Mobility Effects and Cumulative Mobility Contexts

Abstract

As the last chapter demonstrated, findings on mobility effects have been mixed, largely because of differences in outcome domains, measurements, and the type and frequency of household mobility under study. Researchers who find evidence in support of mobility effects have proposed several explanations for why. This chapter discusses some of the principal mechanisms used to explain mobility effects. A proposed framework highlights the importance of preexisting resources and risk factors, the move context, and the accumulation of context-related stressors. To provide preliminary support for the cumulative context framework, several components of the model are explored using data from the NLSY79 linked mother–child files.

Appendix: National Longitudinal Survey of Youth 1979

Sample

The National Longitudinal Survey of Youth (NLSY79) is a longitudinal study of a representative sample of American men and women aged 14 to 21 in 1979. The children of the female NLSY79 respondents were surveyed biennially starting in 1986, and these NLSY79 Child and Young Adult data files can be linked with the original NLSY79 files to assess intergenerational phenomena and outcomes. The analyses in Chap. 6 draw on data from the 2000, 2002, 2004, and 2006 survey waves.

Dependent Variables

Academic Achievement

Academic achievement was measured using the NLSY79 Child and Youth respondents’ 2000, 2002, 2004, and 2006 Peabody Individual Achievement Test (PIAT) scores. The PIAT is a widely used measure of academic achievement for children (Dunn and Markwardt 1970). Since 1986, the children in this study have been assessed biennially between ages 5 and 15. Each assessment begins with five age-appropriate questions and progresses to more advanced concepts. The reading recognition test measures word recognition and pronunciation ability, and the math test measures basic math skills and concepts. The score is the mean of the child’s age-standardized percentile scores on subsets in mathematics, reading comprehension, and reading recognition.

Behavior Problems

The behavioral problems measure is based on Peterson and Zill’s (1986) Behavior Problems Index. This index consists of twenty-eight questions, drawn primarily from the widely used Child Behavior Checklist (Achenbach and Edelbrock 1981) along with other widely used child behavior scales. The respondent’s mother indicates whether a statement about the child’s behavior is “often true,” “sometimes true,” or “never true.” The composite score is a measure with higher numbers indicating more behavior problems.

Independent Variables

Residential (Local) and Geographic (Distance) Mobility

The act of household mobility is captured by two dummy variables indicating whether a respondent (1) relocated within the same city or (2) relocated to a new city, county, or state between each survey wave. The omitted category for comparison is not moving.

Individual and Household Characteristics

Individual and household characteristics include time-variant variables, such as annual household income (logged) and age. A dummy variable for parent marital status indicated whether or not a respondent’s mother was never married at each survey wave. Time-invariant variables include the child’s sex, birth order, mother’s age at child’s birth, mother’s highest year of education completed (measured once in 2000), family structure (father in household or not), and the number of children in the respondent’s household. Children were assigned to racial groups based on the primary racial identification of their mothers as Black; Hispanic; or non-Black/non-Hispanic. All other variables in the analysis vary across survey waves. Urban residence was measured as whether or not the respondent lived in an urban or suburban (1) versus a suburban area (0).

Social Capital

Parent–community social capital was measured using two variables: how many of the child’s friends the parent knows by sight and name, coded as none of them (0), only a few (1), about half (2), most of them (3), or all of them (4) (Teachman et al. 1996) and, following Coleman (1988), a dichotomous variable marking whether or not a child attends Catholic school.

Child–community social capital was assessed by whether or not the child participates in extracurricular activities (White and Gager 2007); how often he or she attends religious services coded as (0) never, (1) a few times a year, (2) about once a month, and (3) about once a week (Parcel and Dufur 2001); and how often the child feels lonely or wishes he or she has more friends as measured as never or hardly ever (1), sometimes (2), or often (3).

Intergenerational social capital was measured by the level of closeness the respondent feels to his or her mother, reported as being not very close (1), fairly close (2), quite close (3), or extremely close (4) (Pryor 1999).

