The Effects of Unemployment on Fertility: Evidence from England

3.1 Data

The fertility data used in this analysis come from the Office for National Statistics (ONS). It compiles counts of live births and stillbirths by age of mother and area of usual residence in England, years 1995–2011. Age-specific fertility rates (ASFRs) are constructed by dividing the number of births by the relevant female population using mid-year population estimates that are based on the censuses, in which female ages range between 16 and 44.^[10] ASFRs provide an appropriate measure of varying fertility rates since they are unaffected by changes in population age distribution and are well suited for comparing fertility rates across age groups. In this analysis, ASFRs are based on age intervals of 16–24, 25–34 and 35–44.^[11] As the best available measure of the labor market conditions prevailing at the time of the conception, births in calendar year, t, are matched with 1 year lagged, t – 1, data of Labor Force Survey (LFS) in the corresponding county.^[12]

Access to the confidential LFS was vital to conduct this study, as it was used to construct county–year–age group-specific unemployment rates and county–year–age group-specific demographic characteristics. The sample is restricted in the age band 16–44 in order to study females who are of childbearing ages.^[13]Table 1 presents the descriptive statistics for demographic characteristics and unemployment rates from the LFS, birth statistics from the ONS and house prices from the Department for Communities and Local Government.^[14]

Figure 1 illustrates fertility and unemployment patterns among counties. Northumberland, Cheshire and Dorset have the lowest average fertility rates in England and the South East region along with the Greater London experience the highest fertility. Merseyside, Tyne and Wear and West Midlands are the areas with the highest unemployment rate over the sample period.

Figure 1:

Mean unemployment rates and mean birth rates across counties in England.

Figure 2 shows trends in birth rates and unemployment rates by age groups at the national level.^[15] Birth rates follow a decreasing trend for age group 1 after reaching a peak of 53.2 births per 1,000 women in 1995. With regard to the 25–34 band, birth rates rebounded after 2000 and reached its highest point of 112.2 in 2009. The oldest age group, 35–44, has experienced a tremendous rise in fertility rate, a steady increase from 21.3 to 35. Turning to the national time series data for unemployment rates, they notably differ in levels across age groups but follow a similar trend. The total unemployment rate reached its lowest point between 2002 and 2005, and gradually increased afterwards. Overall, substantial variation across counties in Figure 1 and considerable shift in birth trends across age groups in Figure 2 strongly suggest the inclusion of county-specific and age group-specific linear time trends. Additionally, from looking at the figure, birth rates appear to follow a countercyclical pattern over the analyzed time interval.

Figure 2:

Age-specific unemployment rates and age-specific fertility rates in England.

In all, a balanced panel is constructed for the 1994–2010 period, with 47 counties and three age groups. The final version of the dataset contains information on ASFRs, age- and gender-specific unemployment rates, educational attainment, marital status, ethnicity and country of birth.

3.2 Methodology

In order to obtain baseline estimates of the relationship between unemployment and fertility, the following fixed effect specification is employed:

The level of analysis is a county–year–age group cell. Y_gct is the birth rate in county c, age group g, in year t and U_{gc(t – 1)} is the lagged unemployment rate.^[16] The county fixed effects, α_c, and age group fixed effects, θ_g, are included to control for differences in birth rates across counties and age groups owing to time-invariant unobservable factors. The year fixed effects, γ_t, account for movements in fertility rates over time that are shared by all counties. The county-specific linear time trends, ω_c×T, and age group-specific linear time trends, δ_g×T, control for unobserved variables correlated with birth rates that change linearly over time within counties and age groups.^[17]X_{c(t – 1)} indicates lagged time-varying county-level demographic controls (country of birth, ethnicity, educational attainment and marital status) and house prices that account for changes in population composition and changes in the real estate market. The regressions are simultaneously carried out for both male and female unemployment in own age group.^[18] All regression are weighted by the relevant population of women in each cell.

