SUPPLEMENTARY INFORMATION

Transcription

1 This Supplementary Information for Advances in Development Reverse Fertily Declines documents in more detail the data and methods that support Myrskylä et al. s finding of a reversal of the relationship between the human development index (HDI) and the total fertily rate (TFR) at HDI levels around The Supplementary Information also provides various robustness tests indicating that the key findings are not sensive to (i) data on a specific country, single data points or outliers, (ii) the introduction of time-lags in the HDI fertily relationship, and (iii) the inclusion of fertily tempo-adjustments that corrects the TFR for tempo distortions caused by the postponement of childbearing to later maternal ages. 1 The TFR and HDI data used for the analyses in this paper are available as Supplementary Data linked to the online version of the paper at. Data The human development index (HDI) and the total fertily rate (TFR) are widely used measures of, respectively, a country s development stage and period fertily. The TFR reflects the number of children that would be born to a woman during her lifetime if she experienced the age-specific fertily rates observed in a calendar year (= period). Most policy debates about fertily trends focus on the TFR. 2 TFR is a key determinant of the number of children born in a calendar year, and thus of population ageing and population growth/decline. Although some weaknesses of the TFR have been illustrated in the lerature 1, 3, which are addressed below, the TFR is the only behavioural measure of fertily available for a large number of countries and for many years. Supplementary analyses presented below show that the finding of a reversal of the fertily development relationship at advanced levels of the HDI is robust wh respect to the primary limations of the TFR as an indicator of fertily levels in developed countries wh an ongoing postponement of childbearing to later ages. 4 The HDI is the primary index used by the Uned Nations Development Programme (UNDP) to monor and evaluate broadly-defined human development. In particular, the HDI combines three dimensions of socioeconomic progress into a single index for each 1 1

2 calendar year, 5 adding together wh equal weight a country s (i) health condions, as measured by annual life expectancy at birth, (ii) standard of living, as measured by the logarhm of the annual gross domestic product (GDP) per capa at purchasing power pary (PPP) in US dollars, and (iii) human capal, as measured by average of the annual adult leracy rate (wh two-thirds weight) and the combined primary, secondary, and tertiary gross school enrolment ratio (wh one-third weight). Since 1999, the HDI is calculated using a consistent formula that not only ranks countries in terms of their development level, but also provides an intertemporal comparison of HDI trends over time. In particular, in calculating the HDI, each dimension of the HDI is first standardised to the un interval using x standardised t = x t x min x max x min where x t is the HDI measure for eher health condions, standard of living or human capal in year t, and x min and x max are time-invariant scaling values for each HDI dimension (for health condions: x min = 25 and x max = 85; for standard of living: x min = ln(100) and x max = ln(40, 000); and for human capal: x min = 0 and x max = 100). The HDI for year t is then obtained by averaging, wh equal weight, across the standardised HDI measures, x standardised t, for health condions, standard of living or human capal in year t. Because, contrary to UNDP s pre-1999 tradion, the scaling values x min and x max are now time-invariant, the human development index obtained via Eq. (S.1) is comparable over time whin each country. Using these constant scaling values, HDI levels can increase beyond the currently observed highest HDI values as development progresses and health condions, standard of living, and/or human capal levels further improve. In addion, constant scaling values imply that in constructing the longudinal HDI time-series for our analyses, countries do not necessarily cluster near a maximum value of one as they reach very advanced development stages. (S.1) The longudinal consistency of the human development index enables our analyses to identify if and how whin-country changes in development levels during causally affect trends in the total fertily rate during this period. UNDP, however, does not publish an annual HDI time-series based on the current method for calculating the index (see Eq. S.1). Published HDI values based on this method are only available for the years 1975, 1980, 1985, 1990, 1995, 2000 and To create an annual HDI timeseries for the period for as many countries as possible, we have obtained the individual measures underlying the HDI that is, life expectancy for health condions, GDP per capa for living standards, and leracy rates and enrolment ratios for human capal from the World Bank World Development Indicators Online Database. 6 HDI data used for our analyses were then calculated using Eq. (S.1). Time-series for the TFR were also retrieved from the same source. The World Bank World Development Indicators data for are almost complete for the total fertily rate (TFR), life expectancy, and gross domestic product (GDP) 2

