Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
54,487 characters · 12 sections · 31 citation commands
Estimating Sibling Spillover Effects with Unobserved Confounding Using Gain-Scores
\pagenumbering{gobble}
A growing area of research in epidemiology is the identification of health-related sibling spillover effects, or the effect of one individual’s exposure on their sibling’s outcome. The health and health care of family members may be inextricably confounded by unobserved factors, rendering identification of spillover effects within families particularly challenging. We demonstrate a gain-score regression method for identifying exposure-to-outcome spillover effects within sibling pairs in a linear fixed effects framework. The method can identify the exposure-to-outcome spillover effect if only one sibling’s exposure affects the other’s outcome; and it identifies the difference between the spillover effects if both siblings’ exposures affect the others’ outcomes. The method fails in the presence of outcome-to-exposure spillover and outcome-to-outcome spillover. Analytic results and Monte Carlo simulations demonstrate the method and its limitations. To exercise this method, we estimate the spillover effect of a child’s preterm birth on an older sibling’s literacy skills, measured by the Phonological Awareness Literacy Screening-Kindergarten test. We analyze 20,010 sibling pairs from a population-wide, Wisconsin-based (United States) birth cohort. Without covariate adjustment, we estimate that preterm birth modestly decreases an older sibling’s test score (-2.11 points; 95% confidence interval: -3.82, -0.40 points). In conclusion, gain-scores are a promising strategy for identifying exposure-to-outcome spillovers in sibling pairs while controlling for sibling-invariant unobserved confounding in linear settings.
\scriptsize
Acknowledgements: This work was supported by the Eunice Kennedy Shriver National Institute for Child Health and Human Development (T32 HD007014-42), the University of Wisconsin-Madison Clinical and Translational Science Award program through the National Institutes of Health National Center for Advancing Translational Sciences (Grant UL1TR00427), the University of Wisconsin-Madison School of Medicine and Public Health’s Wisconsin Partnership Program, and the University of Wisconsin-Madison Institute for Research on Poverty. We thank the Wisconsin Department of Children and Families, Department of Health Services, and Department of Public Instruction for the use of data. We also thank Steven T. Cook, Dan Ross, Jane A. Smith, Kristen Voskuil, and Lynn Wimer for data access and programming assistance. We thank Michael Sobel for methodological discussions and Deborah B. Ehrenthal for feedback on this manuscript. The content is solely the responsibility of the authors and does not necessarily represent the official views of supporting agencies. Supporting agencies do not certify the accuracy of the analyses presented. Conflicts of interest: none.
\pagenumbering{arabic} \setcounter{page}{1}
A sibling spillover effect (i.e., “interference” or “carryover effect”) is the effect of an individual’s exposure on their sibling’s outcome.ogburn2014, vwbook, sjolander2016 The past two decades of epidemiologic research witnessed a burgeoning interest in the role of family environments in childhood health, calling attention to the importance of spillovers within families.kuh2003, lawlorbook, liu2010, feinberg2012, viner2015, benshlomo2016, deneve2017, morris2017 Yet, sibling spillovers are largely unexamined in the epidemiologic literature, as most field-specific advancements in spillover identification have been restricted to infectious diseases.halloran1991, halloran1995, longini1998, hudgens2008, vw2011, clemens2011, halloran2012, tt2012, vw2012, halloran2016, benjaminchun2018 With growing interest on the familial interdependence of health,lawlorbook, liu2010, feinberg2012, viner2015, benshlomo2016, deneve2017, morris2017 the need for analytical tools to identify sibling spillovers is apparent.