This study uses the identification assumption that local unemployment rates are conditionally exogenous to household fertility outcomes. However, there are certain concerns associated with the use of unemployment rates as exogenous regressors. One of these is that local unemployment rates might be correlated with changes in other unobserved variables that may affect the fertility decision of individuals. Second, there may be a positive correlation between fertility and local labor supply. If birth rates increase due to the changes in local labor supply, then the unemployment measure may be picking up this relationship rather than the effect of local labor market demand. Lastly, the International Labour Organization (ILO) definition of unemployment may not be able to capture the full extent of the local labor market conditions, causing a measurement error. The estimation strategy to deal with these problems is to specify a variable that can account for demand-induced variation in unemployment, and can thus be used to obtain an unbiased estimate of the effect of unemployment on fertility. To this end, the IV approach was adopted to explore the robustness of the OLS results. The predicted unemployment rates were built based on the work of Bartik (1991), Blanchard and Katz (1992), Schaller (2015) and Anderberg et al. (2015) for the UK case, where the initial industry composition of employment is interacted with the corresponding national industry-specific trends in unemployment. In particular, the local industry composition by gender and age group at baseline, defined as the calendar year 1993, is combined with industry-specific unemployment rates by gender, age group and time at the national level over the sample period.^[19] For each county, age group, gender and year industry-predicted unemployment rates are constructed as follows:

where ψ_ghjk is the share of industry k among employed individuals of age group g, gender h, county j at baseline, and where UNEMP_ghkt is the unemployment rate, at the national level, in industry k for individuals of age group g, gender h and in time period t. Given that the predicted unemployment measure is a weighted average of the national industry-specific unemployment rates, these weights reflect the baseline local industry composition in the relevant gender and age group.

The IV approach has a number of attractive features. Most importantly, the estimates cannot be affected by contemporaneous omitted variables since the only local input into the predicted unemployment rates is the industry structure at baseline and these rates cannot be related to any contemporaneous (during the sample period) omitted variables. Furthermore, as mentioned earlier, birth rates in a county are a function of both local labor supply and labor demand. It is for this reason that the use of observed changes in local labor market confounds the results. Instead, the IV uses labor demand shocks as an identifying source of variation and act as an exogenous change in local labor demand. Additionally, the predicted unemployment rate during the next period relies only on initial local industry composition and national-level industry-specific unemployment rates that influence the gender composition of employment opportunities. However, one might be concerned for the earlier time periods of the panel. Later in this paper, this issue is investigated by dropping some of the years at the beginning of the sample period from the estimated model. This estimation does not send any warning signals that the main results are substantially affected by underlying serial correlation in county-specific circumstances.

4.1 OLS Specifications

Table 2 presents the results from the OLS estimation. Column 1 reports the estimation with all fixed effects included; column 2 adds basic demographic characteristics (education, ethnicity, country of birth and partnership status) and an additional control variable (house prices); column 3 adds county-specific linear time trends; and column 4 adds age group-county-specific linear time trends.

The specification in the first column yields a positive and statistically insignificant coefficient of 0.008. After adding more controls and time trends, the results consistently show that the overall unemployment rate is positively associated with fertility rates. To assess whether different age groups are more likely to move in response to an economic shock, I include age group-county-specific linear time trends in column 4. The coefficient of the main interest remains similar to the ones in columns 1–3, implying that age groups do not systematically move into different counties in response to adverse labor market conditions. According to the estimate from the fully saturated model in column 4, a 1 percentage point increase in the unemployment rate is associated with a 1.3 % increase in birth rates, which is significant at the 1 % level. This finding suggests that fertility moves countercyclically, and, accordingly the substitution effect dominates any negative income effect over the sample period. The main reasons for this finding are twofold: On the LFP side, the difference between the LFP of men and women has shrunk considerably since the mid-1990s. On the earnings side, the gender pay gap has been following a downward trend.^[20] Consequently, improvements in women’s job opportunities and in wages have increased the opportunity cost of childbearing, causing fertility to move in the opposite direction to that of the business cycle. In order to address omitted variables biases, the rest of the analysis is carried out based on the specification in column 3, in which I control for observable demographic characteristics and house prices. I also exploit the panel aspects of the data by including county and year fixed effects as well as the county-specific linear time trends.

