Item Parceling in Structural Equation Models for Optimum Solutions
Rationale Regarding Item Parceling
Bandalos and Finney (2001) report that the three most common reasons researchers cite for using item parceling are to: increase the stability of the parameter estimates (29%) improve the variable to sample size ratio (22.6%), and to remedy small sample sizes (21%). The empirical evidence that parceling is a desirable correction to these data problems is mixed, however. In the majority of studies assessing the effectiveness of item parceling to solve these data problems, item parceled solutions have been compared to disaggregated analyses without item parcels. Bagozzi and colleagues, in a series of studies, reported that parceling was actually preferred to disaggregated analyses in most cases because the measurement error is reduced with parceled sets of items (Bagozzi & Heatherton, 1994; Bagozzi & Edwards, 1998). Yet, they recommend careful consideration to validity, unidimensionality, and level of specificity when constructing item parcels.
A rationale for item parceling that is often stated is to reduce the effects of nonnormality and likelihood of forming difficulty factors in factor analyses with binary items. Thompson and Melancon (1996) demonstrated that using item parceling with nonnormal data did result in more normally distributed item parcels and improved model fit. Their method of item parceling was to create parcels of items with opposite skew in an iterative procedure that resulted in parcels with less skew than the original items.
Nasser and Takahashi (2003) examined the behavior of various fit indices as they varied both the number of parcels and number of items per parcel Using Sarason’s Reactions to Tests instrument. Their results support the use of parcels rather than individual items and using a strategy to construct item parcels in which there are fewer parcels but more numbers of items per parcel. They indicate that solutions from parceled data with more items per parcels results in more normality, validity, continuity, and reliability than from parceled data with fewer items per parcel. Although, they indicated that some indices (i.e., χ2:df and RMSEA) were less consistent and generally had better fit when more parameters in the model were estimated. As MacCallum, Widaman, Zhang, and Hong (1999) point out, parceled solutions can be expected to provide better models of fit because a) they have fewer parameters to estimate, b) they have fewer chances for residuals to be correlated, and c) they lead to a reduction in sampling error. Little, Cunningham, Shahar, and Widaman (2002) list three reasons that parceling can be advantageous over using the original items: 1) estimating large numbers of items is likely to result in spurious correlations, 2) subsets of items from a large item pool will likely share specific sources of variance that may not be of primary interest, and 3) solutions from item-level data are less likely to yield stable solutions than solutions from parcels of items. However, if the latent construct is not unidimensional, it is likely that the item parcels will also be multidimensional that that it will be difficult to define what the latent construct actually is because the structure will be a confounding of the primary factor and systematic variance that is shared across parcels. In sum, when parceling with multidimensional structures, the parceling can mask many forms of model misspecification. The other caveat of item parceling that Little et al. suggest is that the unstandardized parameters may be meaningful in clinical practice and norms may be established based on the scale of the original items and these norms may not translate to the reparameterized model with item parcels.
Studies with Known Population Structure
The aforementioned studies have relied on analyses of actual data and theoretical explanations, yet, fewer studies have been conducted in which the population structure of the underlying model was known. This is critical, as it is not known in applied studies if better model fit is necessarily a desired goal, as would be the case with a misspecified model. In simulation studies with known population parameters, it can be determined if the increase in model fit with parceling methods is due to the increased sensitivity of the parceling method to a fully specified model, or whether the increase in fit is in err in a misspecified model.
Marsh, Hau, Balla, and Grayson, (1998) and Yuan, Bentler, & Kano, (1997), in separate simulation studies, demonstrated that it was advantageous to parcel rather than to use the same number of individual items; the fit indices were higher and results were more likely to yield a proper solution when parcels were used, rather than the same number of individual items (e.g., six parcels vs. six items). However, if the total number of individual items were used (e.g., 12 items instead of six 2-item parcels), the individual items were more likely to result in a proper solution.
