Does Unemployment Worsen Babies' Health? A Tale of Siblings, Maternal Behaviour and Selection

We study the effect of unemployment on birth outcomes by exploiting geographical variation in the unemployment rate across local areas in England, and comparing siblings born to the same mother via family fixed effects. Using rich individual data from hospital administrative records between 2003 and 2012, babies' health is found to be strongly pro-cyclical. A one-percentage point increase in the unemployment rate leads to an increase in low birth weight and preterm babies of respectively 1.3 and 1.4%, and a 0.1% decrease in foetal growth. We find heterogenous responses: unemployment has an effect on babies' health which varies from strongly adverse in the most deprived areas, to mildly favourable in the most prosperous areas. We provide evidence of three channels that can explain the overall negative effect of unemployment on new-born health: maternal stress; unhealthy behaviours - namely excessive alcohol consumption and smoking; and delays in the take-up of prenatal services. While the heterogenous effects of unemployment by area of deprivation seem to be explained by maternal behaviour. Most importantly, we also show for the first time that selection into fertility is the main driver for the previously observed, opposite counter-cyclical results, e.g., Dehejia and Lleras-Muney (2004). Our results are robust to internal migration, different geographical aggregation of the unemployment rate, the use of gender-specific unemployment rates, and potential endogeneity of the unemployment rate which we control for by using a shift-share instrumental variable approach.