The analysis continues with a stratification of the regressions with age-specific characteristics so as to gain further insight into the demographic basis of this result. The upper panel of Table 3 suggests that the fertility is the most responsive to unemployment rates between the ages of 25 and 34. In this age group, a 1 percentage point increase in the unemployment rate leads to a 1.32 % increase in fertility. The incidence of youth unemployment also has a positive and marginally significant impact on fertility. The older group, 35–44, by contrast, shows no significant impact of unemployment on birth rates. Altogether, the results in this table suggest that age groups react differently when they are exposed to local unemployment shocks. The next table, therefore, proceeds to further examine the age group–gender characteristics.

The differences between age- and gender-specific unemployment rates are highlighted in Table 4 in which I expect to find that female unemployment will be positively associated with fertility, whereas male unemployment will have the opposite impact. Overall, the results are in line with this notion. For example, a 1 percentage point increase in female unemployment leads to a 1.29 % increase in births for the prime age group. For men, the negative and significant effects are concentrated among the younger cohorts, 16–24 and 25–34. However, for the age group 35–44, there are insignificant coefficients on unemployment for both males and females. This finding could be because there is a weak relationship between older cohorts’ fertility decision and their labor market status. The following section of the paper is concerned with endogeneity of unemployment, and IV estimation results are presented.

4.2 IV Estimation

As discussed in Section 3.2, the potential sources of bias in the OLS results are the possibility of reverse causality and/or some unobservables that are affecting fertility rates other than unemployment. I start by considering the full sample and estimate the effect of overall unemployment (aged 16–44) on fertility. Confirming the OLS findings in Table 2, columns 1–4, I find that the IV coefficients have the same sign; however, they are larger in magnitude. In the first stage, the predicted unemployment rates are significantly correlated with the endogenous variable and in the expected direction. Since the IV estimates can be interpreted as the impact of a fall in local labor demand on fertility, finding higher coefficients suggest that these estimates reflect the local spillovers in unemployment. In other words, they capture: (1) the main effect due to being unemployed, (2) the risk of being in unemployed in the near future and (3) expectations about future wage growth.

In order to quantify the different effects of the predicted unemployment rate, I proceed to separately estimate the relationship across age groups and age group–gender cohorts. In Table 3, the point estimates on age groups are all positive and statistically significant at conventional levels. Column 2, in bottom panel of Table 4, strongly confirms previous findings that male and female unemployment have different impacts on fertility and reveals the importance of sub-group characteristics. In parallel to the main hypothesis of this paper, the results imply that a one percentage point increase in male unemployment leads to a 2.18 % decrease in birth rates whereas same amount of increase in female unemployment leads to a 6.26 % increase in birth rates for the age group 25–34. This shows that the positive effect of female unemployment on fertility is much larger for prime aged women. The results for men again show that unemployment has a significant negative effect on fertility, with the effect being stronger at later ages. Taken together, these results suggest that unemployment is an important determinant of fertility behavior. The following part of the results proceeds to explore whether the responses are homogenous across demographic characteristics.

4.3 Analysis by Demographic Characteristics

Having detailed information in UK LFS enables a filtering out of the effects of unemployment on fertility by education, country of birth and marital status. In order to allow the point estimates on county and year fixed effects to vary across subgroups, I construct covariates and estimate the relationship for each age group. In Table 5, columns 1–3 present estimates for the OLS specifications and columns 4–6 present estimates for the IV specifications. Because the cell size in industry employment composition by demographic characteristics is small at baseline, some IV estimates have lower number of observations. However, this does not cause excluded variables to show a weak partial correlation with unemployment. Although not reported due to space restrictions, the values for first-stage F-statistics are consistently higher than 10. Nonetheless, one caveat is that the measurement error of group-specific unemployment rates may lead to biased estimates, so these results should be interpreted with caution.