Unobserved confounding is particularly salient with sibling spillovers. Siblings often share experiences that cultivate their development, which may be unmeasured even in data-rich contexts.deneve2017 Fixed effect (FE) designs that control for unobserved time-invariant confounding are immediately appealing,sjolander2016, gunasekara2014, imai2019 but there is little precedent for their use to identify sibling spillovers. Sjölander et al. (2016) investigated FE models with sibling pairs for identifying targeted effects of one siblings’ exposure on their own outcome in the presence of spillover, noting that spillover may be identifiable if only one sibling’s exposure affects the other’s outcome.sjolander2016 Black et al. (2020) employed a difference-in-differences model with three-sibling clusters for identifying a lower-bound estimate of a child’s disability on an older sibling’s academic performance.black2020
In this paper, we demonstrate a method for the identification of one- and two-sided exposure-to-outcome spillovers in sibling pairs with gain-scores (i.e., difference-in-differences, or difference scores), a staple of FE estimation that removes shared confounding by differencing outcomes.kim2019 We evaluate the gain-score estimator in identifying spillover effects across various models with one-sided or two-sided spillovers. Consistent with the applied FE literature, we focus on linear models with homogenous effects.gunasekara2014, imai2019, kim2019
This paper is organized as follows. First, we briefly introduce causal directed acyclic graphs (DAGs), which illustrate our models. Second, we discuss various two-sibling models with one-or two-sided spillover and explain how and when gain-score methods can identify spillover effects. Third, we illustrate our results with simulations. Fourth, we apply the method to identify the effect of a younger sibling's preterm birth on an older sibling's literacy test performance.
Causal DAGs are useful for explaining the identification of causal effects. We review necessary terminology for this exposition. Causal DAGs are diagrams consisting of nodes (variables) and directed edges (direct causal effects) that represent the assumed data-generating process (causal model).greenland1999, pearlbook, shpitser2012, elwert2013, pearl2013, morganbook Paths are sequences of adjacent edges, regardless of the arrows’ directions. On causal paths between exposure and outcome, all arrows point from the exposure to the outcome. On non-causal paths between an exposure and an outcome, at least one arrow points away from the outcome. Causal paths “transmit” causal effects, whereas non-causal paths may transmit spurious associations. \textit{Colliders} are variables that receive two inbound arrows on a path (a given variable may be a collider on one path but not on another). Pearl’s \textit{d-separation} criterion determines which variables in data generated by the assumed DAG are conditionally or unconditionally independent: two variables are independent if all paths between them are closed; and a path is closed if it includes a non-collider as intermediate variable that is conditioned on, or if it includes a collider as intermediate variable that is not conditioned on.pearlbook, shpitser2012, morganbook Conversely, two variables may be associated if at least one path between them is open (\textit{d-connected}); and a path is open if it is not closed.
Typically, health researchers attempt to identify causal effects by adjusting for observed variables via regression analysis, matching, or inverse-probability weighting, so that all causal paths between exposure and outcome are open and all non-causal paths between them are closed.pearlbook, shpitser2012, morganbook However, researchers may worry about open non-causal paths with unobserved confounders that cannot be closed by covariate adjustment. In multilevel analyses in which observations are clustered into groups (e.g., children in sibling pairs), FE methods can sometimes identify causal effects by subtracting out certain types of group-level unobserved confounding.gunasekara2014, imai2019, kim2019 Next, we describe several sibling spillover models with unobserved confounding and show when gain-score estimation—a FE approach—can identify the spillover effect.
We first present our baseline sibling spillover model and subsequently introduce variations on this model. For illustration, we discuss the spillover effect of a child’s early health shock (e.g., serious illness) on their sibling’s later academic achievement (e.g., test scores). This example is purposefully generic but broadly applicable, and it draws upon prior work of health-related spillover effects on academic performanceblack2020 while motivating our empirical application.
Our baseline model is a linear two-sibling comparison design with one-sided spillover (Figure 1A). Subscript $i=1,\ldots,N$ indicates cluster (family) and subscript $j=1,2$ indicates sibling. $T_{ij}$ represents a binary or continuous exposure (e.g., the health shock), $Y_{ij}$ represents a continuous outcome (e.g., academic performance), $U_i$ represents unobserved family-level confounding (i.e., the FE), and $D_i$ represents the gain-score, $D_i=Y_{i2}-Y_{i1}$. Causal effects in this model include the spillover effect, $\theta$ ($T_{i1}\rightarrow Y_{i2}$), of sibling 1’s exposure on sibling 2’s outcome; the targeted effects, $\delta$ ($T_{ij}\rightarrow Y_{ij}$), of each sibling’s exposure on their own outcome; and confounding effects of the unobserved family-level confounders on each sibling’s exposure and outcome, $\psi\chi$ ($T_{i1}\leftarrow U_i\rightarrow Y_{i1}$) and $\psi\gamma$ ($T_{i2}\leftarrow U_i\rightarrow Y_{i2}$).
