replication of each result. On the other hand, CF A and MANOVA were conducted to see the validity of original measure. The results of CFA was compared with original study, and MANOV A tested theoretical "known groups" validity, thus these procedures were conducted on the full sample.
Second, PCA was performed to va lidate the model structure among the manifest and latent variables in the DBQ. This analysis was performed on both the first and the second samp le to see the measurement reliability among different samples . To determine the number of factors, thi s study looked at both Kaiser's rule (i.e. retaining the factors having eigenvalues greater than 1.00) and parallel ana lysis (PA) (Lautenschlager, 1989). Also, to interpret factor structures, Varimax with Kaiser normalization was used as the rotation method. This analysis invest igates how many factors exist and how variables and factors relate. The factor structure among manifest and latent variables was compared with the original model structure proposed in the initial study (Rossi et al, 2008).
After validating the factor structures, CF A was performed to test the actual model fit. The CFA was conducted on a full sample (N=1366) using AMOS6 and maximum likelihood estimation. The result of the CFA procedure was used not only to see model fit, but for Modify cooking of model struct ures to improve model fit. This study evaluated multiple fit indices: x2 value, the comparative fit index (CFI) and the rootmean-square of approximation (RMSEA). However , since overall x 2 value is highly influenced by sample size (Byrne, 2001;Kline, 1998) , not much weight was attributed to the actual values of x2-The CFI evaluated the fit between the null model and the proposed model CFI values greater than .90 indicate a reasonably good fit (Hu & Bentler, 1999). Also, a value of RMSEA less than .05 indicates a good fit, RMSEA between .05 through .08 indicates fair fit, and RMSEA between .08 through .10 indicates a mediocre fit (Browne & Cudeck, 1993). After the se procedures, the initial model from the original study presented in Figure 1 (Rossi et al., 2008) was compared with the new proposed model from this analysis.
Next, this study investigated the measure's stability among several different groups, based on gender and the stage of change (Di Clemente et al., 1991 ). A factorial invariance comparison (Byrne, Shavelson, & Muthen, 1989) also known as Multi Group Confirmatory Factor Analysis (MGCFA) was performed on these groups . This procedure tested four models; the con generic model ( con figural invariance) , the lambda-invariant model (metric invariance) , the tau-equivalent model, and the parallel model. The con generic model ( con figural invariance) is the least restrictive model assuming the two samples have the same factor structure , but different loadings , factor variances, and error variances. The lambda-invariant model (metric invariance) is the same as the congeneric model , but also assumes factor loadings to be equal across both samples. The tau -equivalent model is the same as lambda-invariant model, except it also assumes the same factor variances and co-variances across the two sample s.
The parallel model is the most restrictive model , assuming every parameter estimate to be equal across both samples. These models represent a sequence of progressively more restricted tests of structural invariance across the two samples. This study utilized the bottom-up procedure by starting with the least restricted model and progressivel y moving to more restrictive models using x 2 difference test. Significant x 2 indicates that the more restricted model is a significantly worse fit than the less restricted model should be retained , and shou ld retain the less restricted model. This st udy also utilized goodness of fit indexes: CFI, unbiased GFI (gamma-hat) (Steiger, 1989) , and McDonald's ( 1989) Non-Centrality Index (NCI) , to test measurement invariance (Cheung & Rensvold, 2002 procedures also included Tukey ' s multiple comparison procedure to test wh ich levels of independent variable (the stages of chan ge) contained significant mean differences.
Effect sizes were also computed for each model. In the MANOV A procedure, multivariate 17 2 was computed and was interpreted as a multivariate effect size. In ANOV A procedures, 17 2 was computed and was interpreted as a univariate effect size .

Split Half
The original 1366 participants were split into one of two groups randomly based on SPSS split half function. The first group was set as an "Exploratory Group" and mainly analyzed in the study. The second group was set as a "Validating Group" and analyzed only for validating purposes. Table 3 presents the descriptive statistics for both samples. As expected , all the demographic information between the 2 groups was similar. The reliability analysis was conducted for both samples .

Principal Components Analysis (PCA)
The purpose of this analysis was to investigate the model structure among the manifest and the latent variables in the dietary fat intake measure (DBQ). The first PCA was performed without setting the number of components. Using Kaiser's rule, the result showed that 5 components had eigenvalues greater than 1, and those 5 components explained 52.09% of variance. Using parallel analysis (PA) approach (Lautenschlager , 1989), the results showed that 5 components had eigenvalues greater than the PA's criteria.

Confirmatory Factor Analysis (CFA)
The purpose of this analys is was to investigate actual model fit using a full sample (N= 1366). First CFA was performed based on the 4 constructs that resulted from the PCA result (see table 6 & 8). The model fit summary show ed the x 2 value was significant, x 2 (203) = 1239 .85, p<0.001. Th is indicated that the model did not adequately fit the data . However, as mentioned ear lier, the x 2 test is highly influenced by the samp le size , and there is a high probabilit y of getting statistical significance with large samp le sizes (Byrne, 200 1;Kline, 1998) . The model fit summary also showed the Com parative fit index (CFI) = 0.85 and RMSEA = 0.06 1. This CFI was not impressive as approaching 0.9 is considered as good fit. This RMSEA indicated the model fit was fair.
A second CF A was performed based on the original factor structu re proposed by Rossi et al (2008) except that the second model used correlated factors instead of a higher order factor. Figure I presents the orig inal CF A model. The result showed that the x2 value was significant, x2(203) = 1514.27, p<0.001, CFI = 0.8 I , and RMSEA = 0.069. As expected, the model fit was worse than the first model. However, since the main focus of this study is va lidating the original measure, using the second model was theoretically more rational than using the first model that resulted from the PCA.
Therefore, decision was made to utilize the second model as a base model to improve the model fit.
A third CF A was performed based on the second model and its modification indices.