Introduction
Prenatal events which influence the health of new-born babies are known to affect a large set of outcomes throughout the lifecycle (Almond and Currie, 2011a,b;Almond et al., 2018). These findings heighten the significance of studies looking at the effect of business cycles on birth outcomes, initiated by Dehejia and Lleras-Muney (2004). The results of this literature are mixed, but broadly suggest that in developed countries downturns increase child health. 1 In this paper we study the effect of unemployment on new-borns' health by exploiting geographical variation in the unemployment rate across all local areas in England for a period of ten years, 2004-13, that includes the Great Recession, and by comparing siblings to adjust for selection into fertility. 2 In fact, a major difficulty when studying economic downturns on birth outcomes is that fertility decisions are influenced by the cycle, so that the socio-economic composition of mothers may differ in a recession. Selection is normally accounted for by studying how unemployment is associated with different maternal and birth's characteristics (e.g., Dehejia and Lleras-Muney, 2004;Aparicio and González, 2020). A more direct and rigorous approach is to compare siblings born to the same mother, but who in-utero are exposed to different points of the cycle. This approach of using maternal fixed effects is uncommon because siblings are not usually identifiable from birth registers, which are the standard data source. Our data instead allow us to adopt this identification strategy (see also Salvanes, 2013;Olafsson, 2016;van den Berg et al., 2018).
Besides selection, a second relevant aspect in this type of studies is that the influence of unemployment on new-born health is expected to depend on the families' socio-economic status (SES). For example, parents from low SES might be more susceptible to economic 1 The pioneer work by Dehejia and Lleras-Muney (2004) use administrative birth register data for the United States, and find that new-borns' health is counter-cyclical. Aparicio and González (2020) and (van den Berg et al., 2018) have recently reached similar conclusions for respectively Spain and Sweden. In contrast, (Olafsson, 2016) and (Alessie et al., 2018) have instead found new-born's health to be procyclical in Iceland and the Netherlands. In low-middle income countries, the consensus tends towards pro-cyclicality of babies' health (e.g., Bhalotra, 2010;Bozzoli and Quintana-Domeque, 2014). The same holds true for the medical literature (e.g., Eiríksdóttir et al., 2013;Varea et al., 2016;Kana et al., 2017;Finch et al., 2019). Related economic research shows that layoff announcements (Carlson, 2015), or job displacement (Lindo, 2011) are negatively associated with health at birth.
2 In 2003 there were 1.52 million of unemployed individuals, which increased to 2.63 million in 2012 (ONS). The Great Recession is considered the worst recession since the 1930s. Previous papers such as Dehejia and Lleras-Muney (2004), Salvanes (2013) or van den Berg et al. (2018), have focused on years prior to the Great Recession, which include milder recessions. uncertainties and be less able to cope with the shocks compared to high SES parents (Currie and Duque, 2016). More disadvantaged parents, having fewer resources, might suffer from worse physical and mental health, and encounter more high risk behaviors (Currie et al., 2015b). Our paper builds on research to date by exploring how far recessions impact unequally on the health of babies from different SES groups, and how the mechanisms which drive this impact differ among those groups. This is particularly relevant in the UK where stark differences in infants' health by social class exist (Marmot et al., 2010;Weightman et al., 2012).
By using a unique administrative dataset of 4.8 million births delivered in National Health Service (NHS) hospitals, we find that the health of English new-born babies is negatively associated with an increase in the unemployment rate. Specifically, we show that overall a sibling born in a recession will on average be less healthy, ceteris paribus, on a range of metrics: a one-percentage point increase in the unemployment rate leads to an increase in low birth weight babies of 1.3%, an increase of preterm babies of 1.4%, and a 0.1% decrease in foetal growth. 3 We find that these adverse overall effects of higher unemployment are greater the more deprived is the residential area of the mother, leading to a decrease in birthweight and foetal growth by respectively 0.15 and 0.14%.
Conversely, for babies conceived in the richest areas, a one-percentage point increase in the unemployment rate is associated with an increase in birthweight and foetal growth by respectively 0.09% and 0.08%, but these latter estimates are only marginally statistically significant. We expose these results to several robustness checks by considering maternal internal migration; gender specific unemployment rate; a higher level of geographical aggregation of the unemployment rate, and potential endogeneity of the unemployment rate which we address by using a shift-share instrument (or "Bartik instrument", Bartik (1991)).
We then investigate empirically three mechanisms that may explain both our overall result and the heterogenous effects of recession by using unique medically measured indicators. We explore how high unemployment, by creating financial distress, may affect the pregnancy outcome by increasing maternal stress proxied by stillbirth and baby's gen-der. We find a positive and significant association between unemployment and stillbirth, and unemployment and the probability of having a female baby which might indicate that more male births result in miscarriages given that males are more sensitive to foetal stress. A further mechanism that helps to explain not only the overall effect, but also the heterogenous effects of unemployment on health by degree of deprivation, is a differential change in unhealthy maternal behaviour. We advance related research (e.g., Dehejia and Lleras-Muney, 2004;van den Berg et al., 2018;Aparicio and González, 2020), by using medically verified healthy behaviors (diagnosis registered at the time of the delivery) instead of self-reported measures. We find that mothers living in the poorest areas are more likely to be diagnosed for alcohol and smoking-related health problems, while mothers living in the richest areas undertake less of these risky behaviours as unemployment increases. Lastly, we explore whether conceiving during an economic crisis results in a postponement of the first prenatal visit as registered by the hospital. We find that it does, suggesting that a mother's opportunity cost of time to attend prenatal checks increases in a recession.
Finally, we contrast our results with what we would have found had sibling information not been available and we had not exploited variation between siblings in exposure to unemployment. We show that when pooling all births (all babies' sample), babies' health is counter-cyclical, as in previous studies (Dehejia and Lleras-Muney, 2004;Aparicio and González, 2020). We discuss how the composition of new-borns is influenced by the unemployment rate in our sample and provide suggestive evidence that an increase in unemployment alters in different ways the composition of new-borns in the siblings' sample, and in the all babies' sample. In particular, the economic recession seems to encourage the selection into the siblings' sample of mothers less likely to come from a prosperous area, more likely to have first born instead of higher-order births which are generally heavier, and as a consequence, be less likely to have healthy babies. The opposite appears true for the make-up of the all babies' sample which may explain the contrasting findings. This paper is structured as follows. Section 2 presents the data and the empirical specification. Section 3 presents our main results, sensitivity analysis, and analysis of heterogenous effects. Section 4 shows evidence on three potential mechanisms that drive the socio economic pattern of child health. Section 5 discusses selection into fertility and compositional analysis, while Section 6 concludes.
2 Data and empirical strategy 2.1 Hospital Episode Statistics Data This paper uses the administrative Hospital Episode Statistics (HES) data. HES data provide individual information concerning all inpatients and outpatients admitted to NHS hospitals from 1989-90. Each patient record contains detailed clinical information, patient characteristics, such as age, gender, residence, method of admission, and hospital of treatment. Since our focus is birth outcomes, we restrict our analysis to delivery admissions from 2003 until 2012. Each episode of delivery reports the following information: mother's age, mother's Lower Super Output Area (LSOA) of residence, 4 mother's ethnicity, length of gestation, gestation period in weeks at first antenatal assessment, result of the pregnancy (live birth or stillbirth), number and type of diagnoses, delivery information (method, type of doctors attending the delivery, etc.), date of admission to the hospital, date of discharge, as well as new-born's gender and birth weight. The HES data contain also information about the hospital where the delivery took place, the Primary Care Trust (PCT) of reference, and the GP practice. 5 The full sample, called "all babies' sample", counts 4,778,918 singleton live births. 6 We exclude 22,777 stillbirths, which will be considered in our Mechanism section 4 to study maternal stress. For our main analysis which compares siblings, we focus on mothers who had at least two live births in the period 2003-2012, leading to a sample of 2,384,935 births. About 77% of those are sibling pairs, 19% sibling triplets, and the rest mothers who delivered more than three offspring. Using the baby's birth weight and the length of gestation we create the following outcome variables: birth weight (BW) in grams, low birth weight babies (LBW -dummy variable equal to 1 if birth weight is below 2.5 kg), very low birth weight babies (VLBW -dummy variable equal to 1 if birth weight is below 1.5 kg), length of the pregnancy (in weeks), preterm (dummy variable equal to 1 if the length of pregnancy is below 37 weeks), and foetal growth in grams (ratio between BW and length of the pregnancy).
We also consider the following control variables: baby's gender, month of birth, and mother's ethnicity. 7 To capture the supply of health care that might have changed in our period, we consider the following variables derived from NHS digital data and created at PCT level for each year: number of GPs per 1k population, the number of specialists per 1k population, the number of GP practices, and the number of hospitals. We also incorporate the number of midwives working in each hospital in each year.

