Caesarean section and children’s health: A quasi-experimental design

Abstract

The prevalence of inflammatory child health conditions—such as asthma, eczema, and food allergy—and their associated costs have increased rapidly over the last 30 years. While environmental factors likely underpin these increases, recent studies explain only a fraction of the trend and rely on associational methods. Caesarean (or C-) section rates increased dramatically in the period of interest, and this method of delivery is an understudied environmental factor linked to child health outcomes via the gut microbiome. We fuse 22 years of birth cohort data from the United States National Surveys of Children’s Health with C-section rates from the National Vital Statistics System generated for subgroups based on state, sex, race, Hispanic origin, and birth year. Then, we model the effects of C-section on rates of asthma, eczema, and food allergy using a quasi-experimental fixed effects design. We find that C-section significantly predicts food allergy, with qualitatively significant implications.

Background

Infants born in the United States (US) today are no longer faced with the pandemic threats of infectious disease and correspondingly high mortality rates of their predecessors, but rather are born into an ‘age of man-made and degenerative diseases’ (Omran 2005). At the turn of the twentieth century, 30.4 per cent of all deaths occurred among children under five, and infectious diseases, such as pneumonia, tuberculosis, and diarrhoea and enteritis, were the leading causes of mortality. By 1997, only 1.4 per cent of all deaths were in children under five, and chronic diseases, such as heart disease and cancer, replaced infectious diseases as the leading causes of death in the overall population (CDC 1999a, 1999b). Although infant and child mortality rates are now quite low, the prevalence of chronic health conditions among children in the US has increased substantially; for instance from 12.8 per cent in 1994 to 26.6 per cent in 2006 (Van Cleave et al. 2010).

In particular, the US has seen a dramatic rise in the prevalence of chronic health conditions among children, including the inflammatory conditions—asthma, eczema, and food allergy—that we are concerned with in this study. Among US children, the estimated prevalence of asthma tripled between 1980 and 2006, and stood at 9.5 per cent (7 million children) in 2010 (Moorman et al. 2012). More recently, there has also been a phenomenal increase in broader categories of allergies and atopic illnesses, with childhood prevalence estimates reaching 5.1 per cent for food allergies and 12.5 per cent for eczema as of 2009–11, up from 1997–99 estimates of 3.4 and 7.4 per cent, respectively (Jackson et al. 2013). The estimated costs of these illnesses to society are now upwards of $7 billion per year (Lapidus et al. 1993; Trasande and Liu 2011; Gupta et al. 2013).

These increases in childhood health conditions are concerning, not only for their immediate impact on the lives of children and the associated costs to society, but also for their life course implications. For example, early-life health is one path through which parents transmit social class positions to their children (Palloni 2006). Childhood health conditions are also associated with later-life health and socio-economic achievement (Blackwell et al. 2001; Case and Paxson 2010), and childhood health problems are tied to future labour market disadvantage (Currie 2009). The increasing prevalence of chronic health conditions in recent cohorts of children, combined with their serious and lasting consequences, makes understanding their underlying causes more pressing.

Possible explanations for increases in children’s chronic conditions

Many child health conditions, such as asthma and allergies, have strong genetic components, with the majority of heritability estimates greater than 50 per cent and some as high as 90 per cent (Sicherer et al. 2000; Thomsen et al. 2010). Further, genetic association studies have identified specific genetic variants associated with childhood asthma and allergies (Ober and Yao 2011; Torgerson et al. 2011; Arshad et al. 2012; Tan et al. 2012; Bonnelykke et al. 2014). While this evidence suggests that genes may explain a predisposition toward these health conditions in the population, the recent explosion of these conditions in children is inconsistent with genetic changes in the population as a key contributor to the changing prevalence of health conditions. Person-to-person network multiplier effects are similarly unlikely to be a major explanation, due to the non-communicability of the diseases and the rapidity of their increases in prevalence. Instead, patterns of substantial variation across geography and rapid increases over a short time period suggest that environmental factors are driving the increases.

Indeed, a broad set of environmental changes in many developed countries, summarized in the literature on the ‘hygiene hypothesis’, has largely coincided with the shift to chronic and inflammatory child conditions. The hygiene hypothesis, in its simplest form, claims that modern Western lifestyles have led to reduced exposure to both pathogenic and commensal microbes, leading to aberrant immune system functioning and an increase in associated chronic health conditions (Versini et al. 2015; Lambrecht and Hammad 2017). Epidemiologic evidence supports the hypothesis, demonstrating an inverse relationship between trends in infections—such as helminthic infections and malaria—and trends in immune-related disease (Versini et al. 2015). While changes in microbial exposures may not be the only environmental drivers of the increasing prevalence of child health conditions, and some part of the increase may be due to heightened awareness and surveillance (Branum and Lukacs 2009), linkages made between microbial diversity and the functioning of the human immune system are a compelling reason to further explore this relationship.

