Spillover effects of gestational age on sibling’s literacy

Abstract

Adverse health events within families can harm children’s development, including their early literacy. Using data from a longitudinal Wisconsin birth cohort, we estimated the spillover effect of younger siblings’ gestational ages on older siblings’ kindergarten-level literacy. We sampled 20,014 sibling pairs born during 2007–2010 who took Phonological Awareness Literacy Screening-Kindergarten tests during 2012–2016. Exposures were gestational age (completed weeks), preterm birth (gestational age <37 weeks), and very preterm birth (gestational age <32 weeks). We used gain-score regression—a fixed effects strategy—to estimate spillover effect. A one-week increase in younger siblings’ gestational age improved the older siblings’ test score by 0.011 SD (95% confidence interval: 0.001, 0.021 SD). The estimated spillover effect was larger among siblings whose mothers reported having a high school diploma/equivalent only (0.024 SD; 95% CI: 0.004, 0.044 SD). The finding underscores the networked effects of one individual’s early-life health shocks on their family members.

INTRODUCTION

In recent decades, research in epidemiology and economics has drawn attention to health-related spillover effects within families (Ben-Shlomo et al., 2016; De Neve & Kawachi, 2017; Dirks et al., 2015; Kuh et al., 2003; Lawlor & Mishra, 2009; Liu et al, 2010; McHale et al., 2012), including the harm of children’s adverse health shocks on siblings’ educational outcomes. This literature examined how younger siblings’ health events or conditions—at birth or in early childhood—disrupt or impair development for older siblings. Recent evidence suggests that a child’s disability harms their older sibling’s cognitive development and educational outcomes in childhood and adolescence (Black et al., 2021; Breining et al., 2014; Fletcher et al., 2012). Similarly, early-life medical interventions for infants may improve their older siblings’ educational outcomes (Daysal et al., 2022). This scholarship is motivated by a resource reallocation hypothesis: parents shift investments to affected children to compensate the health shock’s deleterious effects (Behrman et al., 1994; Yi et al, 2015). Investments may be material (e.g., educational goods), financial (e.g., early interventions), or interpersonal (e.g., informal teaching) (Bharadwa et al., 2018; Cunha et al., 2010; Heflin, 2016;). While resource reallocation benefits the affected child, resource reallocation may come at the expense of their siblings’ educational outcomes.

Presently, it is uncertain whether adverse birth outcomes exert harmful spillovers onto siblings’ educational outcomes. Further, previously used sibling spillover estimation methods—cousin fixed effects (FE) (Breining, 2014; Fletcher et al., 2012), difference-in-differences with sibling triads (Black et al., 2021), and regression discontinuity designs (Daysal et al., 2022)—are either limited in their ability to control for family-level unobserved confounding or by the complexity of data required for analysis.

We investigated the effect of younger siblings’ gestational age on older siblings’ literacy skill at entry to kindergarten using a population-based birth cohort from Wisconsin, United States. Consistent with prior research (Black et al., 2021; Breining et al., 2014; Daysal et al., 2022; Fletcher et al., 2012), we focus on younger-to-older sibling spillover effects to understand how a child’s sudden health shock—in this case, shorter gestation at birth or preterm birth—affects the development of their siblings. This analysis advances knowledge on two fronts. First, this is the first study that analysed spillover effects of gestational age on siblings’ educational outcomes. Despite ample evidence that shorter gestation harms individual health and development (Flood & Malone, 2012; Romero et al., 2014), including educational outcomes (Kirkegaard et al., 2006; Lipkind et al., 2012; Mallinson et al., 2019; Mathiasen et al., 2010; Richards et al., 2016), it is unknown whether shorter gestation harms siblings’ academic outcomes. Second, we applied a robust and flexible sibling FE method to estimate spillovers using gain-score estimation to control for unobserved family-level confounders (Mallinson & Elwert, 2022). Our results indicated that a child’s shorter gestation harmed their older sibling’s readiness for kindergarten-level literacy instruction, highlighting the relevance of early-life health shocks and their spillover effects on other children’s development.

METHODS

Sibling spillover estimation

Controlling for unobserved confounding is a central challenge in sibling spillover analyses. Shared family history and endowments that affect child health and development are often unmeasured, even in data-rich contexts (De Neve and Kawachi, 2017). The developing literature on health shocks and siblings’ educational outcomes employed various methods to estimate spillovers and allay confounding bias.

Two early studies executed cousin FE linear regressions, which control for sibling-invariant confounding at the extended family-level, such as genetics from shared grandparents. Analysing different cohorts with in-sample cousins, Fletcher et al. (2012) found that children scored lower on a cognitive ability screener if their sibling had a developmental disability, and Breining (2014) concluded that students earned lower 9^th grade exit exam scores if their younger sibling had Attention-Deficit/Hyperactivity-Disorder (ADHD). However, cousin FE methods cannot account for confounding at the nuclear family-level (e.g., by parental genetic history or parental educational aspirations).

To control for unobserved family-level and sibling-invariant confounding, Black et al. (2021) developed a difference-in-difference approach with sibling triads: regressing a child’s outcome on a binary sibling indicator (first-born vs. second-born) and on an interaction between the sibling indicator and a binary indicator for a third-born’s health shock. Under maintained assumptions, the interaction term’s partial regression coefficient is shown to underestimate the spillover of the third-born’s health shock on the second-born’s outcome. With population-based data from Denmark and Florida, United States, the authors found that a third-born’s disability decreased the second-born’s grades. Their method is novel, as conventional sibling FE techniques collapse in the presence of such spillovers (Sjölander et al., 2016). The method, however, necessitates sibling triads—which may be unavailable with smaller cohorts—and it only estimates bounds for spillovers.