Unemployment rate
We use unemployment data from the UK Department of Work and Pensions (DWP). 8 Persons above 17 years, who register at the local employment office, and are actively seeking work are called 'registered unemployed', and receive an unemployment benefit called Jobseeker's Allowance (JSA). We use quarterly data on all claimants, from 2003 until 2012. We calculate a yearly average of claimants by year and Middle Layer Super Output Area (MSOA), 9 and divide this figure by the population at-risk aged 18-64 (working-age population) by year and MSOA, to obtain a measure of the proportion of people claiming unemployment-related benefit. We refer to this as the unemployment rate (UR). For our main analysis we consider unemployment as a whole, but as a robustness check we focus also on gender-specific unemployment, as it could affect babies' health in a different way (see for example van den Berg et al., 2018). Figure 1 reports two maps representing the average unemployment rate pre-recession (2003)(2004)(2005)(2006)(2007), and during and post-recession (2009)(2010)(2011)(2012). As can be seen there is clear geographical variation in UR across local areas (MSOAs). 7 Unfortunately, ethnicity is missing for many individuals, but we include in the ethnicity variable the category ethnicity unknown.
8 The data can be downloaded from http://tabulation-tool.dwp.gov.uk/NESS /BEN/iben.htm 9 There are about 7,200 MSOAs in England and Wales, with an average population of about 7,800 per MSOA. We focus on MSOA instead of LSOA because LSOA are small areas that can be considered more the neighbourhood than the labour market area.

Descriptive statistics
The individual HES data are linked to the unemployment rate, and the health care services variables through the mother's area of residence 10 and the year at conception.
The date of conception is estimated from the HES using the date of admission, the length of gestation, and the week when the first antenatal health assessment is conducted. Table   1 presents descriptive statistics of the main variables considered. In columns (1)-(4), the all babies' sample can be compared with the sample of mothers that are observed at least twice in our period or siblings' sample, and the sample of mothers that delivered only once or one-baby's sample. There are not major differences across the three samples.
The prevalence of LBW babies is 5-6%, the one of VLBW is about 0.7-0.8%, while there are about 6% preterm births. The unemployment rate is slightly higher in the siblings' sample (3.49) compared to the one for the one-baby's sample (3.46). As expected in the siblings' sample, families have bigger size with on average 1.38 children, compared to 0.93 in the one-baby's sample. 11 In the one-baby's sample we also observe a slightly higher prevalence of teen mothers (6.5%) and of old mothers over 35 (22%) compared to the same groups in the siblings' sample (respectively 5.4% and 16%). The opposite is true for the prevalence of 20-24 and 25-29 mothers.
Our main empirical analysis focuses on the siblings' sample (Sections 2.4, 3, and 4), but the all babies' and one-baby's samples will be used to compare our findings with previous work in Section 5.

Empirical strategy
To study the effect of unemployment on babies' health, we compare outcomes of siblings born in different years, and adopt a mother fixed effect approach. This is possible as the hospital data allow us to identify mothers who gave birth more than once in the period observed. We can then control for unobserved time-invariant characteristics of the mother, which may be correlated with selectivity into fertility at times of high unemployment. We estimate the following equation using in turn (i) pooled OLS, and (ii) mother fixed effects (FE): 10 The mother's area of residence is at LSOA level. We link every LSOA with its respective MSOA or PCT to merge the HES with the other data.
11 Parity is the order of the pregnancy including also stillbirths. For example, if a mother's first pregnancy results in a stillborn then parity is 1, and if the second is a live birth then parity is 2.
where Y ijt corresponds to a birth outcome (birth weight, LBW, VLBW, length of the pregnancy or preterm, or fetal growth) for infant of mother i living in MSOA j and conceived in year t. U R jt corresponds to the unemployment rate for MSOA j and year t, and the coefficient of interest is β and indicates the effect of unemployment on the birth outcome. The mother fixed effect is captured by α i and implicitly incorporates an MSOA fixed effect assuming that mothers do not migrate between MSOAs. Year fixed effects are captured by θ t , while γ l represents Local Authority (LA) (indexed with l), specific trends where δ l corresponds to LA fixed effects and t to the time trend. 12 The year dummies control for any year specific factors that could affect both infant's health and the economy. The LA specific trends allow for omitted trends that vary by LA. The standard errors are clustered at the mother level. 13 A variety of sensitivity analyses is discussed in Section 3.2.
Equation 1 is estimated with and without extra individual controls (baby's gender, month of birth, and mother's ethnicity) and PCT-or hospital-level supply care controls. Table 2 reports the results obtained using the sample of siblings to estimate Equation (1) with pooled OLS without (Panel A), and then with (Panel B) the extra controls of month of birth, new-born's gender, mother's ethnicity, and health care supply characteristics. We then introduce maternal FE, without (Panel C) and with (Panel D), the extra controls.

Main results
The pooled OLS estimates (Panels A-B) show that an increase in UR worsens newborn health: a one-percentage point increase in UR leads to a decrease in birthweight by 0.08%, and in foetal growth of 0.07%, with all the controls. If unobserved maternal characteristics are controlled for by adding maternal fixed effects, unemployment is estimated to have a slightly increased influence, and a broader range of health measures 12 In England there are about 150 LAs that most of the times overlap with PCTs. Health care and social services are usually provided at the LA/PCT level.
13 Results do not change if we cluster the standard errors at MSOA level. This analysis is available upon request. are statistically significant, as shown in Panels C and D. From Panel C, a one-percentage point increase in the unemployment rate leads to a decrease of about 0.11% in both birthweight and foetal growth, together with an increase of 1.1% in LBW, and of 0.9% in preterm births. Also, the probability of VLBW increases by 1.7%, but only at a 10% level of significance. If extra controls are added (Panel D) the results reported in Panel C are slightly strengthened. The effect of unemployment on length of gestation is negative as expected in Panel D, but not statistically significant, while a one-percentage point increase in UR leads to an increase of preterm babies by 1.4%. Overall we find that siblings born in adverse points of the cycle are less healthy, and that omitting the family fixed effects reduces the estimated loss of health due to the recession.
The effects reported from the existing literature where mother fixed effects are included are mixed. Olafsson (2016) finds that having being exposed to the crisis in the first trimester leads to an increase in LBW by 1.9 percentage points in Iceland, van den Berg et al. (2018) show that an increase in the unemployment rate by one percentage points results in a decrease in VLBW by about 10% in Sweden, while Salvanes (2013) find no effect for Norway.

