- Email: [email protected]
Cross-cultural equivalence of feeding beliefs and practices: The psychometric properties of the child feeding questionnaire among Blacks and Hispanics
Preventive Medicine 41 (2005) 521 – 531 www.elsevier.com/locate/ypmed
Cross-cultural equivalence of feeding beliefs and practices: The psychometric properties of the child feeding questionnaire among Blacks and Hispanics Cheryl B. Anderson, Ph.D.T, Sheryl O. Hughes, Ph.D., Jennifer O. Fisher, Ph.D., Theresa A. Nicklas, Dr. PH., L.N. Department of Pediatrics, Baylor College of Medicine, The Children’s Nutrition Research Center, 1100 Bates, Houston, TX 77030, USA Available online 10 March 2005
Abstract Background. Psychometrically sound measures are considered a necessary condition for valid research. This study used structural equation modeling to examine the cross-cultural equivalence of a widely used measure of parental beliefs and practices regarding child feeding, the Child Feeding Questionnaire [Birch L.L., Fisher J.O., Grimm-Thomas K., Markey C.N., Sawyer R., Johnson S.L. Confirmatory factor analysis of the child feeding questionnaire: a measure of parental attitudes, beliefs and practices about child feeding and obesity proneness. Appetite 2001;36:201–10]. Methods. Low-income parents of 101 Black and 130 Hispanic pre-school children (126 girls, 105 boys) completed a reduced version of the CFQ. Results. Confirmatory factor analyses using LISREL 8.51 supported the hypothesized factor structure but revealed cross-cultural conceptual problems on the perceived child weight factor and problematic items on the restriction factor that were addressed in a modified model. Invariance analyses demonstrated invariance of factor structure, loadings, and covariances in the modified model across ethnic groups. MANCOVA, that controlled for parent BMI and marital status, revealed ethnic differences on the child feeding responsibility, child weight concern, and perceived weight of child factors that were moderated by parent education and child BMI. Conclusions. Results supported the use of a modified version of the CFQ among Blacks and Hispanics and revealed no ethnic differences on factor scores, except on interactions with parent education and overweight status of child. D 2005 Elsevier Inc. All rights reserved. Keywords: Invariance; Measurement; Confirmatory factor analysis; Child feeding; Eating behavior; Ethnic differences
Introduction The prevalence and health burdens of childhood obesity are greater among ethnic minorities, and the origins of these ethnic differences are not well understood. Recent increases in the observed prevalence of childhood overweight and obesity point to the role of the social and physical environment as causal agents in promoting positive energy balance [1]. The development of culturally appropriate survey instruments to identify factors that may be related to childhood overweight is essential to more fully understand T Corresponding author. Fax: +1 713 798 7098. E-mail address: [email protected] (C.B. Anderson). 0091-7435/$ - see front matter D 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.ypmed.2005.01.003
the ethnic differences in overweight that have been observed. For most young children, the family is the first and most fundamental socio-environmental context in which eating patterns are established. Parents influence children’s eating by selecting foods for the family diet, modeling eating behavior, and by providing direct instruction on when, where, what, and how much to eat [2,3]. Recent studies have begun to explore the relationship between parental beliefs and behaviors related to child feeding and child weight. The research, however, has focused almost exclusively on middle-class European–American populations [4]. Little is known about parental beliefs and feeding practices within specific ethnic minority cultures. It is also unclear
522
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
how social class indicators may interact with ethnicity to influence parental beliefs and practices that may impact children. Although ethnic differences on some relevant parental beliefs have been reported, such as perception of an overweight child as not overweight [5], the modifying effects of social class variables, such as parental education and income, are not often pursued. The current study examined the cross-cultural equivalence of the Child Feeding Questionnaire (CFQ [6]), one of few existing measures assessing child feeding and perhaps the most widely used. The purpose of the study was to extend Birch et al.’s initial work on authoritarian-type child feeding perceptions, concerns, and practices by evaluating the factor structure of the CFQ among samples of Hispanic and African-American parents and the invariance of the questionnaire across these ethnic groups. Research on eating behaviors has been limited by a lack of attention to the psychometric properties of measurement instruments, especially across ethnic groups. Factorial invariance assumes that a set of items is measuring the same constructs in all groups, and it is typically evaluated using confirmatory factor analysis within the framework of structural equation modeling [7]. Sets of parameters are tested in an orderly sequence of steps and in an increasingly restrictive manner to assess levels of invariance, primarily three levels that are most relevant to cross-group measurement: (1) whether the number of factors is the same across groups, (2) whether the factor loadings are equivalent across groups, and (3) whether the structural relationships between factors (i.e., factor variances/covariances) are equivalent across groups [7]. Testing for the equality of the measurement errors can be included, however, differences in measurement error are usually expected between groups and this inequality is not considered a serious issue. Without evidence of factorial invariance for at least the first two levels, the number of factors and size of factor loadings, there is no assurance that respondents in various groups similarly interpret the items of a particular measurement instrument and that differences in mean scores represent true group differences (i.e., comparing mean scores between ethnic groups may be completely invalid) [7,8]. The present study tested the factorial validity of the Child Feeding Questionnaire by testing a tentative measurement model based on the model developed by Birch et al. and testing subsequent modifications in two ethnic minority groups. The factorial invariance of the final measurement model was tested across groups and four hypotheses were considered: (1) that the number of underlying factors would be the same in both groups, (2) that the pattern and value of the factor loadings would be equivalent, (3) that the measurement errors or unique variances would be equivalent, and (4) that the structural relations among the factors (i.e., factor variances/covariances) would be equivalent. Next, where invariance across the subgroups was established, we compared means across groups to determine whether there were significant differences between Hispanics and Blacks in child feeding beliefs and practices and
examined whether ethnic differences were moderated by demographic and social class indicators.
Methods Participants Participants were part of a study to investigate eating practices of low income, Black and Hispanic preschool children attending Head Start programs in greater metropolitan Houston, Texas. At eight centers, 231 primary caregivers (130 Hispanic, 101 Black) completed questionnaires. The primary caregiver was defined as the person caring for the preschooler most of the time when the child was not in school. Of these caregivers, 98% were female (85% mothers, 13% grandmothers) and 2% were male. There were 231 children in the study (55% female, 45% male), ranging from 3 to 5 years in age (M = 4.15, SD = 0.71). Participant characteristics by ethnic group are shown in Table 1. Measures Child feeding The Child Feeding Questionnaire (CFQ [6]) is a 31-item self-report questionnaire that measures three aspects of parental control in child feeding and four aspects of parental perceptions and concerns about child obesity using a 5-point Likert scale. The parental control subscales include restriction (8 items), pressure to eat (4 items), and monitoring of Table 1 Characteristics of participants by ethnic group
Parent gender-female Child gender-female Age, mean in years (SD) Parent Child Education of parent High school diploma or less Some college or more Marital status, parent married Household income, less than $2500/month Parent BMI Underweight (BMI b 18.5) Normal (18.5 V BMI b 25) Overweight (25 V BMI b 30) Obese (BMI z 30) Child BMIa Underweight or at risk (b15th percentile) Normal (z15th to b85th percentile) Overweight or at risk (z85th percentile) a
African-American (n = 101)
Hispanic (n = 130)
97.0% 57.4%
99.2% 52.3%
32.6 (9.5) 4.0 (0.7)
30.3 (8.3) 4.3 (0.7)
50.0% 50.0% 52.5% 92.1%
78.4% 21.6% 76.9% 96.9%
0.99% 17.8% 28.7% 52.5%
0.77% 23.1% 40.8% 35.4%
3.0%
0.8%
64.4%
61.5%
32.7%
37.7%
From age- and gender-specific cut points of the CDC growth charts.
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
eating (3 items). The parental perceptions and concerns subscales include responsibility for feeding (3 items), perceived weight of parent (current and retrospective, 4 items), perceived weight of child (current and retrospective, 1 to 6 items), and concern about child weight (3 items). The perceived child weight items are linked to current age of child such that the number of items used is based on the age of the children studied. Birch et al. reported coefficient alphas of .88 (Responsibility), .71 (Parent weight), .83 (Child weight), .75 (Concern about child weight), .70 (Pressure to eat), .73 (Restriction), and .92 (Monitoring) for the CFQ subscales. Items for the CFQ are shown in the Appendix. Its items are presented consecutively to subjects by factor (i.e., not scrambled). In the current study, the perceived parent weight factor (4 items) was not measured due to a clerical error discovered after data collection was completed, and three of six perceived child weight items were not used since they were not applicable due to age of child. In the Spanish translation, anchors for the child weight items were changed from bmarkedly underweight/ overweightQ to bvery underweight/overweightQ for both ethnic groups. Anthropometrics Height and weight were measured by staff in duplicate, averaged, and used to compute child and parent BMI (kg/m2). For adults, boverweightQ was defined as BMI of 25.0 to 29.9 and bobesityQ as BMI z 30.0, according to the World Health Organization criteria [9]. For children, BMI z scores and percentiles were calculated using the age- and gender-specific
523
cut points from the revised growth charts produced by the Centers for Disease Control and Prevention [10]. bAt risk for overweightQ was defined as at or above the 85th percentile but less than the 95th percentile, and boverweightQ as BMI z 95th percentile. Procedure The Institutional Review Board of Baylor College of Medicine approved the study and all participating parents provided informed consent. Participants were recruited at Houston area Head Start centers where data collection took place. Parents completed questionnaires in groups in a room separate from children and received $25 (gift certificate or cash) for participation. Statistical analyses The primary goal of the analyses was to compare the functional organization of the Child Feeding Questionnaire in two minority groups. This was accomplished in four phases. First, we evaluated the factor structure identified by Birch et al., as shown in Fig. 1, in the combined minority groups and separately by race. Second, we evaluated the model again, but without three item parcels that were utilized by Birch et al. on the Restriction factor (that is, we conducted an all item-level analysis). The rationale for forming the parcels (ratings from two or more items are averaged together to function as a single item; see Appendix A for content of the three parcels) was not reported in the
Fig. 1. Factor structure of the Child Feeding Questionnaire ([6]).
