CHAPTER II. METHODS AND DESIGN 61
I. METHODS FOR THE QUANTITATIVE STUDY (STUDY 1) 67
I.4 Quantitative data analysis procedures 75
The quantitative study starts by describing the sample, with the mean, standard deviation and range of different components of the HLLE (such as the expectations of the parents of the child’s schooling and the beliefs they held in relation to language and literacy development, the number of children’s books in the home, frequency of shared book reading, other print matter shared with the child, frequency of letter and word identification activities at home and time watching TV). The frequencies of these variables are then compared to the results obtained by other Chilean and international studies that have analysed the same or similar components.
Following the descriptive analyses, exploratory factor analyses and correlations were then used to provide a preliminary idea of which variables, combined in which composites should be included in a predictive model of the HLLE for Chilean low SES families.
Factor analyses have been amply and successfully used in previous study for example to determine if different subsets of literacy skills represent the same underlying ability (Lonigan, Burguess & Anthony, 2000) and to see if different measures of oral language grouped together (Senechal et al., 1998).
In order to do the factor analyses, variables from the family questionnaire were grouped according to their theme and four theoretical scales were built with related variables. The first one included variables on the language and literacy resources available in these low SES homes. The second group included variables related to language and literacy beliefs and expectations. The third group included variables which assessed the frequency and variety of child reading, writing and decoding experiences in the home. Finally the fourth group included items which measured the frequency of decontextualized conversations in the home.
Four separate factor analyses were then conducted, one on each of the four groups. The second, third and fourth groups of variables mentioned above included between 10 and 13 variables. Through the use of principal axis factoring, the exploratory factor analysis permitted the selection of the least factors that could account for the common variance. The factor analyses thus served to reduce the number of variables, which was necessary to obtain a certain degree of parsimony in the HLLE model.
The exploratory factor analyses were performed using the Mplus 6.11 program. More than 5% of data was missing for each variable so missing data treatment was used. Mplus treats missing data by analysing the frequency of the missing data patterns, after which it imputes data for the missing data and then checks that the assumptions are met. Since the assumption of normal distribution was not met for many of the variables, the MLR estimator was used because it was considered the most robust estimator to deal with violation of the assumption of multivariate normality. Varimax rotation was used because as argued by Cohen, Manion & Morrison (2007) this type of rotation allows for a clearer interpretation of the data, where factors are more clearly distinguished from each other.
The dichotomous variables within each of the theoretical dimensions were excluded from the factor analysis. Since Mplus does not calculate Cronbach´s Alpha, it was
calculated manually with the data from the correlation matrix using the following formula:
𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛𝑠 𝑎𝑣𝑒𝑟𝑎𝑔𝑒 ∗ 𝑁° 𝑖𝑡𝑒𝑚𝑠 𝑁° 𝑖𝑡𝑒𝑚𝑠 − 1 ∗ (𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛𝑠 𝑎𝑣𝑒𝑟𝑎𝑔𝑒 + 1)
The rotated factor loadings are thus presented as well as the composites that were created with the resulting items.
Then, correlational analyses looked at the relationship between background variables (such as SES or family demographics) and the development of language and literacy skills as well as the relationships between caregivers´ expectations, language and literacy beliefs and home literacy resources and the four outcome variables
(vocabulary, letter-‐word identification, spelling and text comprehension). Correlational analyses have been used in previous HLE studies such as the one by Leseman & de Jong (1998) as a preliminary step to path analyses and to improve the understanding of the relationship between background characteristics and home literacy practices and between home literacy practices and language and literacy skills. The present research however goes a step beyond by also including correlations between these components and caregivers´ language and literacy beliefs.
Taking into consideration the results of these analyses, a hypothetical model of the HLLE of Chilean low SES urban families with preschoolers was developed and is presented. As a first step, four separate path analyses models were tested, one for each of the four outcomes. These models were then compared and their fit indices discussed. The fit of the path models for each individual outcome supported the plausibility of an overall model with all four outcomes. This “overall” model includes distal and proximal components and presents a hypothesis of their direct or mediated influence on the four language outcomes as measured by some tests of the Woodcock Muñoz Language Survey revised (WMLS-‐R). The results of this overall model, and of the four individual models are compared and commented.