Human microbiome, environment, and host interactions

The human gut microbiome consists of trillions of microbes of a variety of species living in the human gastrointestinal tracts. These microbes synthesize vitamins, are involved with human metabolism, affect gene regulation, and interact with human immune systems. The early postnatal environment determines essential elements of the structure of the microbiome, and diet affects its composition throughout the life course (Kau et al. 2011). Recently, the gut microbiome has been identified as an important mediator of the hygiene hypothesis because it regulates the immune system. However, modern lifestyle practices, such as increases in the use of C-section and antibiotics, as well as changes in dietary regimes, have led to a reduction in the diversity of the microbiome, which may translate to aberrant immune system functioning (Rook et al. 2014). Indeed, lifestyle exposures modulate the gut microbiome and have been linked to immune-mediated disorders. For example, early environmental exposures associated with allergies and asthma include day-care attendance in early childhood, exposure to pets, breastfeeding, antibiotic use, mode of delivery, diet and nutrition, and others (Ober and Yao 2011; Lambrecht and Hammad 2017). Some of these exposures, such as breastfeeding, increased over our study period (Wright and Schanler 2001; CDC 2013), while others, such as childhood antibiotic use, showed mixed trends (Finkelstein et al. 2003; Vaz et al. 2014). Infant feeding guidelines also became more restrictive during our study period, with the aim of limiting infants and toddlers at high risk of food allergy from exposure to highly allergenic foods, such as peanuts (American Academy of Pediatrics 2000). Some of these factors increase microbial diversity, while others restrict it (Lambrecht and Hammad 2017). A benefit of examining method of delivery is that it represents a child’s initial extensive microbiome colonization (Ferretti et al. 2018). The other exposures occur post-delivery, and to the extent they are linked with C-section, they would mediate rather than confound the relationship with child health outcomes. Moreover, broad trends in environmental exposures can be captured by state and year fixed effects.

While the microbiome is not static, early microbiota set the course for immune development (Gronlund et al. 1999; Penders et al. 2006; Fallani et al. 2010). For example, Bifidobacteria are less common in infants delivered by C-section, while colonies of C. difficile, found commonly in hospital rooms, are more common. These early differences are connected to subsequent health: Bifidobacteria are protective for allergy (Enomoto et al. 2014; Azad et al. 2015) while C. difficile is associated with atopic disease, such as eczema, allergic rhinitis, and asthma (Sepp et al. 1997; Bottcher et al. 2000; Penders et al. 2006; Kalliomaki et al. 2008). The microbiome becomes increasingly complex and stable over time, based on host and environmental factors, but the initial microbial communities acquired from infant environments are important early sources of either protective or pathogenic flora. Infant delivery is, thus, an important variable that can alter early microbial communities when they begin to define a trajectory leading to the adult microbiome (Gronlund et al. 1999; Biasucci et al. 2008; Kalliomaki et al. 2008; Biasucci et al. 2010; Fallani et al. 2010; Kaplan et al. 2011; Collado et al. 2012).

Mechanisms linking the microbiome with children’s chronic conditions

While research into the complexity of interactions among lifestyle factors, microbial species, and the human immune system is still in its infancy, evidence points to an immune system aetiology of a multitude of chronic childhood conditions, including those on which our study is focused. For example, childhood asthma and allergies are associated and share a chronic inflammatory nature (Visness et al. 2010; Rance and O’Laughlen 2011). With regard to allergy and asthma, there is evidence suggesting that the conditions may develop sequentially; food allergies appear in infancy and are subsequently followed by asthma and allergic rhinitis (Thomsen et al. 2010; Ober and Yao 2011). Collectively, this suggests that reduced exposure to commensal microbes during early development could lead to lasting effects on children’s health.

C-section links the hygiene hypothesis to increases in children’s chronic conditions

In proposing environmental factors that are conceivably responsible for the large changes in allergic and respiratory conditions in children, we consider factors that: (1) have experienced large changes over a short time; and (2) are conceptually plausible through biological mechanisms implicated by the hygiene hypothesis. One environmental factor that meets these criteria is caesarean delivery. First, the prevalence of caesarean delivery increased by 60 per cent over the 13 consecutive years between 1996 and 2009, and involved around 32 per cent of all births in the US by 2015 (Martin et al. 2017). Second, the extraordinary rise in caesarean delivery rates is associated with children’s health outcomes via a key biological mechanism: the gut microbiome. In particular, caesarean delivery and vaginal birth are two different environments contributing to the development and modification of the infant microbiome. In the sterile caesarean environment, the mother’s skin flora influences the infant microbiome, as do the environments with which the infant subsequently comes into contact, including the hospital room and medical personnel. During a vaginal delivery, however, the infant acquires maternal vaginal and faecal flora (Penders et al. 2006), with the contribution of the mother’s faecal flora, which contains beneficial bifidobacterial species, making up a greater portion of the infant’s microbiome over time (Ferretti et al. 2018). We know such species are less abundant in infants delivered by C-section, thus, caesarean delivery can serve as a measure for adverse gut microbial colonization of the infant.

Associational evidence linking C-section and children’s health

Several meta-analyses and review articles summarize epidemiological studies, finding support for positive associations between C-section and asthma, eczema, and allergy (Bager et al. 2008; Thavagnanam et al. 2008). However, these reviews also convey the strong limitations that studies face in their ability to address confounders. Indeed, while the conceptual steps linking the rapid rise and geographic dispersion of C-sections in the US with the coincident rise in asthma, eczema, and food allergies via the children’s gut microbiome are relatively straightforward, there are few studies that can rule out a large number of confounding factors. C-sections may be related to other attributes of families and local environments that independently affect childhood illness, thus confounding associations made from observational studies in humans. Similar issues abound when examining child health outcomes in observational data. In part, this lack of strong evidence, especially evidence that may suggest a common set of causes, is due to the lack of research designs that are capable of controlling for a large number of important confounding variables while retaining enough variation to observe an association.