Robustness checks
In this section we show that our main results are robust to internal migration of the mother, different geographical aggregation of the unemployment rate, the use of genderspecific unemployment rates, and potential endogeneity of the unemployment rate which we control for by using a shift-share instrumental variable approach.
Mothers' migration In our sample, about 36% of mothers change MSOA of residence from one birth to the next. Given that moving is a choice, our estimates may be biased due to selective migration: namely that mothers with specific characteristics, correlated with health outcomes, choose to move. We estimate Equation (1) by focusing on the sub-sample of mothers who remain residents in the same MSOA. Results are reported in Panel A of Table 3 including the full set of controls, and clustering the standard errors at MSOA level. A one-percentage point increase in UR is associated with a 0.3% decrease in birthweight and foetal growth, which are larger reductions compared to the analysis reported in Panel D of Table 2. While the effects of unemployment on LBW and preterm are higher than the ones reported in Panel D of Table 2, they are statistically significant only at the 10% level. Overall, our main findings are confirmed when focusing on the sub-sample of mothers who never moved even if larger in magnitude. 14 Gender-specific unemployment Here we test whether male or female unemployment affects new-born health in the same way. Male unemployment is usually more sensitive to the business cycle because men tend to work more. However, participation of women in the English labour market is high and the majority of women have a job. 15 Results are reported in Panels B1 and B2 of Table 3. Panel B1 shows that a one percentage point increase in male unemployment leads to a decrease of 0.06% in birthweight, and in foetal growth, with an increase of 0.9% in LBW and 0.8% in preterm babies. These effects are slightly smaller in magnitude compared to the main ones reported in Panel D of Table 2 but qualitatively similar. Panel B2 shows that a one percentage point increase in female unemployment leads to a decrease of 0.2% in birthweight, and foetal growth, an increase of 1.9% in LBW, and of 2.6% in preterm babies. The estimates of the effects of unemployment using female unemployment are more than twice as large as the estimates using overall unemployment rate, suggesting that mothers unemployment has a particularly relevant and detrimental effect on babies' health.

Different geographical aggregation of unemployment Economic conditions might
have different effects on health depending on their level of geographical aggregation (Lindo, 2015). Estimates based on more disaggregated analysis (MSOAs) can be more precise and improve power because they consider variation in unemployment that is masked by more aggregated (LAs) measures. However, unemployment rates at LA include possible spillover effects across MSOAs within a LA. Panel C of Table 3 reports the analysis at LA level. The estimates are slightly larger than those at MSOA level, in line with spillover effects from neighbouring areas that are not taken into account in MSOA level analysis.
14 Any potential form of endogeneity in the unemployment rate, including endogeneity due to selective migration, is taken into account at the end of Section 3.2. 15 In 2014 for example, almost as many women with children (74.1%) participated in the labour force as women with no children (75%) (ONS, 2017). Aparicio and González (2020) only consider male unemployment in the Spanish context, and van den Berg et al. (2018) show that in Sweden, the effect of unemployment on infants' health is mainly driven by male unemployment.
Endogeneity of unemployment Maternal fixed effects control for any time-invariant maternal characteristics. However, unobservable time-varying characteristics could be correlated with both unemployment and the health of babies, leading to biased estimates. Moreover, UR might be measured with error. To overcome these problems, we construct an instrumental variable called the Bartik IV (Bartik (1991) andBlanchard et al. (1992)). 16 We instrument for the MSOA-level UR using a predicted MSOA-level UR, which we create as the weighted average of the national-level URs across industries, where the weights are the MSOA-level fraction of the employed working-age population in each industry a few years before the start of our sample period, i.e. 2001. 17 The instrumental variable estimates (2nd stage) are reported in Panel D of Table 3. The F test is always above 10 indicating that the instrument is relevant. The results are similar even if slightly larger in magnitude compared to the main ones reported in Panel D of

