ABSTRACT
We analyze the long-term effects of potentially avoidable cesarean sections on children’s health. Using Finnish administrative data, we document that physicians perform more unplanned C-sections during their regular working hours on days that precede a weekend or public holiday and use this exogenous variation as an instrument for C-sections. We supplement our instrumental variables results with a differences-in-differences estimation strategy that exploits variation in birth mode within sibling pairs and across families. Our results suggest that avoidable, unplanned C-sections increase the risk of asthma, but do not affect other immune-mediated disorders previously associated with C-sections.
I. Introduction
There is little doubt that prenatal health and early childhood circumstances can have long-term effects on mortality, morbidity, and human capital development. The theory of the developmental origins of adult health and disease has proven to describe a surprisingly general phenomenon. The effects of prenatal health conditions and early-life events extend to a wide spectrum of educational, cognitive, behavioral, and demographic outcomes (Almond, Currie, and Duque 2018).
In human development, the transition from fetal to newborn life at birth is an abrupt event that represents major physiological challenges for the neonates. There is accumulating evidence that many medical and operative interventions at birth are associated with long-term health. Most notably, cesarean delivery for low-risk pregnancies is associated with a wide variety of adverse short- and long-term health outcomes. However, the causal nature of these relationships has received little attention.
The most prominent mechanism thought to mediate the long-term effects of cesarean sections on health and disease emphasizes the importance of early exposure to a diverse range of microbes that adjust the human immune system to appropriately react to extrauterine environment. This general class of mechanisms is often dubbed either as the hygiene hypothesis (Strachan 1989) or the old friends hypothesis (Scudellari 2017). According to these hypotheses, children born by cesarean section lack the beneficial exposure to their mother’s vaginal microbiome and are more prone to develop immune-mediated diseases.
The cesarean section is the most commonly performed major surgery in many countries. Understanding the consequences of cesarean sections on later-life health and human capital development is important from a number of perspectives varying from clinical decision making to economic and health policy. The rapidly growing incidence of cesarean sections across the globe suggests that even small increases in mortality and morbidity due to C-sections would lead to large reductions in life expectancy and substantial losses of human welfare.1
This paper provides new evidence on the effect of potentially avoidable cesarean sections on several relevant health outcomes. To identify the causal effect and abstract from cases where C-sections respond to a clear medical indication, we exploit variation in medical decision-making depending on the type of day and time of birth. We show that the probability of unscheduled C-section increases substantially during the normal working hours (8 a.m.−4 p.m.) on working days that precede a leisure day. Importantly, we find that these excess C-sections are not driven either by selection of different mothers giving birth at these times or by advancing births that would have been cesarean deliveries in any event.
Using fine-grained data on birth times and intrapartum diagnoses, we show that the increased likelihood of cesarean sections during the normal working hours on days that precede a leisure day is coupled with the increased use of more discretionary diagnoses. Moreover, we observe that this change in doctors’ assessment does not affect mothers who work in a medical profession. Our data lend significant support for the contention that the excess numbers of unplanned cesarean deliveries observed during the normal working hours on days that precede a leisure day are largely supply driven. We use this time variation as an instrument for C-section birth. We provide a detailed discussion and numerous robustness checks to support the validity of the required identification assumptions.
We investigate the effects of cesarean sections on infant and child outcomes using a comprehensive and precise administrative data resource that includes birth and health records for all children born in Finland between 1990 and 2014. We follow entire birth cohorts from birth to teenage years and use detailed diagnosis data to study the causal effects of cesarean sections on children’s health. We focus on outcomes whose onset is hypothesized to be influenced by cesarean delivery: asthma and other atopic diseases, type 1 diabetes, and obesity. These are among the most common chronic conditions in childhood (Torpy 2010).2 Our instrumental variable estimates suggest that avoidable C-sections increase the probability of asthma diagnosis from early childhood onward. This effect is clinically and economically relevant. However, we do not find consistent evidence that cesarean sections affect the probability of developing atopic diseases at large, type 1 diabetes, or obesity.
We complement our instrumental variables estimates using a differences-in-differences model with family fixed effects that compares the health gap between siblings in families where the second child was born by unplanned C-section with the health gap between siblings who were born by vaginal delivery. The results from our supplementary empirical strategy support our main findings. These estimates suggest that unplanned C-sections increase the risk of childhood asthma and enable us to rule out meaningful effects on other atopic diseases, type 1 diabetes, and obesity. We provide several sensitivity checks that suggest that the effect on asthma is unlikely to be explained by negative selection.
Our results are consistent with the hypothesis that the mode of delivery may influence the development of the immune system and have long-term effects on health and disease. However, our results paint a more nuanced picture about the long-term effects of cesarean deliveries than existing evidence based mostly on associations. We show that controlling for the observable characteristics that most previous studies have accounted for is not enough to deal with the endogeneity of the type of birth. Our findings suggest that C-sections cause a much narrower spectrum of diseases than currently hypothesized and call for a careful analysis on the relationships between the delivery mode and long-term health.
Our paper relates to an important literature estimating the effects of early interventions on long-term health and human capital development. Moreover, we contribute in at least three ways to a nascent economics literature on the effects of treatment choices at birth. First, we investigate the long-term effects of unplanned C-sections on children. To evaluate the costs and benefits of C-sections, it is crucial to investigate long-term effects, as potential alterations of the immune system and long-run consequences of C-sections are not necessarily visible at birth and in early childhood. Moreover, we report age-by-age estimates for entire cohorts from birth to teenage years and provide evidence about the effects of early-life events during the middle childhood, thus expanding our knowledge about the “missing middle” years.3 Existing papers investigating the effects of potentially avoidable C-sections have concentrated on neonatal outcomes or short-term effects.4 Costa-Ramón et al. (2018) investigate the effects of cesarean sections on neonatal health using time variation in unplanned C-section rates. Card, Fenizia, and Silver (2019) study the short-term health effects of hospital delivery practices using relative distance from a mother’s home to hospitals with high and low C-sections rates.5
Second, we study the effects of discretionary unplanned C-sections that could potentially be avoided, while existing papers have not been able to separate planned (elective) and unplanned C-sections or have concentrated on C-sections with a clear medical indication. Hannah et al. (2000), Jensen and Wüst (2015), and Mühlrad (2017) show that breech babies can benefit from C-section delivery. However, these results concern medically necessary C-sections in a specific high-risk group and do not readily generalize to avoidable unplanned C-sections or cesarean deliveries in general. While C-sections are often life-saving at the top of the risk distribution (Currie and Macleod 2017), more evidence is required about the effects of discretionary C-sections that could be potentially avoidable.
Third, to evaluate the causal effects of C-sections, we use two different identification strategies based on somewhat different assumptions. Our instrumental variable strategy builds on previous work using time variation in C-section rates in combination with high-quality administrative data. Moreover, we employ a differences-in-differences research design that has not been used in previous studies on C-sections. For both methods we provide several pieces of evidence that support the credibility of the identification assumptions. Thus, by using two different strategies, we hope to provide more reliable evidence on the causal effects of avoidable unscheduled interventions at birth on children both in the short and long run. Furthermore, we reconcile our findings with those from previous associational studies by showing that controlling for even a very rich set of observable characteristics is not enough to deal with the endogeneity of delivery mode.
Section II provides background information about the biological mechanisms hypothesized to mediate the effects of mode of delivery on infant outcomes, the different types of cesarean sections, and the institutional context of our analysis. Section III introduces the data, provides key descriptive statistics, and lays out our econometric approach. Section IV reports our main results. Section V presents robustness checks and additional evidence to support our main conclusions. The last section concludes.
II. Background
A. Mechanisms
A large body of literature documents the developmental origins of health and disease. The process of labor can be seen as one crucial step in adaptation to the extrauterine environment. The prevailing evidence highlights the role of vaginal delivery as an important early programming event with potentially life-long consequences (Hyde et al. 2012). While there is strong consensus that medically indicated cesarean sections decrease the risk of fetal death at birth, the absence or modification of vaginal delivery has been linked to several adverse health outcomes and anomalies in human development. In the following, we summarize some of the most widely acknowledged findings to understand how C-sections might have long-lasting effects on health and human development.