Veering from FE-based methods, Daysal et al. (2022) used a regression discontinuity design to estimate the effect of a medical intervention for very low birth weight (<1,500 grams) younger siblings on older siblings’ test scores. Their approach assumed that children born around the birth weight cutoff were exchangeable except for the intervention—that is, children born at 1,499 grams are indistinguishable from those born at 1,501 grams, but only the former received the intervention. This design can estimate spillover effects when a categorical treatment assignment depends on an eligibility threshold, for individuals that are near the threshold. While the method excels at confounder control in such contexts, it is not applicable in more general settings where interest is in spillovers from continuous treatments (e.g., birth weight itself).

Most recently, Mallinson & Elwert (2022) proposed a novel two-sibling FE approach for estimating spillover effects: gain-score regression. Their strategy involves regressing a gain-score (i.e., “change score”)—the difference between siblings’ continuous outcomes—on siblings’ exposures (continuous or binary) and then summing the partial regression coefficients for both exposures. In the presence of one-sided spillover (i.e., only one sibling’s exposure affects the other sibling’s outcome), this sum identifies the spillover effect precisely. In the presence of two-sided spillover (i.e., both siblings’ exposures affect the other’s outcome), this sum identifies a lower bound of one of these spillover effects.

We used the gain-score strategy (Mallinson & Elwert, 2022) for several reasons: it controls for unobserved sibling-invariant confounding at the nuclear family-level; it requires only two siblings per family; and it is compatible with binary or continuous exposures. We explain gain-score estimation in the “Statistical models” subsection, and we argue that our approach estimates a lower-bound spillover effect of the younger sibling’s gestational age on their older sibling’s PALS-K score.

Sample

We analysed data from Big Data for Little Kids (BD4LK), a cohort of birth records for all live in-state resident deliveries in Wisconsin during 2007–2016 (N>660,000). BD4LK’s birth records link to multiple administrative data sources with unique maternal and child identifiers, including paid Medicaid claims (2007–2020) and Phonological Awareness Literacy Screening-Kindergarten (PALS-K) test scores from Wisconsin public schools (2012–2016 school years). BD4LK’s deterministic linkage was robust as described elsewhere (Larson et al., 2019).

We sampled sibling pairs born sequentially to the same mother between January 1, 2007-September 1, 2010, in which both siblings took the English-language PALS-K test. We restricted sampling on birthdate because children had to be at least 5 years old by September 1, 2015, for PALS-K testing eligibility (Wisconsin Department of Public Instruction, n.d.-a). Additionally, our identification strategy (described in the “Statistical models” subsection) required both siblings’ test scores. PALS-K was offered in English and in Spanish, but test versions have different scoring structures (Ford & Invernizzi, 2014; Invernizzi et al., 2015). Although siblings were from different deliveries, they could be plural-born (e.g., a sibling could have a twin). Among plural-born clusters, we randomly sampled one child.

BD4LK includes 252,883 unique deliveries during January 1, 2007-September 1, 2010. We identified 806 records (0.32%) with multiple maternal or child identifiers that imperfectly matched across data sources. We randomly selected one match and excluded duplicates. We linked 177,863 records (70.33%) to PALS-K scores, of which 46,745 records were in-sample siblings (26.28% of test-linked records). We then pulled eligible sibling pairs: sequentially-born from different deliveries, took the English-language PALS-K test, and complete information on relevant variables, which we list in the next subsection. Our final sample included 40,028 children (85.63% of tested siblings), or 20,014 sibling pairs.

PALS-K

The PALS-K test evaluates students’ readiness for kindergarten-level literacy instruction (Invernizzi et al., 2004; Invernizzi et al., 2015), and Wisconsin public schools administered PALS-K to identify students who were at risk of developing reading difficulties (Wisconsin Department of Public Instruction, n.d.-b). The Wisconsin Department of Public Instruction (DPI) required Wisconsin public schools to administer PALS-K tests to 5-year-old kindergarteners from the 2012–13 school year through the 2015–16 school year, and then subsequently allowed school districts to use different screeners for literacy assessments (Wisconsin Department of Public Instruction, n.d.-b). In June 2023, PALS assessments (including PALS-K) and the PALS Online System officially retired (Renaissance, n.d.).

PALS-K comprised six domains of early literacy skills: rhyme awareness, beginning sound awareness, alphabet knowledge, letter sounds, spelling, and concept of word (Invernizzi et al., 2004; Invernizzi et al., 2015). The sum of scores across domains generated a composite score with a range of 0–102 points. Scoring below the testing benchmark (<28 points) signalled an elevated risk of reading difficulty. PALS-K also had a seventh optional domain, “word recognition in isolation,” but this domain was optional and was therefore not included in the composite score. The DPI required schools to report test scores to students’ parents or guardians. While the DPI did not mandate specific interventions for students who scored below the benchmark, the Wisconsin Statute required interventions to be scientifically based and to address the student’s deficiencies and needs (Wisconsin Department of Public Instruction, n.d.-c.; Wisconsin State Legislature, 2012).

Variables

Our outcome was the school year-standardized PALS-K score (mean 0, standard deviation [SD] 1). We standardized within school years to account for annual scoring variations. Birth records provided clinical estimates of gestational age (pregnancy duration) based on the first accurate ultrasound and last date of menstruation. Our three exposures were gestational age (completed weeks), preterm birth (gestational age <37 completed weeks), and very preterm birth (gestational age <32 completed weeks). Continuous gestational age facilitated spillover estimation on a finer gradient, whereas dichotomous exposures permitted estimation at commonly used thresholds.

Maternal characteristics included age (years), education (no high school diploma; high school diploma/equivalent; 1–3 years college; 4+ years college), race/ethnicity (Asian/Pacific Islander non-Hispanic [NH]; Black NH; Hispanic; Native American/Alaska Native NH; White NH; other NH), nativity (United States-born; foreign-born), and Medicaid delivery coverage (Medicaid; non-Medicaid). Medicaid is government-funded medical insurance for low-income residents, and eligibility varies by state (U.S. Centers for Medicare & Medicaid Services, n.d.). Children’s characteristics were sex (female; male), plurality (singleton; plural), birth order (first; second; third; fourth or later), sibling age difference (years), PALS-K test year, and age at PALS-K testing (years). In absence of the actual date of PALS-K testing, we assumed that schools administered tests on September 1 for each school year. We then estimated the child’s age at PALS-K testing by subtracting their birth date (indicate on the birth record) from the assumed PALS-K testing date.

