Introduction

Educational attainment has been linked with many neighborhood, family, and individual characteristics. Children living in low socioeconomic neighborhoods tend to have significantly lower levels of educational attainment.[1] Family variables, such as structure changes and mobility, also influence educational attainment. Children who experience family disruptions often have lower educational attainment due to the fewer resources and less time to monitor school work associated with single parent families.[2,3] Residential mobility, particularly if the child is in school, can also negatively affect educational attainment.[4] Children in poor families often move more than children in rich families, and are more likely to attend schools with high student mobility rates.[2] Childhood health can significantly affect educational attainment; health conditions can affect development, and children with health conditions may have to miss school more frequently. The most common reason for seeking medical treatment among children is injuries.[5] Mental health conditions can significantly reduce educational attainment, with externalizing disorders, such as Attention-Deficit/Hyperactivity Disorder (ADHD) and conduct disorders having the largest effect.[6]

As previous research indicates, inequalities in education and health that emerge in late adolescence and adulthood are often rooted in childhood or even in the fetal period.[7] Using a developmental perspective to examine such inequalities assesses important processes across the life course by focusing on the effects of physical or social exposures at several points in time.[8,9] Accounting for these processes depends on very large numbers of cases with many variables measured over substantial time periods. Such observational research has been restricted by the limited availability of datasets adequate to characterize life course trajectories and distinguish among various possible models.[9] Existing studies generally focus on the temporal pattern of influence of a single time-varying predictor.[10–12] Such work provides insight into the pathway of influence for that measure but does not address how pathways may be affected by other important variables. The linkable administrative databases being created across the developed world are uniquely valuable in assembling a wide range of temporal variables starting at birth and continuing well into adulthood.[13,14] At the same time, disadvantages associated with using these data need to be acknowledged. Examining several time-varying predictors simultaneously facilitates a more complete understanding of the dynamic effects of social and biological risk factors on later outcomes.

We used administrative data to generate birth cohorts and to study pathways of social and health inequality in childhood and early adolescence leading to differences in educational achievement.[15] Determining those periods in which specific events have more prominent effects on later life outcomes is crucial for life course epidemiology.[16] Researchers have proposed several conceptual frameworks for the different processes involved in the first two decades of life.[9] The ‘accumulation of risk’ model refers to the adverse effects on later health of exposures accumulating over the life course.[9] Sensitive period models postulate that exposures exert their greatest effects during particular times but are modifiable by other influences. Critical period models refer to “exposures acting during a critical window of development” affecting later development.[9] Work on the fetal origins of adult disease provides an important example of such a “critical period” approach.[17]

Operationalizing and testing different life course models requires data which are seldom available. This research examines the patterns by which five time-varying predictors—three social (Living in a Low Income Neighborhood, Residential Mobility, Family Structure Changes) and two health (Externalizing Mental Conditions, Injuries)—affect the odds of high school graduation. Given the importance of high school graduation on social and economic opportunities, gaining a better understanding of how and when specific events in childhood and adolescence affect these outcomes is necessary.[16] The following questions are addressed: Which measures best predict high school graduation? Do significant time-varying covariates operate in a cumulative fashion, in sensitive periods, or in critical periods? Do the temporal patterns selected for each time-varying covariate differ when adjusting for measured and unmeasured confounders?

Materials and methods

Results

Almost 25 percent of the individuals in the cohort did not graduate from high school (24.10%). Table 1 provides descriptive statistics of all variables included in the model; differences between those graduating from high school and those not graduating are identified by χ²(for binary variables) and t (for continuous variables) statistics. Individuals not graduating from high school differed significantly from those who did on almost all covariates. S3 and S4 Tables presents the correlations among the time-invariant and time-varying measures.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics (N = 89,763).

https://doi.org/10.1371/journal.pone.0188976.t001

Model including all variables

Table 2 highlights the temporal pattern of events associated with failure to graduate from high school and the time-invariant predictors included for adjustment. Being born to an unmarried mother was associated with failure to graduate. Living in a low-income neighborhood and residential mobility at any point are associated with higher odds of failure to graduate. These measures were particularly important in the period prior to the outcome. Externalizing mental health conditions impact high school graduation most when the diagnoses occurred in the age group prior to the outcome. Individuals with more frequent injuries had significantly higher odds of failing to graduate high school. The variable most strongly increasing the odds of failing to graduate from high school was having less than average grade 9 achievement, followed by living in a low-income neighborhood in all age groups and by residential mobility in all age groups.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Odds ratios for failure to graduate from high school with time-invariant covariates.