Hall, Snell, and Singer Foust (1999) determined, in a simulation study of item parceling under various structural models with different parceling strategies, that when a secondary factor structure exists parceling can obscure the true factor structure and result in biased parameter estimates and inflated indices of fit. However, this depends on how the secondary factor structure is modeled; if the items that all load on a secondary factor are parceled together (isolated parceling) the fit is not inflated. It is when the items with a secondary influence are parceled with items without the secondary influence (distributed parceling), that the fit is inflated. Hall et al. explained this phenomena by suggesting in the latter case that the influence of the secondary factor is distributed across parcels and this creates common variance among these parcels and hence the loadings on the primary factor increase and partially reflect the influence of the secondary factor (i.e., the influence of the secondary factor confounds the loadings on the primary factor). Yet, when the items with the secondary influence are parceled together, the common variance is not inflated since the secondary structure influence is not across all the parcels; hence the error variance increases to reflect the influence of the secondary structure and fit indices are consequently lower, relative to the distributed parceling strategy. Hall et al. further suggested that the inflated indices of fit with distributed parceling methods helps to explain why past authors in the educational testing literature (e.g., Kishton & Widaman, 1994; Lawrence & Dorans, 1987; Manhart, 1996; Schau, Stevens, Dauphinee, & Del Vecchio, 1995; Thompson & Melancon, 1996) have suggested that distributing like items across different parcels can help improve the model fit and have therefore recommended this method. As Hall et al.’s simulation work indicates, however, this strategy artificially inflates model fit and therefore Hall et al. recommend putting like items together when creating parcels (i.e., putting items that have high loadings on a factor together).
Bandalos (2002) in a comprehensive simulation study that varied number of items per parcel, degree of nonnormality, number of categories in original items, n sizes and correct and misspecified models, investigated the effect of item parceling under conditions of nonnormality, an oft cited reason for item parceling. The findings indicated that item parceling produces higher fit indices, less parameter bias in disturbance terms, and less Type I errors than analyses of the original items when nonnormality is introduced. The higher the level of nonnormality, the larger the discrepancy was between the parceled and non-parceled solutions, favoring the parceled solution. Bandalos concluded that model fit improvement is due to better distributional characteristics (e.g., less nonnormality and less coarse measurements) in the item parcels and a reduction in the number of variance/covariance parameters that are modeled. The caveat is that this only holds true when the factors have underlying unidimensional structures. Both Hall et al. (1999) and Bandalos (2002) have shown that when even small influences of secondary factors are present (e.g., a misspecified model) the method of item parceling makes a great deal of difference in whether the misspecification can be identified or not. Bandalos concluded that item
Item Parceling in Structural Equation Models for Optimum Solutions
Rationale Regarding Item Parceling
Bandalos and Finney (2001) report that the three most common reasons researchers cite for using item parceling are to: increase the stability of the parameter estimates (29%) improve the variable to sample size ratio (22.6%), and to remedy small sample sizes (21%). The empirical evidence that parceling is a desirable correction to these data problems is mixed, however. In the majority of studies assessing the effectiveness of item parceling to solve these data problems, item parceled solutions have been compared to disaggregated analyses without item parcels. Bagozzi and colleagues, in a series of studies, reported that parceling was actually preferred to disaggregated analyses in most cases because the measurement error is reduced with parceled sets of items (Bagozzi & Heatherton, 1994; Bagozzi & Edwards, 1998). Yet, they recommend careful consideration to validity, unidimensionality, and level of specificity when constructing item parcels.
A rationale for item parceling that is often stated is to reduce the effects of nonnormality and likelihood of forming difficulty factors in factor analyses with binary items. Thompson and Melancon (1996) demonstrated that using item parceling with nonnormal data did result in more normally distributed item parcels and improved model fit. Their method of item parceling was to create parcels of items with opposite skew in an iterative procedure that resulted in parcels with less skew than the original items.
Nasser and Takahashi (2003) examined the behavior of various fit indices as they varied both the number of parcels and number of items per parcel Using Sarason’s Reactions to Tests instrument. Their results support the use of parcels rather than individual items and using a strategy to construct item parcels in which there are fewer parcels but more numbers of items per parcel. They indicate that solutions from parceled data with more items per parcels results in more normality, validity, continuity, and reliability than from parceled data with fewer items per parcel. Although, they indicated that some indices (i.e., χ2:df and RMSEA) were less consistent and generally had better fit when more parameters in the model were estimated. As MacCallum, Widaman, Zhang, and Hong (1999) point out, parceled solutions can be expected to provide better models of fit because a) they have fewer parameters to estimate, b) they have fewer chances for residuals to be correlated, and c) they lead to a reduction in sampling error. Little, Cunningham, Shahar, and Widaman (2002) list three reasons that parceling can be advantageous over using the original items: 1) estimating large numbers of items is likely to result in spurious correlations, 2) subsets of items from a large item pool will likely share specific sources of variance that may not be of primary interest, and 3) solutions from item-level data are less likely to yield stable solutions than solutions from parcels of items. However, if the latent construct is not unidimensional, it is likely that the item parcels will also be multidimensional that that it will be difficult to define what the latent construct actually is because the structure will be a confounding of the primary factor and systematic variance that is shared across parcels. In sum, when parceling with multidimensional structures, the parceling can mask many forms of model misspecification. The other caveat of item parceling that Little et al. suggest is that the unstandardized parameters may be meaningful in clinical practice and norms may be established based on the scale of the original items and these norms may not translate to the reparameterized model with item parcels.