Analyses

Descriptive analysis

We tabulated descriptive statistics by sibling status (older; younger). Additionally, we computed mean standardized PALS-K scores by sibling status and mean differences of standardized scores between siblings by sibling-cluster preterm birth and very preterm birth exposure status (neither; older only; younger only; both siblings). Finally, we plotted locally weighted regressions of standardized PALS-K scores against siblings’ gestational ages (older-against-younger; younger-against-older).

Statistical models

We used gain-score estimation, a FE technique, to estimate a lower-bound spillover effect of a child’s gestational age on their older sibling’s PALS-K score (Kim & Steiner, 2021a, 2021b; Mallinson & Elwert, 2022). Assuming two-sided spillover, we could not point-identify the desired spillover effect precisely but could estimate lower bounds (i.e., underestimate) this spillover. A causal directed acyclic graph presents the assumptions of our two-sided spillover model, where $i\in \left\{1,\dots ,N\right\}$ indexes families and $j\in \left\{1,2\right\}$ indexes siblings ($j=1$ is older), and we assume no spillover between families (i.e., partial interference (Sobel, 2006)) (Figure 1). $G{A}_{ij}$ denotes gestational age, $P{K}_{ij}$ denotes PALS-K score, $U_i$ is the vector of unobserved family-level confounders (e.g., environmental causes of preterm birth and neurological development (Ferguson et al., 2013; Gressens et al., 2011), or other shared health shocks), and $D_i=P{K}_{i1}-P{K}_{i2}$ is the gain-score. Greek letters denote linear and homogenous effects: $\delta$ is the targeted effect ($G{A}_{ij}\to P{K}_{ij}$); $\theta$ is the younger-to-older spillover effect ($G{A}_{i2}\to P{K}_{i1}$) and our primary target of inference; $\kappa$ is the older-to-younger spillover effect ($G{A}_{i1}\to P{K}_{i2}$); and $\chi$ ($U_i\to G{A}_{i1}$), $\gamma$ ($U_i\to G{A}_{i2}$), and $\psi$ ($U_i\to P{K}_{ij}$) are unobserved family-level confounding effects.

Figure 1.

Causal directed acyclic graph of the relationship between gestational age and Phonological Awareness Literacy Screening-Kindergarten scores within sibling pairs. Subscript $i$ denotes cluster (family) and subscript $j=\left\{1,2\right\}$ denotes sibling ($j=1$ is the older sibling). $G{A}_{ij}$ is gestational age in completed weeks, $P{K}_{ij}$ is the Phonological Awareness Literacy Screening-Kindergarten test score, $U_i$ is the vector of sibling-invariant confounders (the fixed effect), and $D_i=P{K}_{i1}-P{K}_{i2}$ is the gain-score.

Besides linearity, homogeneity, and partial interference, the model in Figure 1 encodes four causal assumptions: two-sided treatment-to-outcome spillover between siblings; equi-confounding of family-level unobservables on siblings’ outcomes; identical targeted effects for both siblings; and older siblings’ outcomes do not affect younger sibling’s exposures or outcomes. We permitted two-sided spillover, as parents may reallocate investments to a less healthy sibling, regardless of birth order. Equi-confounding and constant targeted effects are standard FE assumptions (Gunasekara et al., 2014). Study design precluded outcome-to-treatment spillover, as all observed births occurred before observed PALS-K tests. Regarding outcome-to-outcome spillover, prior research indicates that siblings’ literacy skills are correlated through their shared home environment, socioeconomic status, parental reading practices, and genetic background (Puglisi et al., 2017; Segal et al., 2018). Consequently, the association between siblings’ PALS-K scores are subsumed into the unobserved FE $U_i$ rather than the older sibling’s PALS-K score directly affecting the younger sibling’s PALS-K score. Our assumptions are widely used in conventional FE analyses to control for unobserved confounding (Gunasekara et al., 2014; Sjölander et al., 2016); nonetheless, they are strong.

Under these assumptions, regressing $P{K}_{i1}$ on $G{A}_{i2}$ cannot identify $\theta$ because of unobserved confounding, covariate adjustment notwithstanding. Gain-score regression, however, identifies the difference between siblings’ spillover effects (Mallinson & Elwert, 2022). We first regress the gain-score on siblings’ gestational ages,

$$D_i=b_{1}G{A}_{i1}+b_{2}G{A}_{i2}+e_i,$$

where $b_j$ denotes the partial regression coefficient and ${\epsilon }_i$ denotes the residual. Computing the difference score subtracts out all unobserved confounding that equally affects both siblings (or, graphically, offsets the spurious paths from $G{A}_{ij}$ via $U_i$ to $P{K}_{ij\prime }$, $j\ne j\prime$) (Kim & Steiner, 2021a, 2021b; Mallinson & Elwert, 2022). The first coefficient evaluates to $b_1=\delta -\kappa$ and recovers the difference between the targeted effect and the older-to-younger spillover; analogously, $b_2=\theta -\delta$ recovers the difference between the younger-to-older spillover and the targeted effect. Summing the coefficients thus identifies the difference between spillover effects: $b_1+b_2=\theta -\kappa$. This sum does not point-identify $\theta$. However, it provides a lower bound for $\theta$ under the added assumption that longer gestation does not harm a sibling’s PALS-K score, $\theta \ge 0$ and $\kappa \ge 0$. When applied, this method will underestimate the spillover effect of the younger sibling’s gestational age on their older sibling’s PALS-K score.