It is well recognized that early exposure to microbes is necessary to train the human immune system to react appropriately to environmental stimulation. The original formulation of the theory, dubbed the hygiene hypothesis, states that the lack of early childhood exposure to infectious agents and symbiotic microbes increases susceptibility to multiple autoimmune diseases by suppressing the natural development of the immune system (Strachan 1989). Lately, refinements to the original formulation, known as the old friends hypothesis, have challenged the role of infectious pathogens and highlight the importance of early exposure to a diverse range of harmless microbes to strengthen the human immune system and combat the threat of environmental pathogens (Scudellari 2017).
The mode of delivery may affect early exposure to microbes through several channels. First, bacteria from the mother and the surrounding environment colonize the infant’s gut during birth (Neu and Rushing 2011). Exposure to the maternal vaginal microbiota is interrupted in a cesarean birth and externally derived environmental bacteria play an important role in the infant’s intestinal colonization. Consequently, infants delivered by C-sections acquire a microbiota that differs from that of vaginally delivered infants (Dominguez-Bello et al. 2016). Second, the transfer of microbiota continues through breastfeeding after birth. Breast milk contains a number of bioactive components that can have an important impact on infant’s microbiota composition and health (Collado et al. 2015). The negative association between cesarean sections and the initiation of breastfeeding provides an additional mechanism to explain the differences in microbiota by type of birth (Prior et al. 2012).
The potential biological mechanisms are consistent with the reported associations between cesarean delivery and adverse infant outcomes. These studies relate cesarean deliveries to a marked increase in the susceptibility of multiple immune and metabolic conditions. Even though cesarean deliveries have been associated with a broad array of immune-mediated diseases, recent meta-analyses conclude that C-sections are most robustly related to asthma, atopic diseases, type 1 diabetes, and obesity (Blustein and Liu 2015; Keag, Norman, and Stock 2018; Cardwell et al. 2008; Thavagnanam et al. 2008; Peters et al. 2018; Bager, Wohlfahrt and Westergaard 2008).6 However, the causal nature and clinical relevance of these relationships remains largely unknown.7
B. Classification of Cesarean Sections
Cesarean sections are performed for several indications at different stages of the pregnancy. Cesarean sections are classified either as scheduled (elective) or unscheduled operations. Scheduled C-sections occur without attempted labor and are agreed upon in advance. The large majority of scheduled C-sections are performed during the regular working hours (8 a.m.−4 p.m.) from Monday to Friday. Medical indications that make scheduled C-sections advisable include, among others, multiple pregnancies with noncephalic presentation of the first fetus or placenta previa. We exclude all scheduled C-sections from our sample.
Most C-sections are not scheduled and happen after spontaneous or medically induced onset of labor. Unscheduled C-sections are surgeries where an attempt of vaginal birth is transformed to a cesarean delivery after the mother has been admitted to a hospital. Unscheduled C-sections are classified by urgency. Emergency C-sections are performed within 30 minutes of the decision, due to an immediate threat to the life of the mother or the baby (NICE 2011). However, most unscheduled C-sections are performed without such immediate threat. The optimal timing and indication for these operations are imprecise and give large discretion to the clinician. Slow progression of labor or cephalopelvic disproportion are examples of diagnoses that may require an unplanned nonurgent cesarean section. There is wide variation among clinicians in the use of discretionary diagnoses that justify C-sections (Barber et al. 2011; Fraser et al. 1987). Our data contain the registered diagnosis linked to the C-section for a subsample of births. These observations enable us to verify that the peaks in unplanned C-sections are coupled with the use of more discretionary diagnoses.
C. Institutional Context
Finland has universal public health coverage. Comprehensive pre- and postnatal care services are publicly provided. There are no private medical institutions running maternity wards. Consequently, all deliveries take place in public hospitals. All medical expenses related to prenatal care, delivery, and postnatal care are fully covered by the public healthcare system.
Pregnant women usually give birth in the nearest hospital. Only high-risk pregnancies are systematically directed to a higher-level hospital for obstetric care and delivery. Expectant women do not have pre-assigned midwives or physicians for the delivery. Midwives take care of the delivery in all hospitals, while physicians have the ultimate responsibility for obstetric care, decide on the type of delivery, and perform C-sections. Physicians’ financial incentives to perform a C-section are negligible in Finland. The total income of doctors working in public hospitals consists of basic salary (63 percent), on-call and overtime payments (28 percent), personal supplements (for example, years of service) (7 percent), and performance-based pay components (for example, fees per appointment or procedure) (3 percent) (Finnish Medical Association 2016). The C-section rate (15.5 percent in 2015) is relatively low from an international perspective (OECD 2017).
The regular working shifts for physicians are from 8 a.m. to 4 p.m. from Monday to Friday. The on-call hours for physicians may not exceed 24 hours during the regular working week and last typically from 8 a.m. to 8 a.m. On weekends, the on-call hours for physicians are from 8 a.m. to 9 a.m. on the next day.8 Midwives follow the same rotation regardless of the type of day and work in three eight-hour shifts.9
While our data do not contain details about the organization of hospital resources across different shifts, we have collected data on the aggregate number of midwives and physicians by weekday and work shift for three different hospitals: two university hospitals and one small hospital. The qualitative evidence from these three hospitals, described in Online Appendix Table A1, shows that staff availability does not vary between working days.
III. Data and Methods
A. Data
The two main data sources used in our analysis are the Finnish Medical Birth Register and the Hospital Discharge Register. The Finnish Medical Birth Register was established in 1987. This administrative data resource includes data on all live births and on stillbirths of fetuses with a birth weight of at least 500 grams or with a gestational age of at least 22 weeks. The register includes information on maternal background, healthcare utilization, and medical interventions during pregnancy and delivery. It also includes mothers’ diagnoses during delivery (ICD-10 codes) and newborn outcomes until the age of seven days. From 1990, the register contains detailed information about the type of C-section (scheduled vs. unscheduled). These data are collected at all delivery hospitals.10
We exclude from our sample planned C-sections and multiple pregnancies. For our instrumental variable strategy, we focus only on first births.11 Our analysis sample includes 392,560 deliveries that took place from 1990 to 2014. For the differences-in-differences analysis, we focus on both first and second births from families where the first child was born by vaginal delivery (more details are provided in Section III.B.2). The analysis sample consists of 645,292 children from 322,646 sibling pairs. There are 43 hospitals in our sample. Online Appendix Table A2 shows summary statistics for all births in Finland between 1990 and 2014.
We match the Finnish Medical Birth Register to the Finnish Hospital Discharge Register, which contains information about the diagnosed medical conditions, medical operations, and the date of diagnoses. This hospital register contains all inpatient consultations in Finland from 1990 to 2013. From 1998, the data include all outpatient visits to hospitals. All diagnoses are coded using the International Classification of Diseases (ICD) tool.12
We explore two sets of outcome variables. First, to test whether unplanned C-sections have an impact on neonatal health, we analyze indicators of neonatal health included in the birth register. We study Apgar scores one minute after birth, admission to intensive care unit (ICU), need of assisted ventilation, and early neonatal mortality (defined as neonatal death in the first week of life).13 Second, we study longer-term outcomes using detailed inpatient and outpatient diagnosis data from the Finnish Hospital Discharge Register. We use primary diagnoses.14 To maintain a relatively large sample size, we follow individuals from birth until age 15. We focus on the four metabolic and immune-related conditions that have been most robustly associated with cesarean delivery: asthma, atopic diseases (atopic dermatitis and allergic rhinitis), type 1 diabetes, and obesity. Online Appendix Table A3 provides more detail about each of these diagnoses.
B. Empirical Strategy
We aim to estimate the impact of a cesarean delivery on child’s health at birth and older ages. We define a binary variable CSi that takes value of one if the delivery is an unplanned C-section and zero if it is a vaginal delivery. Thus, we aim to estimate the following equation:
1
where Yi is the health outcome of infant i, Xi is a vector of covariates, and δm, λy, ϕh are fixed effects for the month, year, and hospital of birth, respectively.15
The estimation of Equation 1 is, however, likely to provide biased estimates of β1 due to potential selection into cesarean birth.16 To study the causal effects of cesarean delivery on children’s health, we exploit two different empirical strategies.
1. Instrumental variable strategy: Variation by time and type of day
Our instrumental variable strategy exploits the higher likelihood of being born by unplanned C-section during the normal working shift on pre-leisure days compared to regular working days. We use the interaction between the type of day and work shift as an instrument for the mode of delivery.