https://doi.org/10.1371/journal.pone.0188976.t002

Additional models

We have controlled for a series of time-invariant background measures in modelling the temporal pattern which best describes the influence of each time-varying predictor. To determine whether model selection differed without these background variables, several additional strategies were examined: a) a model controlling only for covariates at birth (this excludes the grade 9 achievement variable) (selected model identified by ○ in Table 3 and in Table 4; c-statistic = 0.790); b) a model that only included the time-varying predictors, not controlling for any of the background variables (selected model identified by ◆ in Table 3 and in Table 5; c-statistic = 0.734).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Difference in temporal pattern selected when adjusting for different confounders.

https://doi.org/10.1371/journal.pone.0188976.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Odds ratios for failure to graduate from high school with time-invariant covariates (excluding grade 9 achievement).

https://doi.org/10.1371/journal.pone.0188976.t004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Odds ratios for failure to graduate from high school without time-invariant covariates.

https://doi.org/10.1371/journal.pone.0188976.t005

Without including Grade 9 achievement, the externalizing mental health conditions—particularly those diagnosed in the 9–13 age period—show dramatically higher odds ratios (Table 4). Compared with Table 2, somewhat higher odds ratios are seen for those living in a low income neighborhood and experiencing residential mobility. Table 5 demonstrates the effects of removing all the time invariant measures. The importance of income is highlighted. The role of the externalizing mental health conditions changes, with the accumulation of risk model providing a better fit.

Omitted variable bias is a common concern in multi-variate analysis.[5,18,37] Sibling analyses address this by capitalizing on common characteristics within the family; our fourth analysis included all time-invariant covariates in a sample incorporating children having at least one sibling in the cohort (n = 102, 869). A multilevel logistic regression model accounted for unmeasured shared family-level characteristics (selected model identified by ∎ in Table 3 and in S5 Table; c-statistic = 0.866). Results were generally similar to those in the other models, indicating that omitted variable bias was not a major influence on the results.

Table 3 highlights consistency (social variables)–and some differences (health variables)–among the strategies; excluding time invariant covariates significantly decreased model fit. Temporal patterns associated with externalizing mental health conditions became more important, but the confidence intervals of many of the odds ratios are wide.

Discussion

The temporal patterns by which time-varying predictors affected high school graduation differed among predictors. Tables 2–5 have shown that predictors including the event in early adolescence generated the highest odds ratios. Others have suggested that baseline and early childhood measures tend to become less predictive of late adolescence and adult outcomes.[37] This paper also highlights the transmission of risk across developmental periods; exposure to adversity in one developmental period had significant implications for adversity in subsequent life stages.[38]

These Canadian results differ somewhat from American findings based on the long-running Panel Study of Income Dynamics (PSID). Although some PSID analyses regarding education were ambiguous, “family economic conditions in early childhood appear to matter more for shaping later development than economic conditions during adolescence.”[39] Methodological differences between approaches might contribute to differences in results. Thus, Duncan et al. used a much smaller sample (effective N = 1,254), incorporating various time invariant measures (including race) and a single time-varying measure, childhood income (the average annual household income in three periods—prenatal to age 5, ages 6–10, and ages 11–15).[39]

Additionally, Canadian-American differences in child development may build on differences in inequality between Canada and the United States. Inequality in the United States is relatively high compared with other OECD countries; Canadian provinces and metropolitan areas both have lower income inequality than their American counterparts and stronger social programs directed toward this inequality.[40,41]

Previous life-course studies using similar methods have only rarely included time-invariant confounders.[42–46] This raises problems. First, while excluding background variables facilitates distinguishing among model fit for different temporal patterns, failing to include important covariates is likely to overestimate the association between the exposures and outcome.[47] Second, including background variables may change the temporal pattern determined to be the best fit for the time-varying predictor. As seen in Tables 2–5, the models selected differ somewhat when controlling for background variables. The very good fit of our models suggests that generally ‘correct’ temporal patterns are identified. Given our analyses of omitted variable bias, unavailable measures (such as parental education and race) seem unlikely to explain a great deal more variance, although ‘aboriginal status’ would probably be a significant time-invariant measure.