Heterogeneous effects by the Income Deprivation Domain
Given our finding that overall unemployment is bad for babies' health, we now consider whether there are heterogeneous effects across social classes. Disadvantaged women are more vulnerable to changes in economic conditions, and they are more likely to work in low-quality jobs where they are more likely to encounter higher physical and mental health risks (Kim et al., 2008). They might also experience worse health compared to advantaged women (Currie et al., 2015b).
To study heterogeneous effects of unemployment, we group mothers by the level of poverty of their area of residence, as measured by the Economic Deprivation Index (EDI) provided by the Ministry of Housing, Communities and Local Government. The EDI is a measure of deprivation constructed by the UK government at LSOA level and comprises 16 The Bartik instrument has also been recently used in the English context in a paper that investigates the effect of economic conditions on intimate partner violence (Anderberg et al., 2016).
17 More formally, our estimated industry-predicted unemployment rate is: where w jk are the weights corresponding to the fraction of employed working-age individuals in each industry k and MSOA j in 2001; U RU K kt is the national level UR in each industry, k, two domains: Income and Employment. The two domains are given equal weighting within the overall EDI. The EDI is available for the years 1999-2009. 18 For our analysis we consider only the Income Deprivation Domain (IDD), which represents the proportion of people aged under 60 in an area that are living in low income households claiming certain means-tested social security benefits. We construct an index of deprivation for every MSOA using the Income Deprivation Domain of the EDI data available at LSOA level. For our analysis we consider the 2002 IDD, one year before our period, so that it is not endogenously affected by changes to the unemployment rate during the years studied. 19 We create a categorical variable that takes five values corresponding to the five quintiles of IDD, where the first quintile corresponds to the least deprived (IDD1), and the fifth quintile to the most deprived areas (IDD5).
In columns (2)-(3) of Table A.1 in the Appendix we report descriptive statistics of the main variables for the least and most deprived areas -IDD1 and IDD5. The health outcomes are worse for the poorest quintile of IDD, with higher proportions of LBW, VLBW and preterm babies. The unemployment rate is four times higher in the most deprived MSOAs. The least deprived areas have a higher percentage of White population, while in the poorest areas there are more Indians, Bangladeshi, Pakistani and Black mothers. The health care supply is slightly higher in the most deprived areas.
The effects of unemployment on birth outcomes by level of deprivation are depicted in Table 4. We estimate Equation (1) by adding dummies for the IDD quintiles, and the interaction terms between unemployment rate and those dummies. 20 We find that the more an area is deprived, the greater is the effect of unemployment in worsening a babies' health. Thus, babies born to mothers who live in the poorest areas (IDD5) are the ones whose health suffers most from an increase in unemployment: for these babies, a onepercentage point increase in UR is associated with a decrease in birthweight and foetal growth by 0.15%. Conversely, for babies conceived in the richest areas, a one-percentage 18 The index is constructed in a consistent manner over time and can be used to track the progress of deprived neighbourhoods. A different index is the Index of Deprivation (IMD) that includes also other domains such as education, skills and training deprivation, health deprivation and disability, crime, barriers to housing and services, and living environment deprivation. This index is only available for 2004, 2007 and 2010. The EDI was produced to overcome the difficulties in comparing the different IMDs as different methodologies were used. 19 We also consider the Income component of the 1999 EDI and use the 2001 population to create the index at MSOA level (population estimates at MSOA level do not exist pre-2001). Results do not change and are available upon request. 20 The IDD quintiles are time-invariant and are identified because about 36% of the mothers have migrated from one MSOA to another. See Robustness analysis in Section 3.2. point increase in UR is associated with an increase in birthweight and foetal growth by 0.1%, but this is only significant at the 5 and 10% levels, respectively. This suggests that the influence of the cycle on new born health has an important social class gradient.

Potential mechanisms
In this section we provide evidence on three potential channels through which unemployment affects infant's health: maternal stress, maternal health behaviour, and prenatal care. We also explore whether for each channel there is an appropriate response by social class, that could explain the heterogeneous response of babies health to unemployment.

Mechanism I: Stress in pregnancy
There is evidence showing a negative relationship between foetal exposure to maternal stress and both birth weight and length of the pregnancy (Wadhwa et al., 1993;Persson and Rossin-Slater, 2018). Financial stress due to high unemployment may also affect the health of unborn children, and particularly the levels of miscarried, stillborn and aborted children (Hogue et al., 2013). We test this channel in two ways. 21 First, we study if the probability of having a stillborn depends on the unemployment rate. A stillbirth is registered by every hospital when a baby is born dead after 24 completed weeks of pregnancy, and our data allow us to observe mothers who delivered one or more babies in the period of observation, whether live or dead. Second, we study whether the probability of having a female baby is influenced by the unemployment rate. Male conceptions are more sensitive to foetal stress and are more likely to miscarry, so that a higher probability of having a female might indicate that male pregnancies are more likely to result in miscarriages (Low, 2015).
We create a new outcome variable, stillborn, equal to one if the baby is stillborn, and zero if he/she is a live birth. In Table 5, we estimate Equation (1) by regressing stillborn on UR in column (1), and on UR interacted with the 5 quintiles of IDD in column (2), and including always the full set of controls. A one-percentage point increase in UR leads to a highly statistically significant increase of 2.7% in stillborn. Vlachadis and Kornarou there are no heterogeneous effects in the unemployment by IDD on stillborn. In columns (3) and (4), we regress a dummy variable equal to one if the baby is female and zero otherwise on UR, and UR interacted with the 5 quintiles of IDD, and including always the full set of controls. Column (3) shows that a one-percentage point increase in UR leads to an increase of 0.14 percentage points, or 0.3%, in the probability of having a female baby. This is consistent with Olafsson (2016) who finds a 3.3 percentage points reduction in sex ratio at birth (less boys) among first-trimester children exposed to the 2008 financial crisis in Iceland. 22 Column (4) shows that mothers who live in the richest areas tend to have more female babies (1.1% increase) than the mothers who live in the poorest areas (0.4% increase). This difference is statistically significant at the 5% level suggesting that there are mildly statistically significant heterogeneous effects in the unemployment by IDD on infants' gender. Overall the results reported in Table 4 indicate a selection in-utero that may be linked to the financial stress experienced during the pregnancy, but do not help explain the greater effect of unemployment on poorer babies' health.