Figure 1 presents the predicted probability of unplanned C-section delivery by hour and type of day. We adjust for hospital, month, and year-of-birth fixed effects. Figure 1A plots the distribution of C-sections over a 24-hour cycle for working days that precede a leisure day compared to other working days.17 We find that substantially more C-sections are performed during regular working hours on days that precede a leisure day compared to the rest of working days. Figure 1B presents the predicted probability of having an unplanned C-section by work shift and type of day. We find that the gap in C-section rates between a day that precedes a leisure day and the rest of working days emerges only during the regular working hours (8 a.m.−4 p.m.).18
Notes: Panel A presents the predicted probability of unplanned C-section by hour and type of day. Panel B shows the predicted probability of unplanned C-section by shift and type of day. Both figures adjust for hospital, month, and year-of-birth fixed effects. Pre-leisure days include working days that precede a Finnish public holiday or a weekend, while working days include the rest of working days. Sample is restricted to singleton first births that are either unscheduled C-sections or vaginal births.
Importantly, we find that the excess C-sections performed on days that precede a leisure day are not driven by advancing births that would have been cesarean deliveries in any event. We do not observe any relative fall in C-sections during the evening hours preceding a leisure day compared to the evenings of regular working days (Figure 1A) or during the leisure day (Online Appendix Figure A2).19 These observations suggest that physicians perform C-sections during the regular working hours on pre-leisure days that would not have been performed otherwise.
The time pattern of C-sections is consistent with previous work by Brown (1996) and Halla et al. (2020) that documents an increase in C-section rates on days that precede a leisure day. Halla et al. (2020) exploit this variation in an instrumental variable framework to study the impact of delivery mode on maternal fertility and labor supply. The existing literature attributes the pre-leisure anomaly in the time pattern of C-sections to physicians’ leisure incentives that arise from the higher time cost and uncertainty of vaginal births. A cesarean section takes on average 30–75 minutes and is perceived as a relatively easy surgical intervention with low complication rates (NICE 2011). The average duration of labor for first-time mothers who have a vaginal birth is 11 hours (NICE 2014).
We provide two pieces of complementary evidence to validate that the excess rate of C-sections is not driven by medical factors. First, we build on previous evidence that some medical diagnoses linked to a cesarean birth are more discretionary than others. Dystocia (prolonged or obstructed labor), one of the most common indications for primary cesarean section, is believed to provide the greatest room for diagnostic discretion (Fraser et al. 1987). The number of dystocia diagnoses has been shown to strongly respond to physician incentives (Evans et al. 1984; Fraser et al. 1987; McCloskey, Petitti, and Hobel 1992). We examine whether there is an excess number of dystocia diagnoses during regular working hours on pre-leisure days. Our results (Online Appendix Table A5) show that giving birth during the regular hours on a pre-leisure day increases the probability of having a dystocia diagnosis compared to other working days. Importantly, we do not find this temporal pattern for medical emergencies, for which there should not be any room for discretion. In particular, we find that our instrument does not predict additional examinations of the fetus during labor, which doctors should perform if there are any signs of fetal suffering.20
Our second piece of evidence builds on the literature showing that physician mothers are less likely to receive C-sections driven by financial incentives (Johnson and Rehavi 2016). Consequently, we expect that the change in medical criteria during the normal shift on pre-leisure days would not affect physician mothers and other medical professionals. Our results (Online Appendix Table A6) support this hypothesis. We do not find that medical professionals have an increased risk of having a C-section during the regular shift on pre-leisure days, while we do find this increase for mothers with an equivalent level of education who are not employed in a medical profession.21
Overall, these pieces of evidence suggest that the observed increase in C-section rates is supply-driven and consistent with the interpretation of previous studies that have attributed similar increases to physicians’ leisure incentives (Brown 1996; Halla et al. 2020).
We exploit the variation in the probability of unplanned C-sections by time and type of day and adopt an instrumental variable approach. We first estimate a standard two-stage least squares (2SLS) with the following first stage:
2
and the corresponding second stage:
3
where NSi is a dummy that takes a value one for births that take place during the normal shift (8 a.m.–4 p.m.) and zero otherwise, and Preleisurei takes a value one for Fridays or working days preceding a Finnish public holiday and zero for other working days, Xi is the vector of individual controls,22 and δm, λy, ϕh are month, year, and hospital of birth fixed effects, respectively. 
in Equation 3 are the predicted C-sections from the first stage. The interaction between regular working hours and a day preceding a leisure day will serve as an instrument. As a result, we will be comparing mothers who give birth in the same hospital during the same shift, but on different types of days (working days preceding a leisure day or other working days).
Our instrumental variables estimation needs to meet three conditions to yield valid estimates. First, the instrument should strongly influence the probability of C-section (first stage). Second, there should be no selection of mothers who give birth during the regular shift on different types of days. Finally, being born during the regular shift on pre-leisure days, compared to other working days, should only affect child outcomes through the increased probability of being born by C-section (exclusion restriction).
Table 1 shows the results from the estimation of the first stage. Column 1 shows the first-stage estimates including month, year, and hospital fixed effects. Column 2 includes a richer set of controls. These estimates show that being born during the normal shift increases the probability of unplanned C-section for all working days. Moreover, being born during the normal shift on pre-leisure days increases the probability of unplanned C-section by 1.4 percentage points.23 This implies a 9.6 percent increase with respect to the average unplanned C-section rate of 14.5 percent in our sample. This estimate is in line with, but larger than, that by Halla et al. (2020), who document a 3.6 percent increase in C-sections on pre-leisure days, taking into account both planned and unplanned cesareans. The first-stage F−statistics are larger than 25 in both specifications. Following the common critical values for weak instruments (Stock and Yogo 2005), we can reject the null hypothesis that the instrument is weak.
First Stage
Figure 2 shows that our instrument does not predict a large set of maternal and pregnancy characteristics, including medical conditions that could predict a C-section. This indicates that mothers giving birth during the regular shift on pre-leisure days compared to other working days are similar in observable characteristics, suggesting that the observed increase in C-sections at these times cannot be explained by selection.24
Notes: The figure represents the coefficients and 95 percent confidence interval from separate regressions of each predetermined variable (standardized to have mean 0 and standard deviation 1) on the instrument (Normal shift * Preleisure), controlling for normal shift time, pre-leisure day, and hospital, year, and month-of-birth fixed effects. Sample is restricted to singleton births that are either unscheduled C-sections or vaginal births that take place on working days.
Finally, regarding the exclusion restriction, we focus on births that take place on working days, when hospital resources and quality of care should be constant. Moreover, to compromise our empirical strategy, any change in the quality of care would need to happen on pre-leisure days only during the regular working hours. The information about the organization of hospital staff discussed in Section II.C (Online Appendix Table A1) suggests that this kind of change is unlikely in our setting, as there is no evidence of differences in the distribution of hospital resources across shifts between different types of working days.25 In Section V.A, we show that the planned activity in the maternity wards during the regular shift is very similar on pre-leisure days compared to the rest of working days and provide numerous supplementary analyses that further reinforce the credibility of the exclusion restriction.
The 2SLS estimator enables us to identify a local average treatment effect (LATE). This is the effect of C-sections for infants whose mothers’ mode of delivery is sensitive to the subjective assessment of the physician. More accurately, we capture births where the type of day affects the decision of the doctor to perform a C-section during the normal shift. The counterfactual for these births is unlikely to be exclusively a cesarean section later on, given that we do not find a relative drop in C-sections on pre-leisure days after the normal shift or during the following day.
The LATE will not be informative of the effect of medically indicated C-sections, as those will be performed regardless of physician’s incentives. Moreover, the LATE does not capture the effect of unplanned C-sections for babies who had a very fast delivery, leaving no room for physician discretion.
Our primary health outcomes and the endogenous variable are binary. Consequently, besides the 2SLS models we estimate (recursive) bivariate probit models. These specifications mirror Equations 2 and 3 and assume that cesarean delivery (CSi) and the binary indicator of health Yi are determined by the following latent indexes:
4
5
where (νi, ξi) follow a bivariate standard normal distribution with unknown correlation. These equations can be estimated through maximum likelihood. Identification in this setting relies on the same assumptions that are needed to estimate the 2SLS model together with an additional assumption about the joint normality of the error terms. In the Results section we report marginal effects for both estimators.26 Online Appendix Table A8 presents the first-stage results of the bivariate probit estimation, which show very similar marginal effects to those from the linear model in Table 1.