Variation in C-section unrelated to individual need or preference

We aim to take a step towards overcoming the issues posed by such confounding variables by conducting our analysis at the population level and applying a quasi-experimental research design. Using ‘natural experiments’, which take advantage of the large temporal and spatial variations in the US in the likelihood of having a C-section (unrelated to individual need or preference for C-section), we determine the aggregate contributions to the increases in childhood health outcomes attributable to the large changes in C-section practices over the last two decades. The World Health Organization has concluded that C-section rates above 10 per cent do not result in any reductions to infant or maternal mortality rates, and notes that the ideal C-section rate at a population level is between 10 and 15 per cent (Betran et al. 2016), suggesting that the current US C-section rate (around 32 per cent of deliveries) is unnecessarily high. These facts suggest that part of the increase in C-section rates over this period is due to factors that are unrelated to individual-level medical need for C-section. Indeed, there is evidence of hospital variation in C-section rates even when individual-level medical and socio-demographic risk factors are controlled (Bailit et al. 1999). It has also been shown that provider preferences explain more of the variation in medical treatment and spending than do patient preferences (Cutler et al. 2013). Additional studies have linked other mechanisms unrelated to maternal–infant medical need or preference to increases in C-sections. Currie and MacLeod (2008) found that caps on medical malpractice lawsuits led to an increase in marginal C-sections (i.e., those for which risk is low), providing doctors with greater profitability and little fear of liability. Additionally, Dubay et al. (1999) found evidence that physicians’ fears of medical malpractice lawsuits led to a small increase in overall C-section rates. Further, Gruber et al. (1999) have demonstrated that Medicaid reimbursement fee differentials between vaginal and caesarean births, where C-sections are reimbursed at higher levels than vaginal births, despite similar costs being incurred, led to higher C-section rates. Grant (2009) replicated this study and found that a $1,000 increase in reimbursement for C-section increased the C-section rate by one percentage point.

Policy changes unrelated to individual-level characteristics are also likely to interact with provider preferences and hospital practices to produce temporal changes in the variation in C-section rates. For instance, changing guidelines around vaginal birth after C-section (VBAC) have led to declining rates of VBAC and, thus, increasing rates of repeat C-section, which vary depending on hospital type (public, private, teaching, etc.) and provider type (obstetrician, family practitioner, midwife, etc.) (Gregory et al. 2010). We exploit these sources of variation to generate group-level C-section measures. We then measure the association between group-level exposure to C-section and individual health outcomes in the spirit of an ‘intent-to-treat’ analysis, while controlling for individual-level confounders.

Data

To our knowledge, a nationally representative, publicly available data set that contains both information on method of delivery and child inflammatory health outcomes at the individual level for the US does not exist. While such a data set would be ideal, we instead leverage variation in C-section rates by state, year, and demographic subgroup by linking method of delivery data from the National Vital Statistics System (NVSS) Natality Files (National Center for Health Statistics 2018) to child health indicators measured by the National Surveys of Children’s Health (NSCH) (Health Resources & Services Administration 2005; 2009; 2013). We merged these data sets based on state, birth year, race, Hispanic origin, and sex of the child. The NVSS contains method of delivery and demographic measures based on birth certificate information for nearly all US births in our study period. The NSCH is a state-representative, repeated, cross-sectional survey that contains health information on almost 300,000 children aged 0–17. We use the 2003, 2007, and 2011 survey waves.

From the NVSS data, we generate C-section rates (i.e., exposure) based on child’s state of birth, birth year, race, Hispanic origin, and sex. The NVSS does not contain full data on child’s race, but does have data on parents’ race, which we use to construct an NVSS child race variable that matches the race variable in the NSCH. We also construct a Hispanic origin measure based on NVSS data by assigning a child to a binary Hispanic origin variable if either or both parents designated Hispanic origin on the child’s birth certificate. After organizing the individual data from the Vital Statistics, we generate C-section rates for subgroups based on the various combinations of state, birth year, race, Hispanic origin, and sex.

The NSCH data enable us to generate measures of food allergy, eczema, and asthma. With regard to asthma, the survey asked, ‘Has a doctor or health professional ever told you that [CHILD] has any of the following conditions?’ For food allergy and eczema, the survey asked, ‘During the past 12 months … have you been told by a doctor or other health care professional that [CHILD] had any of the following conditions?’ Each survey wave measured asthma prevalence, but measures of food allergy and eczema were not included in the 2011 survey. The child’s parent or guardian, typically the mother, reported the information. There are some missing health outcome data in the NSCH, although sensitivity checks wherein missing data were changed to zero to indicate a negative response to the health questions suggest that the missing data do not tend to alter results.

Although the NSCH covers the 1986–2011 birth cohorts, we are limited to using cohorts from 1990 to 2011 because the NVSS birth data before 1990 do not contain information on method of delivery. This reduces our sample size from almost 300,000 children to a total of approximately 250,000 children with reported asthma measures captured across three survey waves. For food allergy and eczema measures, which were captured in 2003 and 2007 only, our sample size is approximately 150,000. The sample sizes for food allergy and eczema are disproportionately smaller than the sample size for asthma because the lack of NVSS birth data before 1990 affected only the 2003 survey year. We match individual-level NSCH data to state-by-sex-by-race-by-Hispanic-by-birth-year C-section rates in order to conduct statistical analyses.

Table 1 presents summary statistics based on the NSCH data for cohorts born 1990–2011. The table shows mean health outcome rates and mean demographic statistics for individuals in each survey wave. Of note are the large increases in the rates of asthma over an eight-year span, and modest increases in the rates of food allergy and eczema over a four-year period. Our variation spans the larger window of 1990–2011, but we show the summary statistics by survey year rather than birth year because each birth year represents only two or three age groups, so rates of health outcomes by birth year would be confounded with rates of health outcomes by age. We control for both age and birth year in our analysis to eliminate this confounding. Average age is lowest in the 2003 survey year due to the limitations in data before 1990, which curtail representation of earlier birth cohorts in this survey year. We control for child’s age in our analysis to address this issue, as well as any association it may have with the onset and diagnosis of a disorder.