524
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
original investigation, and the statistical properties (e.g., factor loadings) of the individual items used in the parcels were also not reported. Third, we modified the measurement model with all item-level data to identify a model that provided the best account of the data in the minority groups. Fourth, we tested the invariance of the final modified model across the Hispanic and Black groups. Confirmatory factor analyses The models proposed were analyzed using maximum likelihood estimation in LISREL 8.51 [11]. PRELIS, the preprocessor of LISREL, was used to generate the covariance matrices using listwise deletion for missing data. The tentative measurement model hypothesized that (a) responses could be explained by six factors, (b) each item would have a non-zero loading on the factor it was intended to measure and a zero loading on all other factors, (c) all factor covariances were freely estimated, and (d) measurement error would be uncorrelated, except for the one error covariance specified by Birch et al. between PCW1 and PCW2. Subsequent models that were run did not allow correlated errors. Model fit Multiple indices were used to assess model fit. The Chisquare statistic provides an overall test of model fit and evaluates the absolute fit between the hypothesized model and the data. A nonsignificant Chi-square indicates that the model fits the data. However, since Chi-square assumes multivariate normality and is sensitive to sample size, significant Chi-squares that reject the model can occur even when the model fit is relatively good [12,13]. Therefore, additional indices of fit have been suggested. Additional indices that were used to evaluate goodness of fit included the Non-Normed Fit Index (NNFI, also called the Tucker– Lewis Index), Bentler’s Comparative Fit Index (CFI), the root mean square error of approximation (RMSEA), and the standardized root mean square residual (SRMR). Traditionally, NNFI and CFI values greater than or equal to .90 have been accepted as indicators of good fit. Recent work, however, by Hu and Bentler [14] has indicated that values near .90 are only minimally acceptable and that values of .95 or higher are necessary to indicate a good fit of the model. They also recommend a RMSEA cutoff value of .06 in conjunction with a SRMR value of .08 or less. The RMSEA has been recognized as one of the most informative criteria available, highly sensitive to models with misspecified factor loadings, while the SRMR is most sensitive to models with misspecified factor covariances or latent structures [14]. Factorial invariance The invariance analyses were performed in a hierarchical ordering of constraints [7,15]. Four, nested, multigroup analyses were performed using LISREL 8.51: (Model 1) the factor pattern was constrained to be equal across groups, (Model 2) equal factor loadings were added to the constraint of equal factor pattern, (Model 3) the error variances were
next added to the loading constraints across groups, and (Model 4) the factor variances and covariances were added to the constraints across groups. If a more restrictive model fits the data as well as the less restrictive model, then there is support for the invariance of the constrained parameters. The difference between the two models is tested by examining significant differences in the Chi-square values for the change in degrees of freedom. No substantial change of model fit (i.e., p z .05) would be indicative of model invariance. As previously stated, Model 2 or equal factor pattern and loadings is considered the minimal evidence of factorial invariance, and Models 3 and/or 4 demonstrate increased evidence of invariance across groups [7,15]. Mean group differences Multivariate analysis of covariance (MANCOVA) was used to determine if the subscales from the modified CFQ differed by ethnic group, social class, and child demographics. The CFQ factor scores (mean of items per factor) were used as dependent variables, and parent ethnicity (Hispanic, Black), parent education (high school diploma or less, some college or more), child gender (male, female), and child BMI (underweight, normal, at risk or overweight) were used as independent variables. Parent income, as an indicator of social class, was not included as a predictor, since all families were Head Start participants and thus, matched on low income. Parental weight (underweight, normal, overweight, obese) and marital status (married, not married) have been associated with child weight and were treated as covariates [16,17]. Significant multivariate effects (for main effects and all interactions) were followed by univariate analyses (ANOVA). A Bonferroni correction (a = .01) was used in the ANOVAs to control for experimentwise alpha. Least squares means (LSM) were computed in the post hoc comparisons.
Results Differences in participant characteristics There were significant ethnic differences in parent education, marital status, and parent BMI. African-Americans had higher education levels, v 2(2, n = 223) = 20.07, P b .0001, and were less likely to be married, v 2(1, n = 231) = 15.19, P b .0001. More African-American parents were obese, while more Hispanic parents were overweight, v 2(2, n = 229) = 6.96, P = .03. There were no significant ethnic differences in child weight status, child age, or parental age. Testing the fit of the Birch et al. model Ten participants with missing data were deleted from the combined sample of 231 (4.3% missing data). The replication of the six-factor Birch et al. model that included the item parcels and one error covariance resulted
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
in a mixed fit in the combined ethnic group that was minimally to moderately acceptable, based on the indices of fit, v 2(154, N = 221) = 223.12, p = 0.002, RMSEA = .045 (90% CI = .031–.057), NNFI = .91, CFI = .93, SRMR = .063. The data did not support the error covariance on child weight, and this covariance was dropped (D df = 1, Dr2 = 6.32, p b .02). By ethnic group, the model fit was better in the Hispanic than the African-American sample. Addressing the problem areas in the original model The fit indices in the combined and subgroup analyses did not fully reflect two important weaknesses in the model: (1) several non-significant or low item loadings and (2) the use of item parcels without a clear rationale or evaluation of the parcels’ components. First, two of the three perceived child weight item loadings were non-significant or below .36 (PCW2 and PCW3), as were two loadings on the restriction factor (RST2 and item parcel RST3). Although there is no rule of thumb on factor loading size in structural equation modeling, it is generally accepted that all factor loadings should be significant and loadings of .60 or greater are usually recommended. The low loadings of the perceived child weight items and a lack of covariance among them made it clear that these items were problematic. Ratings of 3 (that is, child is of baverageQ weight) on the 1 to 5 response scale were given by 77% of the combined sample for their child during its first year of life (PCW1), 81% of the sample for their child at 1–2 years old (PCW2), and 80% of the sample for their child at 3–5 years old (PCW3). Furthermore, further analysis of the three PCW items showed that their correlations with actual current child weight increased substantially with increases in child age, such that the largest correlations were between parental perceptions of the child’s current weight and the child’s actual current weight (birth to 1 year: r = .12, age 1 to 2: r = .29, current age: r = .46). These results suggest that averaging parental perceptions of child’s weight over time and relating this average to current weight (or other variables) may not make conceptual sense. Three constructs (one for each time interval), not just one, appear to exist. If the subscale is to be related to current child weight, as is routinely being done in the field of nutrition, then it may be conceptually best to use only items that refer to current weight. Since only one item reflected current weight and this item was problematic (three or more items are desired to load on a latent factor), the modified model was fit without the Perceived Child Weight factor. However, the single item for current weight (PCW3) was used in the MANCOVA analysis to represent the perceived current weight of child factor. Secondly, parcels were used in conjunction with itemlevel data on the restriction factor in the Birch et al. model. In the published article, neither the fit of the model nor the item loadings were shown without the item parcels, and there was no rationale given for their use. Thus, the reliabilities of the individual items that comprised the parcels on the restriction factor were not known and required evaluation.
525
Modifications to the CFQ model Regarding model modification, we felt the problems found could be best addressed by fitting the Birch model completely at the item level, without the child weight factor (since further scale development work is needed on this factor), re-examining the fit indices and factor loadings, and modifying the model as necessary. Before dropping the child weight factor, the model was fit using all item-level data. The fit of the Birch model was considerably worse when the item parcels were not used and the data were fit only at the item level, v 2(237, N = 221) = 452.32, p = 0.0000, RMSEA = .064 (90% CI = .055–.073), NNFI = .81, CFI = .83, SRMR = .075. Four of the eight restriction items (the factor that utilized the item parcels) loaded at or below .35. The item loadings on the child weight factor were similarly unacceptable. A modified model was fit to the data in which the perceived child weight factor was dropped. The resulting five-factor model did not provide an improved fit, and further modified models were fit that also sequentially dropped five low-loading items on the Restriction factor. Two items involved reward behaviors, one involved keeping bsome foodsQ out of the child’s reach, and two involved the child eating too many bfavorite foods.Q In dropping these items, it was reasoned that the reward items did not theoretically reflect the construct of restriction and that the other three items could be interpreted by subjects in a variety of positive or negative ways. The three items retained were more clearly indicators of a construct that represented restriction of foods that could contribute to overweight (i.e., sweets, high fat foods, junk foods). The loading of RST4a, which dealt with restriction of bjunk foods,Q was not desirable at .37, likely reflecting imprecision in wording. However, it was retained in the model because the loading was significant, the LISREL modification indices indicated that this item had no appreciable loadings on any other factors, and it was theoretically relevant as an obesity-related food. The final model (Fig. 2) had an adequate fit to the data, v 2(94, N = 221) = 128.97, p = 0.0097, RMSEA = .041 (90% CI = .021–.058), NNFI = .95, CFI = .96, SRMR = .055. As in the previous models that were run, the fit of the final modified model was better in the Hispanic than in the African-American sample (Table 2). Testing model equivalence across ethnic groups As shown in Table 3, the test of equal factor patterns (Model 1) for the modified model was not rejected ( p = .1800), indicating that a model of five factors was equivalent across race, and this was supported by the fit indices. The test of equal factor loadings (Model 2) was not significant ( p = .1200) and the Chi-square difference between Models 1 and 2 was not significant ( p N .10; see Table 3), indicating equivalence of the factor loadings
526
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
Fig. 2. Final modified measurement model fit to Hispanic and African-American subgroups and tested for invariance across groups.
shown in Table 4. All parameter estimates were significant and positive. Final loadings ranged from .37 to .90 (M = .70), with only 10 of the 16 loadings greater than .60. The low correlations observed between factors (range of .00– .46) supported that the factors were distinct.