Bivariate probit estimation is expected to present substantial advantages in the context of this work, as it has been shown to be more efficient and less biased than 2SLS when treatment and outcome probabilities are close to zero or one (Chiburis, Das, and Lokshin 2012; Bhattacharya, Goldman, and McCaffrey 2006; Nielsen, Smith, and Celikaksoy 2009). Given that we work in a low C-section rate setting and examine relatively rare outcomes, we expect bivariate probit to outperform 2SLS in terms of efficiency. In the Online Appendix we show, based on Monte Carlo simulations for our particular context, that this is indeed the case. The results from this exercise reveal that, with a C-section rate around 15 percent and having diseases with a prevalence equal or lower than 5 percent as outcomes, 2SLS results are largely uninformative due to their lack of precision. Bivariate probit results, in turn, are much more efficient, while being comparable to 2SLS in terms of unbiasedness.
2. Differences-in-differences
Our second empirical strategy applies a differences-in-differences approach to a sample of sibling pairs. We restrict the sample to families where the older sibling was born by vaginal delivery and compare the health gap between siblings in families where the second child was born by an unplanned C-section with families where the second child was born by vaginal delivery. This enables us to control for all time-invariant unobserved heterogeneity at the family level and the effect of birth order. Our empirical strategy builds on numerous studies that have used siblings fixed effects to estimate the impact of health shocks while in utero or after birth (for example, Oreopoulos et al. 2008; Almond, Edlund, and Palme 2009; Almqvist et al. 2012; Aizer, Stroud, and Buka 2016) and extends the model to a differences-in-differences specification with family fixed effects. Black et al. (2017) used a related approach to study the impact of child disability on sibling outcomes.
We estimate the following equation:
6
where Yif is the health outcome of child i in family f, Secondbornif is a dummy variable equal to one for the second child and zero for the first child, CSif is an indicator equal to one for unplanned C-section and zero for vaginal delivery, Xif is a vector with the same pregnancy and maternal controls of Equation 3, except for maternal characteristics that are time-invariant, and diagnoses during pregnancy and delivery (like prolonged and obstructed labor),27 γf, δm, λy, and ϕh are family, month, year, and hospital of birth fixed effects, respectively.28 We cluster standard errors at the family level. Our parameter of interest is ψ2, which identifies the change in the health gap between siblings in families where the first child was born by vaginal delivery and the second child by C-section compared to families where both children were born by vaginal delivery.
We do not include families whose older child was born by C-section for two reasons. First, mothers who have a C-section in the first delivery and vaginal birth in the second delivery are a very selected sample, given the very high probability of having a repeat C-section.29 Second, some studies find that having a C-section is associated with lower fertility (Halla et al. 2020; Keag, Norman, and Stock 2018). We abstract from these concerns by focusing on mothers whose first birth was a vaginal delivery.
Even though our rich data sources make it possible to control for a large set of observable characteristics, it could be that there are sibling-specific unobservable differences that vary within family. In particular, younger siblings born by C-section could be negatively selected compared to their vaginally delivered older siblings if the cesarean delivery is caused by complications, either during the pregnancy or delivery, which we cannot observe in our data. These unobservable complications could cause our estimates to be upwardly biased. Thus, our differences-in-differences estimates could overestimate the impact of C-sections on the different diagnoses. In Section V.B, we assess the magnitude of the potential bias and provide evidence that it is relatively small. We will nonetheless keep the direction of this bias in mind when interpreting the results from this strategy.
IV. Results
A. Neonatal Outcomes
We first estimate the impact of C-sections on neonatal outcomes. We study one-minute Apgar scores, admission to the intensive care unit (ICU), assisted ventilation, and neonatal (seven-day) mortality. Table 2 shows our ordinary least squares (OLS), 2SLS, bivariate probit marginal effects, and differences-in-differences estimates. For each estimation method, we report in square brackets Romano–Wolf p-values adjusted for multiple hypothesis testing (Clarke, Romano, and Wolf 2020).30
Neonatal Outcomes
We find that the OLS results replicate existing findings. Cesarean sections are associated with adverse outcomes at birth and higher neonatal mortality.31 Our 2SLS estimates are not significant for any of the outcomes. However, the magnitude of coefficients and large standard errors suggest that we cannot reject that there is a (potentially large) effect on neonatal outcomes. As discussed in Section III.B.1, 2SLS estimates are expected to be particularly uninformative with low treatment and outcome probabilities.
Bivariate probit marginal effects are substantially more precisely estimated than the 2SLS coefficients, yet all point estimates from the bivariate probit models are within the confidence intervals of the 2SLS estimates. The bivariate probit results suggest that unplanned C-sections increase the probability of having a low Apgar score (Apgar lower than seven), being admitted to the intensive care unit, and receiving assisted ventilation. The magnitudes of the bivariate probit marginal effect estimates are similar to OLS estimates. However, we do not find significantly increased mortality risk within seven days after birth. The results from the differences-in-differences models give support to these findings with similarly sized and more precise coefficients. Overall, our results suggest that unplanned C-sections have a negative impact on neonatal health. However, these adverse effects do not translate into a higher probability of early neonatal mortality.
B. Later Child Health
We now turn to the long-run effects of C-sections on health outcomes. Table 3 shows the OLS, 2SLS, bivariate probit, and differences-in-differences marginal effect estimates at ages five and ten. We analyze health conditions that have been extensively documented in the literature as being positively associated with cesarean deliveries: type 1 diabetes, obesity, asthma, and other atopic diseases (atopic dermatitis and allergic rhinitis). Romano–Wolf p-values adjusted for multiple hypothesis testing are reported in square brackets (Clarke, Romano, and Wolf 2020). We will show year-by-year bivariate probit and differences-in-differences estimates up to age 15 in Figures 4 and 5, respectively. Online Appendix Figure A4 shows OLS estimates. Given that we study health outcomes for children who are born from 1990 to 2014, the sample size decreases as we consider older ages.
Child Diagnoses by Age
Notes: These figures plot the cesarean section coefficients and their 95 percent confidence interval from the OLS estimation of the association between C-section and the probability of being diagnosed with asthma, atopic diseases, type 1 diabetes, or obesity before age five. In each figure, each coefficient corresponds to a different specification with varying sets of controls that are sequentially added: no controls ; the list of controls included in at least 50 percent of medical studies; controls included in at least 20 percent of them (more information in Online Appendix Table A9); the rest of maternal, pregnancy, and delivery characteristics included in our usual group of controls (described in Section III.B); hospital fixed effects; year fixed effects, month fixed effects; excluding planned C-sections, and finally including family fixed effects.
Notes: The figure plots the marginal effects from the bivariate probit estimation of the effect of unplanned C-section on the probability of each diagnosis by age, with our usual specification. All regressions include hospital, year, and month-of-birth fixed effects and the full set of controls as described in Section III.B.1.
Notes: The figure plots the coefficient of unplanned C-section for each diagnosis by age in family fixed effects models. All regressions include family, hospital, year, and month-of-birth fixed effects and the full set of controls as described in Section III.B.2.
1. OLS results
The OLS estimates (Table 3 and Online Appendix Figure A4) suggest that cesarean sections are associated with a higher probability of asthma, obesity, and atopic diseases. These findings are consistent with existing studies that have documented significant associations between cesarean sections and metabolic and immune-related conditions. However, we do not detect that C-sections are associated with a higher probability of a type 1 diabetes diagnosis.
To compare our findings with the results from previous studies, we repeat in Figure 3 the OLS estimation for the probability of having each disease by age five and show how the coefficient changes as we vary the set of included controls. We reviewed the literature on the relation between cesarean delivery and these diseases and for each outcome made a list of the most common control variables.32 In these regressions, we include all C-sections (planned and unplanned), given that most studies are not able to control for this.33
The first coefficient in each panel of Figure 3 shows the estimate when no additional covariates are included in the regression. We then start to cumulatively add sets of controls. First, we add those included in at least 50 percent of the reviewed papers, followed by those included in at least 20 percent of them (see Online Appendix Table A9 for a detailed list). In the next specification, we add the rest of the variables included in our usual group of controls (described in Section III.B) that previous literature was not taking into account.34 Finally, in the last five columns, we sequentially add hospital fixed effects, year-of-birth fixed effects, month-of-birth fixed effects, we restrict the analysis to unplanned C-sections, and we include family fixed effects.