Mechanism II: Maternal health behaviour
Other channels that might explain the negative effect of unemployment on infant health include mothers' health behaviour during the pregnancy. Cigarette smoking and nutrition affect the intrauterine growth, while length of gestation is influenced by smoking and stress (Torche, 2011;Koppensteiner and Manacorda, 2016). Heavy prenatal alcohol exposure has been found to have adverse effects on new-borns by crossing the placenta and passing to the foetal, and by decreasing the supply of oxygen and food (Jones and Smith, 1973;West et al., 1994;Goodlett and Horn, 2001).
The HES data do not contain information about the mother's behaviours, but they report diagnoses identified at the time of the delivery. 23 In particular, the first 20 diagnoses are reported using ICD-10 codes (up to three-digit code). In our sample of live 22 On the contrary, van den Berg et al. (2018) find a positive association between unemployment and abortion, while no statistically significant effect on stillbirth; Aparicio and González (2020) instead find insignificant associations with miscarriages, abortion, or stillbirths. 23 There are no longitudinal UK surveys with information on pregnancies representative at the MSOA level that allow an alternative test of these behavioural mechanisms. The British Household Panel Survey (BHPS) and the Understanding Society surveys are longitudinal surveys representative of the UK and they cover our period of observation, but the sample size for siblings is very small and not representative at the MSOA level. births siblings, the most common first few diagnoses are related to the delivery, while the remaining diagnoses refer to any other type of disease or health problem. 24 Following the medical literature (Currie et al., 2015a;Dietz et al., 2010; US Department of Health and Human Services and others, 2014), we identify diagnoses associated with alcohol consumption and smoking. 25 Even if these diagnoses represent only very serious cases of either smoking or drinking addiction, or smoking or drinking related-diseases, they are objectively measured, hence more reliable than self-reported behavioural information. We create four new outcome variables. The first one, Alcohol, is a dummy equal to one if the mother has at least one diagnosis related to alcohol use, where the diagnosis made is reported in Panel A, column (1) of   Reduced smoking in columns (7)-(8). In columns (1), (3), (5) and (7) we regress the outcomes on UR, while in columns (2), (4), (6) and (8) we regress the outcomes on UR and UR interacted with the five quintiles of IDD, controlling for the full set of controls.
Column (1) shows that a one-percentage point increase in UR is associated with an increase of 2.8% in alcohol-related diagnoses. Column (2) shows that the probability of being diagnosed for alcohol-related health problems is lower for the mothers who live in the least deprived areas (-9.6%), while the opposite is true for the ones living in the middle to most deprived areas (2.6-3.6%). From column (3) we can see that a onepercentage point increase in UR is associated with an increase of 5.8% in smoking-related diagnosis. Column (4) shows that this effect is driven by the poorest mothers (living in areas where IDD is 4 or 5) where a one-percentage point increase in UR leads to an increase in Smoking by 8.2-9.3%, while a decrease by 2.5 for the richest mothers. Once 24 The most frequent first diagnosis is 'Perineal laceration during delivery' (26%, ICD-10 O70) followed by 'Single spontaneous delivery' (15%, ICD-10 O80), and 'Labour and delivery complicated by foetal stress' (14%, ICD-10 O68). 25 For further details about these diagnosis see https://nccd.cdc.gov/dph ardi/info /icdcodes.aspx and https://www.aaaai.org 26 Our results for smoking are the same if we exclude diagnosis for cancer (ICT-10 codes starting with C) from the list. Results are available upon request. the outcomes are constructed using a restricted set of diagnosis (columns (5) and (7)), the positive association between unemployment and alcohol-or smoking-related diagnosis is confirmed, and it is mainly driven by the poorest mothers (columns (6) and (8)).
Overall, these results are consistent with, and help to explain, the heterogenous health response to cyclical unemployment (Table 4). The poorest mothers are more likely to engage in risky behaviour following an economic downturn, leading to worse new-born health outcomes. Viceversa, there is some indication that the richest mothers smoke and drink less alcohol compared to the poorest ones (this effect disappears when we use a reduced set of diagnoses), even if the differences are only marginally statistically significant.
These results are similar to Currie et al. (2015b)'s work who find that unemployment increases smoking and drug use, and decreases self-reported health status particularly so for the disadvantaged mothers -African American, Hispanic, less educated or unmarried.
Contrarily, White, married women and highly educated mothers report better mental health and physical health.

Mechanism III: Prenatal care
We finally test whether unemployment has an effect on prenatal care, where prenatal care is measured as the gestation period in weeks at first antenatal assessment. This visit corresponds to the dating scan to verify how far along in the pregnancy the woman is and to check the baby's development. All pregnant women in England are offered this ultrasound scan at around 8 to 14 weeks of pregnancy free-of-charge, as all recommended prenatal visits. Table 7 presents the results of two regressions where the outcome is Week 1st visit and corresponds to the week when the pregnant mother has her first antenatal appointment.
In column (1) we regress Week 1st visit on UR, while in column (2) we regress Week 1st visit on UR and UR interacted with the quintiles of IDD, including the full set of controls in both regressions. Column (1) shows that a one-percentage point increase in UR leads to a postponement of the first visit by half a day, or an increase of 0.6% in the number of weeks before the first visit. Column (2) showS that an increase of onepercentage point in UR leads women from the least deprived areas to postpone their first visit by 2.2 days or 2.3%, while women from the most deprived areas delay by 0.9 days or 0.9%. In recessions, mothers might have less time to attend check-ups because the opportunity cost of time that might otherwise be used to find a job is higher, thus leading to a postponement of the first visit. 27 At the same time, mothers living in the poorest areas postpone prenatal care less than the ones living in the richest areas. Nevertheless, this does not explain our heterogeneous findings, and suggests that the poorest mothers' health behaviour discussed in 4.2 might be the main channel.

