Multiple Regression Models
Multiple least squares regression models were used to test the prediction of children's functioning by the set of risk indexes. Separate regressions were conducted for each outcome variable, and thus, the number of participants included in each regression model varied (Ns ranged from 332 to 337). Risk indexes were entered into the regression model in a predetermined, hierarchical order, with specific demographic variables in the first step; occupation, education, and race entered in the second step; family psychosocial and mother's report of depressive symptoms entered in the next two steps; quality of the neighborhood environment entered in the next; and study site entered in the last step. This order of entry addressed the questions discussed in the introduction. The variables comprising these categories are described in the left-hand column of –. The findings on parent report of behavior are found in , teacher-reported outcomes are found in , and child achievement test scores are found in . In each table, the first column shows the zero-order correlations of the risk factors with outcomes, the second column reports standardized regression coefficients at entry into the model, the third column lists the total R2 that indicated the cumulative proportion of variance for the model at the particular step (cumulative R2), and the fourth column (last step) is the standardized beta for the individual risk factors that remain significant when controlling for the entry of all other risk factors. The last column (Unique) includes a calculation of the proportion of unique variance accounted for by each major domain when all other domains are controlled in the model. The results are discussed in terms of the significance of the risk factors across the different domains of functioning.
Total R2 and Standardized Regression Coefficients for Variables at Each Step and the Final Step of the Hierarchical Regressions Predicting Parent Report of Children's Symptoms
Total R2 and Standardized Regression Coefficients for Variables at Each Step and the Final Step of the Hierarchical Regressions Predicting Children's Achievement Test Scores
Total R2 and Standardized Regression Coefficients for Variables at Each Step and the Final Step of the Hierarchical Regressions Predicting Teacher Report of Children's Symptoms
In general, between 18% and 29% of the variance in outcomes was predicted by the entire set of variables. Specific questions of this study were examined by steps in the analysis that revealed the predictive strength of different categories of context, either independently or above and beyond the significance of particular other categories. The first question addressed was whether the neighborhood variable predicted outcomes after demographic and family psychosocial variables had been taken into account. Neighborhood added significantly to the prediction of parent-report externalizing and the teacher report of authority acceptance and social competence, after demographic and family context variables had been entered. However, the absolute increment in prediction by neighborhood context was quite small across all child outcome variables. This would suggest that neighborhood, as a distal factor, accounted for a small but unique portion of the variance in young children's externalizing problems.
Second, parent education significantly predicted teacher-report social competence and cognitive concentration and reading achievement scores after the specific demographic variables had been entered. Parent occupation significantly predicted parent-report externalizing and internalizing. Race contributed to the prediction of all outcome measures after the specific demographic variables had been entered. The SES–race domain remained significant as a predictor category at the last step when all other categories had been entered, for parent-report externalizing and internalizing, teacher-report cognitive concentration, and reading achievement However, in response to our third question, the significance of race as a factor disappeared when all other variables were entered in the equations; except for reading achievement, whereas the SES variables remained as significant predictors for parent-report externalizing and internalizing, teacher-report cognitive concentration, and reading achievement The specific demographic variables were not significant at the last step of the analysis for any of the outcome measures; however, as a predictor category, they predicted significant unique variance in children's math achievement ().
The fourth question dealt with whether family psychosocial risk variables account for variance in children's symptomatology over and above commonly studied demographic risk variables. This category of variables added significantly to the prediction of all the outcome variables when the demographic factors and SES and race were already entered and remained significant when neighborhood was entered, suggesting that these measures had a unique relation to various measures of child adjustment at first grade.
Fifth, mothers' depression significantly predicted only parent-report externalizing and internalizing after the family psychosocial variables and the demographic variables were entered and continued to predict those outcomes significantly when neighborhood context had been entered in the model.
Site effects were assessed by three dummy-coded (four sites) variables entered in the last step of the regression. Significant site effects were indicated by a significant increment in the total R2 at this step. Site accounted for significant variance in the last step of model only for teacher-report social competence—although none of the sites differed significantly from the others—and reading achievement, with the Washington state site showing higher reading scores than all of the other sites.