We estimate two gain-score models for each exposure (gestational age; preterm birth; very preterm birth): one without covariate adjustment, and one adjusted for select maternal characteristics at the time of the older sibling’s delivery (age; education; Medicaid delivery coverage) and for the older sibling’s plurality and birth order. We estimated spillover effects by summing the partial regression coefficients for each siblings’ exposures. Non-zero sums indicated spillover. We computed 95% confidence intervals and two-tailed t-tests. All gain-score models controlled for unobserved family-level characteristics, $U_i$. Additionally, we executed models that further controlled for sibling-varying confounders that preceded the older sibling’s birth. We did not include covariates that were measured after the older sibling’s delivery to avoid collider bias (Elwert & Winship, 2014; Greenland, 2003). Our models did not explicitly control for maternal race/ethnicity and nativity, since they were subsumed into $U_i$.

Effect heterogeneity

Effect heterogeneity analyses interrogated possible spillover mechanisms. Prior evidence suggests that the effect of parental investment is most impactful in early childhood (Cunha et al., 2010). Furthermore, socioeconomic advantage may buffer health shock-induced externalities, as the penalty of shorter gestation on individual PALS-K performance was lesser in highly educated or non-Medicaid families (Mallinson et al., 2019). Thus, we stratified adjusted gain-score models on sibling age difference (<2 years; 2+ years), Medicaid delivery coverage, and maternal education. We measured Medicaid coverage and maternal education at older siblings’ deliveries to prevent post-treatment collider bias.

Robustness checks

We repeated overall and stratified adjusted gain-score models in two sub-populations. First, we analysed singleton-born siblings only (19,525 pairs). This informed whether unobserved plural-born siblings impacted estimates. Second, we analysed siblings of native-born (United States) mothers only (18,548 pairs). Children of foreign-born parents are more likely to speak and read English as a second language relative to their peers (Sullivan et al., 2016). Whether the English-language PALS-K test accurately evaluated literacy among these students is uncertain. Moreover, it is unclear how effectively Wisconsin public schools directed students to Spanish-language tests, and PALS-K was not offered in other non-English languages. We conducted analyses in Stata statistical software, release 16 (StataCorp, 2019). The University of Wisconsin-Madison minimal risk institutional review board approved our project.

RESULTS

Average standardized and raw PALS-K scores were similar between older siblings (0.021 SD; 63.58 points) and younger siblings (−0.021 SD; 64.22 points) (Table 1). The average gestational age (~38.9 weeks), preterm incidence (~7%), and very preterm birth incidence (~1%) were also consistent by sibling status. The mean age difference in sibling pairs was 1.88 years (SD 0.59 years), and 60.46% of siblings were born <2 years apart. Approximately 57% of mothers reported at least some college education at the older sibling’s delivery, and roughly 40% of observed deliveries were Medicaid-covered.

Table 1.

Baseline sample characteristics by sibling status (N=20,014 sibling pairs)

	Older sibling	Younger sibling
PALS-K Test Outcomes
Raw score (points; range 0–102 points), mean (SD)	63.58 (24.11)	64.22 (23.84)
Standardized score, mean (SD)	0.021 (1.000)	−0.021 (0.999)
Exposures
Gestational age (completed weeks), mean (SD)	38.98 (1.76)	38.79 (1.67)
Gestational age (categorical), N (%)
Preterm birth (<37 weeks)	1,358 (6.79)	1,331 (6.65)
Very preterm birth (<32 weeks)	137 (0.68)	140 (0.70)
Maternal Characteristics
Age at delivery (years), mean (SD)	26.01 (5.12)	27.89 (5.20)
Age at delivery (categorical), N (%)
<20 years	2,263 (11.31)	809 (4.04)
20–24 years	5,627 (28.12)	4,921 (24.59)
25–29 years	7,184 (35.89)	6,629 (33.12)
30–34 years	3,806 (19.02)	5,426 (27.11)
35+ years	1,134 (5.67)	2,229 (11.14)
Education at delivery, %
No high school diploma	2,867 (14.32)	2,324 (11.61)
High school diploma	5,791 (28.93)	5,923 (29.59)
1–3 years college	5,037 (25.17)	5,372 (26.84)
4+ years college	6,319 (31.57)	6,395 (31.95)
Race/ethnicity, %
Asian/Pacific Islander NH	951 (4.75)	--
Black NH	1,715 (8.57)	--
Hispanic	1,282 (6.41)	--
Native American NH	375 (1.87)	--
White NH	15,679 (78.34)	--
Other NH	12 (0.06)	--
Delivery coverage, N (%)
Medicaid	7,531 (37.63)	8,431 (42.13)
Non-Medicaid	12,483 (62.37)	11,583 (57.87)
Child Characteristics Sex, N (%)
Female	9,728 (48.61)	9,978 (49.86)
Male	10,286 (51.39)	10,036 (50.14)
Plurality, N (%)
Singleton birth	19,855 (99.21)	19,679 (98.33)
Plural birth	159 (0.79)	335 (1.67)
Birth order, N (%)
First	11,614 (58.03)	--
Second	4,890 (24.43)	--
Third	2,145 (10.72)	--
Fourth or later	1,365 (6.82)	--
Sibling cluster age difference (years), mean (SD)	1.88 (0.59)	--
Sibling cluster age difference (categorical), N (%)
<2 years	12,100 (60.46)	--
2+ years	7,914 (39.54)	--
PALS-K Testing Year, N (%)
2012	8,658 (43.26)	1 (0.00)
2013	8,978 (44.86)	1,268 (6.34)
2014	2,344 (11.71)	7,025 (35.10)
2015	34 (0.17)	11,720 (58.56)
Age at PALS-K Test (years), mean (SD)a	5.51 (0.29)	5.46 (0.30)

^aIn absence of the date of the PALS-K test, we assumed that schools administered tests on September 1 of the school year (e.g., if a child took the PALS-K test in 2012, then we assumed that the child took the test on September 1, 2012). We then estimated the child’s age at PALS-K testing by subtracting their birth date as indicated on the birth record from the assumed testing date.