3 per capa at purchasing power pary. The data on leracy rates and school enrolment ratios are often missing for early time periods, particularly in developing countries at low HDI levels. GDP per capa and life expectancy jointly account for two thirds of the HDI, and these two dimensions account for an even larger fraction of the year-to-year variation in the HDI because leracy rates and school enrolment ratios are not subject to large short-term fluctuations. The relative completeness of the time-series data for these variables is fortunate, and contributes to the qualy of the cross-sectional and longudinal HDI data. In the rare cases where a country s data on life expectancy includes gaps, we used linear interpolation to impute missing values. Because life expectancy generally evolves gradually, a linear interpolation of life expectancy data is unlikely to distort our analyses. Linear interpolation was also used to impute missing values for enrolment ratios and leracy rates. School enrolment ratios and leracy rates also evolve slowly because they are primarily cohort-driven, and linear interpolation of missing data is therefore relatively innocuous in terms of data qualy. Because GDP per capa and TFR are subject to large annual fluctuations, no imputations were conducted for these variables to fill in missing data. For calendar years in which a country s GDP per capal or TFR was missing, no HDI value was calculated. These missing data problems, however, occur primarily early in the period that is used for our analyses, and are concentrated among developing countries wh relatively low HDI levels. Our analyses focusing on the reversal of the HDI TFR relationship at very advanced development stages are thus unlikely to be substantively affected by missing data in the HDI and TFR time-series. The list of countries included in our analyses, and the years for which HDI data are available for each country, is reported in Table S.1. For 2005, information is available for 140 countries; in 1975, for 107 countries. In all our analyses, sovereign and non-sovereign cy-states such as Hong Kong, Macao, Monaco and Singapore are excluded because their predominantly urban status implies peculiar population dynamics when compared to nations that have a more balanced rural-urban composion. While this is our preferred specification, our main findings continue to hold if these cy-states are included in the analyses. A complete HDI and TFR time-series is available for all advanced countries wh a 2005 HDI.90, wh the exception of Slovenia for which only post-1990 data are available. Among the countries wh a 2005 HDI.85 the data is complete for 32 out of 37 countries (Table S.1). Comparisons between the published UNDP HDI values (only available for the years 1975, 1980,..., 2000, 2005) and the HDI values used for our analyses reveal a very close correspondence between these two measures, wh a correlation of 0.99 or higher for all years. Our analyses are therefore unlikely to be affected by any remaining small differences between our calculations of HDI values based on World Development Indicators database and the published UNDP HDI values. 5, 6 In the graphical analyses in Figure 1 of our main paper, HDI and TFR values are 3

4 Table S.1: List of countries, and years for which HDI and TFR time-series are available (sorted descending by 2005 HDI) Time- Time- Time- Time- HDI TFR series HDI TFR series HDI TFR series HDI TFR series Country years Country years Country years Country years Australia Slovak Rep Tunisia Congo, Rep Norway Bahrain Iran Nepal Iceland Croatia Ecuador P. N. Guinea Ireland Lhuania Sri Lanka Sudan Luxembourg Latvia Georgia Madagascar Sweden Uruguay Paraguay Cameroon Canada Costa Rica Belize Yemen Finland Bulgaria Algeria Lesotho France Mexico Syria Kenya Netherlands Trinidad El Salvador Maurania USA Romania Cape Verde Togo Denmark Russia Jamaica Uganda Japan Tonga Indonesia Zimbabwe Swzerland Saudi Arabia Vietnam Senegal Belgium Oman Moldova Angola N. Zealand Panama Kyrgyz Rep Guinea Spain Malaysia Mongolia Rwanda UK Belarus Egypt Benin Austria Brazil Nicaragua Tanzania Italy Macedonia Bolivia Cote d Ivoire Israel Maurius Honduras Zambia Greece Albania Guatemala Congo, DR Germany Venezuela Vanuatu Malawi Slovenia Colombia Tajikistan Burundi S. Korea Ukraine South Africa Mozambique Cyprus Kazakhstan Gabon Ethiopia Portugal Thailand Morocco Sierra Leone Czech Rep Armenia Eq. Guinea C. African Rep Kuwa China India Burkina Faso Malta Azerbaijan Namibia Chad Hungary Samoa Cambodia Mali UAE Turkey Botswana Niger Poland Suriname Pakistan Argentina Jordan Bangladesh Estonia Peru Swaziland Chile Philippines Ghana Note: Time-series years is the period for which a complete HDI and TFR time-series is available for the longudinal analyses. 4