Finally, we examined the last-step significance of each individual measure by type of adjustment outcome as an index of the unique contribution of each contextual factor. Behavior problems were predicted by quite different variables depending on the source of information on behavior. Parent ratings of externalizing problems on the CBCL were predicted uniquely by parent occupation, life stress, maternal depression, and neighborhood risk. Mothers were the predominant source of information for these contextual variables (except for neighborhood risk), as well as for the outcome measure. On the other hand, teacher ratings of behavior problems, as assessed by the Authority Acceptance subscale, were predicted by family expressiveness and neighborhood risk. The two other teacher-report measures, cognitive concentration and social competence, were uniquely predicted by Family Life Stress and Family Expressiveness. The two standardized measures of achievement were predicted uniquely by entirely different factors. Parent Education and Race predicted reading achievement, whereas Life Stress and Family Expressiveness predicted math achievement.
The patterns of unique variance in child outcomes predicted by the major risk domains mirror the preceding findings for individual risk factors but provide more succinct conclusions about the kinds of contextual risk related to different types of child outcome. In general the overall set of family psychosocial risk factors made the strongest unique predictions across all of the child outcomes. The one exception was the child's reading achievement, in which SES made the strongest contribution. SES and race predicted unique variance in parent-report outcomes,teacher-report cognitive concentration, and reading achievement, accounting for 1% to 6% of unique variance. Maternal depression uniquely predicted variance in mothers' ratings of children's behavior problems, and neighborhood risk uniquely predicted parent- and teacher-report behavior problems.
The goal of this article was to examine how risk domains in kindergarten predicted children's adjustment at the end of first grade. A separate, but related, question is how change in children's status between kindergarten and the end of first grade can be predicted by the domains of risk. To explore this question, the main analyses were conducted again, controlling for the Time 1 outcomes by entering the relevant variable in the first step of the regression. These analyses were not possible for teacher report of cognitive concentration or social competence, as they were not assessed in kindergarten. The pattern of results of these analyses was largely the same as that presented in the Results section with the following exceptions. First, the models accounted for a larger proportion of variance in the Time 2 outcomes, with the same variable at Time 1 accounting for an additional 30% to 50% of the variance in Time 2 outcomes, indicating that there was a relatively high degree of stability over one year. Second, there were several differences in the variables that were significant at entry into the equation but few differences in the variables that remained significant at the last step. Specifically, the pattern of significance and magnitudes of standardized regression coefficients for CBCL Externalizing and Internalizing and WJ–R Letter-Word Recognition were the same. For Authority Acceptance, only neighborhood quality remained significant at the last step. For WJ–R Calculations, the number of siblings and race become significant in the last step, whereas life stress became nonsignificant Taken together, these findings suggest that the relative impact of specific and broad demographic variables, family context variables, maternal depression, and neighborhood context were largely equivalent whether the analyses were predicting level of Time 2 outcomes or change in outcomes from Time 1 to Time 2.
To test whether the relation between the risk factors and outcomes differed for boys and girls, exploratory interactions of each risk factor with gender were tested using multiple regression. The order of entry of predictor variables was (a) the single order terms of all of the risk factors, (b) children's gender (coded as boys = −1, girls = 1, as recommended by Aiken & West, 1991
), and (c) the interaction terms of each risk variable by gender. If there were significant interactions, a second multiple regression was tested eliminating all nonsignificant interactions from Step 3 in order to determine whether-the interactions remained significant (i.e., were spurious as a result of numerous variables in the model). If an interaction remained significant, it was interpreted and the simple slopes of the interaction were probed using the technique suggested by West, Aiken, and Krull (1996)
. The models were retested twice. First, gender was recoeded as boys = 0, girls = 1. The regression coefficients that resulted from this regression reflected the relation of each risk factor with the outcomes for boys. Second, gender was recoded as girls = 0, boys = 1, with the resulting regression coefficients reflecting the relation of each risk factor with the outcomes for girls.
There was a unique main effect of gender (coded boys = −1, girls = 1) only for teacher-report Authority Acceptance (B = −.41.p < .01), with behavior problems being higher for boys than for girls. There were few significant Gender × Risk Factor interactions. The interaction of Neighborhood Risk × Gender was significant for parent-report CBCL Externalizing (boys: β = 3.35, p < .01; girls: β = −0.80, ns) and teacher-rated Authority Acceptance (boys: β = 0.43, p < .01; girls: β = 0.03, ns). In both cases, the relation between neighborhood risk and the outcomes was stronger for boys than for girls.