Looking at the estimates presented by educational attainment in Panel (A) of Table 5, I find that the coefficients on unemployment are negative and significant for the 16–24 age group, but they are positive for older cohorts who are highly educated. In the further education category, unemployment also seems to be positively associated with fertility. The fact that the effect is more pronounced on the degree level may be attributable to the fact that women take advantage of the low opportunity cost of childbearing to prevent future career interruption which occurs due to the transition to motherhood. The IV estimates again are larger in magnitude and tend in the same direction.

Panel B shows results separately for UK-born and non-UK-born cohorts. Coefficients on unemployment for non-UK borns in both OLS and IV estimations are negative and mostly larger in magnitude. This result may be explained by the fact that immigrants are more sensitive to cyclical increases in unemployment than those of natives. For UK borns, the effects of unemployment on fertility tend to be positive.

Turning now to the evidence on partnership in panel (C), I detect clear differences by marital status. Contrary to my expectations, I find that unemployment has a only negligible influence on fertility at younger ages, but it exhibits stronger association among older cohorts. Indeed, the coefficient of unemployment is negative and significantly different from zero for singles at more advanced ages. Looking at the results for married cohorts, unemployment is positively associated with fertility and the effect is mostly concentrated at the prime age cohort.

4.4 Robustness Checks

Additional robustness checks are conducted in order to detect whether the main findings remain stable to different specifications. The first column in Table 6 presents results for teenage fertility. Results in the upper panel of column 1 indicate that unemployment has a positive and marginally significant impact on birth rates. As can be seen from the bottom panel, both male and female unemployment maintain the expected signs of direction. Although the effects are small in magnitude, the results are still in line with the main findings of this paper.

Between columns 2 and 4, I further discuss the important issue of the validity of my instrument, which is introduced in Section 4.2. The strong correlation of my instrument with the endogenous variable is apparent from all the first-stage results presented in Appendix Table 7. Although the IV results seem robust to a number of alternative specifications, one may argue that unemployment is not necessarily a purely demand-driven measure and is also affected by the changes in labor supply. Finally, I build on my instrument and use employment growth index as an alternative measure to assess the robustness of my results. The measure reflects exogenous labor demand for females and males in each age group and is constructed as follows: for each county, age group and year, I start with a variable measuring the proportion of employment based on local industry structure at baseline. Next, similar to Aizer (2010), I construct an annual employment growth index for each gender and age group, by interacting baseline measure with national trends in employment growth rates in industries dominant in the county, then collapsing over industries within each county–year–age group cell:

where ψ_ghjk is the share of industry k among employed individuals of age group g, gender h, county j at baseline, and where EMPGROWTH_ghkt is the industry employment growth rate, at the national level, in industry k for individuals of age group g, gender h and in time period t. The first-stage results show that all measures of unemployment are well correlated with the instruments. However, the Kleibergen–Paap F statistic points to a weak identification issue for those aged 35–44, so that coefficients may suffer from bias and should be interpreted with caution. On the whole, the IV estimates presented in Table 6 alleviate concerns about the biased estimates induced by the changes in labor supply and are consistent with the baseline results. Reassuringly, I find that: (1) fertility moves countercyclically over the business cycle; (2) the prime age group is more responsive to changes in unemployment; (3) male unemployment is negatively and female unemployment is positively associated with births, with these effects often being smaller and mostly significant. Finally, the results introduced in columns 5–7 show that the outcome is not driven by the serial correlation and present estimates after dropping the first 4 years of the sample. Across all age groups, the coefficients consistently remain qualitatively unchanged. Taken together, I obtain results that are very similar to the baseline estimates and provide meaningful insights on the countercyclical nature of births along with group-specific differences across social strata in England.