Notes: The PALS-K test evaluates kindergarten-level literacy. The standardized score is standardized to mean 0 (SD 1) within testing year (2012, 2013, 2014, and 2015). Each sibling pair included one older sibling and one younger sibling. The full sample of 20,014 sibling pairs comprised 40,028 children. Abbreviations: “NH” non-Hispanic; “PALS-K” Phonological Awareness Literacy Screening-Kindergarten; “SD” standard deviation.

The mean standardized PALS-K score range was 0.002–0.050 SD among older and younger siblings if neither was preterm (Table 2). Older siblings’ mean scores were consistently lower if only they were born preterm (−0.131 SD), if only the younger sibling was born preterm (−0.202 SD), or if both siblings were born preterm (−0.382 SD). Likewise, younger siblings’ mean scores were lower if only they were born preterm (−0.272 SD), if only the older sibling was born preterm (−0.090), or if both siblings were born preterm (−0.306 SD). Similar patterns with larger differences emerged when tabulating scores by siblings’ very preterm birth statuses. Locally-weighted linearity plots exhibit that the relationship between gestational age and siblings’ PALS-K performance were relatively linear (Figures S1 and S2, Appendix).

Table 2.

Standardized Phonological Awareness Literacy Screening-Kindergarten test scores by siblings’ birth outcomes (N=20,014 sibling pairs)

Birth Outcome in Sibling Pair	Sibling Pairs	Standardized PALS-K Scores, Mean (SD)
		Older Sibling	Younger Sibling	Difference (Older-Younger)
Preterm birth (gestational age <37 weeks)
No sibling	17,656	0.050 (0.993)	0.002 (0.991)	0.048 (0.991)
Older sibling only	1,027	−0.131 (1.019)	−0.090 (1.006)	−0.042 (1.049)
Younger sibling only	1,000	−0.202 (1.027)	−0.272 (1.060)	0.070 (1.049)
Both	331	−0.382 (1.066)	−0.306 (1.032)	−0.076 (1.091)
Very preterm birth (gestational age <32 weeks)
No sibling	19,746	0.027 (0.999)	−0.016 (0.997)	0.043 (0.997)
Older sibling only	128	−0.475 (1.137)	−0.246 (0.989)	−0.229 (1.093)
Younger sibling only	131	−0.295 (0.953)	−0.589 (1.073)	0.294 (1.098)
Both	9	−0.685 (0.804)	0.245 (0.597)	−0.928 (0.763)

Notes: The PALS-K test evaluates kindergarten-level literacy. The standardized score is standardized to mean 0 (SD 1) within testing year (2012, 2013, 2014, and 2015). Each sibling pair included one older sibling and one younger sibling. The full sample of 20,014 sibling pairs comprised 40,028 children. Abbreviations: “PALS-K” Phonological Awareness Literacy Screening-Kindergarten; “SD” standard deviation.

Gain-score model results indicated that shorter gestation among younger siblings penalized their older siblings’ PALS-K scores (Table 3). Unadjusted, a one-week increase in the older sibling’s gestational age was associated with a 0.024-SD increase in the gain-score (95% CI: 0.015, 0.032 SD), and a one-week increase in the younger sibling’s gestational age was associated with a 0.009-SD decrease in the gain-score (95% CI: −0.017, 0.000 SD). We computed a lower-bound estimate of the younger-to-older spillover effect by summing these coefficients, and a one-week increase in the younger sibling’s gestational age improved their older sibling’s PALS-K score by 0.015 SD (95% CI: 0.005, 0.025 SD). Covariate adjustment modestly attenuated this spillover estimate to 0.011 SD (95% CI: 0.001, 0.021 SD).

Table 3.

Gain-score regression estimates for the spillover effect of gestational age on siblings’ Phonological Awareness Literacy Screening-Kindergarten test scores (N=20,014 sibling pairs)

	Standardized PALS-K Score, Unadjusted		Standardized PALS-K Score, Adjusteda
	Coefficient	95% CI	Coefficient	95% CI
Model 1: Gestational age (weeks)
Siblings’ gestational age coefficients
Older sibling	0.024	0.015, 0.032	0.022	0.014, 0.031
Younger sibling	−0.009	−0.017, 0.000	−0.011	−0.020, −0.002
Younger-to-older spillover estimateb	0.015	0.005, 0.025	0.011	0.001, 0.021
Model 2: PTB (gestational age <37 weeks)
Siblings’ PTB coefficients
Older sibling	−0.101	−0.157, −0.045	−0.08	−0.142, −0.028
Younger sibling	0.010	−0.046, 0.067	0.025	−0.032, 0.081
Younger-to-older spillover estimateb	−0.091	−0.163, −0.019	−0.061	−0.133, 0.012
Model 3: VPTB (gestational age <32 weeks)
Siblings’ VPTB coefficients
Older sibling	−0.331	−0.499, −0.163	−0.293	−0.461, −0.124
Younger sibling	0.193	0.027, 0.359	0.211	0.045, 0.377
Younger-to-older spillover estimateb	−0.138	−0.368, 0.091	−0.082	−0.312, 0.148

^aAdjusted for maternal characteristics at the time of the older sibling’s birth (age; education; Medicaid delivery coverage), older sibling’s plurality, and older sibling’s birth order.

^bUnder maintained assumptions, we interpret the spillover effect estimate as a lower-bound estimate of the spillover effect of a younger sibling’s gestational age, PTB, or VPTB on their older sibling’s standardized PALS-K score.

Notes: The PALS-K test evaluates kindergarten-level literacy. The standardized score is standardized to mean 0 (SD 1) within testing year (2012, 2013, 2014, and 2015). Each sibling pair included one older sibling and one younger sibling. The full sample of 20,014 sibling pairs comprised 40,028 children. Abbreviations: “CI” confidence interval; “PALS-K” Phonological Awareness Literacy Screening-Kindergarten; “PTB” preterm birth; “VPTB” very preterm birth.