Table 1.

Summary statistics for US children born 1990–2011, by survey year

	Mean/proportion (standard deviation)
Survey year:	2003	2007	2011
Food allergy	0.042	0.052	–
	(0.201)	(0.222)	(–)
Asthma	0.110	0.132	0.142
	(0.313)	(0.339)	(0.349)
Eczema	0.111	0.125	–
	(0.315)	(0.331)	(–)
Male	0.511	0.519	0.515
	(0.500)	(0.500)	(0.500)
Female	0.489	0.481	0.485
	(0.500)	(0.500)	(0.500)
Age	6.520	9.168	8.866
	(4.141)	(5.340)	(5.242)
Black	0.108	0.109	0.102
	(0.310)	(0.312)	(0.302)
White	0.794	0.784	0.739
	(0.405)	(0.412)	(0.439)
‘Other race’	0.098	0.107	0.159
	(0.298)	(0.309)	(0.366)
Hispanic	0.091	0.087	0.132
	(0.287)	(0.282)	(0.338)
Birth year range	1990–2003	1990–2007	1994–2011
Observations	70,789	85,971	92,825

Notes: Each survey wave includes children aged 0–17, but because method of delivery data were not collected before 1990, the 2003 survey year includes a narrower range of birth years. We do not include observations with missing race, Hispanic origin, or C-section data in the summary statistics. Summary statistics report the maximum number of observations across all variables in each survey year, thus hiding minor differences in the number of missing observations by health outcome. For this reason, the summed number of observations across survey years does not perfectly correspond to the total observations in Table 2, which reports observations for each health outcome across multiple survey waves.

Source: NSCH waves 2003, 2007, and 2011.

Figure 1 shows the change over time in national C-section rates from the NVSS data for the years 1990–2011. The prevalence of C-sections in the US was about 21 per cent in 1990, decreasing to 20 per cent in 1997, before increasing to 32 per cent by 2011, a substantial upsurge in 14 years. The sharp increase in C-section rates after 1997 followed a change in guidelines around VBAC (Barber et al. 2011). Our data only show a snapshot of a dramatic long-term increase in C-section rates, which were documented at about 4.5 per cent in 1965 (Taffel et al. 1987) and had climbed to 32 per cent by 2015 (Martin et al. 2017).

Figure 1.

C-section as a proportion of births, 1990–2011, US

Source: NVSS.

Methods

Many of the studies linking C-sections to child health outcomes suffer from small sample sizes, inadequate adjustment for confounders, or both. Control variables are typically limited to those that are measurable at the individual level (Bager et al. 2008; Thavagnanam et al. 2008). Such controls are likely to be inadequate, given the context of the massive changes in health practices surrounding C-section that have occurred over time and across space (states and localities), which are in part independent of maternal–child characteristics and medical indications. While a true randomized control trial for caesarean vs. vaginal delivery would be the ideal way to eliminate confounding, it would be impractical and unethical. Instead, we use data fusion techniques common to the life sciences (Castanedo 2013) combined with quasi-experimental strategies (i.e., natural experiments) to approximate the randomization of treatment by leveraging the dramatic and recent state-specific changes in the prevalence of C-sections in the US. We fuse large-scale representative data from the NVSS and NSCH to understand the extent to which the probability of exposure to caesarean delivery during a critical period of immune development influences later childhood health outcomes.

Although purposeful randomization between caesarean and vaginal delivery is unavailable, we design our study to leverage time- and geographically based variation unrelated to maternal–child need or maternal preference for C-section, such as variation in hospital practices, physician delivery preferences, and changes in practice guidelines and policies. To approximate an experiment with our data, we parse the data into 600 groups based on the observable characteristics of state, sex, race, and Hispanic origin, and achieve ‘quasi-random’ assignment to groups by eliminating any variation between the groups based on these observed differences through the use of fixed effects controls. These observable characteristics are measured in both the NVSS and NSCH data sets, and present a mechanism for fusing the data in a way that enables more precise estimation than an analysis at the state level alone.

We link individuals in these groups in the NSCH data to their risk of C-section through the data fusion technique described, where the universe of births in the US (NVSS data) is used to construct the proportion of births delivered by C-section in each time/place/demographic group. Group fixed effects are used so that we leverage only idiosyncratic variation in C-sections that is unrelated to the maternal–child characteristics influencing delivery. Because we are not using individual-level information on caesarean delivery for each child in our data, our analysis is in the spirit of an intent-to-treat effect rather than ‘treatment-on-the-treated’. We ask whether individuals in our constructed groups who faced a higher idiosyncratic likelihood of exposure to C-section also experienced higher likelihoods of childhood health conditions.

In addition to the group-level fixed effects, we also control for trends by sex, race, and Hispanic origin in the full sample, and trends at the state and national levels. This means that the variation in our study can only be generated from within-group changes over time that are not absorbed by our controls. Thus, the variation in C-sections and children’s health outcomes are generated by small state- and time-specific divergences from the average full sample trends for the groups. Such divergences are less likely to be systematic, and hence they represent the quasi-random variation we seek to exploit. Our assumption, given our controls, is that this variation is unrelated to individual choice or individual maternal–child medical need for C-sections and, thus, can be used to generate natural experiments in exposure to C-section practices to estimate effects on child health outcomes. The remaining variation we exploit is likely to represent idiosyncratic variation in provider preferences, hospital practices, and policy and guideline changes of the sort outlined in the ‘Background’ section.