5 rescaled. This rescaling is implemented so that the graphical representation in Figure 1 particularly reflects HDI differences at intermediate and advanced levels of development, and the distance between different TFR levels is proportional to the different long-term population growth rates implied by the respective TFR levels. Specifically, rescaled values are obtained using HDI = log(1 HDI) and TFR = log(.4886 TFR)/µ, where µ = 31 approximately equal to the mean age at childbearing in developed countries. Israel, which is an outlier among countries wh HDI > 0.9 in Figure 1 wh a HDI level of and a TFR level of 2.82 in 2005, has been widely recognised as having relatively high fertily levels given s development stage, possibly related to minories and religious factors specific to the Israeli population. 7 Table S.2 reports the list of countries that are used for the longudinal analyses in Figure 2 of our main paper, along wh their HDI and TFR levels in 1975, the reference year and The inflection points, or reference years in the HDI fertily relationships in Figure 2 are as follows: In the U.S., fertily reversed in 1976 (=reference year) at an HDI of (GDP/capa US$20,670, life expectancy 72.9 years, education index 0.95); the reversal in Norway occurred in 1983 at a HDI of (GDP/capa US$21,239, life expectancy 76.1 years, education index 0.93); in Italy, the reversal occurred in 1994 at an HDI of (GDP/capa US$22,965, life expectancy 77.7 years, education index 0.92), and in Israel, the reversal occurred in 1992 at a HDI of (GDP/capa US$18,812, life expectancy 76.5 years, education index 0.91). Longudinal Analyses: Statistical Model and Estimation The graphical analyses in Figures 1 and 2 of our main paper suggest that the HDI TFR relationship reverses from negative (increases in HDI are associated wh lower TFR) to posive (increases in HDI are associated wh higher TFR) at an intermediate HDI level in the range We augment this graphical result wh a statistical model for the effect of HDI increases on fertily change in order to (i) estimate the crical level HDI cr at which the HDI TFR relationship reverses in the longudinal data, and (ii) ascertain the causal interpretation of our results as a reversal in the HDI fertily relationship. The starting point of our statistical model linking development to fertily is the relationship TFR = (α pre + β pre HDI ) B pre + (α post + β post HDI ) B post + ε, (S.2) where TFR and HDI are the TFR and HDI for country i in year t. The terms B pre and B post are indicator variables for whether country i s HDI level is below or above the crical HDI level, HDI cr, at which the HDI TFR relationship is hypothesised to reverse. B pre and change their value, respectively, from 1 to 0 and from 0 to 1 as country i s HDI level B post increases above the crical level HDI cr. The coefficients β pre and β post are, respectively, the 5

6 Table S.2: Data used for longudinal analyses in Figure Reference Year 2005 Country Label HDI TFR Year HDI TFR HDI TFR Countries ending in the top right quadrant of Fig. 2 in 2005 Norway NL USA Denmark (1) Germany (2) Spain (3) Belgium (4) Luxembourg (5) Finland (6) Israel (7) Italy (8) Sweden (9) France (10) Iceland (11) Uned Kingdom (12) New Zealand (13) Greece (14) Ireland (15) Countries ending in the bottom right quadrant of Fig. 2 in 2005 Japan Austria (16) Australia (17) Swzerland (18) Canada (19) S. Korea (20) Notes: Reference year is the year in which a country attains s lowest TFR level whin the HDI range effects of development (as measured by the HDI) on the total fertily rate at HDI levels below and above the crical level HDI cr. The residual ε is defined as ε = η i + γ t + υ, where η i and γ t are respectively country- and time-specific parameters (fixed effects), and υ is a normally-distributed residual wh zero mean and constant unknown variance. The inclusion of country and time fixed-effects controls for unobserved time-invariant countryspecific factors and common time trends that may affect TFR and HDI trends across and whin countries. Preliminary analyses using the Woolridge 8 and Baltagi-Wu LBI 9 tests indicated that the residual υ has a un root (the estimated autocorrelation parameter for υ was 0.99). To adequately adjust for this un-root process in the residual υ, our preferred analyses are based on the differences-in-differences model 8 that is obtained by differencing TFR and HDI trends in Eq. (S.2) over time: TFR = α B post + β pre HDI pre + β post HDI post + γ t + υ, (S.3) where is the difference operator wh x t = x t x t 1, HDI pre = B pre HDI, and 6