Gain-score models with dichotomous exposures yielded analogous results. Lower-bound spillover effects estimates of younger siblings’ preterm birth on older siblings’ PALS-K scores were −0.091 SD (95% CI: −0.163, −0.019 SD) without covariate adjustment and −0.061 SD (95% CI: −0.133, 0.012 SD) with covariate adjustment. Younger siblings’ very preterm birth reduced older siblings’ PALS-K scores by ~0.3 SD in unadjusted and adjusted models, although corresponding CIs overlapped zero.

Spillover estimates did not differ by siblings’ age difference but varied by socioeconomic measures (Table 4). Estimating the effect of younger siblings’ gestational age on older siblings’ PALS-K score, the lower-bound estimate was greater in non-Medicaid clusters (0.014 SD; 95% CI: 0.002, 0.027 SD) than in Medicaid clusters (0.008 SD; 95% CI: −0.009, 0.025). Stratifying continuous gestational age models on maternal education, the lower-bound spillover estimate among siblings with college-educated mothers paralleled sample-wide results (~0.11 SD). However, the lower-bound estimate doubled among children whose mothers only had a high school diploma/equivalent (0.024 SD; 95% CI: 0.004, 0.044 SD), while there was no evidence of spillover in sibling clusters with higher or lower levels of maternal education. We observed similar patterns in models with dichotomous exposures.

Table 4.

Heterogeneity analysis: gain-score regression estimates for the spillover effects of younger siblings’ birth outcomes on older siblings’ Phonological Awareness Literacy Screening-Kindergarten test scores (N=20,014 sibling pairs)

Stratification Variable	Sibling Pairs	Gestational Age (Weeks)		PTB (Gestational Age <37 Weeks)		VPTB (Gestational Age <32 Weeks)
		Standardized PALS-K Score	95% CI	Standardized PALS-K Score	95% CI	Standardized PALS-K Score	95% CI
Siblings’ age difference
<2 years	12,100	0.010	−0.002, 0.022	−0.047	−0.135, 0.041	−0.055	−0.328, 0.219
2+ years	7,914	0.009	−0.008, 0.027	−0.055	−0.181, 0.071	−0.156	−0.586, 0.273
Medicaid delivery coverage
Medicaid	7,531	0.008	−0.009, 0.025	−0.034	−0.153, 0.084	−0.060	−0.410, 0.291
Non-Medicaid	12,483	0.014	0.002, 0.027	−0.088	−0.180, 0.003	−0.104	−0.416, 0.208
Maternal education
No HS diploma	2,867	−0.010	−0.036, 0.017	0.054	−0.134, 0.242	0.017	−0.510, 0.544
HS diploma/equivalent	5,791	0.024	0.004, 0.044	−0.125	−0.263, 0.013	−0.057	−0.507, 0.392
1–3 years college	5,037	0.012	−0.008, 0.032	−0.064	−0.211, 0.082	−0.153	−0.636, 0.331
4+ years college	6,319	0.010	−0.007, 0.026	−0.048	−0.172, 0.076	−0.131	−0.550, 0.287

Notes: Under maintained assumptions, we interpret the spillover effect estimate as a lower-bound estimate of the spillover effect of a younger sibling’s gestational age, PTB, or VPTB on their older sibling’s standardized PALS-K score. All models adjusted for maternal characteristics at the time of the older sibling’s birth (age; education; Medicaid delivery coverage), older sibling’s plurality, and older sibling’s birth order. Aside from siblings’ age differences, stratifying variables were measured at the time of the older siblings’ birth. The PALS-K test evaluates kindergarten-level literacy. The standardized score is standardized to mean 0 (SD 1) within testing year (2012, 2013, 2014, and 2015). Each sibling pair included one older sibling and one younger sibling. The full sample of 20,014 sibling pairs comprised 40,028 children. Abbreviations: “CI” confidence interval; “HS” high school; “PALS-K” Phonological Awareness Literacy Screening-Kindergarten; “PTB” preterm birth; “VPTB” very preterm birth.

Robustness checks reinforced our main findings. We repeated the overall and stratified analyses with adjusted gain-score models after removing sibling clusters with plural-born children (Table S1, Appendix). Generally, results were consistent with a few exceptions. Lower-bound spillover estimates indicated that younger siblings’ preterm birth reduced older siblings’ PALS-K scores in the full singleton-only subsample (−0.079 SD; 95% CI: −0.154, −0.004) and in the subsample of non-Medicaid, singleton-only sibling clusters (−0.103 SD; 95% CI: −0.199, −0.008 SD). We then repeated the overall and stratified analyses with adjusted gain-score after removing siblings with foreign-born mothers (Table S2, Appendix). These results were nearly identical to the main findings. A post-hoc analysis standardized PALS-K scores to the whole sample and not within school year. Qualitative findings did not change (results not shown).

DISCUSSION

We estimated the effect of younger siblings’ gestational age on older siblings’ kindergarten-level literacy using gain-score models to eliminate fixed family-level unobserved confounding. Results aligned with our hypothesis that shorter gestation harmed siblings’ educational outcomes: a one-week decrease in a child’s gestation lowered their older sibling’s PALS-K score by 1% of a SD, and a child’s preterm birth lowered their older sibling’s PALS-K score by 6–8% of a SD. Although these estimates suggest small effects, they likely underestimated spillovers, as our method identified the difference in siblings’ spillover effects given two-sided spillover. Further, comparing our findings to related work indicates that our estimates are not trivial. A recent analysis with BD4LK cohort data found that a one-week decrease in a child’s gestation was associated with a 3% of a SD reduction in their own PALS-K score (Mallinson et al., 2019). This indicates that sibling spillover effects are fully one-third as large as the individual targeted effect. These results are also comparable to related findings. Using a sibling FE design in a population-based sample from Florida, Black et al. (2021) estimated that a younger sibling’s disability may decrease an older sibling’s score on a state-administered standardized test by 5% of a SD—a small, but measurable, spillover effect.