The pattern of this figure points out that controlling for a richer set of observables, as well as including fixed effects, does not significantly change the magnitude and sign of the coefficients. Three types of controls, however, seem to have a somewhat larger effect on the estimates. First, we see for most outcomes that estimates become smaller when controlling for the more extensive list of covariates included in previous papers (the 20 percent list) compared to controlling for the shorter list of controls included in 50 percent of them. A key difference between these two groups is that the 20 percent list includes controls for whether the mother had been diagnosed, before the child’s birth, with that particular disease. The second substantial difference in estimates arises when we restrict the sample to exclude scheduled C-sections. Finally, estimates also become closer to zero when family fixed effects, which control for unobserved family heterogeneity, are included.
In any case, even with the richest set of controls, our main conclusion holds: C-sections are strongly associated with an increased risk of being diagnosed with asthma, obesity, and atopic diseases. Controlling for observable characteristics is not enough to alter this finding. For type 1 diabetes we only see a significant association when the smallest list of controls is included. This association vanishes as we include a stricter set of controls and fixed effects. This reconciles the findings from our OLS model in Online Appendix Figure A4 with the results from previous literature.
2. Instrumental variables results
The 2SLS results suggest that unplanned C-sections increase the probability of having a type 1 diabetes diagnosis before age five, even though the effect is not significant by age ten. The effect size of the estimate is large, but very imprecise. Our results suggest a nine percentage point increase in the probability of type 1 diabetes, but are consistent with an increase ranging from 6.3 to 12.5 percentage points. The 2SLS estimates for asthma are not significant. However, the lack of precision does not enable us to rule out even very large (positive or negative) effects. For instance, the estimates by age five suggest that the impact of C-sections may range from −4.2 percentage points to 18.4 percentage points. Finally, the 2SLS estimates for obesity and atopic diseases are not significant, but also too imprecise to rule out very large effects.
Similarly to our results for neonatal outcomes, the bivariate probit estimates (marginal effects) are substantially more precisely estimated than the 2SLS coefficients. Yet, practically all point estimates from the bivariate probit models are within the confidence intervals of the 2SLS estimates. For type 1 diabetes, the marginal effect is much smaller than the coefficient from the linear model and not significant. For asthma, the results suggest a significant increase in the probability of a diagnosis by age five of 0.031 (95 percent confidence interval 0.022–0.04). Even though estimates are noisier and no longer significant by age ten, the results in Figure 4 show that unplanned C-sections significantly increase the probability of an asthma diagnosis for children as young as two years old. The effect is statistically significant up to age nine. For obesity, the bivariate probit results are precisely estimated at zero at age five (0.001, 95 percent confidence interval 0.000–0.002) and age ten (0.003, 95 percent confidence interval 0.000–0.006). However, the results in Figure 4 show a statistically detectable effect from age 11. Finally, we do not find a significant impact on atopic diseases at age five or ten.
3. Differences-in-differences results
The differences-in-differences results are very similar to the bivariate probit results. We find that the second-born child has substantially greater risk of having an asthma diagnosis by age five than the firstborn child in families where the second child is born by C-section. Similarly to the bivariate probit estimates, Figure 5 shows that this effect is significant from ages one to eight. Despite the fact that our differences-in-differences estimates could be upwardly biased (Section III.B.2), we do not find any significant effects on obesity, atopic diseases, or type 1 diabetes. These results reinforce the conclusion that C-sections do not have impact on these outcomes.
4. Overview of results
Overall, our results suggest that unplanned C-sections increase the probability of suffering from asthma during childhood. The magnitude of this effect differs slightly depending on the estimation method. The bivariate probit estimates indicate a slightly larger but more imprecisely estimated impact (around two percentage points on average for ages five to ten) than the estimates based on differences-in-differences analysis (1.3 percentage points). By comparing these estimates to the sample mean, we find that the less precise bivariate probit estimates suggest a 36 percent increase in the probability of having asthma diagnosis (compared to the sample mean of 5.5 percent over ages 5–10), while the differences-in-differences estimates suggest a 21 percent increase (compared to the sample mean of 5.8 percent). The latter is closer to the 20 percent increase in the risk of asthma that is documented in recent meta-analyses (Thavagnanam et al. 2008; Keag, Norman, and Stock 2018).
Our analysis indicates that C-sections do not increase the probability of type 1 diabetes or atopic diseases. For diabetes, we can rule out effects larger than 0.7 percentage points at age five using the bivariate probit model and larger than 0.1 percentage points using the differences-in-differences model. For atopic diseases, in turn, our results discard effects larger than 1.2–1.3 percentage points with both methods. Finally, bivariate probit results suggest there might be an effect of C-sections on obesity after age 11. This observation is consistent with the evidence that puberty is a vulnerable period for the development of overweight and obesity (Lobstein, Baur, and Uauy 2004). However, our analysis is not conclusive in this regard because the results from the differences-in-differences estimation do not corroborate this finding. For younger ages, all methods suggest that there is no impact on obesity. For instance, estimates at age five enable us to rule out effects larger than 0.3 percentage points.
One potential limitation of our analysis is that we study diagnoses made at inpatient or outpatient visits to a hospital. For some outcomes, these diagnoses may be a good approximation to the true prevalence of the disease, while for other diseases hospital diagnoses may lead to underestimation. A previous study documents that in Finland practically all new type 1 diabetes diagnoses are made in a hospital and listed in the Hospital Discharge Register (Harjutsalo 2008). This evidence implies that we are able to observe practically all type 1 diabetes diagnoses in our population of interest. However, since 1994, diagnoses for asthma in Finland are often made by general practitioners (Tuomisto et al. 2010). Thus, we are likely to trace only the most severe cases of asthma. The same might be true for atopic disease and obesity.35 In any case, an OLS estimation with the control variables included in most existing studies, or even with a richer set of controls, still yields a significant association of cesarean birth with these hospital diagnoses.
V. Validity Checks
A. Exclusion Restriction and Sensitivity Checks
Our instrumental variables strategy relies on the assumption that the interaction of regular working hours and days that precede a weekend or public holiday affects health outcomes only through its impact on the likelihood of cesarean sections. We argue that, in this setting, this is likely to hold, since a violation would require other changes to happen on days that precede a public holiday but only during the regular shift. In the following, we provide several pieces of evidence that support the credibility of this assumption.
First, we show that, conditional on the type of birth, being born during the regular working hours on pre-leisure days does not have any significant correlation with health outcomes. Online Appendix Figure A6 shows the reduced-form relationship between having an asthma diagnosis at different ages and the instrument, conditional on being born by C-section (in the first panel) or being born by vaginal delivery (in the second one). The fact that there is no significant reduced-form effect within type of birth is consistent with the instrument being related to child health only through the increased probability of C-section.
Second, we explore the overall activity at maternity wards across the different types of days. The first panel of Online Appendix Figure A7 shows the proportion of planned cesarean sections by time of birth and type of day. We find that scheduled activity is organized very similarly during all working days. The same can be seen in the first column of Table 4, where we regress an indicator for planned activity (equal to one for planned C-section or induction) on our instrument and do not find evidence of any significant difference. Moreover, we compare the number of births by type of day and weekday (Online Appendix Figure A7, second panel) and do not find any evidence of maternity ward crowding during the days that precede a public holiday.
Exclusion Restriction Validity Checks
Third, we explore the quality of care provided during different weekdays. The first panel of Online Appendix Figure A8 shows that the probability of having a low Apgar score (below 7) does not differ between weekdays or type of day, suggesting that the quality of care during labor and delivery does not differ by type of day. The second panel of Online Appendix Figure A8 shows the probability of early neonatal mortality, defined as death of a live-born baby within the first seven days of life, by weekday and type of day. We expect that this measure would capture changes in the quality of care after birth. We do not find evidence that early neonatal mortality is higher for babies born on days that precede a public holiday compared to other weekdays. We further explore this issue in Columns 2 and 3 in Table 4, where we analyze if during the normal shift on preleisure days the likelihood of suffering a medical negligence during childbirth is higher. To proxy for this, we analyze if there is an increased risk of shoulder dystocia or brachial plexus injury.36 We do not find any evidence that our instrument predicts either of these outcomes, suggesting that the likelihood of medical negligence is not higher during the regular shift on these types of days.