7 HDI Figure S.1: Profile of the log-likelihood function for Model (S.3) wh respect to the threshold HDI cr at which the HDI fertily relationship is hypothesised to reverse from negative to posive (calculations includes all countries wh a HDI 0.85 in 2005.) HDI post = B post HDI. The coefficients β pre and β post continue to measure the effects of development (as measured by the HDI) on the total fertily rate at HDI levels below, and at or above, the crical level HDI cr. Differencing implicly controls for the country fixed-effects γ i and removes the un root from the residual autocorrelation. Our preferred specification in Eq. (S.3) therefore allows us to test whether the reversal in the HDI fertily relationship, which is graphically documented in Figure 2, is statistically significant and persists after controlling for potentially confounding factors such as unobserved time-invariant country-specific factors and common time trends. Only such longudinal statistical analysis wh control for unobserved country-level characteristics can warrant a causal interpretation of the effect of HDI changes on fertily. In particular, the hypothesis of a reversal of the HDI fertily relationship at HDI cr implies that β pre < 0 and β post > 0. The first step in testing this hypothesis is to estimate HDI cr via an erative search process, using all countries that attained a development level of HDI 0.85 by The exclusion of countries wh a 2005 HDI of less than 0.85 is appropriate as our preliminary analyses (Figures 1 2) suggest that HDI cr is whin the range of The statistical estimate of HDI cr is obtained using maximum likelihood, by including HDI cr as a parameter in the likelihood function of Eq. (S.3). The likelihood function is maximised (after a logarhmic transformation) by using a two-stage grid-search algorhm that in the first stage varies the value HDI cr from.800,.805,.810,...,.910, and in the second stage refines the search wh a step size of.0001 in the neighbourhood of the best-fting firststage HDI cr value. The profile of the log-likelihood for Eq. (S.3) in the range is shown in Figure S.1. The log-likelihood in Figure S.1 is maximised when HDI is at , or approximately We therefore use 0.86 as our preferred value for HDI cr in our final 7

8 estimation of Eq. (S.3). Table S.3 shows the estimated parameters ˆβ pre and ˆβ post for our statistical model of the HDI fertily relationship (Eq. S.3). Model M.1 in Table S.3 reports our preferred estimates that are based on the differences-in-differences relationship in Eq. (S.3), controlling explicly for time fixed-effects and implicly for country fixed-effects and un-root autocorrelation. The model is estimated using all countries wh HDI 0.85 in The most important conclusion obtained from these estimation results is the reversal in the effect of HDI on fertily, wh the estimated coefficients changing from a statistically significant negative value for ˆβ pre = 1.59 to a statistically significant posive value for ˆβ post = The coefficient of development for HDI levels below 0.86, ˆβ pre = 1.59, is consistent wh previous knowledge on the negative effect of development on fertily. 2, The coefficient ˆβ post = 4.07 for the effect of development on fertily for HDI levels at or above 0.86 confirms that this relationship has been reversed as countries have attained advanced stages of development. Moreover, this statistical model confirms that this reversal not only occurs when using cross-sectional country-level data (Figure 1), but also in longudinal analyses of HDI and TFR whin countries that control for unobserved time-invariant country-specific factors and common time trends that may affect TFR and HDI trends. The estimated coefficient of ˆβ post = 4.07 implies that, above a HDI level 0.86, a five point increase in human development index is linked, on average, to a = un increase in the TFR. This increase in the TFR of approximately 0.2 children per woman for a 0.05 increase in HDI is sizable, and corresponds closely to the graphical analyses presented in Figure 2. For example, once the Uned States, the Netherlands and Norway had attained their lowest TFR level whin the HDI range of , further increases in HDI by 0.05 were associated wh, on average, TFR increases of 0.25 (USA), 0.26 (NL) and 0.13 (N). Our preferred estimates in Model M.1 of Table S.3 therefore strongly support our hypothesis that the empirical HDI fertily relationship has changed from negative to posive as countries attained very advanced levels of development. Robustness Tests To assess the robustness of our preferred results (Model M.1 in Table S.3) wh respect to possibly influential single data points or single countries, we have conducted extensive influence analyses using the leverage, Cook s D, and DFBETA statistics associated wh each observation. Based on these statistics we identified four countries (Kuwa, South Korea, Estonia and Malta) and one single observation (Belgium, year 2004) that looked like outliers. The one observation for Belgium is likely to be an outlier because of a change in reported school enrolment rates, which affects the development index. Excluding these four countries and the single outlying observation for Belgium did not change the finding on the reversal of the relationship. For example, the coefficients for Model 1 whout these data were (p<0.001) for ˆβ pre and 3.81 (p<0.001) for ˆβ post. 8