Siblings, maternal behaviour and selection
Given that our main pro-cyclical findings contrast with many of the results available for other developed countries, albeit usually not using sibling samples (e.g., Dehejia and -Muney, 2004;van den Berg et al., 2018;Aparicio and González, 2020), in this section we explore what could be the reason. In particular, we perform some analysis using different samples and methods to compare our results with previous work. We first study the influence of unemployment in the all babies' sample. In line with the literature we estimate the following equation:

Lleras
where the main differences with respect to Equation (1)  influenced by a change in unemployment. In Table 8, Panel B, we report estimates of Equation 2 which show that unemployment has a positive effect on the health of these babies, and with a substantially larger elasticity than the one found in the all babies' sample, as reported in Panel A of the same Table. In Table A.3 we run the same analysis of Panel D (Table 8) on the one-baby sample, but excluding teen and old mothers' first births. The results are very similar, hence, they are not due to an artifact of the 10-year window which allows to observe only the first born for the old mothers (at the beginning of the period) or the teen mothers (at the end of the period). Our data allow us to study the association between unemployment and different maternal or birth characteristics. We first focus on the siblings' sample. In Table 9 we estimate Equation (1) where the outcome of interest is parity (the number of times that a woman has given birth) regressed on UR in column (1) and on UR interacted with the quintiles of IDD in column (2) always including all the controls. The results in column (1) show that a one-percentage point increase in UR is associated to a decrease of 0.17% in parity. From column (2), we see that this decrease is accentuated for women who live in the richest areas (-0.73%), while it is less for women who live in the most deprived areas (-0.15%). As a consequence, the change in the births composition where more babies are born from deprived areas could explain our overall finding that unemployment reduces new-born health.

Unemployment and composition of new-borns
We then study the association between the unemployment rate (UR) and the probability of having first born instead of higher order births. We find that a one-percentage 28 In particular, Dehejia and Lleras-Muney (2004) find that white mothers who give birth during a recession are typically less educated than those giving birth in a boom, while the reverse is true for Black mothers. Salvanes (2013) show that low-educated women in Norway are disproportionately represented when there is an economic downturn, while for Spain, Aparicio and González (2020) find that married and older parents are more likely to have children when unemployment is high. Finally, for Sweden, van den Berg et al (2019) does not find notable compositional changes. point increase in UR leads to an increase by 3% of first-births (column (3)), but this change differs depending on the area where the mon lives. In fact, a one-percentage point increase in UR leads to a decrease by 4.9% of first-births among rich moms, and an increase by 6.8% of first-births among poor moms (column (4)). Given that first born are usually found to be lighter than higher-order births (Brenøe and Molitor, 2018), this compositional change supports our main findings.
Finally, recessions commonly lead to a postponement of childbearing, which is often later compensated during times of economic prosperity (Currie and Schwandt, 2014; Sobotka et al., 2011). Table 9 presents the results of two regressions of maternal age on UR, without (column (5)) and with (column (6)), the interaction with the IDD quintiles, always including all the controls. There are no statistically significant associations indicating that maternal age does not change with unemployment, so we do not find a postponement of fertility following economic downturn.
In Panel B of Table 9 we replicate the analysis of Panel A focusing on the one-baby's sample. We find that a one-percentage point increase in the UR leads to a decrease in parity by 0.94% (column (1)) compared with the 0.17 found in the siblings' sample.
More interestingly, column (2) shows that when UR goes up by one-percentage point, then parity increases by 2.16% among mothers who live in the richest areas, while it decreases by 1.52% among the mothers who come from the poorest areas. In contrast to Panel A, Panel B shows a significant increase in maternal age by 0.15% (column (3)), where a one-percentage point increase in the unemployment rate leads mothers' age to go down by 0.55% when living in the least deprived MSOAs, and to go up by 0.18% when living in the most deprived MSOAs. Finally, a one-percentage point increase in the UR leads to a decrease by 1.2% of first-births, but there are no statistically significant differences among maternal socio-economic status.
These results indicate that economic downturns might lead to a different composition of births in the two samples. In particular, recession may indeed encourage the selection into the siblings' sample of mothers less likely to come from a prosperous area, more likely to engage in risky behaviours, and as a consequence, be less likely to have healthy babies. While our data do not allow us to investigate how fertility decisions are made, we believe it is important to also document the composition of the sibling sample which is usually not identifiable with birth register data and might mask crucial differences with respect to the all babies' sample from which it is drawn.