Additionally, we do not find that mothers who have a C-section during the normal working hours on a day that precedes a public holiday have a longer length of stay than mothers who have a C-section at other times. We explore this in Column 4 in Table 4, where we regress mother’s length of stay on the instrument for the sample of mothers who delivered by C-section. Finally, we explore in Column 5 if our instrument predicts a higher risk of postpartum complications for the mother. This is an indicator equal to one if, within one week after the delivery, the mother is diagnosed with any complication related to the puerperium.37 We do not find any significant difference in this risk on preleisure days in general, and, during the normal shift in particular, the results suggest if anything a reduction in complications. We interpret all these findings as evidence that the quality of care remains constant across different types of working days during the normal shift.
Fourth, since babies born on days that precede a public holiday or weekend stay in the hospital during the following nonworking days, one could argue that their quality of postnatal care is worse compared to children born on other working days. This would be constant for both babies born during the regular shift and at other times and, hence, would not necessarily compromise the exclusion restriction. Yet, we assess this concern. Table 5 shows the marginal effects from bivariate probit regressions that restrict the sample to babies born on Thursdays or Fridays.38 We find, despite the reduced sample size, that the results from this estimation are consistent with our main results.
Validity Checks
Finally, we report in Figure 2 that mothers who give birth during the regular working hours on days that precede a public holiday do not have higher probability of having induced labor. However, the induction of labor is likely to offer more room for discretionary behavior, in which case the decision to perform a C-section might be more sensitive to physician’s subjective assessment.39 In other words, we expect that mothers whose labor has been artificially induced are more likely to be part of the complier population. Column 3 in Table 5 shows that our coefficients remain about the same if we exclude mothers whose labor was induced from our sample. The same conclusion holds if we exclude inductions from our differences-in-differences estimation. These results suggest that our findings are not driven by mothers whose labor has been induced after an admission to the maternity ward.
B. Differences-in-Differences Validity Checks
The results from our differences-in-differences model with family fixed effects could be biased if there are unobservable characteristics correlated with the mode of delivery that vary within family and across siblings. Under this scenario, this methodology would yield upwardly-biased estimates. However, as shown in Section IV.B.3, our differences-in-differences results suggest that C-sections do not increase the risk of developing various immune-mediated diseases that have previously been associated with cesarean births.
To assess the extent to which our results could be influenced by selection, we first run a regression using birth weight as a placebo outcome, given that it cannot be affected by unplanned C-sections. Table 5 shows that our differences-in-differences model with family fixed effects does not predict birth weight. This result supports the validity of this strategy: family fixed effects, jointly with the large set of controls, seem to be taking into account general health differences between siblings born by C-section and vaginal delivery.
Second, we compare our differences-in-differences estimates to those from other samples of sibling pairs where we expect the second child to be negatively selected with respect to their older sibling, but where none of them were born by C-section. These samples include: (i) a sample of siblings where the first child is born by eutocic birth and the second child is born either by eutocic or by instrumented birth and (ii) a sample of siblings where the firstborn had a low-risk pregnancy and the second-born had either a low- or a high-risk pregnancy, while all children in the sample were born by vaginal delivery.40 Consequently, we assess the health gap between siblings across families that had a complication during the second birth or during the second pregnancy, compared to families where none of the siblings encountered any of these complications during pregnancy or birth.
Table 6 shows our differences-in-differences estimates using these samples of siblings. The first four columns show that, compared to families where both siblings were born by eutocic birth, second children born by instrumented vaginal delivery have worse neonatal health than their older siblings who had an eutocic birth. We find a significantly higher probability of having low Apgar scores and of being admitted to the ICU (top panel). In the bottom panel, we can see that children who experienced a high-risk pregnancy do not have significantly worse neonatal health by any of the indicators, even though all coefficients have a positive sign. In the last four columns, we explore if negative selection leading to instrumented birth or high-risk pregnancy is associated with a higher probability of having any of the diagnoses we analyze in Section IV. We do not find evidence that siblings born by instrumented vaginal delivery or those who had a high-risk pregnancy have an increased risk of type 1 diabetes, asthma, atopic diseases, or obesity at age five. These observations suggest that our differences-in-differences results for asthma are unlikely to be explained by negative selection.
Validity of Differences-in-Differences
C. Placebo Regressions
In this section, we perform a falsification test and compare the performance of the OLS estimator with that of our instrumental variable strategy and differences-in-differences model. In particular, we analyze, as a long-term placebo outcome, the probability that the child has a diagnosis related to an injury, poisoning, or other consequences of external causes at different ages.41 Results of this exercise can be found in Online Appendix Figure A9. Using OLS, we find that C-sections are associated with an increased risk of the child suffering a hospital admission due to external causes. However, as we would expect, we do not find a causal relationship between C-sections and these diagnoses with either our instrumental variables or our differences-in-differences approach. This result reinforces the validity of both approaches to deal with omitted variable bias when examining the impact of C-sections on children’s health.
VI. Conclusions
This work provides new evidence on the effects of avoidable cesarean sections on various short- and long-term health outcomes. We use a novel instrumental variable estimation strategy to overcome the potential endogeneity of birth mode and abstract from cases in which C-sections respond to a clear clinical indication. Our empirical strategy builds on the finding that unplanned C-sections are more common during regular working hours on Fridays and working days preceding public holidays. We complement this empirical strategy by estimating a differences-in-differences model with family fixed effects that compares the health gap between siblings in families where the second child was born by unplanned C-section with the health gap between siblings who were both born by vaginal delivery.
Our results suggest that C-sections have a substantial negative impact on neonatal health. However, these adverse effects are not severe enough to translate into a higher probability of increased neonatal mortality. Our long-run analysis follows children from birth to age 15 and investigates the impact of C-sections on four health outcomes that have been consistently associated with C-sections: type 1 diabetes, asthma, obesity, and atopic diseases. In contrast to the OLS estimates, our instrumental variable and differences-in-differences estimates show that unplanned C-sections do not have a significant effect on the probability of having a type 1 diabetes, obesity, or atopic disease diagnosis. However, we do find that being born by an unplanned C-section increases the probability of having asthma. This effect is detectable from ages one and two and of similar size to the associations reported by previous studies (Thavagnanam et al. 2008; Keag, Norman, and Stock 2018).
Our results are consistent with the hypothesis that mode of delivery can affect the development of immune-related conditions, but suggest more nuanced effects of C-sections than previous work. We try to reconcile our results with those in the literature, and we find that controlling for the observable characteristics that most previous studies were including is not enough to deal with the endogeneity of birth mode. We illustrate the importance of such endogeneity by performing a placebo test and showing that, even with a rich set of controls, OLS results suggest that children born by C-section are more likely to suffer unrelated health problems, like external injuries. Both of our identification strategies, in turn, survive this falsification check and suggest that there is no causal effect of cesarean sections on this placebo outcome.
We provide novel evidence on the long-term effects of unplanned C-sections that do not respond to a clear medical indication, using inpatient and outpatient data for all children born in Finland from 1990 to 2014. Although we are able to observe most of the cases of type 1 diabetes, for some diagnoses (asthma, atopic disease, and obesity), we might be only able to trace the most severe cases, given that these conditions are often treated by general practitioners. Future work should focus on analyzing the impact of C-sections on obesity and other metabolic disorders using primary-care data and anthropometric measurements.
We make use of the detailed diagnosis data to show that variation by time and type of day can be a valid source of variation to investigate the impact of avoidable C-sections. First, we show that mothers who give birth during regular working hours on pre-leisure days are comparable in terms of an extensive list of pregnancy, health, and sociodemographic characteristics to mothers who give birth during these times on the rest of working days. Second, we show that during the normal shift on these pre-leisure days, physicians make greater use of more discretionary diagnoses as justification for the C-section. We also show that these additional C-sections are not performed on mothers who work in a medical profession and whose mode of delivery has been shown by the literature not to respond to doctors’ incentives (Johnson and Rehavi 2016).
All in all, our results suggest that the additional C-sections performed during regular working hours on pre-leisure days are not driven by medical factors. A simple back-of-the envelope calculation can shed some light on the potential gains that could result from reducing this practice. Being born during the normal shift on pre-leisure days increases the probability of C-section by 1.4 percentage points.42 Given that the overall C-section rate in Finland from 1990 to 2014 is 16.45 percent, removing these excess cesareans that occur during pre-leisure days would lower the C-section rate by 1.86 percent.43 In 2014, there were 9,534 C-sections in total (out of 57,323 deliveries). This implies that, in absence of this practice, around 180 fewer C-sections would have been performed. Taking into account that the cost of a C-section without complications is 1,815 euros higher than that of a vaginal delivery,44 eliminating this variation would result in about 326,700 euros of savings in just one year.