across race. The errors or uniquenesses were not equivalent across race, as indicated by the significant v 2 for Model 3 (p = .0000), the significant Chi-square difference between Models 2 and 3 (p b .001), and the poorer fit indices associated with this model. This was not unexpected, since rarely do two samples share the same amount of measurement error. The test of equal factor variances and covariances (Model 4) was not rejected (p = .06) and the Chi-square difference test between Models 2 and 4 was not significant (p N .05), indicating that the correlations between factors were similar across race. The factor loadings from the modified model are shown in the Appendix A. The LISREL factor intercorrelations, means, and Pearson correlations with child weight are
Evaluation of ethnic differences and moderating effects of demographics and social class The multivariate analysis of covariance (MANCOVA) assessed mean differences in the feeding subscales by ethnic group and interactions with three possible moderators. Two levels of child age- and gender-specific BMI (bor z85th percentile), child gender (male, female), parent ethnicity
Table 2 Results of factorial validity analyses Model Birch et al. model replication (6-factor model, 3 parcels on Restriction, 24 items, no error covariances) Combined sample Hispanic Black Birch et al. model, without parcels (6-factor model, all item-level data, 24 items, no error covariances) Combined sample Modified model (5-factor model, all item-level data, 16 items, no error covariances) Combined sample Hispanic Black
df
v2
p
RMSEA (90% CI)
CFI
NNFI
SRMR
155 155 155
216.80 162.27 196.53
0.0008 0.330 0.013
0.043 (0.028–0.055) 0.020 (0.000–0.047) 0.053 (0.025–0.074)
0.93 0.97 0.90
0.92 0.96 0.88
0.063 0.066 0.091
237
452.32
0.0000
0.064 (0.055–0.073)
0.83
0.81
0.075
94 94 94
128.97 70.82 166.93
0.0097 0.9600 0.0550
0.041 (0.021–0.058) 0.000 (0.000–0.000) 0.050 (0.000–0.077)
0.96 1.00 0.93
0.95 1.04 0.92
0.055 0.056 0.087
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
527
Table 3 Results of ethnic invariance analyses for the modified model Invariance model
df
v2
p
RMSEA (90% CI)
CFI
NNFI
SRMR
Ethnic invariance model 1—test of same factor pattern Ethnic invariance model 2—test of same pattern and loadings Ethnic invariance model 3—test of same pattern, loadings, and errors Ethnic invariance model 4—test of same pattern, loadings, and factor variances and covariances
188
206.00
0.1800
0.030 (0.000–0.053)
0.96
0.95
0.056
199
223.00
0.1200
0.033 (0.000–0.055)
0.95
0.94
0.060
215
337.37
0.0000
0.072 (0.057–0.082)
0.83
0.81
0.076
214
247.16
0.0600
0.038 (0.000–0.057)
0.95
0.94
0.0081
Ethnic invariance comparisons
df diff
v 2 diff
p value
Interpretation
Ethnic invariance model 1 vs. 2 Ethnic invariance model 2 vs. 3 Ethnic invariance model 2 vs. 4
11 16 15
17.00 114.37 24.16
N0.10 b0.001 N0.05
Factor pattern and loadings not different across race Measurement errors different across race Factor pattern, loadings, variances and covariances not different across race
(Black, Hispanic), and parent education (high school or less, some college, or more) were used, with parental BMI and marital status as covariates. Results of the MANCOVA (child BMI child gender parent ethnicity parent education: all main effects and interactions), indicated a significant main effect for child BMI, Wilk’s K = .83, F(6,192) = 6.32, p b .0001. Two two-way interactions were significant: one for ethnicity child BMI (Wilk’s K = .93, F(6,192) = 2.43, p = .0273) and one for ethnicity parent education (Wilk’s K = .90, F(6,192) = 3.40, p = 0.0032). There was one significant three-way interaction (ethnicity
parent education child BMI), Wilk’s K = .89, F(6,192) = 3.97, p = .0009. Subsequent univariate analyses for the child BMI main effect indicated that perceived current child weight was the only variable significant at the .01 level (p b .0001). Post hoc comparisons indicated that parents of children at or above the 85th percentile for weight perceived their child as significantly more overweight (LSM = 3.25) than parents of children below the 85th percentile (LSM = 2.82). The mean for overweight children, however, did not indicate an accurate parental perception of the child’s
Table 4 Descriptive statistics by factor, interfactor correlationsa, correlations with child BMIb Factor 1. 2. 3. 4. 5. 6.
Perceived responsibility Concern about child weight Restriction Pressure to eat Monitoring Perceived child weightc
Modified model M (SD)
1
Hispanic
African-Am
4.79 2.30 4.19 3.70 4.05 3.01
4.61 2.06 4.31 3.57 4.27 2.92
(0.56) (1.22) (1.06) (0.90) (1.10) (0.33)
(0.67) (1.15) (0.84) (0.86) (0.98) (0.58)
– 0.10 0.12 0.28 0.04 –
2
3 0.02 – 0.00 0.19 0.16 –
4 0.25 0.17 – 0.59 0.42 –
5 0.40 0.12 0.29 – 0.22 –
Child BMI percentileb Hispanic 1. 2. 3. 4. 5. 6. a
Perceived responsibility Concern about child weight Restriction Pressure to eat Monitoring Perceived child weightc
0.06 0.14 0.01 0.12 0.02 0.42TTT
Black 0.05 0.27TT 0.02 0.17 0.01 0.38TTT
LISREL interfactor correlations or phi values. Hispanic correlations are below the diagonal, African-American correlations are above the diagonal. Pearson correlation with age- and gender-specific percentiles per the revised growth charts from the Centers for Disease Control and Prevention. c Perceived child weight, one item, which was not included in the modified LISREL model. TTp b 0.01. TTTp b 0.0001. b
0.36 0.18 0.25 0.25 – –
528
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
weight (that is, a scale rating of 3 was baverageQ weight). Further analysis revealed that 78% of parents of overweight or at risk children regarded the child as average or underweight. In the parent ethnicity child BMI interaction, only perceived current child weight was significant (p = .0008). Among children at or above the 85th percentile, Black parents perceived their child as significantly less overweight (LSM = 3.07) than did Hispanic parents (LSM = 3.44), although as stated earlier, the means for both ethnic groups demonstrate an underestimation of child weight by parents. The parent ethnicity parent education interaction was significant for concern (p = .0105) and for responsibility (p = .0048). Lower educated Black parents expressed significantly less concern about child weight (LSM = 1.79) than did higher educated Black parents (LSM = 2.67), but this difference was not present among Hispanics, where low and higher educated parents held similar levels of concern (LSM = 2.21low ed, 2.50higher ed). The mean values indicate generally low levels of concern about weight among both ethnic groups. Regarding responsibility, Hispanic parents with low education rated themselves as more responsible for the feeding of the child (LSM = 4.86) than did higher educated Hispanic parents (LSM = 4.30), but there were no significant differences in perceived responsibility for feeding among low and higher educated Blacks (LSM = 4.59low ed, 4.71higher ed). In the ethnicity parent education child BMI threeway interaction, only perceived current child weight was significant at the .01 level, p = .0027. Low educated Hispanic parents of children at or above the 85th percentile perceived their child as less overweight (LSM = 3.06) than did higher educated Hispanic parents (LSM = 3.83), but this education effect within race was not present among Blacks (LSM = 3.11low ed, 3.02high ed). As the mean values show, higher educated Black parents of overweight children perceived their child as significantly less overweight than did the higher educated Hispanics.
Discussion This study evaluated the suitability of the Child Feeding Questionnaire (CFQ) for use with low-income Black and Hispanic parents, analyzing six of its seven hypothesized dimensions. Overall, the minority data supported the factor structure found in the White and Hispanic samples used in the CFQ’s original development. However, substantial modifications were required in the item pool comprising the final minority model as a result of primarily conceptual and methodological issues that appeared to be unrelated to ethnicity. The final model of five factors and 16 items demonstrated invariance of factor structure, factor loadings, and factor variances– covariances across the two minority subgroups, indicating that the subscales of Responsibility, Concern about
Weight, Restriction, Pressure to Eat, and Monitoring assess similar underlying constructs across the Hispanic and African-American groups. Subsequent mean analyses revealed no mean differences between Hispanics and African-Americans in parental attitudes and practices, except those modified by parent education and child BMI. Our results showed that the modified CFQ functioned adequately in both ethnic minority groups but two adjustments were required. First, five of the eight restriction items were removed due to non-significant factor loadings. The two reward items may not conceptually reflect restriction. Reward and restriction likely represent two constructs instead of one and may differentially predict behavior, similar to findings in the addiction literature on the impact of positive and negative expectancies [18]. Three other restriction items had imprecise wordings that may have fostered widely different interpretations. Although the bjunk foodQ item was significant and retained in the model, its less than optimal loading reflects imprecision in wording and a need for item revision (i.e., what are bjunk foodsQ to parents? Are they different in minority populations?). The modified restriction factor resulted in a construct representing restriction of negative, obesity-related foods rather than the restriction of food per se. The modified restriction factor also was fit without item parcels. Item groupings or parcels are routinely used to reduce the number of items in a large item pool to allow estimation of a desired model. Items can be combined on a conceptual basis or by some selected pattern. Their use requires caution and proper methodology, however, because they tend to make the model fit better and can hide model misfits or poorly performing items [19,20], as was found in the current data. The use of parcels is acceptable with a stated rationale, after individual subfactors have been fit at the item level and there is good evidence that the individual items that comprise the parcels are sufficiently strong or reliable. Second, a lack of conceptual clarity was found in the perceived weight construct. Child weight over time may or may not be consistent, actually or in the parent’s mind. Therefore, parental evaluations of child weight at previous time points may not be at all related to the current weight of the child. For example, parents may remember their child as a chubby baby, but perceive the child’s current weight as normal. Furthermore, parents may apply different standards for age related weight (for example, it may be OK to be a chubby baby). These issues make the interpretation of perceived child weight as an average score derived from multiple time points problematic. The same conceptual problems likely exist on the perceived parent weight factor, whose four items were not included in the current data. Further research is needed on parental beliefs about child (current and retrospective) over time and their relation to child age and weight. Pending such research, we recommend focusing only on the perceived current weight of child and further item development to reflect this construct.