The existence of a substantial amount of variation in C-section rates is necessary to extract the exogenous variation we use in our analysis. There is considerable variation in C-section rates in our data, and Figure 2 presents an example of a portion of that variation. The chart represents variation in C-section rates for mothers in California and Wisconsin from 1990 to 2011, broken down by race and sex of the child. It demonstrates that these two states broadly follow the national trend, but with varying trajectories. We exploit this geographical variation, as well as differences in the prevalence of C-section by race and sex. For example, in California, the incidence of C-section was generally higher among blacks than whites, varying a bit over time, but this was the opposite in Wisconsin, which saw lower C-section rates among blacks than whites over most of the period. C-section rates were frequently higher for boys than girls in both states up until 2004, when this pattern reversed. Also notable is that some subgroups broke with the national trend in some years, for instance, with black female babies in California following a different trajectory in and just after 1995.

Figure 2.

C-section as a proportion of births, by race and sex, 1990–2011, California (CA) and Wisconsin (WI)

Source: As for Figure 1.

Our estimation strategy uses ordinary least squares (OLS) regression, implemented using Stata Version 15 (StataCorp 2017), to estimate the association between C-section rates and child health outcomes controlling for likely demographic confounders and building in several fixed effects to eliminate time-invariant confounding. We first estimate equation (1) using the Stata command ‘regress’ with the cluster option, in which we link aggregated NSCH health data by state, s, birth year, t, sex of child, g, child race, r, and child Hispanic origin, h, to C-section practices, MOD, for similarly parsed subgroups (s, t, g, r, h), and cluster by state, race, sex, and Hispanic origin. In this basic model (Model 1), we control for age trends through age-specific fixed effects ($A_a$), for time trends through year-specific fixed effects ($T_t$), and for covariates ($X$), which are the dummy variables for male, black, ‘other race’, and Hispanic origin. Error terms are cluster-robust.

$$Healt{h}_{istgrh}= {\beta }_0+ {\beta }_{1}MO{D}_{stgrh}+ X{\beta }_2+ A_a+ T_t + {\epsilon }_{istgrh}$$ (1)

We build an extended number of fixed effects into Equation (2), using the Stata command ‘areg’ with the ‘cluster’ and ‘absorb’ options to isolate quasi-experimental variation. In addition to age- and year-specific fixed effects, we include fixed effects interactions between time and state to control for average state trends over time, and fixed effects interactions between time and sex, time and race, and time and Hispanic origin to control for differential changes in these groups’ C-section rates over time at a national level. Additionally, equation (2) includes a fixed effect for the state-by-sex-by-race-by-Hispanic-origin groups generated when linking the NSCH and NVSS data sets, and clusters by state, sex, race, and Hispanic origin. We use OLS regression in our statistical analysis, making these models linear probability models because health outcomes are binary at the individual level. We also tested models in which we generated average state-by-sex-by-race-by-Hispanic-by-birth-year rates of health outcomes, a continuous measure, and achieved similar results. We present the individual-level results to retain as much variation as possible, along with cluster-robust standard errors.

$$\begin{gathered} Healt{h}_{istgrh}= {\beta }_0+ {\beta }_{1}MO{D}_{stgrh}+A_a+T_t+\left(State \times T_t\right)+\left(Male \times T_t\right)+ \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\left(Black\times T_t\right)+\left(Other Race\times T_t\right)+\left(Hispanic \times T_t\right) + \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\left(S_s\times G_g\times R_r\times H_h\right)+{\epsilon }_{istgrh} \end{gathered}$$ (2)

The intuition behind the final model (Model 2) is that we are able to measure the relationship between within-group variation in C-section rates and variation in health outcomes among individuals in these groups, controlling for age trends, time trends for the nationally representative sample, and time trends interacted with sex, race, and Hispanic origin for the nationally representative sample. For example, we are able to examine the relationship between C-section rates among mothers of white, Hispanic girls born in Wisconsin in 1990 and individual-level health outcomes of white, Hispanic girls born in Wisconsin in 1990, controlling for their age, average changes in C-section rates in that year nationally, and average changes in C-section rates for all whites, for Hispanics, and for girls across the nation. These fixed effects sweep away the variation due to differences between groups, which dramatically decreases the chances of residual confounding compared with the other models.

Results

Table 2 presents the results of the two models from the equations outlined in the ‘Methods’ section, for each of the three health outcomes. With regard to the covariates in Model 1, males are significantly more likely to be diagnosed with food allergy and asthma, and equally likely to be diagnosed with eczema when compared with females. Food allergy, asthma, and eczema are significantly more likely to be diagnosed among blacks than whites in the sample, and children of Hispanic origin are significantly more likely to receive a diagnosis of asthma compared with non-Hispanics, but slightly less likely to be diagnosed with food allergy or eczema. Finally, children categorized as ‘other race’ receive significantly more diagnoses of food allergy, asthma, and eczema than whites.

Table 2.