In addition, the overall model was tested separately for boys and girls to assess whether the risk factors accounted for a different proportion of variance between boys' and girls' outcomes. The model R2s for each outcome tested separately for boys and girls are presented in . A greater proportion of variance was predicted for boys than for girls in parent-rated externalizing and internalizing problems and teacher-rated behavior problems, whereas a greater proportion of variance was predicted for girls than for boys in teacher-rated cognitive concentration. The gender difference between R2s for Social Competence, WJ–R Letter–Word Recognition, and WJ–R Calculations were nonsignificant.
Overall R2 of the Model (Including All Risk Factors) Tested Separately for Boys and Girls and z Test for the Difference Between Independent R2s
It was recognized that one explanation for differences in variance accounted for boys and girls was that there might have been less variance for girls in some of the outcome variables. Thus, tests for mean differences and homogeneity of variance were conducted for each outcome variable. Although there were significant mean differences on all the outcome variables (except achievement scores), with boys demonstrating higher levels of problems, there was a significant difference in the variance for boys and girls only for teacher-rated Authority Acceptance. Thus, the difference in variance did not account for differences in the relations between the risk factors and outcomes for boys and girls.
To further explore the direct and mediated effects of the broad demographic and neighborhood environment variables, nested path models were tested using PROC CALIS in SAS. In each set, a path model was specified that included the direct effects of education, occupation, and neighborhood risk but specified no relations (i.e., paths were set to zero) among these predictors and mediators or mediators and outcomes. This model was compared with the equivalent model, with the mediating effects estimated (i.e., paths from the predictors to the mediators and from the mediators to the outcomes freed to be estimated). Each set of models tested the mediating effects of one of three sets of mediators: specific demographics, family context, or maternal depression. The effects of the full models were compared with the direct-effects only models that examined the roles of education, occupation, and neighborhood risk. Using the criteria recommended by Baron and Kenny (1986)
for investigating mediation, variables were retained is the model if they met the following criteria: Predictors were significantly related to both the mediators and outcomes, and mediators were significantly related to the outcomes. A subset of the study outcome variables was selected a priori on the basis of theoretical interest and to limit the number of tests conducted. Thus, parent-report externalizing problems, teacher-report Authority Acceptance and Social Competence, and the WJ–R Letter–Word Reading Achievement scale were used.
Specific demographic mediators
In the direct-effects model, education risk (higher score = lower education) was significantly related to lower teacher-report Social Competence (β = .13, z = 2.05, p < .05) and reading achievement (β = −.28, z = 4.60, p < .001). Neighborhood risk was related to higher levels of parent-report externalizing (β = .16, z = 2.82, p < .01) and teacher-report Authority Acceptance (β = .29, z = 5.16, p < .001), lower levels of teacher-report Social Competence (β = .24, z = 4.32, p < .001), and reading achievement (β = −.23, z = 4.33, p < .001). As occupation risk was not significantly related to any of the outcome variables, it was dropped from subsequent models. The direct-effects-only model demonstrated a poor fit to the data, χ2(29, N = 332) = 242.93, p < .001, Bentler's Comparative Fit Index (CFI) = .76.
In the mediation model, the direct effects of education and neighborhood risk, as well as their indirect effects through single-parent status, number of children in the family, and mother's age at target child's birth on outcomes were tested. Neighborhood risk significantly predicted single-parent status (β = .37, z = 6.53, p < .001), and single-parent status predicted both teacher-report Authority Acceptance (β = .11, z = 2.07, p < .05) and lower reading achievement (β = −.19, z = 3.68, p < .001). Neither number of children in the family nor mother's age at target child's birth was significantly related to the outcome variables. The mediational model demonstrated an adequate fit, χ2(8, N = 332) = 59.93, p < .001, CFI = .94. In addition, the chi-square difference was significant, indicating that the estimation of the mediating paths significantly improved the fit of the model, χ2 difference (21) = 183.00, p < .001. The direct effects of education and neighborhood risk on the outcome variables remained largely the same in this model as compared with the direct-effects-only model. The magnitude of the paths was similar with all effects within .02 of those in the direct-effects-only model, with the exception of the path from neighborhood risk to reading achievement, which decreased in magnitude from −.23 (direct effects) to −.14 (z = 2.48, p < .05) in the mediated model. Thus, the effect of neighborhood risk on reading achievement was partially mediated through the effect of single-parent status.