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
The two problems of focus appear to reflect cross-cultural weaknesses in the scale (that is, they also existed in the White samples) that are conceptual and methodological (for example, items are presented to subjects by factor with many consecutive items highly similar in wording, which can promote method effects), and not specific to any one ethnic group. The modifications made to the item pool in our analyses (i.e., dropping certain items) increased the scale’s psychometric integrity for use across ethnic groups, however continued scale development work appears warranted. Our results corroborate findings from other studies of systematic bias in the perception of child weight among minorities [5,21]. Both minorities underestimated the weight of overweight children, but Blacks more so than Hispanics. Importantly, parental education moderated this effect in our data, over and above the influence of family income, marital status, and parent BMI. Similar to findings by Baughcum et al. [21], an education effect was seen among Hispanics on perceived weight, where overweight children were classified as average or underweight more often among parents with a high school education or less, in contrast to Hispanics with more education. This social class effect was not seen among Blacks, however, where even higher educated Black parents perceived their overweight child as more normal than did higher educated Hispanics. Parental education also modified the ethnic effect on responsibility for feeding and concern about child weight. In contrast to Hispanics, both higher and lower educated Blacks perceived high responsibility for the feeding of their child. While this could be related to a lower prevalence of married parents among the African-Americans, marital status was controlled for in our analyses. Importantly, low levels of concern about child weight were seen in both minority groups. While higher and lower educated Hispanics and Blacks all held low ratings of concern about child weight, markedly low concern ratings were found among lower educated Blacks. In comparing the current data to that of the original Birch study, an interesting contrast was observed. Similar to Birch, moderate positive correlations were seen in our data between restriction and monitoring of foods. However, there was not a moderate positive correlation between restriction and concern about weight, as would be expected and as Birch et al. found in their predominately White samples. High mean ratings on restriction, as well as monitoring, were countered with low ratings on concern about weight in the African-American subgroup but not the Hispanic. This implies that motives for restricting foods may differ between cultures. For example, restriction may be regarded as important and practiced because eating foods like sweets could interfere with the child’s consumption of a main meal, not because such foods lead to weight gain. Except for perceived child weight, there were no significant relationships found in the MANCOVA analysis between measured child BMI and the six constructs that
529
were measured from the CFQ. As in Birch’s previous work in predominately White samples, we found a positive Pearson correlation between concern and child BMI in both minority subgroups, but the relationship was significant only among Blacks. The lack of relationship shown in the MANCOVA, however, does not make the other parental attitudes and practices irrelevant. Child eating behaviors, which have been related to parental attitudes and practices [4], represent important modifiable influences on child weight. Future research, including continued work on scale development and eating behaviors, will enable better conclusions about the importance of these constructs to child weight. The results of this study should be interpreted in light of several limitations. The findings are limited to Hispanics and African-Americans, and no White sample was available for further ethnic comparisons. Furthermore, the questionnaire did not include the four items for perceived parent weight such that a reduced version of the CFQ was tested. We acknowledge the small sample size of the groups, which may have facilitated the invariance in parameters found across groups. In addition, the primary mediator of the parental attitudes/practices and child weight relationship, child eating, was not measured in this study. However, using rigorous psychometric analyses to evaluate ethnic differences in two populations at high risk for obesity, our results support the use of a modified version of the Child Feeding Questionnaire (CFQ), a widely used measure in nutrition research, among two minority subgroups. The findings also support further item development and psychometric work on the instrument to address conceptual and methodological weaknesses that appear to be crosscultural.
Acknowledgments The authors wish to thank Sandra Lopez and Janet Bonner for their help in data collection. We also extend a special thanks to the children and parents who participated in the study. This research was supported by funds from the United States Department of Agriculture grant, bEnvironmental Influences on Children’s Food Consumption,Q Grant No. 2001-35200-10659 (PI: Theresa A. Nicklas). Partial support was also received from Dairy Management Inc. This work is a publication of the United States Department of Agriculture (USDA/ARS) Children’s Nutrition Research Center, Department of Pediatrics, Baylor College of Medicine, Houston, Texas, and had been funded in part with federal funds from the USDA/ARS under Cooperative Agreement No. 58-6250-6001. The contents of this publication do not necessarily reflect the views or policies of the USDA, nor does mention of trade names, commercial products, or organizations imply endorsement from the U.S. government.
530
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
Appendix A. Modified model factor structure from original CFQ questionnaire, completely standardized loadings, standard errors, t values
Latent variable
Variable name
Item description and presentation order
Modified model Standardized loading
Perceived responsibility
Perceived parent weight
Perceived child weight
Concern about child weight
Restrictionc
PR1 PR2 PR3 PPW1 PPW2 PPW3 PPW4 PCW1 PCW2 PCW3 PCW4 PCW5 PCW6 CN1 CN2 CN3 RST1a RST1b RST1c RST2 RST3a RST3b RST4a
Pressure to eat
Monitoring
RST4b PE1 PE2 PE3 PE4 MN1 MN2 MN3
1. for feeding 2. for how much served 3. for healthy foods 4. childhood (5–10 years old) 5. adolescence 6. your 20s 7. at present 8. during first year of life 9. between age 1 and 2 10. between age 3 and 5 11. kindergarten through 2nd grade 12. 3rd through 5th grade 13. 6th through 8th grade 14. about eating too much when parent not w/child 15. about child having to go on diet 16. about child becoming overweight 17. make sure does not eat too many sweets 18. make sure does not eat too many high fat foods 19. make sure does not eat too many favorite foods 20. keep some foods out of reach 21. offer sweets as reward for good behavior 22. offer favorite foods for good behavior 23. would eat too many junk foods without guidance 24. would eat too much of favorite foods 25. should eat all food on plate 26. make sure child eats enough 27. try to make child eat when not hungry 28. child would eat much less without guidance 29. keep track of sweets 30. keep track of snack food 31. keep track of high fat foods
SEa
t value
0.59
b
___
___b
0.84 0.68 Omitted Omitted Omitted Omitted Dropped Dropped Dropped but used in MANCOVA Omitted Omitted Omitted 0.60
0.22 0.13 – – – – – – – – – – ___b
7.00 7.23 – – – – – – – – – ___b
0.88 0.72 0.87 0.76
0.20 0.18 ___b 0.13
7.49 7.99 ___b 8.03
Dropped
–
–
Dropped Dropped Dropped 0.37
– – – 0.12
– – – 4.88
Dropped 0.42 0.49 0.55 0.56
– ___b 0.28 0.36 0.38
– ___b 3.93 4.09 4.10
0.90 0.92 0.67
___b 0.07 0.08
___b 15.77 11.35
a
Standard error. Indicates a parameter fixed at 1.0 in the LISREL solution. c Item parcels used within restriction factor in original Birch et al. model: RST1a, b, and c made up Parcel 1; RST3a and b made up Parcel 2; RST4a and b made up Parcel 3. b
References [1] Hill JO, Peters JC. Environmental contributions to the obesity epidemic. Science 1998;280(1371):1374. [2] Johnson SL, Birch LL. Parent’s and children’s adoposity and eating style. Pediatrics 1994;94:653 – 61. [3] Costanzo PR, Woody EZ. Domain-specific parenting styles and their impact on the child’s development of particular deviance: the example of obesity proneness. J Soc Clin Psychol 1985;4:425 – 45. [4] Birch LL, Davison KK. Family environmental factors influencing the developing behavioral controls of food intake and childhood overweight. Pediatr Clin North Am 2001;48(4): 893 – 907. [5] Myers S, Vargas Z. Parental perceptions of the preschool obese child. Pediatr Nursing 2000;26(1):23 – 30.
[6] Birch LL, Fisher JO, Grimm-Thomas K, Markey CN, Sawyer R, Johnson SL. Confirmatory factor analysis of the child feeding questionnaire: a measure of parental attitudes, beliefs and practices about child feeding and obesity proneness. Appetite 2001;36:201 – 10. [7] Byrne BM. Structural equation modeling with LISREL, PRELIS, and SIMPLIS: Basic Concepts, Applications, and Programming. Mahwah, NJ7 Erlbaum; 1998. [8] Vandenberg RJ, Lance CE. A review and synthesis of the measurement invariance literature: suggestions, practices, and recommendations for organizational research. Organ Res Methods 2000;3(1):4 – 70. [9] World Health Organization Expert Committee. Physical status: the use and interpretation of anthropometry. WHO Technical Report Series: 854. Geneva7 World Health Organization; 1995.
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531 [10] Centers for Disease Control and Prevention. CD growth charts: United States. Hyattsville, MD7 National Center for Health Statistics, U.S. Dept of Health and Human Services; 2000. [11] LISREL 8.51. Chicago, IL: Scientific Software International, Inc.; 2001. [12] Hu L, Bentler PM. Evaluating model fit. In: Hoyle RH, editor. Structural equation modeling: Concepts, Issues, and Applications. Thousand Oaks, CA7 Sage; 1995. p. 76 – 99. [13] Marsh HW, Balla JR, McDonald RP. Goodness-of-fit indexes in confirmatory factor analysis: the effect of sample size. Psychol Bull 1988;103:391 – 410. [14] Hu L, Bentler PM. Cutoff criteria for fit indexes in covariance structure analysis: conventional criteria versus new alternatives. Struct Equ Model 1999;6(1):1 – 55. [15] Joreskog KG, Sorbom D. LISREL 8: User’s Reference Guide. Chicago, IL7 Scientific Software; 1996.
531
[16] Whitaker RC, Wright JA, Pepe MS, Seidel KD, Dietz WH. Predicting obesity in young adulthood from childhood and parental obesity. N Engl J Med 1997;337:869 – 73. [17] Strauss RS, Knight J. Influence of the home environment on the development of obesity in children. Pediatrics 1999;103(6):1 – 8. [18] Anderson CB, Pollak KI, Wetter D. Relations between self-generated positive and negative expected smoking outcomes and smoking behavior: an exploratory study among adolescents. Psychol Addict Behav 2002;16(3):196 – 204. [19] Bandalos DL. The effects of item parceling on goodness-of-fit and parameter estimate bias in structural equation modeling. Struct Equ Model 2002;9(1):78 – 102. [20] Plummer BA. To parcel or not to parcel: the effects of item parceling in confirmatory factor analysis. Diss Abstr Int 2001;61(10-B):5617. [21] Baughcum AE, Chamberlin LA, Deeks CM, Powers SW, Whitaker RC. Maternal perceptions of overweight preschool children. Pediatrics 2000;106(6):1380 – 6.