\afterpage{
}
All models embed several simplifying assumptions. First, the targeted effects, $\delta$, of $T_{ij}$ on $Y_{ij}$ and the confounding effects, $\psi$, of $U_i$ on $Y_{ij}$ are equal for both siblings.gunasekara2014, imai2019 Second, all effects are linear and homogenous. Third, there is only partial interference (i.e., spillovers within sibling clusters but not between sibling clusters).hudgens2008, sobel2006 Aside from partial interference, these assumptions align with conventional FE models.sjolander2016, gunasekara2014, imai2019, kim2019
Notably, our presentation abstracts from sibling-specific observed baseline covariates, $\textbf{\textit{C}}_{ij}$. Covariates may be added to our baseline and subsequent models as long as one can condition on $\textbf{\textit{C}}_{ij}$ without loss of generality.
This subsection details the gain-score estimation strategy and demonstrates when our estimator point-identifies spillover effects (i.e., recovers the estimand precisely) for nine sibling spillover models that differ by whether spillover is one- or two-sided and by whether additional spillovers originate from outcomes.
We investigate the ability of a gain-score estimator to identify exposure-to-outcome spillover effects. First, we regress the gain-score on both siblings’ exposures,
where $b_1$ and $b_2$ are partial regression coefficients for $T_{i1}$ and $T_{i2}$, respectively, and $e_i$ is an error term. We then sum the partial regression coefficients to compute a “spillover coefficient” ($SC$),
We will now interrogate whether the SC identifies causal spillover effects in each of several commonly assumed data generating processes in health research.
The object of interest (estimand) is $\theta$, or the direct spillover effect of sibling 1’s health shock on sibling 2’s academic performance. Under the baseline model (Figure 1A), three open paths connect $T_{i1}$ and $Y_{i2}$. The first path, $T_{i1}\rightarrow Y_{i2}$, is the causal spillover effect of interest. The other two paths are non-causal paths that may transmit spurious association. The first non-causal path, $T_{i1}\leftarrow U_i\rightarrow T_{i2}\rightarrow Y_{i2}$, can be closed by adjusting on $T_{i2}$. However, the second non-causal path, $T_{i1}\leftarrow U_i\rightarrow Y_{i2}$, cannot be closed by covariate adjustment because it only contains the unobserved variable $U_i$.
Nonetheless, we can identify $\theta$ through gain-score regression. Under the assumptions of Figure 1A, it can be shown that $b_1=\theta-\delta$ and $b_2=\delta$, using elementary regression algebra. Therefore, the spillover coefficient equals $SC=b_1+b_2=\theta$.
The intuition for this result is that first-differencing exactly offsets confounding biases involving $U_i$,26 and that the $SC$ corrects for the contamination of the spillover estimate in $b_1$. Specifically, the coefficient $b_1$ on $T_{i1}$ captures the association flowing along the open paths from $T_{i1}$ to $D_i$. There are five paths from $T_{i1}$ to $D_i$ (listed together with their corresponding path coefficients):
The first path is closed because the regression conditions on $T_{i2}$. The second and third paths cancel each other out exactly. The fourth path transmits the spillover effect. The fifth path transmits the negative of the targeted effect. Hence, the regression coefficient $b_1=\theta-\delta$ identifies the difference between the spillover and targeted effect.