Conclusions
In this paper we study the effects of unemployment on the health of babies' born in England between 2003 and 2012, a period of considerable cyclicity, which includes the Great Recession. To address the problem of selective fertility over the cycle, we compare the health of siblings whose in utero experiences occur at different points of the cycle. We find the health of new-born babies to be strongly pro-cyclical, i.e. babies born during more prosperous times tend to be healthier. In addition, the babies born to mothers who live in the poorest areas are found to be the ones whose health is most damaged by recession.
We provide robust evidence of three channels that can explain the overall negative effect of unemployment on new-born health: maternal stress; unhealthy behaviours -namely excessive alcohol consumption and smoking; and delays in the take-up of prenatal services.
The heterogenous effects of unemployment on babies health by socio-economic status appear to be explained by differences in maternal health behavior, where recession widens the gaps between the behaviors of poor and rich women towards opposite consumption of alcohol and smoke.
We also discuss if studies that consider all babies' samples, the data basis for most studies (e.g., Dehejia and Lleras-Muney, 2004; Aparicio and González, 2020), provide similar findings than when siblings data are instead used. We provide suggestive evidence that the two samples are populated by different mothers who behave in different ways once pregnant, leading to the opposite counter-or pro-cyclical findings.
Given that inequality begins before birth, and that large differences in health at birth have important consequences for later outcomes, such as education, earnings and disability (Currie, 2011), our evidence that poorer babies are most damaged by recessions suggests that policy interventions or safety net programs that target the most vulnerable individuals, and especially so in recessions, may help reduce adult inequality. Further research into fertility choices, to better understand how people from different social classes make the decision to conceive during booms or busts, would be valuable.    Note: * The unemployment rate corresponds to the the job-seekers allowance rate. All the statistics refer to the sample of live births, except the proportion of stillbirth. LA indicates Local Authority. † Alcohol diagnosis (Reduced alcohol diagnosis ) is a dummy equal to one if the mother has at least one diagnosis related to alcohol use where the diagnosis are listed in Panel A, column (1) (column (2)) of  Note: Standard errors in parentheses. Every coefficient corresponds to a separate regression. Birthweight (in grams), length of pregnancy (in weeks), and foetal growth are continuous variables, while the other outcomes are dummy variables. In Panel A and B, every regression is estimated with pooled OLS and it includes year FE, MSOA FE and LA-specific trends.
In Panel C and D, every regression is estimated with mother FE, and it includes year FE, and LA-specific trends. In Panel B and D, we control for the following extra variables: month of birth, new-born's gender, mother's ethnicity, number of GP practices in the PCT, number of hospitals in the PCT, number of GPs in the PCT, number of specialists doctors in the PCT and the number of midwives working in each hospital where the delivery occurred. The standard errors are clustered at the MSOA Level in Panel A and B, and at the mother level in Panel C and D. *** p<0.01, ** p<0.05, * p<0.1 Table 3: Robustness checks on the effects of unemployment on different infants outcomes. Siblings' sample.
(1) Note: Standard errors in parentheses. Every coefficient corresponds to a separate regression. Birthweight (in grams), length of pregnancy (in weeks), and foetal growth are continuous variables, while the other outcomes are dummy variable. Every regression is estimated with mother FE and it includes year FE, LA-specific trends, month of birth, new-born's gender, number of GP practices in the PCT, number of hospitals in the PCT, number of GPs in the PCT, number of specialists doctors in the PCT and the number of midwives working in each hospital where the delivery occurred. Panel D reports the second stage of a 2SLS estimation using the Bartik instrument. In Panel A standard errors are clustered at the MSOA level, while in Panel B1, B2, C and D at the mother level. *** p<0.01, ** p<0.05, * p<0.1  Note: Standard errors in parentheses. Every column corresponds to a separate regression. Stillbirth is a dummy equal to 1 if the pregnancy resulted in a stillbirth, and 0 in a live birth. Female baby is a dummy equal to one if the infant is a female, and 0 if a male. In columns (1) and (2) we consider siblings who are either live or stillbirth, while in columns (3) and (4) we only consider live births. Every regression is estimated with mother FE and it includes year FE, LA-specific trends, dummies for the IDD quintiles, month of birth, number of GP practices in the PCT, number of hospitals in the PCT, number of GPs in the PCT, number of specialists doctors in the PCT and the number of midwives working in each hospital where the delivery occurred. In columns (1) and (2) we also control for new-born's gender. Standard errors are clustered at the mother level. *** p<0.01, ** p<0.05, * p<0.1 Table 6: Maternal health behaviour and heterogeneity by IDD quintiles. Siblings' sample. (1) (3)  Note: Standard errors in parentheses. Every column corresponds to a separate regression. Week 1st visit corresponds to the pregnancy week when the mother had her first hospital visit. Every regression is estimated with mother FE and it includes year FE, LA-specific trends, dummies for the IDD quintiles, month of birth, new-born's gender, number of GP practices in the PCT, number of hospitals in the PCT, number of GPs in the PCT, number of specialists doctors in the PCT and the number of midwives working in each hospital where the delivery occurred. Standard errors are clustered at the mother level. *** p<0.01, ** p<0.05, * p<0.1    Note: * The unemployment rate corresponds to the the job-seekers allowance rate. All the statistics refer to the sample of live births, except the proportion of stillbirth. † Alcohol diagnosis (Reduced alcohol diagnosis ) is a dummy equal to one if the mother has at least one diagnosis related to alcohol use where the diagnosis are listed in Panel A, column (1) (column (2)) of   I20-I25  I20-I25  Other heart disease  I00-I09, I26-I28, I29-I51  I51  Cerebrovascular disease  I60-I69  I60-I69  Atherosclerosis  I70  I70  Aortic aneurysm  I71  I71  Other arterial disease  I72-  Note: Standard errors in parentheses. Every coefficient corresponds to a separate regression. Birthweight (in grams), length of pregnancy (in weeks), and foetal growth are continuous variables, while the other outcomes are dummy variables. Every regression is estimated with pooled OLS and includes year FE, MSOA FE and LA-specific trends. The extra controls are month of birth, new-born's gender, mother's ethnicity, number of GP practices in the PCT, number of hospitals in the PCT, number of GPs in the PCT, number of specialists doctors in the PCT and the number of midwives working in each hospital where the delivery occurred. Standard errors are clustered at MSOA level. *** p<0.01, ** p<0.05, * p<0.1