Cross-cultural equivalence of feeding beliefs and practices: The psychometric properties of the child feeding questionnaire among Blacks and Hispanics Cheryl B. Anderson, Ph.D.T, Sheryl O. Hughes, Ph.D., Jennifer O. Fisher, Ph.D., Theresa A. Nicklas, Dr. PH., L.N. Department of Pediatrics, Baylor College of Medicine, The Children’s Nutrition Research Center, 1100 Bates, Houston, TX 77030, USA Available online 10 March 2005
Abstract Background. Psychometrically sound measures are considered a necessary condition for valid research. This study used structural equation modeling to examine the cross-cultural equivalence of a widely used measure of parental beliefs and practices regarding child feeding, the Child Feeding Questionnaire [Birch L.L., Fisher J.O., Grimm-Thomas K., Markey C.N., Sawyer R., Johnson S.L. Confirmatory factor analysis of the child feeding questionnaire: a measure of parental attitudes, beliefs and practices about child feeding and obesity proneness. Appetite 2001;36:201–10]. Methods. Low-income parents of 101 Black and 130 Hispanic pre-school children (126 girls, 105 boys) completed a reduced version of the CFQ. Results. Confirmatory factor analyses using LISREL 8.51 supported the hypothesized factor structure but revealed cross-cultural conceptual problems on the perceived child weight factor and problematic items on the restriction factor that were addressed in a modified model. Invariance analyses demonstrated invariance of factor structure, loadings, and covariances in the modified model across ethnic groups. MANCOVA, that controlled for parent BMI and marital status, revealed ethnic differences on the child feeding responsibility, child weight concern, and perceived weight of child factors that were moderated by parent education and child BMI. Conclusions. Results supported the use of a modified version of the CFQ among Blacks and Hispanics and revealed no ethnic differences on factor scores, except on interactions with parent education and overweight status of child. D 2005 Elsevier Inc. All rights reserved. Keywords: Invariance; Measurement; Confirmatory factor analysis; Child feeding; Eating behavior; Ethnic differences
Introduction The prevalence and health burdens of childhood obesity are greater among ethnic minorities, and the origins of these ethnic differences are not well understood. Recent increases in the observed prevalence of childhood overweight and obesity point to the role of the social and physical environment as causal agents in promoting positive energy balance [1]. The development of culturally appropriate survey instruments to identify factors that may be related to childhood overweight is essential to more fully understand T Corresponding author. Fax: +1 713 798 7098. E-mail address: [email protected] (C.B. Anderson). 0091-7435/$ - see front matter D 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.ypmed.2005.01.003
the ethnic differences in overweight that have been observed. For most young children, the family is the first and most fundamental socio-environmental context in which eating patterns are established. Parents influence children’s eating by selecting foods for the family diet, modeling eating behavior, and by providing direct instruction on when, where, what, and how much to eat [2,3]. Recent studies have begun to explore the relationship between parental beliefs and behaviors related to child feeding and child weight. The research, however, has focused almost exclusively on middle-class European–American populations [4]. Little is known about parental beliefs and feeding practices within specific ethnic minority cultures. It is also unclear
522
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
how social class indicators may interact with ethnicity to influence parental beliefs and practices that may impact children. Although ethnic differences on some relevant parental beliefs have been reported, such as perception of an overweight child as not overweight [5], the modifying effects of social class variables, such as parental education and income, are not often pursued. The current study examined the cross-cultural equivalence of the Child Feeding Questionnaire (CFQ [6]), one of few existing measures assessing child feeding and perhaps the most widely used. The purpose of the study was to extend Birch et al.’s initial work on authoritarian-type child feeding perceptions, concerns, and practices by evaluating the factor structure of the CFQ among samples of Hispanic and African-American parents and the invariance of the questionnaire across these ethnic groups. Research on eating behaviors has been limited by a lack of attention to the psychometric properties of measurement instruments, especially across ethnic groups. Factorial invariance assumes that a set of items is measuring the same constructs in all groups, and it is typically evaluated using confirmatory factor analysis within the framework of structural equation modeling [7]. Sets of parameters are tested in an orderly sequence of steps and in an increasingly restrictive manner to assess levels of invariance, primarily three levels that are most relevant to cross-group measurement: (1) whether the number of factors is the same across groups, (2) whether the factor loadings are equivalent across groups, and (3) whether the structural relationships between factors (i.e., factor variances/covariances) are equivalent across groups [7]. Testing for the equality of the measurement errors can be included, however, differences in measurement error are usually expected between groups and this inequality is not considered a serious issue. Without evidence of factorial invariance for at least the first two levels, the number of factors and size of factor loadings, there is no assurance that respondents in various groups similarly interpret the items of a particular measurement instrument and that differences in mean scores represent true group differences (i.e., comparing mean scores between ethnic groups may be completely invalid) [7,8]. The present study tested the factorial validity of the Child Feeding Questionnaire by testing a tentative measurement model based on the model developed by Birch et al. and testing subsequent modifications in two ethnic minority groups. The factorial invariance of the final measurement model was tested across groups and four hypotheses were considered: (1) that the number of underlying factors would be the same in both groups, (2) that the pattern and value of the factor loadings would be equivalent, (3) that the measurement errors or unique variances would be equivalent, and (4) that the structural relations among the factors (i.e., factor variances/covariances) would be equivalent. Next, where invariance across the subgroups was established, we compared means across groups to determine whether there were significant differences between Hispanics and Blacks in child feeding beliefs and practices and
examined whether ethnic differences were moderated by demographic and social class indicators.
Methods Participants Participants were part of a study to investigate eating practices of low income, Black and Hispanic preschool children attending Head Start programs in greater metropolitan Houston, Texas. At eight centers, 231 primary caregivers (130 Hispanic, 101 Black) completed questionnaires. The primary caregiver was defined as the person caring for the preschooler most of the time when the child was not in school. Of these caregivers, 98% were female (85% mothers, 13% grandmothers) and 2% were male. There were 231 children in the study (55% female, 45% male), ranging from 3 to 5 years in age (M = 4.15, SD = 0.71). Participant characteristics by ethnic group are shown in Table 1. Measures Child feeding The Child Feeding Questionnaire (CFQ [6]) is a 31-item self-report questionnaire that measures three aspects of parental control in child feeding and four aspects of parental perceptions and concerns about child obesity using a 5-point Likert scale. The parental control subscales include restriction (8 items), pressure to eat (4 items), and monitoring of Table 1 Characteristics of participants by ethnic group
Parent gender-female Child gender-female Age, mean in years (SD) Parent Child Education of parent High school diploma or less Some college or more Marital status, parent married Household income, less than $2500/month Parent BMI Underweight (BMI b 18.5) Normal (18.5 V BMI b 25) Overweight (25 V BMI b 30) Obese (BMI z 30) Child BMIa Underweight or at risk (b15th percentile) Normal (z15th to b85th percentile) Overweight or at risk (z85th percentile) a
African-American (n = 101)
Hispanic (n = 130)
97.0% 57.4%
99.2% 52.3%
32.6 (9.5) 4.0 (0.7)
30.3 (8.3) 4.3 (0.7)
50.0% 50.0% 52.5% 92.1%
78.4% 21.6% 76.9% 96.9%
0.99% 17.8% 28.7% 52.5%
0.77% 23.1% 40.8% 35.4%
3.0%
0.8%
64.4%
61.5%
32.7%
37.7%
From age- and gender-specific cut points of the CDC growth charts.
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
eating (3 items). The parental perceptions and concerns subscales include responsibility for feeding (3 items), perceived weight of parent (current and retrospective, 4 items), perceived weight of child (current and retrospective, 1 to 6 items), and concern about child weight (3 items). The perceived child weight items are linked to current age of child such that the number of items used is based on the age of the children studied. Birch et al. reported coefficient alphas of .88 (Responsibility), .71 (Parent weight), .83 (Child weight), .75 (Concern about child weight), .70 (Pressure to eat), .73 (Restriction), and .92 (Monitoring) for the CFQ subscales. Items for the CFQ are shown in the Appendix. Its items are presented consecutively to subjects by factor (i.e., not scrambled). In the current study, the perceived parent weight factor (4 items) was not measured due to a clerical error discovered after data collection was completed, and three of six perceived child weight items were not used since they were not applicable due to age of child. In the Spanish translation, anchors for the child weight items were changed from bmarkedly underweight/ overweightQ to bvery underweight/overweightQ for both ethnic groups. Anthropometrics Height and weight were measured by staff in duplicate, averaged, and used to compute child and parent BMI (kg/m2). For adults, boverweightQ was defined as BMI of 25.0 to 29.9 and bobesityQ as BMI z 30.0, according to the World Health Organization criteria [9]. For children, BMI z scores and percentiles were calculated using the age- and gender-specific
523
cut points from the revised growth charts produced by the Centers for Disease Control and Prevention [10]. bAt risk for overweightQ was defined as at or above the 85th percentile but less than the 95th percentile, and boverweightQ as BMI z 95th percentile. Procedure The Institutional Review Board of Baylor College of Medicine approved the study and all participating parents provided informed consent. Participants were recruited at Houston area Head Start centers where data collection took place. Parents completed questionnaires in groups in a room separate from children and received $25 (gift certificate or cash) for participation. Statistical analyses The primary goal of the analyses was to compare the functional organization of the Child Feeding Questionnaire in two minority groups. This was accomplished in four phases. First, we evaluated the factor structure identified by Birch et al., as shown in Fig. 1, in the combined minority groups and separately by race. Second, we evaluated the model again, but without three item parcels that were utilized by Birch et al. on the Restriction factor (that is, we conducted an all item-level analysis). The rationale for forming the parcels (ratings from two or more items are averaged together to function as a single item; see Appendix A for content of the three parcels) was not reported in the
Fig. 1. Factor structure of the Child Feeding Questionnaire ([6]).