Which confounders should be controlled for in life-stage modelling? Our models include several measures of circumstance at birth and one at early adolescence (the Grade 9 Achievement Index). The Grade 9 Achievement Index is by far the strongest predictor of high school graduation and probably “soaks up” at least part of the impact of measures from earlier ages. Tables 2–5 and S5 Table show that strong associations between age category and educational attainment remain with and without this index.

In analyses excluding grade 9 achievement, externalizing mental health conditions expressed in ages 9–13 were very important predictors. Income in the neighborhood of residence consistently showed high correlations with such income in an earlier or later period. Despite these relationships, the odds ratios associated with different income/neighborhood combinations varied substantially. Residence in such low-income neighborhoods over all three time periods showed strong, regular relationships. In other circumstances, the temporal patterns selected for the time-varying covariates might vary, given the complexities of life course epidemiology.[48]

Administrative data provide major advantages, both generally and specifically for life course research. The very large number of cases allows examining variables in considerable detail across age groups. In contrast, cohort studies (such as PSID) using a much smaller N typically rely on high cost surveys. This smaller N results in lower precision when such techniques as sibling difference models (capitalizing on within family variations) are used to deal with such bias. This is important; adjustments for omitted variable bias sometimes produce large reductions in the estimated effects of income. (37) Surveys can be unreliable, may suffer from higher nonresponse rates and report events “as occurring more recently than they did in fact occur”.[37] Administrative data both allow more accurate timing of events (such as diagnosis and family structure changes) essential for assessing risk in different age groups and facilitate the inclusion of important confounders.[9]

Administrative data clearly have limitations. Research using such data often has to be opportunistic in relying on (or not relying on) certain variables. Checks on unavailable variables can help; pilot studies, different statistical approaches, and comparisons with other work are often feasible.[21] Specific study limitations include the use of binary exposure and outcome variables for simplification (but the models are already complicated). Symptoms may occur well before an official diagnosis and family instability often precedes a formal change in family structure (such as divorce or remarriage). We defined externalizing mental health diagnoses as at least two ADHD and/or CD/ODD diagnoses in a four-year time period. Defining this variable differently may change its “effect” on high school graduation. However, linkable data bases allow ‘expansion’ through the creation of new variables. Over time, additional measures (such as child maltreatment, parental welfare use, prescription use, and maternal education) will facilitate long-term cohort research and might conceivably change model selection. Analyses of temporal patterns can be useful in studying such other outcomes as educational achievement, teen pregnancy, and workforce participation.

This research presents an important way to build on the strengths of large population-based data. Complicated time-dependent models are dependent on such data sets. The alternative seems to rely exclusively on simpler models based on having more information per person, but many fewer cases. Temporal patterns have significant implications on the timing of the best intervention. An exposure important at a specific period of time (critical period) suggests an early intervention.[7] Exposures operating in sensitive periods warrant intervention at “strategic junctures in the life course”, whereas exposures accumulating over time require interventions “consistently over time, building on models of social change”.[7] Because exposures operate through different processes, the timing of an intervention for a specific exposure may vary.

The relationships among several measures of social behavior, cognitive development, and health have been explored.[9] Life course models that best fit each time-varying predictor have been identified for one geographic area; replication elsewhere is called for. A stronger focus on adolescence as a “second sensitive developmental period” appears warranted. The rapid brain maturation associated with early adolescence can significantly modify childhood trajectories. The powerful and consistent role of income has been established. Efforts to support families remaining in the same homes or in the same schools (despite residential mobility) seem appropriate.

Supporting information

Acknowledgments

Data used in this study are from the Population Health Research Data Repository housed at the Manitoba Centre for Health Policy, University of Manitoba and were derived from data provided by Manitoba Health, Seniors, and Active Living and by Manitoba Education and Training under project #2013/2014-04. The results and conclusions are those of the authors and no official endorsement by the Manitoba Centre for Health Policy, Manitoba Health, Seniors and Active Living or other data providers is intended or should be inferred. All data management, programming and analyses were performed using SAS^® version 9.4.