We provide this evidence in the context of Finland, a country with one of the lowest C-section rates in the world (OECD 2017). We would expect this variation to provide an even stronger source of identification in other countries with higher rates of medical interventionism during childbirth. Thus, we hope this paper provides a solid base upon which future research on the effects of avoidable cesarean sections can be built.
Footnotes
The authors thank Tania Barham, Pilar García-Gómez, Libertad González, Kristiina Huttunen, Ajin Lee, Matilde Machado, Ciaran Phibbs, Marcos Vera-Hernandez. and participants at UPF Health and Applied seminar series, HGSE Labor and Public Economics seminar, and at ESPE 2018, AES 2018, EuHEA 2018, PAA 2019, SEHO 2019, ASHEcon 2019 and iHEA 2019 conferences for their comments and suggestions. They are extremely grateful to Ritva Hurskainen and Soile Kivijärvi from Hyvinkää Maternity Hospital for the introduction to the daily routines of a modern labor ward. This study has received financial support from the Finnish Institute for Health and Welfare for the data access charges. Costa-Ramón, Kortelainen and Rodríguez-González have nothing to disclose. Sääksvuori has received research funding over the past three years from Yrjö Jahnsson Foundation and the Academy of Finland. The final data provided to the authors are de-identified; thus, the research does not constitute human subjects research, and Institutional Review Board (IRB) approval is not required. This paper uses administrative healthcare and employment data maintained by the Finnish Institute for Health and Welfare and Statistics Finland. Healthcare data are regulated under the Act on the Secondary Use of Health and Social Data (552/2019) and can be obtained by sending a direct request to the Finnish Institute for Health and Welfare (https://thl.fi/en). The Finnish Longitudinal Employer-Employee Data can be obtained by sending a direct request to Statistics Finland (https://www.stat.fi). The authors are willing to assist in making data access requests.
Supplementary materials are freely available online at: http://uwpress.wisc.edu/journals/journals/jhr-supplementary.html
↵1. Cesarean section rates have increased in the United States from 20.7 percent in 1996 to 32.9 percent in 2009 (Currie and Macleod 2017). In OECD countries, the rate of cesarean sections has increased from 20 percent in 2000 to 25 percent in 2013 (OECD 2013). Currently, the highest rates of cesarean sections are reported in many of the world’s most populous countries including, among others, China (41.3 percent in 2016) and Brazil (55.6 percent in 2015). Boerma et al. (2018) review the disparities in C-section use around the world.
↵2. Understanding and quantifying the potential contribution of C-sections to the development of these diseases is not limited to medical practice and health policy. Chronic health conditions cause an immense financial burden to households and public healthcare financing. The total cost of asthma in the working age population was estimated to be $24.7 billion during 1999–2002 in Europe (Global Asthma Network 2018). The two other atopic diseases we investigate imply high costs. Atopic dermatitis has been estimated to cost at least $5.3 billion (in 2015 USD) in the United States (Drucker et al. 2017). The estimated annual cost of allergic rhinitis is in the range of $2–5 billion (in 2003 USD) (Reed, Lee, and McCrory 2004). Type 1 diabetes has been found to cost $14.4 billion a year in medical costs and lost income in the United States (Tao et al. 2010). Finally, childhood obesity, which has been on the rise in recent years, has been calculated to imply $19,000 per child in lifetime medical costs in the United States (Finkelstein, Graham, and Malhotra 2014).
↵3. Almond, Currie, and Duque (2018) discuss that, due to data availability, most of the literature analyzes the effect of early-life events on birth or adult outcomes. This implies that we have little knowledge about how developmental trajectories are affected by policies or shocks experienced over the life course. They refer to this gap in the literature as the “missing middle.”
↵4. To our knowledge, the only paper looking at longer-term effects is by Jachetta (2015). This paper explores the relationship between cesarean deliveries and hospitalizations using regional variation in medical malpractice insurance premia in the United States as an instrument for C-sections. However, the instrument used in that work does not necessarily allow for credible causal inference, since the author finds that higher premia also predict delayed prenatal care, lower birth weight, and reduced gestational age.
↵5. A few studies have also examined the effects of cesarean sections on mothers. Halla et al. (2020) study the effects of C-sections on fertility and maternal labor supply. Tonei (2019) studies the impact on mental health for mothers with breech babies who undergo a C-section. Our findings on children health complement these maternal results and contribute to obtaining a more complete picture of the effect of cesarean sections.
↵6. In addition to health outcomes, the literature has associated cesarean sections with worse cognitive and emotional development (Bentley et al. 2016).
↵7. Hyde et al. (2012) summarize evidence from 14 RCTs that compare the effects of cesarean and vaginal deliveries on infant health. All these studies are small RCTs conducted in populations of at-risk babies (for example, breech delivery). These studies have had exceptionally large problems to achieve target recruitment and do not include long-term follow-ups. Overall, there exist no RCTs to date that would enable one to investigate the long-term effects of cesarean sections on infant health. Hyde and Modi (2012) report evidence from survey studies that investigate the perceived acceptability of randomizing the mode of delivery to address long-term health outcomes in low-risk pregnancies. The perceived acceptability of randomizing the mode of delivery in healthy, term, cephalic, and singleton pregnancies remains low among obstetricians and mothers, suggesting that adequately powered large-scale RCTs to compare the effects of cesarean and vaginal deliveries on long-term outcomes may remain unrealized in the near future.
↵8. The head of the department in each hospital has large discretionary power to assign physicians to their working schedules. However, physicians may often independently change their work shifts among themselves. Physicians working in public hospitals are generally not unilaterally able to choose their working schedules. Even though the statutes that govern on-call arrangements have changed in recent years, during most years covered in our data, small hospitals with less than 1,000 annual births could autonomously decide their on-call arrangements. In certain hospitals, physicians were allowed to be at home while on duty, if they could arrive to the hospital within 30 minutes from home.
↵9. An example of midwives’ schedules: (i) 7 a.m.–3 p.m., (ii) 2 p.m.–9.30 p.m., and (iii) 9.15 p.m.–7.15 a.m.
↵10. Home births are also included in the registry but are extremely rare in Finland. From 1996 to 2013, there were 1,053,802 births in total, and only 197 were planned home deliveries (Ovaskainen et al. 2019). We focus on births that took place in hospitals.
↵11. We follow a common practice in the literature and focus on first births, which also allows us to keep just one birth per mother, and abstract from a potential source of correlation between the observations. First-time mothers are also the group of mothers where we find larger variation. Given the faster pace of labor in higher-order births (NICE 2014) and the high risk of repeated C-section, there is less room for discretion in the decision to perform an unplanned C-section in subsequent deliveries. Our results are qualitatively similar but less precise when we include higher order births.
↵12. Diagnoses for years 1990–1995 are recorded using ICD-9 classification. Diagnoses from 1996 onwards are recorded using ICD-10 classification. The quality and completeness of the Finnish Hospital Discharge Register has been assessed in multiple validation studies that have compared recorded data entries with external information. The completeness and accuracy of the data are found to be exceptionally high (Sund 2012). We assess to what extent our data are able to identify the individuals with a certain diagnosis in the Results section.
↵13. Apgar scores result from the examination of the newborn by the midwife or pediatrician one minute after the birth. Five different dimensions are measured and graded from 0 to 2: appearance (skin color), pulse (heart rate), grimace (reflex irritability), activity (muscle tone), and respiration. The resulting score takes values from 1 to 10.
↵14. We replicated all our analysis using both primary and secondary diagnoses. All results remain unchanged. Results are available upon request.
↵15. The vector of covariates includes parity, the gender of the baby, mother’s marital status, nationality, socioeconomic status, age, and smoking status. In addition, we include pregnancy and delivery related indicators that include in vitro fertilization, gestational weeks, high (above 75th percentile) and low (below 25th percentile) number of visits to prenatal clinic, induced labor, prostaglandin pre-induction, epidural use, and laughing gas anesthesia.
↵16. Online Appendix Figure A1 shows that mothers and babies who undergo a C-section are very different from those mothers and babies who undergo a vaginal delivery.