524
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
original investigation, and the statistical properties (e.g., factor loadings) of the individual items used in the parcels were also not reported. Third, we modified the measurement model with all item-level data to identify a model that provided the best account of the data in the minority groups. Fourth, we tested the invariance of the final modified model across the Hispanic and Black groups. Confirmatory factor analyses The models proposed were analyzed using maximum likelihood estimation in LISREL 8.51 [11]. PRELIS, the preprocessor of LISREL, was used to generate the covariance matrices using listwise deletion for missing data. The tentative measurement model hypothesized that (a) responses could be explained by six factors, (b) each item would have a non-zero loading on the factor it was intended to measure and a zero loading on all other factors, (c) all factor covariances were freely estimated, and (d) measurement error would be uncorrelated, except for the one error covariance specified by Birch et al. between PCW1 and PCW2. Subsequent models that were run did not allow correlated errors. Model fit Multiple indices were used to assess model fit. The Chisquare statistic provides an overall test of model fit and evaluates the absolute fit between the hypothesized model and the data. A nonsignificant Chi-square indicates that the model fits the data. However, since Chi-square assumes multivariate normality and is sensitive to sample size, significant Chi-squares that reject the model can occur even when the model fit is relatively good [12,13]. Therefore, additional indices of fit have been suggested. Additional indices that were used to evaluate goodness of fit included the Non-Normed Fit Index (NNFI, also called the Tucker– Lewis Index), Bentler’s Comparative Fit Index (CFI), the root mean square error of approximation (RMSEA), and the standardized root mean square residual (SRMR). Traditionally, NNFI and CFI values greater than or equal to .90 have been accepted as indicators of good fit. Recent work, however, by Hu and Bentler [14] has indicated that values near .90 are only minimally acceptable and that values of .95 or higher are necessary to indicate a good fit of the model. They also recommend a RMSEA cutoff value of .06 in conjunction with a SRMR value of .08 or less. The RMSEA has been recognized as one of the most informative criteria available, highly sensitive to models with misspecified factor loadings, while the SRMR is most sensitive to models with misspecified factor covariances or latent structures [14]. Factorial invariance The invariance analyses were performed in a hierarchical ordering of constraints [7,15]. Four, nested, multigroup analyses were performed using LISREL 8.51: (Model 1) the factor pattern was constrained to be equal across groups, (Model 2) equal factor loadings were added to the constraint of equal factor pattern, (Model 3) the error variances were
next added to the loading constraints across groups, and (Model 4) the factor variances and covariances were added to the constraints across groups. If a more restrictive model fits the data as well as the less restrictive model, then there is support for the invariance of the constrained parameters. The difference between the two models is tested by examining significant differences in the Chi-square values for the change in degrees of freedom. No substantial change of model fit (i.e., p z .05) would be indicative of model invariance. As previously stated, Model 2 or equal factor pattern and loadings is considered the minimal evidence of factorial invariance, and Models 3 and/or 4 demonstrate increased evidence of invariance across groups [7,15]. Mean group differences Multivariate analysis of covariance (MANCOVA) was used to determine if the subscales from the modified CFQ differed by ethnic group, social class, and child demographics. The CFQ factor scores (mean of items per factor) were used as dependent variables, and parent ethnicity (Hispanic, Black), parent education (high school diploma or less, some college or more), child gender (male, female), and child BMI (underweight, normal, at risk or overweight) were used as independent variables. Parent income, as an indicator of social class, was not included as a predictor, since all families were Head Start participants and thus, matched on low income. Parental weight (underweight, normal, overweight, obese) and marital status (married, not married) have been associated with child weight and were treated as covariates [16,17]. Significant multivariate effects (for main effects and all interactions) were followed by univariate analyses (ANOVA). A Bonferroni correction (a = .01) was used in the ANOVAs to control for experimentwise alpha. Least squares means (LSM) were computed in the post hoc comparisons.
Results Differences in participant characteristics There were significant ethnic differences in parent education, marital status, and parent BMI. African-Americans had higher education levels, v 2(2, n = 223) = 20.07, P b .0001, and were less likely to be married, v 2(1, n = 231) = 15.19, P b .0001. More African-American parents were obese, while more Hispanic parents were overweight, v 2(2, n = 229) = 6.96, P = .03. There were no significant ethnic differences in child weight status, child age, or parental age. Testing the fit of the Birch et al. model Ten participants with missing data were deleted from the combined sample of 231 (4.3% missing data). The replication of the six-factor Birch et al. model that included the item parcels and one error covariance resulted
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
in a mixed fit in the combined ethnic group that was minimally to moderately acceptable, based on the indices of fit, v 2(154, N = 221) = 223.12, p = 0.002, RMSEA = .045 (90% CI = .031–.057), NNFI = .91, CFI = .93, SRMR = .063. The data did not support the error covariance on child weight, and this covariance was dropped (D df = 1, Dr2 = 6.32, p b .02). By ethnic group, the model fit was better in the Hispanic than the African-American sample. Addressing the problem areas in the original model The fit indices in the combined and subgroup analyses did not fully reflect two important weaknesses in the model: (1) several non-significant or low item loadings and (2) the use of item parcels without a clear rationale or evaluation of the parcels’ components. First, two of the three perceived child weight item loadings were non-significant or below .36 (PCW2 and PCW3), as were two loadings on the restriction factor (RST2 and item parcel RST3). Although there is no rule of thumb on factor loading size in structural equation modeling, it is generally accepted that all factor loadings should be significant and loadings of .60 or greater are usually recommended. The low loadings of the perceived child weight items and a lack of covariance among them made it clear that these items were problematic. Ratings of 3 (that is, child is of baverageQ weight) on the 1 to 5 response scale were given by 77% of the combined sample for their child during its first year of life (PCW1), 81% of the sample for their child at 1–2 years old (PCW2), and 80% of the sample for their child at 3–5 years old (PCW3). Furthermore, further analysis of the three PCW items showed that their correlations with actual current child weight increased substantially with increases in child age, such that the largest correlations were between parental perceptions of the child’s current weight and the child’s actual current weight (birth to 1 year: r = .12, age 1 to 2: r = .29, current age: r = .46). These results suggest that averaging parental perceptions of child’s weight over time and relating this average to current weight (or other variables) may not make conceptual sense. Three constructs (one for each time interval), not just one, appear to exist. If the subscale is to be related to current child weight, as is routinely being done in the field of nutrition, then it may be conceptually best to use only items that refer to current weight. Since only one item reflected current weight and this item was problematic (three or more items are desired to load on a latent factor), the modified model was fit without the Perceived Child Weight factor. However, the single item for current weight (PCW3) was used in the MANCOVA analysis to represent the perceived current weight of child factor. Secondly, parcels were used in conjunction with itemlevel data on the restriction factor in the Birch et al. model. In the published article, neither the fit of the model nor the item loadings were shown without the item parcels, and there was no rationale given for their use. Thus, the reliabilities of the individual items that comprised the parcels on the restriction factor were not known and required evaluation.
525
Modifications to the CFQ model Regarding model modification, we felt the problems found could be best addressed by fitting the Birch model completely at the item level, without the child weight factor (since further scale development work is needed on this factor), re-examining the fit indices and factor loadings, and modifying the model as necessary. Before dropping the child weight factor, the model was fit using all item-level data. The fit of the Birch model was considerably worse when the item parcels were not used and the data were fit only at the item level, v 2(237, N = 221) = 452.32, p = 0.0000, RMSEA = .064 (90% CI = .055–.073), NNFI = .81, CFI = .83, SRMR = .075. Four of the eight restriction items (the factor that utilized the item parcels) loaded at or below .35. The item loadings on the child weight factor were similarly unacceptable. A modified model was fit to the data in which the perceived child weight factor was dropped. The resulting five-factor model did not provide an improved fit, and further modified models were fit that also sequentially dropped five low-loading items on the Restriction factor. Two items involved reward behaviors, one involved keeping bsome foodsQ out of the child’s reach, and two involved the child eating too many bfavorite foods.Q In dropping these items, it was reasoned that the reward items did not theoretically reflect the construct of restriction and that the other three items could be interpreted by subjects in a variety of positive or negative ways. The three items retained were more clearly indicators of a construct that represented restriction of foods that could contribute to overweight (i.e., sweets, high fat foods, junk foods). The loading of RST4a, which dealt with restriction of bjunk foods,Q was not desirable at .37, likely reflecting imprecision in wording. However, it was retained in the model because the loading was significant, the LISREL modification indices indicated that this item had no appreciable loadings on any other factors, and it was theoretically relevant as an obesity-related food. The final model (Fig. 2) had an adequate fit to the data, v 2(94, N = 221) = 128.97, p = 0.0097, RMSEA = .041 (90% CI = .021–.058), NNFI = .95, CFI = .96, SRMR = .055. As in the previous models that were run, the fit of the final modified model was better in the Hispanic than in the African-American sample (Table 2). Testing model equivalence across ethnic groups As shown in Table 3, the test of equal factor patterns (Model 1) for the modified model was not rejected ( p = .1800), indicating that a model of five factors was equivalent across race, and this was supported by the fit indices. The test of equal factor loadings (Model 2) was not significant ( p = .1200) and the Chi-square difference between Models 1 and 2 was not significant ( p N .10; see Table 3), indicating equivalence of the factor loadings
526
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
Fig. 2. Final modified measurement model fit to Hispanic and African-American subgroups and tested for invariance across groups.
shown in Table 4. All parameter estimates were significant and positive. Final loadings ranged from .37 to .90 (M = .70), with only 10 of the 16 loadings greater than .60. The low correlations observed between factors (range of .00– .46) supported that the factors were distinct.