Path analysis had several features that fit well with the specific purposes and holistic perspective of the current HLLE research such as a) the flexibility regarding the types of relationships that can be specified in the hypothetical model, b) the comprehensive view it provided of the relationship between distal and proximal variables, or c) the fact that it allowed for several outcome variables to be included simultaneously. Section III in chapter IV provides a more detailed explanation of the advantages of path analysis over other types of regression analysis in relation to the purposes of this research.
This research was informed by the experience of previous studies that explored the Home Learning Environment using path analysis. For example, Leseman & de Jong, (1998) used path analysis to assess “the relationships between sociocultural and ethnic-‐ cultural background, home literacy, home language, and early language and literacy learning in school” (p.311).
In a similar fashion to Leseman & de Jong´s study, in the present research background characteristics were considered exogenous variables that predicted home literacy and language practices.
Leseman & de Jong´s conceptualization of home literacy included measures of frequency of literacy-‐related interactions in the home, but also included measures of the quality of these interactions. Similar measures of the quality of interactions were unfortunately not available in the UBC parent questionnaire.
Leseman & de Jong´s conceptualization of home literacy was mostly focused on shared reading. However, since there is evidence that shared reading is not a frequent practice in Chilean homes, for the purpose of the present research it seemed more valid to include measures of word and letter writing and identification in addition to the measures of the frequency of shared reading in the path model because there is evidence that these practices happen more frequently in Chilean homes.
When analysing their sample that included different ethnic subgroups Leseman and de Jong discovered that background characteristics of the families were strongly
associated to home literacy practices and to language and literacy measures. Considering that the present research analyses represents a seemingly more homogeneous sample of mid to low SES Chilean families, it seemed interesting to investigate if minor variations in background characteristics, such as SES, would still be strongly associated with home literacy practices and language and literacy measures.
Finally, the home literacy opportunity facet included in Leseman & de Jong´s study was based on self-‐reports and measured by a questionnaire. Even though they
acknowledged the desirability bias that this might have implied, they argued that none of the analyses indicated that such an effect was present. This supported the use of data from a questionnaire for the present research.
Path analysis is usually considered a confirmatory rather than exploratory type of statistical analysis. However, several studies using path analyses models test the fit of their model, then modify it (by deleting or adding parameters) and retest the new
model. For example, Farver et al. (2006) used a path analysis to first assess the fit of a model that assessed the relations between “parents’ literacy involvement, mothers’ parenting stress, and children’s PPVT-‐R/TVIP scores and social functioning, without considering the children’s literacy interest” and then entered another HLLE scale (parents’ reports of children’s literacy interest) to test how much it mediated between parents literacy involvement and children´s school readiness outcomes. Then they dropped some of the non-‐significant associations that emerged and produced an overall model with acceptable fit.
However, as reviewed by Hox & Bechger (1998, p. 9-‐10) there is evidence that “model modification often fails to find the correct model (Spirtes, Scheines & Glymour, 1991), and that models so achieved cross-‐validate badly (Maccallum, 1986; MacCallum, Roznowskei & Necowitz, 1992)”.
One alternative way of testing the stability of the present research´s model would have been to divide the sample in two groups, test a path analysis model on a first group, modify it until the fit indices were acceptable and then cross validate the model with the second group.
However, in the view of this researcher, path analysis is more a model testing, rather than model producing procedure, i.e. it is a procedure where a model based on theory and knowledge of a specific set of variables and populations is tested. Consequently, a more confirmatory rather than exploratory approach to path analyses was taken in this research.
After the path analysis was conducted, a direct discriminant analysis was performed using the scales from the path analysis model as predictors and a composite of the results of the four outcome tests as a dependent variable. The main purpose of the discriminant analysis was to build an index to help categorize the families according to the quality of HLLE provided. As a result, the homes in the sample were then classified into three groups: high HLLE, medium HLLE and low HLLE. A subsample from each of these three groups is used for the qualitative study.