Our study was motivated by a hypothesis of parental resource reallocation. A large body of literature demonstrates that the home literacy environment—which includes shared reading, teaching, and conversations between parent and child—contributes to children’s early language acquisition and cognitive development (Burgess et al., 2002; Demir-Lira et al., 2019; Schmitt et al., 2011; Sénéchal & LeFevre, 2002). To enrich the home literacy environment, parents must invest their time, money, and skills in their children’s learning. Adverse health shocks in the family may cause the family to reallocate these investments to compensate for the health shock’s deleterious effects (Behrman et al. 1994; Bharadwa et al., 2018; Cunha et al., 2010; Heflin, 2016; Yi et al., 2015). In this case, a child’s shorter gestational age or preterm birth does not directly affect their sibling’s literacy. However, in the face of limited resources, parents may divest these supports from unaffected children and reallocate them toward the affected child. This shifting of resources may consequently buffer the health shock’s harm on the affected child but at the expense of their siblings’ development, including early literacy.

Following this hypothesis, shifting parental investments are most impactful in early childhood (Cunha et al., 2010), and socioeconomic advantage and the opportunities it affords—including material resources and time—may buffer a health shock’s negative externalities (Mallinson et al., 2019). Consequently, we anticipated that spillovers would be greatest among siblings who were closer in age, had Medicaid-insured births, or had less educated mothers. If a mother had a high school diploma/equivalent, each week of the younger sibling’s gestation improved the older sibling’s PALS-K score by 2.4% of a SD—double the estimate for the full sample. There was no evidence of spillover among siblings with less or more educated mothers. Contrasting our hypothesis, there was no effect heterogeneity by siblings’ age difference, and the spillover estimate was greater in non-Medicaid clusters.

However, it is unclear how effect heterogeneity results relate to the hypothesized spillover mechanism. Our method identified the difference between the younger-to-older spillover (our parameter of interest) and the older-to-younger spillover. While we could not point-identify either spillover effect, we recovered a lower bound for the spillover effect of a younger sibling’s gestational age on their older sibling’s PALS-K score. Of course, it is possible for both spillover effects to vary by background. While this makes translating our results challenging, we can nonetheless interpret spillover effect heterogeneity in three settings. Recall the assumption that longer gestation never harms siblings’ literacy.

The first scenario posits that the younger-to-older spillover varies while the older-to-younger spillover is fixed. Under these conditions, the effect of younger siblings’ gestational age on older siblings’ PALS-K scores exhibits a parabolic pattern across levels of maternal education: spillover is greatest among siblings whose mothers earned a high school diploma or equivalent while there is no spillover among siblings whose mothers did not complete high school or whose mothers attended college (notably, Black et al. (2021) observed the opposite pattern). This finding is not entirely incompatible with our hypothesis that younger-to-older spillovers would be greatest among siblings with less educated mothers. Since education is correlated to socioeconomic status (Lawlor & Mishra, 2009; Liu et al., 2010), families with lower maternal education may lack the resources to buffer adverse events, thus increasing the spillover effect of a health shock experienced by a younger sibling on the older sibling’s development. However, there could be a high enough level of socioeconomic disadvantage from poverty, resource deprivation, food insecurity, and housing instability (Burgess et al., 2002; Cunha et al., 2010; Sullivan et al., 2016)—experienced by families with the least formal education (Heflin, 2016)—that washes out any potential spillover effect of the younger sibling’s gestational age onto the older sibling’s literacy test scores. This may explain why we observed harmful younger-to-older spillovers among siblings whose mothers had a high school diploma or equivalent but not among siblings whose mothers did not complete high school. This rationale may extend to effect heterogeneity by Medicaid delivery coverage. Medicaid-covered families may face similar disadvantages to less educated families, so health shock-induced spillovers are washed out by competing threats. Correspondingly, some non-Medicaid children are more advantaged but are subsequently susceptible to harmful spillovers.

The second scenario posits that the older-to-younger spillover varies while the younger-to-older spillover is fixed. Effect heterogeneity results did not completely align with our hypothesis—much like the first scenario—but this discordance changes. The older-to-younger spillover is strongest among Medicaid-covered families and the least educated families (hypothesis affirming) as well as the among most educated families (hypothesis disaffirming). However, the critical takeaway is a less conservatively bounded younger-to-older spillover estimate. Assuming that spillovers share the same sign, the preferred estimate comes from the subgroup analysis in which the older-to-younger spillover is closest to zero and, by extension, in which the younger-to-older spillover estimate is greatest in magnitude. Thus, each week of gestation for the younger sibling improved the older sibling’s PALS-K score by at least 2.4% of a SD. Based on supplementary findings from our singleton-only sample, this also means that a younger sibling’s preterm birth reduced the older sibling’s PALS-K score by at least 10.3% of a SD.

In the third scenario, both spillover effects vary by background. Bearing fewer restrictive assumptions, it also lends little to speculation. We can interpret results generally (i.e., the lower-bound younger-to-older spillover estimate is greatest among siblings of mothers with a high school-level education), but we cannot conclude how these results invoke the hypothesized spillover mechanism (i.e., higher levels of maternal education curtail spillover).

Nonetheless, we found that shorter gestation harmed older siblings’ scores on a standardized test of early literacy, advancing the mounting evidence that negative externalities of children’s health pervade the family (Black et al., 2021; Breining, 2014; Daysal et al., 2022; De Neve & Kawachi, 2017; Dirks et al., 2015; Fletcher et al., 2012; McHale et al., 2012). If shorter gestation harms siblings’ early literacy or other academic outcomes—even if modestly—then the accumulation of additional health shocks may further threaten and impair development. Indeed, other adverse birth outcomes, such as low birth weight, stillbirth, or diagnosis of a congenital anomaly, may also exert harmful spillover effects onto other children in the family. We may also consider additional implications of our study. In the United States, racial disparities in pregnancy outcomes result in Black and Native American mothers having a greater incidence of preterm birth relative to other populations (Purisch et al., 2017; Raglan et al., 2016). This also means that families of Black and Native American mothers disproportionately incur shorter gestation’s harmful spillovers. From a health services perspective, we may consider the detriments averted by targeted medical care. Programs that prevent preterm birth through tailored prenatal services (such as Wisconsin’s Prenatal Care Coordination program (Mallinson et al., 2020)) benefit families by also preventing preterm birth’s negative spillovers.