across race. The errors or uniquenesses were not equivalent across race, as indicated by the significant v 2 for Model 3 (p = .0000), the significant Chi-square difference between Models 2 and 3 (p b .001), and the poorer fit indices associated with this model. This was not unexpected, since rarely do two samples share the same amount of measurement error. The test of equal factor variances and covariances (Model 4) was not rejected (p = .06) and the Chi-square difference test between Models 2 and 4 was not significant (p N .05), indicating that the correlations between factors were similar across race. The factor loadings from the modified model are shown in the Appendix A. The LISREL factor intercorrelations, means, and Pearson correlations with child weight are
Evaluation of ethnic differences and moderating effects of demographics and social class The multivariate analysis of covariance (MANCOVA) assessed mean differences in the feeding subscales by ethnic group and interactions with three possible moderators. Two levels of child age- and gender-specific BMI (bor z85th percentile), child gender (male, female), parent ethnicity
Table 2 Results of factorial validity analyses Model Birch et al. model replication (6-factor model, 3 parcels on Restriction, 24 items, no error covariances) Combined sample Hispanic Black Birch et al. model, without parcels (6-factor model, all item-level data, 24 items, no error covariances) Combined sample Modified model (5-factor model, all item-level data, 16 items, no error covariances) Combined sample Hispanic Black
df
v2
p
RMSEA (90% CI)
CFI
NNFI
SRMR
155 155 155
216.80 162.27 196.53
0.0008 0.330 0.013
0.043 (0.028–0.055) 0.020 (0.000–0.047) 0.053 (0.025–0.074)
0.93 0.97 0.90
0.92 0.96 0.88
0.063 0.066 0.091
237
452.32
0.0000
0.064 (0.055–0.073)
0.83
0.81
0.075
94 94 94
128.97 70.82 166.93
0.0097 0.9600 0.0550
0.041 (0.021–0.058) 0.000 (0.000–0.000) 0.050 (0.000–0.077)
0.96 1.00 0.93
0.95 1.04 0.92
0.055 0.056 0.087
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
527
Table 3 Results of ethnic invariance analyses for the modified model Invariance model
df
v2
p
RMSEA (90% CI)
CFI
NNFI
SRMR
Ethnic invariance model 1—test of same factor pattern Ethnic invariance model 2—test of same pattern and loadings Ethnic invariance model 3—test of same pattern, loadings, and errors Ethnic invariance model 4—test of same pattern, loadings, and factor variances and covariances
188
206.00
0.1800
0.030 (0.000–0.053)
0.96
0.95
0.056
199
223.00
0.1200
0.033 (0.000–0.055)
0.95
0.94
0.060
215
337.37
0.0000
0.072 (0.057–0.082)
0.83
0.81
0.076
214
247.16
0.0600
0.038 (0.000–0.057)
0.95
0.94
0.0081
Ethnic invariance comparisons
df diff
v 2 diff
p value
Interpretation
Ethnic invariance model 1 vs. 2 Ethnic invariance model 2 vs. 3 Ethnic invariance model 2 vs. 4
11 16 15
17.00 114.37 24.16
N0.10 b0.001 N0.05
Factor pattern and loadings not different across race Measurement errors different across race Factor pattern, loadings, variances and covariances not different across race
(Black, Hispanic), and parent education (high school or less, some college, or more) were used, with parental BMI and marital status as covariates. Results of the MANCOVA (child BMI child gender parent ethnicity parent education: all main effects and interactions), indicated a significant main effect for child BMI, Wilk’s K = .83, F(6,192) = 6.32, p b .0001. Two two-way interactions were significant: one for ethnicity child BMI (Wilk’s K = .93, F(6,192) = 2.43, p = .0273) and one for ethnicity parent education (Wilk’s K = .90, F(6,192) = 3.40, p = 0.0032). There was one significant three-way interaction (ethnicity
parent education child BMI), Wilk’s K = .89, F(6,192) = 3.97, p = .0009. Subsequent univariate analyses for the child BMI main effect indicated that perceived current child weight was the only variable significant at the .01 level (p b .0001). Post hoc comparisons indicated that parents of children at or above the 85th percentile for weight perceived their child as significantly more overweight (LSM = 3.25) than parents of children below the 85th percentile (LSM = 2.82). The mean for overweight children, however, did not indicate an accurate parental perception of the child’s
Table 4 Descriptive statistics by factor, interfactor correlationsa, correlations with child BMIb Factor 1. 2. 3. 4. 5. 6.
Perceived responsibility Concern about child weight Restriction Pressure to eat Monitoring Perceived child weightc
Modified model M (SD)
1
Hispanic
African-Am
4.79 2.30 4.19 3.70 4.05 3.01
4.61 2.06 4.31 3.57 4.27 2.92
(0.56) (1.22) (1.06) (0.90) (1.10) (0.33)
(0.67) (1.15) (0.84) (0.86) (0.98) (0.58)
– 0.10 0.12 0.28 0.04 –
2
3 0.02 – 0.00 0.19 0.16 –
4 0.25 0.17 – 0.59 0.42 –
5 0.40 0.12 0.29 – 0.22 –
Child BMI percentileb Hispanic 1. 2. 3. 4. 5. 6. a
Perceived responsibility Concern about child weight Restriction Pressure to eat Monitoring Perceived child weightc
0.06 0.14 0.01 0.12 0.02 0.42TTT
Black 0.05 0.27TT 0.02 0.17 0.01 0.38TTT
LISREL interfactor correlations or phi values. Hispanic correlations are below the diagonal, African-American correlations are above the diagonal. Pearson correlation with age- and gender-specific percentiles per the revised growth charts from the Centers for Disease Control and Prevention. c Perceived child weight, one item, which was not included in the modified LISREL model. TTp b 0.01. TTTp b 0.0001. b
0.36 0.18 0.25 0.25 – –
528
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
weight (that is, a scale rating of 3 was baverageQ weight). Further analysis revealed that 78% of parents of overweight or at risk children regarded the child as average or underweight. In the parent ethnicity child BMI interaction, only perceived current child weight was significant (p = .0008). Among children at or above the 85th percentile, Black parents perceived their child as significantly less overweight (LSM = 3.07) than did Hispanic parents (LSM = 3.44), although as stated earlier, the means for both ethnic groups demonstrate an underestimation of child weight by parents. The parent ethnicity parent education interaction was significant for concern (p = .0105) and for responsibility (p = .0048). Lower educated Black parents expressed significantly less concern about child weight (LSM = 1.79) than did higher educated Black parents (LSM = 2.67), but this difference was not present among Hispanics, where low and higher educated parents held similar levels of concern (LSM = 2.21low ed, 2.50higher ed). The mean values indicate generally low levels of concern about weight among both ethnic groups. Regarding responsibility, Hispanic parents with low education rated themselves as more responsible for the feeding of the child (LSM = 4.86) than did higher educated Hispanic parents (LSM = 4.30), but there were no significant differences in perceived responsibility for feeding among low and higher educated Blacks (LSM = 4.59low ed, 4.71higher ed). In the ethnicity parent education child BMI threeway interaction, only perceived current child weight was significant at the .01 level, p = .0027. Low educated Hispanic parents of children at or above the 85th percentile perceived their child as less overweight (LSM = 3.06) than did higher educated Hispanic parents (LSM = 3.83), but this education effect within race was not present among Blacks (LSM = 3.11low ed, 3.02high ed). As the mean values show, higher educated Black parents of overweight children perceived their child as significantly less overweight than did the higher educated Hispanics.
Discussion This study evaluated the suitability of the Child Feeding Questionnaire (CFQ) for use with low-income Black and Hispanic parents, analyzing six of its seven hypothesized dimensions. Overall, the minority data supported the factor structure found in the White and Hispanic samples used in the CFQ’s original development. However, substantial modifications were required in the item pool comprising the final minority model as a result of primarily conceptual and methodological issues that appeared to be unrelated to ethnicity. The final model of five factors and 16 items demonstrated invariance of factor structure, factor loadings, and factor variances– covariances across the two minority subgroups, indicating that the subscales of Responsibility, Concern about
Weight, Restriction, Pressure to Eat, and Monitoring assess similar underlying constructs across the Hispanic and African-American groups. Subsequent mean analyses revealed no mean differences between Hispanics and African-Americans in parental attitudes and practices, except those modified by parent education and child BMI. Our results showed that the modified CFQ functioned adequately in both ethnic minority groups but two adjustments were required. First, five of the eight restriction items were removed due to non-significant factor loadings. The two reward items may not conceptually reflect restriction. Reward and restriction likely represent two constructs instead of one and may differentially predict behavior, similar to findings in the addiction literature on the impact of positive and negative expectancies [18]. Three other restriction items had imprecise wordings that may have fostered widely different interpretations. Although the bjunk foodQ item was significant and retained in the model, its less than optimal loading reflects imprecision in wording and a need for item revision (i.e., what are bjunk foodsQ to parents? Are they different in minority populations?). The modified restriction factor resulted in a construct representing restriction of negative, obesity-related foods rather than the restriction of food per se. The modified restriction factor also was fit without item parcels. Item groupings or parcels are routinely used to reduce the number of items in a large item pool to allow estimation of a desired model. Items can be combined on a conceptual basis or by some selected pattern. Their use requires caution and proper methodology, however, because they tend to make the model fit better and can hide model misfits or poorly performing items [19,20], as was found in the current data. The use of parcels is acceptable with a stated rationale, after individual subfactors have been fit at the item level and there is good evidence that the individual items that comprise the parcels are sufficiently strong or reliable. Second, a lack of conceptual clarity was found in the perceived weight construct. Child weight over time may or may not be consistent, actually or in the parent’s mind. Therefore, parental evaluations of child weight at previous time points may not be at all related to the current weight of the child. For example, parents may remember their child as a chubby baby, but perceive the child’s current weight as normal. Furthermore, parents may apply different standards for age related weight (for example, it may be OK to be a chubby baby). These issues make the interpretation of perceived child weight as an average score derived from multiple time points problematic. The same conceptual problems likely exist on the perceived parent weight factor, whose four items were not included in the current data. Further research is needed on parental beliefs about child (current and retrospective) over time and their relation to child age and weight. Pending such research, we recommend focusing only on the perceived current weight of child and further item development to reflect this construct.