9 Table S.3: Estimated effects for the HDI on TFR before and after a country reached a development level HDI = Estimates from differences-in-differences model wh controls for time fixed-effects and implic controls for country fixedeffects and whin-country autocorrelation. Model Description Coef. Std. Error p-value (M.1) Preferred estimates: Differences-in-differences model (Eq. S.3) Data: all countries wh HDI.85 in 2005 (N = 37 countries; 1,051 observations). ˆβ pre ˆβ post < (M.2) Differences-in-differences model wh lagged HDI (Eq. S.4) Data: all countries wh HDI.85 in 2005 (N = 37 countries; 1,014 observations). ˆβ pre ˆβ post < (M.3) (M.4) Differences-in-differences model (Eq. S.5), using the tempo-adjusted total fertily rate as dependent variable Data: all countries wh HDI.85 in 2005 for which a time series of the tempo-adjusted TFR is available (N = 25 countries; 705 observations, of which 505 include a tempo-adjusted TFR). Note: Tempo-adjusted TFR is used as the dependent variable starting from the year the tempo-adjusted TFR data become available; the standard TFR is used for the preceding years. The model includes controls for the years when TFR changes to tempo-adjusted TFR. ˆβ pre ˆβ post Differences-in-differences model (Eq. S.6), wh adjustment for changes in mean age of mothers at first birth Data: all countries wh HDI.85 in 2005 for which data on mean age at childbearing is available (N = 26 countries; 736 observations). ˆβ pre ˆβ post

10 Re-estimating our preferred model (Model M.1 in Table S.3) using different values for HDI cr, does not affect our conclusions. In particular, using the alternative values 0.85, 0.87, 0.88, 0.89 and 0.90 for HDI cr in estimating the differences-in-differences model (Eq. S.3), continues to result in parameter estimates wh ˆβ pre < 0 and ˆβ post > 0. Hence, our conclusion about a reversal of the HDI fertily relationship at intermediate development levels is not sensive to other choices of HDI cr whin the range We also explored the extent to which our preferred estimates depend on the set of countries on which the estimates are based. We did this by studying the influence of single countries on our preferred estimates by excluding one country at a time, and re-estimating the model Eq. (S.3). The results were robust. The coefficient ˆβ pre was consistently negative wh the estimates ranging from to -3.33, and the coefficient ˆβ post was consistently posive, estimates ranging from 3.27 to 5.05 wh maximum p-value less than 0.001, confirming the reversal. It is conceivable that changes in HDI affect fertily levels only wh some lag, especially since takes about one year to conceive and give birth to a child. We have therefore reestimated the HDI fertily relationship including lagged values of the human development index. Specifically, we have modified our preferred differences-in-differences model to include 1-year lagged values of the human development index, HDI t 1, instead of HDI t, and estimate the relationship TFR = α B post + β pre HDI pre i(t 1) + βpost HDI post i(t 1) + γ t + υ, (S.4) where β pre and β post continue to measure the effects of development (as measured by the HDI t 1 ) on the total fertily rate. The results of this lagged HDI fertily relationship are reported in Table S.3 (Model M.2). This lagged model continues to yield a statistically significant estimate of β post > 0 that supports our conclusion that a posive HDI fertily relationship exists at advanced stages of development. While the point-estimate for β pre continues to be negative in this lagged relationship, is no longer statistically significant. The failure of these analyses to identify a statistically-significant negative effect of increases in HDI on the total fertily rate at HDI levels below 0.86 is probably due to the smaller sample sizes in this model as compared to our preferred estimates, as well as the fact that, due to the focus on the HDI fertily at advanced development levels, our regression analyses are restricted to countries that have attained a HDI of at least.85 by 2005, thereby excluding countries at lower levels of development for which the negative association between HDI and fertily is particularly strong (Figure 1). Further analyses wh 5-year and 10-year lagged values of the human development index also yield a posive and statistically significant estimate for β post, while the estimates for β pre are not statistically different from zero, as before, most likely as a result of the shorter time-series that are available for these analyses and the exclusion of countries at low HDI levels in