Regression of child health outcomes on exposure to C-section; US children born 1990–2011

	Food allergy		Asthma		Eczema
	Model 1	Model 2	Model 1	Model 2	Model 1	Model 2
C-section rate	0.058*	0.106+	0.132***	0.028	−0.009	0.001
	(0.022)	(0.055)	(0.034)	(0.060)	(0.034)	(0.082)
Male	0.004*	–	0.037***	–	−0.002	–
	(0.002)	–	(0.003)	–	(0.003)	–
Black	0.008***	–	0.082***	–	0.074***	–
	(0.002)	–	(0.004)	–	(0.005)	–
‘Other race’	0.012***	–	0.038***	–	0.026***	–
	(0.002)	–	(0.005)	–	(0.004)	–
Hispanic	−0.005*	–	0.012***	–	−0.007*	–
	(0.002)	–	(0.004)	–	(0.003)	–
Observations	156,492	156,492	249,238	249,238	156,563	156,563
R-squared	0.004	0.017	0.028	0.039	0.013	0.027
Age and time fixed effects	Yes	Yes	Yes	Yes	Yes	Yes
Extended fixed effects	No	Yes	No	Yes	No	Yes

Notes: Cluster-robust standard errors in parentheses. Model 1 reports the regression results from equation (1), the effect of C-section rates on child health outcomes, controlling for demographic covariates and for age and time fixed effects. Model 2 reports the regression results for equation (2), the effect of C-section rates on child health outcomes, controlling for age, time, state-by-time, sex-by time, race-by-time, Hispanic-by-time, and state-by-sex-by-race-by-Hispanic fixed effects. C-section rates were generated from NVSS data for state-by-sex-by-race-by-Hispanic-by-birth-year subgroups and joined to individual-level health outcomes from NSCH data.

^***p<0.001

^**p<0.01

^*p<0.05

⁺p<0.1

Source: Authors’ calculations from NVSS data and NSCH waves 2003, 2007, and 2011.

Results from the first model show that C-section rates are positively and significantly predictive for food allergy and asthma, but not predictive of eczema. In the case of asthma, for instance, a one percentage point increase in the C-section rate is associated with a 0.132 percentage point increase in the condition. With the C-section rate increasing by approximately twelve percentage points over the study period, this translates to an estimated 1.584 percentage point increase in asthma over the same period. However, these results still have two sources of variation at work. The first is the natural experimental variation (i.e., variation within state-demographic groups in exposure to C-section over time) that is our focus. The second is the variation between state-demographic groups over time that conflates demographic group exposure with the natural experiments. For example, if much of the variation is related to demographic differences—such as if black children are exposed to higher rates of C-section and also experience higher rates of poor health—then the analysis is likely to be identified partly from variation between demographic groups. We introduce additional fixed effects to reduce demographic differences and further isolate within-subgroup variation over time.

The results of Model 2, our preferred specification, suggest that a one percentage point increase in C-section rate is associated with a 0.106 percentage point increase in food allergy, significant at the 10 per cent level, as well as a 0.028 percentage point increase in asthma, although the latter association is not statistically significant. The effect size for eczema remains close to zero.

Model validation

To probe the validity of our use of state-by-year-by-subgroup variation in exposure to C-section as a set of natural experiments, we conduct a variety of false hypothesis (or falsification) tests. To do so, we remove the child health variable from our model and replace it with a variable indicating a predetermined individual characteristic. These variables are either fixed before the occurrence of a C-section (and hence are temporally unable to causally impact the health outcomes of interest) or are unlikely to be directly causally related to the health outcomes tested, given the fixed effects in the model. The intuition behind these tests is that C-section should not be predictive of these variables, once the fixed effects variables included in the model have eliminated a large degree of residual confounding. For instance, the percentage of fathers born outside the US is a fixed individual characteristic, which was established before C-section and should not be significantly associated statistically with subgroup level C-section exposure measures, given the fixed effects included in the preferred specification. In all, we test whether C-section predicts mother’s or father’s country of birth, the highest educational attainment of an adult in the household, the primary language spoken at home, and poverty rates. Summary statistics for these five variables are found in Table A1 in the Appendix. Results of the falsification tests are presented in Table A2 in the Appendix and show that our ‘basic’ fixed effects fail to capture the confounding between these predetermined individual characteristics and C-section rates fully, but that our ‘extended’ fixed effects are able to reduce these confounding associations. These results are consistent with our assumption of approximating a set of natural experiments.

Limitations

We are mindful of several potential limitations of our analysis. First, although we are aware of many compelling mechanisms linking variation in C-section rates with longer-term child health outcomes, we do not have data that measure these mechanisms. It is possible that factors unrelated to the microbiome, such as the physical process of labour and vaginal delivery, may be mediating the relationship between C-section and children’s health outcomes. Evidence has demonstrated that labour and vaginal delivery work to expel liquid from the baby’s lungs and that full lung aeration is delayed in infants born by caesarean delivery (Jain and Eaton 2006; te Pas et al. 2008). Without measures of the microbiome and other mechanisms, we are unable to gauge the extent to which different mediators, such as breastfeeding or the physical process of labour, play stronger, weaker, or competing roles in explaining the relationship between method of delivery and children’s health.

Second, while our use of natural experiments is meant to undo potential confounding—if our assumption that our experimental variation (within subgroups and over time) is not related to individual-level confounding factors is true—we cannot directly test this, as we do not have complete historical records for individual subjects over time. Thus, it is possible that confounding variables that influence both C-section and children’s health outcomes and have been changing over time within state-demographic subgroups could influence our results. However, our falsification tests give us confidence that the fixed effects model is adjusting for important confounders. Moreover, the advantage of our analysis is that we are able to control for an important domain in spite of data limitations.