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
The two problems of focus appear to reflect cross-cultural weaknesses in the scale (that is, they also existed in the White samples) that are conceptual and methodological (for example, items are presented to subjects by factor with many consecutive items highly similar in wording, which can promote method effects), and not specific to any one ethnic group. The modifications made to the item pool in our analyses (i.e., dropping certain items) increased the scale’s psychometric integrity for use across ethnic groups, however continued scale development work appears warranted. Our results corroborate findings from other studies of systematic bias in the perception of child weight among minorities [5,21]. Both minorities underestimated the weight of overweight children, but Blacks more so than Hispanics. Importantly, parental education moderated this effect in our data, over and above the influence of family income, marital status, and parent BMI. Similar to findings by Baughcum et al. [21], an education effect was seen among Hispanics on perceived weight, where overweight children were classified as average or underweight more often among parents with a high school education or less, in contrast to Hispanics with more education. This social class effect was not seen among Blacks, however, where even higher educated Black parents perceived their overweight child as more normal than did higher educated Hispanics. Parental education also modified the ethnic effect on responsibility for feeding and concern about child weight. In contrast to Hispanics, both higher and lower educated Blacks perceived high responsibility for the feeding of their child. While this could be related to a lower prevalence of married parents among the African-Americans, marital status was controlled for in our analyses. Importantly, low levels of concern about child weight were seen in both minority groups. While higher and lower educated Hispanics and Blacks all held low ratings of concern about child weight, markedly low concern ratings were found among lower educated Blacks. In comparing the current data to that of the original Birch study, an interesting contrast was observed. Similar to Birch, moderate positive correlations were seen in our data between restriction and monitoring of foods. However, there was not a moderate positive correlation between restriction and concern about weight, as would be expected and as Birch et al. found in their predominately White samples. High mean ratings on restriction, as well as monitoring, were countered with low ratings on concern about weight in the African-American subgroup but not the Hispanic. This implies that motives for restricting foods may differ between cultures. For example, restriction may be regarded as important and practiced because eating foods like sweets could interfere with the child’s consumption of a main meal, not because such foods lead to weight gain. Except for perceived child weight, there were no significant relationships found in the MANCOVA analysis between measured child BMI and the six constructs that
529
were measured from the CFQ. As in Birch’s previous work in predominately White samples, we found a positive Pearson correlation between concern and child BMI in both minority subgroups, but the relationship was significant only among Blacks. The lack of relationship shown in the MANCOVA, however, does not make the other parental attitudes and practices irrelevant. Child eating behaviors, which have been related to parental attitudes and practices [4], represent important modifiable influences on child weight. Future research, including continued work on scale development and eating behaviors, will enable better conclusions about the importance of these constructs to child weight. The results of this study should be interpreted in light of several limitations. The findings are limited to Hispanics and African-Americans, and no White sample was available for further ethnic comparisons. Furthermore, the questionnaire did not include the four items for perceived parent weight such that a reduced version of the CFQ was tested. We acknowledge the small sample size of the groups, which may have facilitated the invariance in parameters found across groups. In addition, the primary mediator of the parental attitudes/practices and child weight relationship, child eating, was not measured in this study. However, using rigorous psychometric analyses to evaluate ethnic differences in two populations at high risk for obesity, our results support the use of a modified version of the Child Feeding Questionnaire (CFQ), a widely used measure in nutrition research, among two minority subgroups. The findings also support further item development and psychometric work on the instrument to address conceptual and methodological weaknesses that appear to be crosscultural.
Acknowledgments The authors wish to thank Sandra Lopez and Janet Bonner for their help in data collection. We also extend a special thanks to the children and parents who participated in the study. This research was supported by funds from the United States Department of Agriculture grant, bEnvironmental Influences on Children’s Food Consumption,Q Grant No. 2001-35200-10659 (PI: Theresa A. Nicklas). Partial support was also received from Dairy Management Inc. This work is a publication of the United States Department of Agriculture (USDA/ARS) Children’s Nutrition Research Center, Department of Pediatrics, Baylor College of Medicine, Houston, Texas, and had been funded in part with federal funds from the USDA/ARS under Cooperative Agreement No. 58-6250-6001. The contents of this publication do not necessarily reflect the views or policies of the USDA, nor does mention of trade names, commercial products, or organizations imply endorsement from the U.S. government.
530
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531
Appendix A. Modified model factor structure from original CFQ questionnaire, completely standardized loadings, standard errors, t values
Latent variable
Variable name
Item description and presentation order
Modified model Standardized loading
Perceived responsibility
Perceived parent weight
Perceived child weight
Concern about child weight
Restrictionc
PR1 PR2 PR3 PPW1 PPW2 PPW3 PPW4 PCW1 PCW2 PCW3 PCW4 PCW5 PCW6 CN1 CN2 CN3 RST1a RST1b RST1c RST2 RST3a RST3b RST4a
Pressure to eat
Monitoring
RST4b PE1 PE2 PE3 PE4 MN1 MN2 MN3
1. for feeding 2. for how much served 3. for healthy foods 4. childhood (5–10 years old) 5. adolescence 6. your 20s 7. at present 8. during first year of life 9. between age 1 and 2 10. between age 3 and 5 11. kindergarten through 2nd grade 12. 3rd through 5th grade 13. 6th through 8th grade 14. about eating too much when parent not w/child 15. about child having to go on diet 16. about child becoming overweight 17. make sure does not eat too many sweets 18. make sure does not eat too many high fat foods 19. make sure does not eat too many favorite foods 20. keep some foods out of reach 21. offer sweets as reward for good behavior 22. offer favorite foods for good behavior 23. would eat too many junk foods without guidance 24. would eat too much of favorite foods 25. should eat all food on plate 26. make sure child eats enough 27. try to make child eat when not hungry 28. child would eat much less without guidance 29. keep track of sweets 30. keep track of snack food 31. keep track of high fat foods
SEa
t value
0.59
b
___
___b
0.84 0.68 Omitted Omitted Omitted Omitted Dropped Dropped Dropped but used in MANCOVA Omitted Omitted Omitted 0.60
0.22 0.13 – – – – – – – – – – ___b
7.00 7.23 – – – – – – – – – ___b
0.88 0.72 0.87 0.76
0.20 0.18 ___b 0.13
7.49 7.99 ___b 8.03
Dropped
–
–
Dropped Dropped Dropped 0.37
– – – 0.12
– – – 4.88
Dropped 0.42 0.49 0.55 0.56
– ___b 0.28 0.36 0.38
– ___b 3.93 4.09 4.10
0.90 0.92 0.67
___b 0.07 0.08
___b 15.77 11.35
a
Standard error. Indicates a parameter fixed at 1.0 in the LISREL solution. c Item parcels used within restriction factor in original Birch et al. model: RST1a, b, and c made up Parcel 1; RST3a and b made up Parcel 2; RST4a and b made up Parcel 3. b
References [1] Hill JO, Peters JC. Environmental contributions to the obesity epidemic. Science 1998;280(1371):1374. [2] Johnson SL, Birch LL. Parent’s and children’s adoposity and eating style. Pediatrics 1994;94:653 – 61. [3] Costanzo PR, Woody EZ. Domain-specific parenting styles and their impact on the child’s development of particular deviance: the example of obesity proneness. J Soc Clin Psychol 1985;4:425 – 45. [4] Birch LL, Davison KK. Family environmental factors influencing the developing behavioral controls of food intake and childhood overweight. Pediatr Clin North Am 2001;48(4): 893 – 907. [5] Myers S, Vargas Z. Parental perceptions of the preschool obese child. Pediatr Nursing 2000;26(1):23 – 30.
[6] Birch LL, Fisher JO, Grimm-Thomas K, Markey CN, Sawyer R, Johnson SL. Confirmatory factor analysis of the child feeding questionnaire: a measure of parental attitudes, beliefs and practices about child feeding and obesity proneness. Appetite 2001;36:201 – 10. [7] Byrne BM. Structural equation modeling with LISREL, PRELIS, and SIMPLIS: Basic Concepts, Applications, and Programming. Mahwah, NJ7 Erlbaum; 1998. [8] Vandenberg RJ, Lance CE. A review and synthesis of the measurement invariance literature: suggestions, practices, and recommendations for organizational research. Organ Res Methods 2000;3(1):4 – 70. [9] World Health Organization Expert Committee. Physical status: the use and interpretation of anthropometry. WHO Technical Report Series: 854. Geneva7 World Health Organization; 1995.
C.B. Anderson et al. / Preventive Medicine 41 (2005) 521–531 [10] Centers for Disease Control and Prevention. CD growth charts: United States. Hyattsville, MD7 National Center for Health Statistics, U.S. Dept of Health and Human Services; 2000. [11] LISREL 8.51. Chicago, IL: Scientific Software International, Inc.; 2001. [12] Hu L, Bentler PM. Evaluating model fit. In: Hoyle RH, editor. Structural equation modeling: Concepts, Issues, and Applications. Thousand Oaks, CA7 Sage; 1995. p. 76 – 99. [13] Marsh HW, Balla JR, McDonald RP. Goodness-of-fit indexes in confirmatory factor analysis: the effect of sample size. Psychol Bull 1988;103:391 – 410. [14] Hu L, Bentler PM. Cutoff criteria for fit indexes in covariance structure analysis: conventional criteria versus new alternatives. Struct Equ Model 1999;6(1):1 – 55. [15] Joreskog KG, Sorbom D. LISREL 8: User’s Reference Guide. Chicago, IL7 Scientific Software; 1996.
531
[16] Whitaker RC, Wright JA, Pepe MS, Seidel KD, Dietz WH. Predicting obesity in young adulthood from childhood and parental obesity. N Engl J Med 1997;337:869 – 73. [17] Strauss RS, Knight J. Influence of the home environment on the development of obesity in children. Pediatrics 1999;103(6):1 – 8. [18] Anderson CB, Pollak KI, Wetter D. Relations between self-generated positive and negative expected smoking outcomes and smoking behavior: an exploratory study among adolescents. Psychol Addict Behav 2002;16(3):196 – 204. [19] Bandalos DL. The effects of item parceling on goodness-of-fit and parameter estimate bias in structural equation modeling. Struct Equ Model 2002;9(1):78 – 102. [20] Plummer BA. To parcel or not to parcel: the effects of item parceling in confirmatory factor analysis. Diss Abstr Int 2001;61(10-B):5617. [21] Baughcum AE, Chamberlin LA, Deeks CM, Powers SW, Whitaker RC. Maternal perceptions of overweight preschool children. Pediatrics 2000;106(6):1380 – 6.