The coefficient $b_2$ on $T_{i2}$ captures the association flowing along the open paths from $T_{i2}$ and $D_i$. There are four paths from $T_{i2}$ to $D_i$:
The first path is closed because the regression conditions on $T_{i1}$; the second and third paths cancel each other out; and the fourth path captures the targeted effect. Thus, $b_2=\delta$ identifies the targeted effect, and $SC=b_1+b_2=\ \theta$ identifies the causal spillover effect.
Many statistical software have functions for summing regression coefficients and obtaining standard errors. Examples include Stata’s lincom command,stataman R’s contrast package,kuhn2016 and SAS’s SCORE procedure.sas
The analysis is only slightly complicated in the presence of exposure-to-exposure spillover ($T_{ij}\rightarrow T_{ij^\prime}$) – for example, when one child’s serious illness increases their sibling’s risk of illness. When $T_{i2}\rightarrow T_{i1}$ (Figure 1B), the analysis does not change. However, if $T_{i1}\rightarrow T_{i2}$ (Figure 1C), then the interpretation of $SC=\theta$ changes from representing the entire spillover effect of $T_{i1}$ on $Y_{i2}$ to capturing only the direct spillover effect, since the indirect component of the spillover effect that operates via the causal path $T_{i1}\rightarrow T_{i2}\rightarrow Y_{i2}$ is closed because the regression controls for $T_{i2}$. See the Supplementary Material for details.
Analysts may also encounter scenarios with two-sided spillover. In our example, each siblings’ health shock could affect the other’s academic performance ($T_{i1}\rightarrow Y_{i2}$ and $T_{i2}\rightarrow Y_{i1}$). Reflecting this possibility, Figure 2A modifies the baseline model of Figure 1A to allow spillover $T_{i2}\rightarrow Y_{i1}$ with effect $\kappa$. The partial regression coefficients in the gain-score approach identify $b_1=\theta-\delta$ and $b_2=\delta-\kappa$, so that $SC=b_1+b_2=\theta-\kappa$. Consequently, with two-sided exposure-to-outcome spillover, the $SC$ does not identify the spillover effect of $T_{i1}$ on $Y_{i2}$ but instead the difference between the two exposure-to-outcome spillover effects. However, if the analyst can defend assumptions about one or more of the signs of the two spillover effects, then the $SC$ remains informative even though it no longer point-identifies $\theta$. Specifically, if $\kappa>0$, the $SC$ underestimates (i.e., gives a lower bound for) $\theta$. By contrast, if $\kappa<0$, then the $SC$ overestimates (gives an upper bound for) $\theta$. One can make additional inferences about $\theta$ depending on the value of the $SC$ and the assumed sign of $\kappa$. For example, if $SC>0$ and $\kappa>0$, then $\theta>0$. Of note, a finding that $SC=0$ is uninformative, because it is compatible with the possibility that the two spillover effects are equal, $\theta=\kappa$, and that they are both zero, $\theta=\kappa=0$.
\afterpage{
}
If $T_{ij}\rightarrow T_{ij'}$ in addition to two-sided spillover (Figures 2B-C), this does not affect the interpretations of $b_1$ and $b_2$, and $SC$ still identifies the difference between siblings’ unmediated spillover effects. However, $SC$ will not capture the mediated part of spillover effect, $T_{ij}\rightarrow T_{ij'}\rightarrow Y_{ij'}$. See the Supplementary Material for details.
Analysts may also encounter settings with outcome-to-outcome spillover ($Y_{ij}\rightarrow Y_{ij'}$) or outcome-to-exposure spillover ($Y_{ij}\rightarrow T_{ij'}$). In our setting, it is reasonable to assume that siblings’ academic outcomes may be causally related by outcome-to-outcome spillover. In contrast, an academic outcome causing a health shock is implausible, but exposure-to-outcome spillovers may be relevant elsewhere.
If outcomes cause future exposures or outcomes (Figure 3), then our estimator does not identify spillovers or simple functions of spillovers. See the Supplementary Material for details.