A third concern is measurement error that may affect our results. First, interstate migration could introduce bias to our results due to the mis-assignment of C-section exposure rates to children’s health outcomes. For example, migration may bias our findings insofar as mothers at high risk of having children with poor health tend to migrate to states with increasing use of this delivery method over time and, conversely, mothers at high risk of having children with good health tend to migrate to states with decreasing use of C-section. We note that interstate migration rates are quite low and that interstate migration declined markedly over the period of study (Kaplan and Schulhofer-Wohl 2017). Also, due to the nature of the data fusion design, treatment is assigned at the group level, rather than the individual level. In this regard, we are deploying an intent-to-treat design rather than a treatment-on-the-treated. Thus, in order to gauge a treatment-on-the-treated effect size, our coefficients would need to be inflated. Additionally, measurement error may exist in our dependent variables. While issues such as inaccurate diagnosis of health conditions or self-reporting bias—which may entail such problems as failure to recall health diagnoses or failure to understand or respond honestly to the questionnaire—are likely to be unrelated to our aggregate-level exposures (e.g., subgroup trends in C-sections), and therefore are unlikely to bias our results, they may reduce the precision of our estimates. Finally, our large set of fixed effects is likely capturing both confounding effects and idiosyncratic variation in C-section rates that would serve as useful variation to identify our effects of interest. ‘Over-controlling’ the idiosyncratic variation reduces our ability to estimate some of our coefficients precisely. These sources of error suggest that our findings for food allergy, asthma, and eczema may be underestimated in terms of both the effect sizes and statistical significance.

Finally, by using group fixed effects based on state, year, sex, race, and Hispanic origin, we generate subgroups based, in some instances, on small numbers of individuals. For example, in states with very little racial diversity, the number of black respondents contributing to the health outcome measures may be quite small. This lack of representativeness might be a concern as it could bias our results. One apparent solution to this issue is to group the data at a less granular level, for instance, by controlling for state-by-sex fixed effects only. However, doing so would also generate a less precise measure of the C-section rate and would reduce the strength of the fixed effects in the model, increasing the likelihood of confounding.

Discussion

Although the relationships between C-section and child health outcomes may seem small, when we consider the dramatic rise in C-section rates and concurrent child health problems over time, the coefficients become more impactful. To put the magnitude of the coefficients in perspective, we return to the food allergy descriptive statistics in Table 1 to estimate the extent to which the increases documented over time might be attributable to C-section. In our sample, the food allergy prevalence stood at 4.2 per cent in 2003 and increased to 5.2 per cent by 2007. Between 2003 and 2007, C-section rates increased by around four percentage points. We calculate that the total increase in food allergy prevalence tied to the increase in the C-section rate during this period is 0.424 percentage points, hypothetically moving the prevalence measure from 4.2 per cent in 2003 to an estimated 4.624 per cent by 2007. This means the increase in C-section rates from 2003 to 2007 could explain 42 per cent of the increase in prevalence of food allergy across this period. When the incidence or prevalence of a health condition is low, as is the case for food allergy, even a small percentage point increase or decrease can mean a substantial percentage change in rates over time. Considering that average C-section rates had increased to 32 per cent nationally by 2015 (Martin et al. 2017), from 5.5 per cent in 1970 (Taffel et al. 1987), it is possible that this increase provides substantial explanatory power for the long-term trends in some child health conditions, such as food allergy.

These results suggest a negative total causal effect of C-section on subsequent child health, at least in the case of food allergy. Causal mediation analysis should be undertaken as a next step in this line of research, in order not only to determine the direct effect of C-sections on children’s health, but also to understand whether certain mediators, such as breastfeeding or antibiotic use, transmit and moderate the relationship to a greater extent than others. We are just beginning to understand how the microbiome modulates children’s health, and because many of the mediators known to link caesarean delivery with children’s health outcomes are also known to change the microbiome or interact with it, future work at this intersection will be important. For example, experimental manipulation of the microbiome at the individual level provides one exciting avenue for future research. Indeed, some studies have shown that probiotic treatment can help reduce or eliminate food allergy (Tang et al. 2015; Hsiao et al. 2017).

Individual-level clinical studies must be replicated and expanded. However, when small sample sizes and a lack of external generalizability hamper such studies, population-level quasi-experimental designs can prove useful. In this regard, our research overcomes a number of limitations found in the previous literature by using new methods and the fusion of multiple data sets. Our use of longitudinal, population-representative data and quasi-experimental methods is a distinct strength, and our results contribute a new perspective to a literature dominated by clinical studies.

Appendix

Construction of child’s age in the NSCH

We calculate birth year by subtracting the child’s age from the survey year. The survey year spans two calendar years in each wave of the NSCH, but we do not have data on the exact date the survey was fielded, so we calculate birth year based on the survey year with the greater number of months represented.

Construction of race measures

The NVSS does not contain full data on child’s race, but does have data on parents’ race, which we use to construct an NVSS child race variable that matches the race variable in the NSCH. We also construct a Hispanic origin measure based on NVSS data by assigning a child to a binary Hispanic origin variable if either or both parents designated Hispanic origin on the birth certificate. After organizing the individual data from the NVSS, we generate C-section rates for subgroups based on the various combinations of state, birth year, race, Hispanic origin, and sex. We consolidate the child’s race measure in the NSCH from four categories (‘white only’, ‘black only’, ‘multiracial’, and ‘other race’) to three categories (‘white only’, ‘black only’, and ‘other race’) to be able to join it to the equivalent measure in the NVSS data, which does not similarly differentiate ‘multiracial’ and ‘other race’. In the NVSS data, we construct the same three-category race measure based on child’s race data. When race data are unavailable for a child, we use parents’ race data to generate the measure. We assign children of parents of different races to the ‘other race’ category. If mother’s race is reported, but not father’s race, we assign the child the mother’s race only. Although this could result in some mis-assignment of race categories, we feel the associated measurement error is low, given that the data show very low rates of interracial childbearing. The NSCH is missing race data for approximately 5 per cent of the sample and is missing Hispanic origin for approximately 2 per cent of the sample, most of whom are also missing race data.