Studies with Known Population Structure
The aforementioned studies have relied on analyses of actual data and theoretical explanations, yet, fewer studies have been conducted in which the population structure of the underlying model was known. This is critical, as it is not known in applied studies if better model fit is necessarily a desired goal, as would be the case with a misspecified model. In simulation studies with known population parameters, it can be determined if the increase in model fit with parceling methods is due to the increased sensitivity of the parceling method to a fully specified model, or whether the increase in fit is in err in a misspecified model.
Marsh, Hau, Balla, and Grayson, (1998) and Yuan, Bentler, & Kano, (1997), in separate simulation studies, demonstrated that it was advantageous to parcel rather than to use the same number of individual items; the fit indices were higher and results were more likely to yield a proper solution when parcels were used, rather than the same number of individual items (e.g., six parcels vs. six items). However, if the total number of individual items were used (e.g., 12 items instead of six 2-item parcels), the individual items were more likely to result in a proper solution.
Hall, Snell, and Singer Foust (1999) determined, in a simulation study of item parceling under various structural models with different parceling strategies, that when a secondary factor structure exists parceling can obscure the true factor structure and result in biased parameter estimates and inflated indices of fit. However, this depends on how the secondary factor structure is modeled; if the items that all load on a secondary factor are parceled together (isolated parceling) the fit is not inflated. It is when the items with a secondary influence are parceled with items without the secondary influence (distributed parceling), that the fit is inflated. Hall et al. explained this phenomena by suggesting in the latter case that the influence of the secondary factor is distributed across parcels and this creates common variance among these parcels and hence the loadings on the primary factor increase and partially reflect the influence of the secondary factor (i.e., the influence of the secondary factor confounds the loadings on the primary factor). Yet, when the items with the secondary influence are parceled together, the common variance is not inflated since the secondary structure influence is not across all the parcels; hence the error variance increases to reflect the influence of the secondary structure and fit indices are consequently lower, relative to the distributed parceling strategy. Hall et al. further suggested that the inflated indices of fit with distributed parceling methods helps to explain why past authors in the educational testing literature (e.g., Kishton & Widaman, 1994; Lawrence & Dorans, 1987; Manhart, 1996; Schau, Stevens, Dauphinee, & Del Vecchio, 1995; Thompson & Melancon, 1996) have suggested that distributing like items across different parcels can help improve the model fit and have therefore recommended this method. As Hall et al.’s simulation work indicates, however, this strategy artificially inflates model fit and therefore Hall et al. recommend putting like items together when creating parcels (i.e., putting items that have high loadings on a factor together).
Bandalos (2002) in a comprehensive simulation study that varied number of items per parcel, degree of nonnormality, number of categories in original items, n sizes and correct and misspecified models, investigated the effect of item parceling under conditions of nonnormality, an oft cited reason for item parceling. The findings indicated that item parceling produces higher fit indices, less parameter bias in disturbance terms, and less Type I errors than analyses of the original items when nonnormality is introduced. The higher the level of nonnormality, the larger the discrepancy was between the parceled and non-parceled solutions, favoring the parceled solution. Bandalos concluded that model fit improvement is due to better distributional characteristics (e.g., less nonnormality and less coarse measurements) in the item parcels and a reduction in the number of variance/covariance parameters that are modeled. The caveat is that this only holds true when the factors have underlying unidimensional structures. Both Hall et al. (1999) and Bandalos (2002) have shown that when even small influences of secondary factors are present (e.g., a misspecified model) the method of item parceling makes a great deal of difference in whether the misspecification can be identified or not. Bandalos concluded that item
การแปล กรุณารอสักครู่..
![](//thimg.ilovetranslation.com/pic/loading_3.gif?v=b9814dd30c1d7c59_8619)