Analysis Notes

Because the purpose of this analysis was to assess the effects of moving on changes in educational achievement and behavior problems, longitudinal data were necessary in order to include measures of the predictors and outcomes in a person-year format. This allows for consideration of social capital for each survey wave in the analysis. Adequately controlling for past behaviors before a move occurs is crucial; otherwise, associated changes in child outcomes after moving cannot be determined confidently. The sample consists of children who completed the PIAT and BPI for the 1998 (baseline), 2000, 2002, 2004, and 2006 survey rounds. The PIAT is administered starting at age 5, and the behavior problems assessment begins at age 4; neither examination is recorded after age 15. Linear mixed modeling (LMM) was used to examine the effects household mobility and social capital have on child educational achievement and behavior problems. Models were run separately for each of these two child outcomes. Descriptive statistics for all measures in 2000 are presented in Table 6.3.

Table 6.3 Descriptive statistics for timing of mobility and social capital

LMM is a flexible and powerful method for the analysis of longitudinal data. In LMM, independent observations are not assumed, meaning that between-subject and within-subject effects are both considered. This modeling structure is also flexible in its use of missing information. Other models use listwise deletion of cases if a complete trajectory is not available for an individual. LMM, on the other hand, accounts for all respondents in the data set and is, therefore, arguably a better model for unbalanced panel data sets like the NLSY where not every respondent is observed in every year. Lastly, LMM allows for the analysis of hierarchically organized data. In this study, four models were tested on three levels. The first of these levels consisted of households, the second was the individual child nested within each household, and the last level, survey wave or “time,” was measured by interview round and nested within each child.

This study applied an upward two-step preliminary modeling procedure employed by Singer and Willett (2003): (a) an unconditional means model and (b) an unconditional growth model. First, the unconditional means model is the preliminary verification for whether this is an appropriate analysis by partitioning the total variation in the outcome variable (BPI and PIAT). The intra-class correlation coefficient (ICC) measures the proportion of variance in the outcome variable that is due to between-children differences rather than differences within children over time. The unconditional growth model was run to (a) assess the effects of aging on child achievement in academics and behavior problems and (b) detect whether there was significant variance to be explained from household-level characteristics. In the models, a considerable decrease in information criterion fit statistics indicates that the behavior and achievement final models are a significantly better fit than the individual- and household-level models. Level-1 and level-2 random effects remain significant in each model, meaning that additional level 1 and 2 predictors may improve model fit. Controls and interactions are included to explore the moderating effect of social capital on the effects of moving.

The Hausman specification test validated the models. LMM assumes that the dependent variable be conditionally normal. Shapiro-Wilkins testing and examination of skewness and kurtosis indicated that both dependent variables were distributed reasonably normally. Variance inflation factors indicated that multicollinearity was not a problem in the models (average VIFs < 1.31).

Limitations and Future Directions

These analyses are subject to several caveats. Measuring child outcomes across only four waves of this longitudinal survey does not allow for analysis of behavior and achievement effects that take longer than 2, 4, or 6 years to develop. In fact, the findings presented in Chap. 5 (Table 5.3) provide evidence that the negative effects of childhood mobility can last into young adulthood. Additionally, despite restricting the data to try to account for some selectivity, as with most empirical studies of correlational data, there is a possibility that unobserved parent and household characteristics account for the geographic mobility. Another limitation is that reverse causation may be present in the models above. For instance, problem behavior may cause children to have distant relationships with their parents, or families may be more or less likely to move because of their child’s preexisting behavior problems or school achievement.

The sample also has limitations. In 1986, when the children of these 21- to 28-year-old mothers were first assessed, the oldest children had been born to very young women. As a result, the sample may exclude some children born to younger women, because they had already left the sample before the 2000 wave. Further, because only the children of NLSY female respondents are surveyed, father–child and father–community interaction (other than what is reported by the mother) cannot be assessed as a component of social capital. Relatedly, because of the NLSY79 child data design, children raised in single-father homes are not included.