Family context mediators
Once again, the direct-effects-only model demonstrated a poor fit to the data, χ2(50, N = 332) = 332.49, p < .001, CFI = .71, In the mediational model, the direct effects of education and neighborhood risk, as well as their indirect effects through life stress, family expressiveness, social support, home environment, and marital distress on outcomes were tested The mediational model demonstrated an adequate fit, χ2(15, N = 332) = 85.80, p < .001, CFI = .93. In addition, the chi-square difference was significant, indicating that the estimation of the mediating paths significantly improved the fit of the model, χ2 difference(35) = 246.69, p < .001. The mediators of social support (β = .24, z = 4.15, p < .001) and home environment (β = .51, z = 10.83, p < .001) were significantly predicted by neighborhood risk.
Parent-report externalizing was related significantly to social support (β = .10, z = 2.04, p < .05) and neighborhood risk (β = .14, z = 2.20, p < .05). However, the effect of neighborhood risk on parent-report externalizing was not notably reduced in this model when compared with the effect in the direct-effects-only model. Teacher-report Authority Acceptance was predicted by home environment (β = .15, z = 2.46, p < .01) and neighborhood risk (β = .19, z = 2.87,p < .01). The effect of neighborhood risk, although remaining significant, decreased .10 in magnitude when compared with the effect in the direct-effects-only model, suggesting that the effect of neighborhood was partially mediated through home environment Teacher-report Social Competence was significantly related to home environment (β = .17, z = 2.70, p < .01), education risk (β = .14, z = 2.22, p < .05), and neighborhood risk (β = .15, z = 2.28, p < .05). The direct effect of education risk remained essentially unchanged when compared with the effect in the direct-effects-only model, whereas the effect of neighborhood risk, although remaining significant, decreased .09 in magnitude, suggesting that the effect of neighborhood risk on social competence was partially mediated through home environment.
Reading achievement was significantly related to home environment (β = −.13, z = 2.20, p <. 05), education risk (β = −.30, z = 4.88, p < .001), and neighborhood risk (β = −.15, z = 2.34, p < .01). The direct effect of education risk remained essentially unchanged when compared with the effect in the direct-effects-only model, whereas the effect of neighborhood risk, although remaining significant, decreased .08 in magnitude, suggesting that the effect of neighborhood risk on reading achievement was partially mediated through home environment.
Maternal depression as a mediator
The direct-effects-only model demonstrated a moderate fit to the data, χ2(12, N = 332) = 132.01, p < .001, CFI = .85, whereas the mediational model, including maternal depression, demonstrated an adequate fit, χ2(5, N = 332) = 4l.59, p < .001, CFI = .95. The chi-square difference was significant, indicating that the estimation of the mediating paths significantly improved the fit of the model, χ2 difference (7) = 90.42, p < .001. Maternal depression was significantly predicted by both education risk (β = .16, z = 258, p < .01) and neighborhood risk (β = .27, z = 4.93, p < .001), and maternal depression was significantly related to higher levels of parent-report externalizing (β = .29, z = 5.43, p < .001), teacher-report problems with social competence (β = .11, z = 2.11;p < 0.5), and lower reading achievement (β = −.13, z = 2.59, p < .01). Although the effect of education on social competence became nonsignificant, the direct effects of education remained largely the same, with magnitudes of effects within .02 of those in the direct-effects model. The direct effects of neighborhood risk on the outcome variables remained largely the same in this model as compared with those in the direct-effects model, with one exception. The effect of neighborhood risk on parent-report externalizing was reduced from .16 to .08 in magnitude and became nonsignificant, suggesting that the effect of neighborhood risk on parent-report externalizing was mediated through maternal depression.