\afterpage{
}
We conducted nine Monte Carlo simulations,adkins2012 one simulation for each of the nine models in Figures 1-3, to demonstrate when the method identifies exposure-to-outcome spillover effects. Our simulation model follows:
\[U_i,\upsilon_{i1},\upsilon_{i2}\sim N(0,1)\] \[T_{i1}=
\] \[T_{i2}=
\] \[Y_{i1}=\delta T_{i1}+\kappa T_{i2}+\lambda Y_{i2}+\psi U_i+\upsilon_{i1}\] \[Y_{i2}=\delta T_{i2}+\theta T_{i1}+\eta Y_{i1}+\psi U_i+\upsilon_{i2}\] \[D_i=Y_{i2}-Y_{i1}\]
We simulated each model with 1000 runs of 5000 observations each, where each observation represented a sibling pair. We set the following parameters at fixed values: $\theta=0.5$, $\delta=1$, $\psi=1$, $\chi=2$, and $\gamma=3$. Parameters distinguishing the models---$\kappa$, $\tau$, $\phi$, $\omega$, $\eta$, and $\lambda$---were set to zero unless otherwise specified. To avoid simultaneity, at least one parameter in each pair ($\tau$, $\phi$), ($\eta$, $\lambda$), and ($\kappa$, $\omega$) was always set to zero. In each sample, we regressed the gain-score on siblings’ exposures and computed the spillover coefficient according to equations (1) and (2). We conducted simulations in Stata Statistical Software: Release 16.statasoft Simulation code is in the Supplementary Material.
Figure 4 displays the simulation results. The first three rows confirm that the spillover coefficient is unbiased in the three settings with one-sided exposure-to-outcome spillover of Figure 1, as the average of estimated spillover coefficient equals the known spillover effect, ${\widehat{SC}}_{Figure1}=0.5$ (empirical 95% CI: 0.42, 0.58). The subsequent three rows demonstrate that the spillover coefficient in the three models of Figure 2 with two-sided exposure-to-outcome spillover identifies the difference between the two spillovers, ${\widehat{SC}}_{Figure2}=0.5-0.3=0.2$ (empirical 95% CI: 0.12, 0.28). Since $\kappa>0$, ${\widehat{SC}}_{Figure2}$ underestimates the spillover effect, $\theta$. The final three simulations show that ${\widehat{SC}}_{Figure3}$ is biased in all models of Figure 3 with spillovers from outcomes. Size and direction of the biases are fairly complicated functions of the coefficients in the data-generating model and can be large. The estimated ordinary least squares standard errors closely resemble the empirical standard errors for each model, indicating that the built-in standard errors in Stata’s lincom command are accurate.stataman
\afterpage{
}
We applied the method to estimating the spillover effect of a child’s preterm birth (gestational age $<$ 37 weeks) on their older sibling’s literacy skills. This analysis builds upon evidence that short gestational age and health shocks within the family may impede a child's own early literacy skills.black2020, mathiasen2010, mallinson2019 If a child is born preterm, parents may reallocate investments (time, financial, or otherwise) from older siblings to support the younger sibling's health, thereby inhibiting the older siblings' development, including early literacy.
For this application, we analyzed Big Data for Little Kids (BD4LK), a longitudinal cohort of birth records for all live in-state resident deliveries in Wisconsin during 2007-2016 (N $>$ 660,000 deliveries) that links to multiple administrative data sources, including Medicaid data (2007-2016) and children’s Phonological Awareness Literacy Screening-Kindergarten (PALS-K) test scores from Wisconsin public schools (2012-2016 school years). BD4LK’s linking process is described elsewhere.mallinson2019, larson2019 PALS-K evaluates readiness for kindergarten-level literacy instruction on six domains (rhyme awareness; beginning sound awareness; alphabet knowledge; letter sounds; spelling; word concept).ford2014, pktech In Wisconsin, children must be five years-old at kindergarten enrollment to qualify for PALS-K testing.widpi Our analysis includes 20,010 sibling pairs (40,020 children) that were sequentially-born from different deliveries to the same biological mother and had non-missing English-language PALS-K test scores and covariates. The Supplementary Material contains the full sampling description.