This study was possible because of its novel data source. BD4LK matches Wisconsin birth records to Medicaid claims data and public school records—thereby allowing researchers to track children’s outcomes from birth and into early childhood—and it groups birth records into family clusters with unique maternal identifiers. This type of longitudinal and population-based data with linked administrative sources is rare (Black et al., 2021; Breining, 2014; Daysal et al., 2022), making BD4LK uniquely poised for answering our particular research question and for investigating other types of health-related spillover effects on children’s development.

The tradeoff of using BD4LK for this study is not negligible: we need to consider how our sample informs the generalizability of our findings. Wisconsin—the population base of our data source—is less racially diverse than the average state. Approximately 80% of Wisconsin’s population is non-Hispanic White, whereas 62% of the entire United States population is non-Hispanic White (United States Census Bureau, 2021a, 2021b). Further, Wisconsin is characterized by high levels of racial geographic segregation and by extreme Black-White disparities in preterm birth, scholastic performance, and educational attainment (Gordon, 2019; March of Dimes, 2022; Wisconsin Department of Health Services, 2022). Because our study examined the impact of gestational age on siblings’ PALS-K scores and its variation by maternal educational background, these results are shaped in part by Wisconsin’s racial homogeneity and inequity.

Still, we should also consider that our findings are analogous to prior studies on the spillover effects of children’s health on siblings’ academic outcomes in other settings. Fletcher et al. (2012) used data from the Panel Study of Income Dynamics, a nationally-representative sample of households in the United States, to test the impact of a child’s developmental disability or externalizing behaviour on their sibling’s score on a standardized test of cognitive ability. Analysing administratively-linked birth registry data from Denmark—a country that is inhabited almost entirely by ethnic Danes (Central Intelligence Agency, 2022)—three studies tested the impact of a child’s ADHD (Breining, 2014), developmental, intellectual, or physical disability (Black et al., 2021), or receipt of an early-life health intervention for very low birth weight delivery (Daysal et al., 2022) on a sibling’s 9^th grade exit exam scores. One of these studies also investigated the spillover effect of a younger sibling’s disability on an older sibling’s standardized test performance in a longitudinal cohort from Florida (Black et al., 2021), which is considerably more racially and ethnically diverse than Wisconsin. Each study found that adverse health events harmed siblings’ academic outcomes (Black et al., 2021; Breining, 2014; Fletcher et al., 2012) or, complementarily, that early-life health care interventions for an infant benefitted an older sibling’s academic performance (Daysal et al., 2022). The consistency of this evidence, alongside the results of our study, suggest that this pattern holds across a variety of settings—be it with different populations, health events, or academic outcomes. While our spillover estimates may not be perfectly generalizable given Wisconsin’s demographic composition, the substantive findings likely apply to other populations.

We also acknowledge that our study has some limitations. Noted previously, we could only investigate parental investment reallocation as a spillover mechanism through effect heterogeneity analyses because we had no direct measure of parental resources. Relatedly, we included only two proxies for socioeconomic status (Medicaid delivery coverage and maternal education) and could not access more direct socioeconomic markers, such as income. Our identification strategy is robust against unobserved equi-confounding on siblings’ outcomes, but it cannot account for unobserved sibling-variant confounding. Additionally, the strategy relies on a set of assumptions that are widely used in conventional FE analyses. These assumptions are strong but presently necessary for pursuing demanding questions regarding spillover effects under the adversarial condition of unobserved confounding. Lastly, PALS-K only measures a single academic domain—readiness for kindergarten-level literacy instruction—and may not generalize to other scholastic abilities.

CONCLUSION

In summary, we found evidence that a younger sibling’s shorter gestation harms their older sibling’s literacy skills at kindergarten entry. These results align with recent literature on adverse health events in childhood and their spillover effects on children’s academic outcomes. Thus, our study advances the mounting evidence that familial health shocks have withstanding effects on children’s development. The implication of this finding is not limited to harmful health-related spillovers; if a child’s adverse health event impairs their sibling’s academic achievement, then intervening on the child’s health may also benefit the sibling’s development (as found by Daysal et al. (2022)). There are several avenues for future research. Following this study, we can estimate spillover effects of adverse neonatal health events on academic outcomes in alternative settings—be it by sampling different populations, using other measures of early academic skill, or tracking academic outcomes at different points in childhood. Similarly, we can also estimate the spillover effects of early health interventions on siblings’ academic outcomes, particularly in populations that are disproportionately affected by preterm birth and other adverse birth outcomes.

Biographies

David Mallinson is a postdoctoral fellow in the Department of Family Medicine and Community Health at the University of Wisconsin School of Medicine and Public Health. His research concerns the effect of pre- and early-life health on the development and wellbeing of children and their family members. He also evaluates the effectiveness of prenatal care interventions in low-income populations.

Felix Elwert is a professor of Sociology, Population Health Sciences, and Biostatistics at the University of Wisconsin-Madison. He conducts research on social and racial inequality in the United States and Europe and develops methods for applied causal inference in the social and health sciences. He is the editor-in-chief of Sociological Methods & Research.

Deborah Ehrenthal is a professor in the Department of Biobehavioral Health and the director of the Social Science Research Institute at the Pennsylvania State University. She specializes in using administrative data to investigate disparities in maternal and child health and to evaluate the impact of prenatal health and health care on the wellbeing of infants, mothers, and their families.

DATA AVAILABILITY STATEMENT

Data used for this project are not publicly available.