Factorial Invariance Comparison (Multi Sample CFA)
The purpose of this analys is was to investigate model stability among different populations (i.e . pre-action vs. post-action, fema le vs. male). This procedure tested 4 models (i.e. Congeneric, Lambda-invariant, Tau -equiva lent, Para llel) for each of two sets of compared groups. First a group in the Precontemplation , Contemplat ion, or Preparation stages was examined against a group in the Action or Maintenance stages.
Second, an exami nation groups by gender was done. Tab le 8 shows the resu lts using the x2 difference test. Table 9 shows the results using Cheung and Ren svo ld' s meth od.

Multivariate Analysis of Variance (MANOVA)
The purpose of thi s analysis was to investigate the known groups validity of the behavioral dietary fat intak e measure (DBQ) using a full samp le (N=l366). Table 11 prese nts descriptive statistic s of 4 means of dependent ( assumption. However, a plot of residual by predicted va lue (see Figure 4) showed the residual varying between + 1 and -3 standard deviation. The variance of precontemp lation stage seems wider than other stage s. This result may happen becau se the unequal sample size among the 5 stages. The residual by quantile plots (see Figure   4) showed fairly linear plotting. The descriptive statistics showed skewness and kurtosis, suggesting the data was fairly normal. Also, this analysis used a large sample size (over 1000), thus this test is assumed to be robust for a violation of assumptions.
Therefore , it was decided not to transform the data. Tukey's pairwise multiple comparison test with using 0.05 alpha level , showed significant mean differences between precontemplation and maintenance, between contemplation and maintenance , and between preparation and maintenance.  Figure 5) as the first ANOVA test, thus decided to utilize original data in this ana lysis without any transformation. Figure 5 presents the residual plot s. Tukey ' s pairwise multiple comparison test showed significant mea n differences across most of the groups except between precontemplation and contemplation, between contemplation and preparation, and between action and maintenance. plotted, see Figure 6) as the first and the third ANOV A test, it was decided original data without any transformation would be used in this analysis. Figure 6 presents the residual plots. Tukey ' s pairwise multiple comparison test showed sign ificant mean differences between precontemplation and action, between precontemplati on and maintenance , between contemplation and action, between contemplation and maintenance, between preparation and action, and between preparation and maintenance.
In addition, Figure 3 presents the mean scores of 4 subcategorie s across the stages of change. This shows that the post -action stage has higher scores than in the preaction stage in all subcategorie s. Also, Table 12 presents means, standa rd deviation, F values, 11 2 and Tukey test result s of the 4 subcategorie s based on the 5 stages of change.

DISCUSSION
The multiple psychometric techniques using 1366 participants supported the 4 factor structure ofDBQ and its vali dity and reliability. The MANOVA suggested that there was a significant mean difference across the stages of change. The assumptions of ANOV A were tested using the descriptive statistics, the Levene's tests and the diagnostic tests, and these results showed that this data met the assumptions of ANOVA . This indicated that the linear comb ination of DBQ scores; Substitute, Moderate fat intake, Modify cooking, and Increase healthful foods, was correctly discriminated by the participant s' stage. The first ANOVA and Tukey's pair wise multiple compari son sugges ted that the mean score of Modify cooking was correctly discriminated by between precontemplation and maintenance , between contemplation and maintenance, and between preparation and maintenance.
The second ANOVA and Tukey's pair wise multiple comparison suggested that the mean score of Substitute was correctly discriminated by most pairs of stage except between precontemplation and contemplation, between contemplation and preparation , and between action and maintenance. The third ANOVA and Tukey' s pair wise multiple compari son suggested that the mean score of Moderate fat intake was correctly discriminated by most pair s of stage except between precontemplation and contemplation, between contemplation and preparation, and between action and maintenance. The fourth ANOVA and Tukey's pair wise multiple comparison suggested that the mean score of Increa se healthful foods was correctly discriminated by between precontemplation and action , between precontemplation and maintenance , between contemplation and action, between contemplation and maintenance , between preparation and action , and between preparation and maintenance. Most multiple comparisons did not show significant mean differences between precontemplation and contemplation, between contemplation and preparation and between action and maintenance. Since the actual behavior of fat consumption should be changed after the action stage, these findings were expected.

CHAPTERS CONCLUSION
This study tried to validate the dietary behavior que stio nnaire (DBQ), a mea sure originally developed by Rossi et al (2008) , from a psychometric perspective. The original purpose of this instrument was to be an outcome measure of TIM based intervention for dietary fat intake. A number of studies have supported the effectiveness of TIM tailored interv entions for this behavior as well as for intervening on multiple risky behaviors (Johnson et al., 2008;Prochask a et al., 2004Prochask a et al., , 2005

LIMITATION
Findings in this study were based on data which came from parents and thus, the generalizability of this study is limited. The original study was conducted on a sample from general population. The minor differences on factor structures and on other results might be due to the difference of sample characteristics. At the same time, however, validating this instrument on sample of parents was also greatly meaningful.
Parents have a high responsibility on managing food consuming habit in daily life.
Therefore, parents are one of the target populations for dietary intervention study.
Moreover, for other populations, their food consuming habits should also be strongly related with their parents' food consuming habits. As mentioned in the previous chapter, this study found multiple items having some potential to improve. If these items were improved, the new measure should be even more powerful. Finally, future studies using more broadly representative and heterogeneous participants and improved items are strongly desirable.        Note: This mode l has 3 correlation paths between the residuals of item 6 and 9 (r= .348) , item 17 and 22 (r= .501) and item 19 and 20 (r=.352).