We estimate the following gain-score regression model,
\[D_i=b_1{PTB}_{i1}+b_2{PTB}_{i2}+\beta_3\textbf{\textit{C}}_{i2}+v_i\]
where $D_i=PALSK_{i2}-PALSK_{i1}$. Subscripts $i=1,\ldots N$ and $j=1,2$ indicate cluster and sibling, respectively, where $j=1$ is the younger sibling. $PTB_{ij}$ is a binary preterm birth indicator (1 if preterm; 0 otherwise), $PALSK_{ij}$ is the continuous PALS-K score (0-102 points), and $\textbf{\textit{C}}_{i2}$ is a vector of covariates measured at the older sibling’s delivery, which may be empty. Covariates include maternal age (years), maternal education (no high school diploma; high school diploma/equivalent; 1-3 years college; 4+ years college), and Medicaid delivery payment. $D_i$ is the gain-score estimator.
We ran the model twice, once with and once without covariates. In each model, we summed the regression coefficients on both siblings’ preterm birth indicators to compute the $SC=b_1+b_2$. Assuming the one-sided spillover model of Figure 1A, $SC$ from the regression without covariates identifies the effect of a younger sibling’s preterm birth on the older sibling’s PALS-K score. Additionally, $b_2$ identifies the effect of each sibling’s preterm birth on their own PALS-K score. We performed all analyses in Stata Statistical Software: Release 16.statasoft The University of Wisconsin-Madison minimal risk institutional review board approved our project.
Supplemental Tables 1 and 2 summarize baseline characteristics of our sample (Supplementary Material). Preterm birth incidence was slightly greater among older siblings relative to younger siblings (6.78% vs. 6.65%). On average, older siblings received slightly lower PALS-K scores (mean 63.58 points; SD 24.12 points) relative to younger siblings (mean 64.22 points; SD 23.83 points). Approximately 10% of sibling clusters had discordant preterm birth exposure. In the regression without covariate adjustment, the older sibling's preterm birth coefficient was ${\hat{b}}_2$ = -2.49 points (95% CI -3.83, -1.15 points), the younger sibling's preterm birth coefficient was ${\hat{b}}_1$ = 0.38 points (95% CI: -0.97, 1.73 points), and the resulting $\widehat{SC}$ was -2.11 points (95% CI: -3.82, -0.40 points) (Table 1). This indicates that a younger sibling’s preterm birth modestly harmed their older sibling’s PALS-K performance. Figure 5 displays these results graphically relative to the assumed data-generating model. However, covariate adjustment attenuated the $\widehat{SC}$ to -1.49 points (95% CI -3.21, 0.22 points).
\afterpage{
}
\afterpage{
}
We described a simple approach to identifying spillovers with gain-scores in sibling pairs. This method can point-identify spillovers if only one sibling’s exposure affects the other’s outcome, and it can identify the difference in siblings’ spillovers in the presence of two-sided spillover. The method leverages the primary benefit of FE estimation: controlling for family-level, sibling-invariant, unobserved confounding. Whereas preceding epidemiologic research on spillover identification primarily considered infectious diseases, our work contributes to the growing literature on spillovers within families.
We acknowledge some limitations. First, we restricted our attention to linear settings. This method does not necessarily apply to contexts with nonlinear relationships, such as those with binary outcomes (see Sjölander et al. (2016)sjolander2016 for binary outcomes in our Figure 1A). Second, we did not consider clusters of three or more siblings. Spillovers that originate from larger sibling clusters may pose unique challenges that are unaddressed here – for example, whether one can identify the effect of a middle child's exposure on the youngest sibling's outcome if an eldest sibling's exposure affects all siblings' outcomes. Lastly, we did not test the method in the presence of shared mediator or collider variables. Sjölander and Zetterqvist (2017) interrogated sibling comparison models with shared mediators and colliders, finding that such factors may induce bias.sjolander2017
Nonetheless, our paper lays groundwork for subsequent research. Specific avenues that advance this method include testing in nonlinear settings or settings with shared mediator variables, as well as expanding models to allow three or more siblings.