NVSS missing data

The state of Oklahoma did not report method of delivery in 1990 on the birth certificate and thus is not included in the analysis until 1991.

In some state-year combinations, Hispanic origin or race information was unstated or unrecorded. Further, some states recorded no births to groups characterized by specific race and Hispanic origin combinations in some years. This resulted in 33 NSCH observations being dropped from the study.

Table A1.

Additional summary statistics for US children born 1990–2011, by survey year

	Mean/proportion (standard deviation)
Survey year:	2003	2007	2011
Poverty level1	5.697	6.003	5.627
	(2.415)	(2.364)	(2.609)
US-born father2	0.904	0.900	–
	(0.294)	(0.299)	(–)
US-born mother2	0.904	0.896	–
	(0.294)	(0.305)	(–)
Highest adult education3	2.734	2.740	2.378
	(0.511)	(0.543)	(0.718)
Primary language at home4	0.950	0.953	0.927
	(0.218)	(0.211)	(0.261)
Observations	70,789	85,971	92,824

¹Poverty level is an eight-category measure ranging from at/below 100 per cent to 400 per cent of the poverty level and above. A poverty level average between categories five and six corresponds to average incomes of 200–300 per cent above the poverty level.

²US-born father and mother measures are the proportion of fathers/mothers born in the US.

³The highest adult education measure has three categories (below high school, high school, above high school); it represents the highest level of education of any adult in the household in survey year 2003 and is calculated as the highest education among mothers, fathers, or non-parent respondents in survey years 2007 and 2011.

⁴The primary language at home is the proportion of children for which English is the primary language spoken at home.

Notes: Each survey wave includes children aged 0–17, but because method of delivery data were not collected before 1990, the 2003 survey year includes a narrower range of birth years. We do not include observations with missing race, Hispanic origin, or C-section data in the summary statistics. Summary statistics report the maximum number of observations across all variables in each survey year, thus hiding minor differences in the number of missing observations by health outcome. For this reason, the summed number of observations across survey years does not perfectly correspond to the total observations in Table A2, which reports observations for each health outcome across multiple survey waves.

Source: As for Table 1.

Table A2.

Results of falsification tests predicting predetermined individual characteristics instead of child health outcomes

	Adult education1		Primary language2		US-born father3		US-born mother3		Poverty level4
	Model 1	Model 2	Model 1	Model 2	Model 1	Model 2	Model 1	Model 2	Model 1	Model 2
C-section rate	−0.606***	−0.127	−0.199***	0.021	−0.340***	0.074	−0.286**	−0.057	1.456	0.400
	(0.173)	(0.127)	(0.048)	(0.053)	(0.098)	(0.110)	(0.097)	(0.101)	(0.933)	(0.435)
Male	0.008	–	0.000	–	0.002	–	0.001	–	−0.026	–
	(0.012)	–	(0.004)	–	(0.006)	–	(0.006)	–	(0.074)	–
Black	−0.212***	–	0.012***	–	−0.063***	–	−0.030*	–	−1.675***	–
	(0.015)	–	(0.003)	–	(0.014)	–	(0.012)	–	(0.105)	–
‘Other race’	−0.076***	–	−0.084***	–	−0.196***	–	−0.196***	–	−0.680***	–
	(0.013)	–	(0.009)	–	(0.018)	–	(0.018)	–	(0.084)	–
Hispanic	−0.436***	–	−0.350***	–	−0.392***	–	−0.372***	–	−1.559***	–
	(0.019)	–	(0.014)	–	(0.021)	–	(0.022)	–	(0.063)	–
Observations	245,537	245,537	249,478	249,478	124,137	124,137	147,632	147,632	230,853	230,853
R-squared	0.114	0.144	0.240	0.303	0.170	0.259	0.162	0.253	0.090	0.145
Age and time fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Extended fixed effects	No	Yes	No	Yes	No	Yes	No	Yes	No	Yes

Notes: Model 1 reports the regression results from equation (1), the effect of C-section rates on outcomes that should be temporally or substantively unrelated, controlling for demographic covariates and for age and time fixed effects. Model 2 reports the regression results for equation (2), the effect of C-section rates on outcomes that should be temporally or substantively unrelated, controlling for age, time, state-by-time, sex-by-time, race-by-time, Hispanic-by-time, and state-by-sex-by-race-by-Hispanic fixed effects. C-section rates were generated from NVSS data for state-by-sex-by-race-by-Hispanic-by-birth-year subgroups and joined to individual-level health outcomes from NSCH data.

Cluster-robust standard errors are in parentheses.

¹The highest adult education measure has three categories (below high school, high school, above high school); it represents the highest level of education of any adult in the household in survey year 2003 and is calculated as the highest education among mothers, fathers, or non-parent respondents in survey years 2007 and 2011.

²The primary language at home is a binary variable capturing whether English is the primary language spoken at home.

³US-born father and mother measures are binary variables capturing whether a child’s father/mother was born in the US.

⁴Poverty level is an eight-category measure ranging from at/below 100 per cent to 400 per cent above the poverty level. A poverty level average between categories five and six corresponds to average incomes of 200–300 per cent above the poverty level.

^***p<0.001

^**p<0.01

^*p<0.05

⁺p<0.1

Source: As for Table 2.