↵17. Working days that precede a leisure day include Fridays and days preceding public holidays. Online Appendix Table A4 documents all public holidays in Finland. Friday is not considered a working day that precedes a leisure day if it is a holiday.
↵18. For our identification strategy we exploit only the variation presented in Figure 1B, that is, the excess C-sections performed during normal working hours on pre-leisure days compared to other working days. Figure 1A shows, however, that within any working day there is also substantial variation in the C-section rate by the time of birth. In particular, more cesareans are performed during normal working hours than during the rest of the day, and we observe the lowest probability of cesarean delivery during the early morning (1 a.m.-7 a.m.). This suggests that physicians try to concentrate most obstetric surgeries during their regular working shift. We do not rely on this variation alone for identification.
↵19. Online Appendix Figure A2 compares the predicted probability of unplanned C-section by hour separately for Saturdays or holidays (the leisure day following the pre-leisure day) and Sundays (a leisure day that is not preceded by a working day). We do not see any relative drop in the C-section rate on Saturdays compared to Sundays at any time of day.
↵20. We examine whether physicians take measurements of intrapartum or fetal scalp pH, which proxies the oxygen saturation of fetal blood during labor.
↵21. Our definition of medical professionals includes physicians, midwives, and nurses. Our observation relates to a large literature on physician-induced demand in healthcare. Since the work of Arrow (1963), it has been recognized that asymmetric incentives between physicians and their patients are a central feature of the medical marketplace. The role of financial incentives on the supply of cesarean sections has been documented by Gruber and Owings (1996). Johnson and Rehavi (2016) observe that financial incentives have a particularly large effect on the probability of having a cesarean section among nonmedical mothers. Our results complement the literature on physician-induced demand and show that the excess rate of C-section on pre-leisure days is restricted to nonmedical professionals.
↵22. Gender of the baby, mother’s marital status, nationality, socioeconomic status, age, smoking status, and the following pregnancy and delivery characteristics: gestational weeks and indicators for in vitro fertilization, high (above 75th percentile) and low (below 25th percentile) number of visits to prenatal clinic, induced labor, prostaglandin pre-induction, epidural use, and laughing gas anesthesia.
↵23. In Online Appendix Figure A3 we explore whether the strength of the first stage varies by type of hospital. We do not find evidence that this variation is driven by any observable characteristic at the hospital level. The first stage is very similar for hospitals of different sizes, for university and nonuniversity hospitals, for hospitals with low and high C-section rate, and for hospitals in high and low populated locations.
↵24. While, ideally, we would like to construct our instrument based on the time of admission to the hospital, this information is not available in our data. One potential concern with using time of birth instead of time of admission is that time of birth might be affected by the type of delivery, as C-sections shorten labor. This could lead to compositional changes in the pool of mothers giving birth during the normal shift on different types of days. However, the comparison of a vast list of maternal characteristics in Figure 2 shows that any such compositional change does not appear to be quantitatively important in our context, as we do not find any significant difference among mothers giving birth during the normal shift on pre-leisure days compared to other working days. Moreover, Costa-Ramón et al. (2018) find that, in a similar context, using an instrument based on the time of birth instead of the time of admission does not affect the results in practice.
↵25. It is important to note, however, that those numbers represent the current situation in a sample of hospitals and may not necessarily be an accurate description of the situation in previous years.
↵26. Bivariate probit models estimate unconditional average causal effects. In contrast, 2SLS estimates the LATE. However, in practice, the average causal effects produced by bivariate probit are likely to be similar to 2SLS estimates (Angrist and Pischke 2009).
↵27. We do not include these diagnoses during labor as controls in the instrumental variables specification, given that we find evidence that they can be an outcome of the time and type of day. The full list of controls is as follows: gender of the child, maternal age, maternal smoking status, maternal weight, and indicators for high or low number of prenatal visits, breech position, hospital visits during pregnancy due to eclampsia, hypertension and placenta previa, abnormal glucose levels, gestational weeks, preterm pregnancy, induced birth, prostaglandin pre-induction, epidural, laughing gas anesthesia, and diagnoses of prolonged and obstructed labor.
↵28. We cannot estimate the baseline effects of the CSif indicator, which are absorbed by the interaction Secondbornif · CSif since by construction only second children have C-sections in our sample.
↵29. In 2010, The American College of Obstetricians and Gynecologists (ACOG) encouraged doctors to allow women to opt for a vaginal delivery after a C-section, but the number of vaginal births after C-section has remained low (American College of Obstetricians and Gynecologists 2010).
↵30. We perform this adjustment for each estimation method including both neonatal and long-run outcomes (all 12 outcome variables explored in Tables 2 and 3) using the RWOLF Stata package (Clarke 2016).
↵31. The OLS estimation is run in a sample that only excludes planned C-sections and births for which we do not observe parity. The specification includes the full set of controls and fixed effects described in Equation 1, as well as controls for birth order.
↵32. We reviewed all the papers cited in the following meta-analyses: Bager, Wohlfahrt, and Westergaard (2008); Cardwell et al. (2008); Darabi et al. (2019); Keag, Norman, and Stock (2018); Li, Zhou, and Liu (2013); and Thavagnanam et al. (2008). Online Appendix Table A9 provides a full list of all papers reviewed for each outcome.
↵33. In Online Appendix Figure A5 we show the number of papers that report the type of C-sections included. Out of 85 papers, 58 include all types of C-sections or do not report which group they include.
↵34. The additional covariates included in our set of controls but not in most previous papers are: mother marital status, high or low number of visits to prenatal clinic during pregnancy, in vitro fertilization, epidural use, laughing gas anesthesia, induced labor, and prostaglandin pre-induction.
↵35. There is some evidence that, among children, ICD-coding underestimates the true prevalence of obesity. ICD-coded cases have a higher BMI and higher healthcare utilization than those not coded (Kuhle et al. 2011).
↵36. Shoulder dystocia is an obstetric emergency that is often unpredictable and is one of the most litigated causes in obstetrics (Politi et al. 2010). This occurs when the baby’s head passes through the birth canal and their shoulders become stuck during labor. If this complication arises, effective and timely clinical management is essential to ensure the well-being of the newborn (Kwek and Yeo 2006). Brachial plexus injury is one of the most important fetal complications of shoulder dystocia (Politi et al. 2010) and consists of an injury to the brachial plexus, that is, the network of nerves that conducts signals from the spinal cord to the shoulder, arm, and hand.
↵37. ICD-10 codes O85–O92.
↵38. The average length of stay in our sample is four days. The majority of babies born on Thursdays and Fridays are hospitalized during the weekend.
↵39. This is despite the fact that recent evidence casts doubt on the commonly held belief that induction of labor increases the risk for cesarean delivery. In particular, recent studies show that inductions at full term do not increase the risk of cesarean delivery (Saccone and Berghella 2015) or even lower it (Mishanina et al. 2014), with no increased risks for the mother and some benefits for the fetus.
↵40. A eutocic delivery is a vaginal delivery with no instrumentation. We define a high-risk pregnancy as a pregnancy where the mother had at least one of these complications: a positive result in the glucose tolerance test, a hospitalization during pregnancy due to blood loss, hypertension, eclampsia, or placenta previa. A low-risk pregnancy is defined as the absence of these issues.
↵41. We exploit the ICD-10 category: Injury, poisoning, and certain other consequences of external causes, ICD-10 Codes: S00–T98.
↵42. This is the first-stage coefficient in Column 2 in Table 1.
↵43. We divide the first stage by the C-section rate and multiply the result by the percentage of days that were pre-leisure days in our sample (21.8 percent).
↵44. These prices are based on the diagnosis-related group (DRG) pricing used by hospitals for internal pricing policies. We take the most recent price list in the largest hospital in Finland (Helsinki University Hospital). In this hospital, the DRG product price for a C-section (without complications) is 4,020 euros and for a vaginal delivery (without complications) 2,205 euros. The information can be found here (in Finnish, accessed on May 14, 2020): https://www.hus.fi/sites/default/files/2020-09/HUS%20Palveluhinnasto%202020%2C%20t[...]0ja%20suoriteperusteiset%20hinnat%20%28osat%201%20ja%202%29.pdf
- Received July 2019.
- Accepted August 2020.
This open access article is distributed under the terms of the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0) and is freely available online at: http://jhr.uwpress.org
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.