11 Adjustment of the Total Fertily Rate for Tempo Effects The recent lerature on low fertily in developed countries has pointed to the important role of delayed childbearing, that is, the ongoing postponement of childbearing to increasingly later ages. 1, 3 In the context of this paper, delayed childbearing is potentially important because the postponement of childbearing can distort the total fertily rate as a measure of the quantum (or long-term level) of fertily. 1 Tempo effects, or the reductions in the total fertily rate resulting from a postponement of childbearing, have been shown to partially explain the very low fertily rates observed in some European countries. 1, 3 We use both a graphical and statistical modelling approach to evaluate the robustness of our key finding about the reversal of the HDI fertily relationship at advanced development stages wh respect to tempo effects. In particular, one could speculate that tempo effects might be at least partially responsible for the observed change in the development fertily association. For example, countries at development levels near the crical level HDI cr = 0.86 might have a more rapid postponement of childbearing than more advanced countries. If this were the case, tempo effects would reduce the TFR more strongly at intermediate than at advanced HDI levels, and the posive association between HDI and TFR in Figures 1 2 could be partially explained by differences in the pace of fertily postponement, rather than by variation in levels among advanced countries. Figure S.2 replicates our earlier cross-sectional analyses in Figure 1, including for 2005 addionally a tempo-adjusted TFR 1 for advanced countries where the postponement of childbearing is most relevant. Tempo-adjusted TFRs were available in 2005 for only 41 countries, including 28 of the 37 countries which had reached a development level of 0.85 by Most importantly, however, Figure S.2 shows that the reversal of the HDI TFR relationship at advanced development stages persists even after adjusting the total fertily rate for tempo effects. Time series of the tempo-adjusted TFR are not readily available from official publications. For 18 of the 35 countries which have reached a HDI of 85 by 2005 we use Bongaarts- Feeney 1 tempo-adjusted TFR time series provided courtesy of Tomáš Sobotka of the Vienna Instute of Demography. For 7 of the 35 countries, we use data on the distribution of births by pary (first and second or higher order births) and mean age of mothers at respective paries to construct comparable time-series of tempo-adjusted TFR. These data were obtained from Eurostat (Eurostat New Cronos Database [Online], accessed at on Oct ). In total, we have time series of tempoadjusted TFR for 25 of the 35 countries that attained a HDI above 0.85 in Frequently, however, the time series are not complete for the period on which preferred longudinal analyses are based. Therefore we use tempo-adjusted TFR as the dependent variable starting from the year the data becomes available (mean year: 1986) and standard 11

12 Total Fertily Rate (TFR) wh adj. TFR Human Development Index (HDI) Figure S.2: Cross-sectional relationship between the total fertily rate (TFR), wh and whout adjustment for tempo effects, and the human development index (HDI) in 1975 and 2005 The 1975 and 2005 data for the HDI and (unadjusted) TFR are identical to Figure 1. Adjusted TFRs are available for 41 countries in The Spearman rank-correlation between HDI and the adjusted TFR in 2005 is.62 (p <.01) for countries wh HDI.9. Countries for which a 2005 adjusted TFR is available include (2005 HDI in parentheses): Australia (0.966), Norway (0.961), Iceland (0.956), Ireland (0.95), Luxembourg (0.949), Sweden (0.947), Finland (0.945), France (0.945), Netherlands (0.945), Uned States (0.944), Denmark (0.943), Japan (0.943), Swzerland (0.942), Belgium (0.94), Spain (0.938), Uned Kingdom (0.936), Austria (0.934), Italy (0.934), Greece (0.918), Germany (0.916), Slovenia (0.913). Data source: World Bank World Development Indicators Online Database 6. Adjusted TFRs were obtained from the European Demographic Data Sheet 2008 (published by the Vienna Instute of Demography, Vienna, Austria) and McDonald P, Kippen R. The Intrinsic Total Fertily Rate: A New Approach to the Measurement of Fertily (Population Association of America Annual Meeting 2007, New York, 2007). Tempo-adjusted TFRs were not available for Canada, New Zealand, Israel, S. Korea, Cyprus, Kuwa, Argentina, Chile and the Uned Arab Emirates 12

13 TFR for the preceding years, and estimate the model TFR = α Bpost + β pre HDI pre + β post HDI post + γ t + υ + c, (S.5) where TFR is the standard TFR up to the year when tempo-adjusted TFR becomes available and after that tempo-adjusted TFR, c is an indicator variable swching from 0 to 1 the year TFR changes to tempo-adjusted TFR, capturing the level shift, and the remaining components in the model are as in model (Eq. S.3). Thus the model (Eq. S.5) is essentially the same model as our preferred model (Eq. S.3) but uses tempo-adjusted data as much as the data is available. The model combines standard and tempo-adjusted TFR because otherwise the data would not be balanced, and estimating a model wh time fixed-effects from unbalanced data would be unstable. Acknowledging these data limations, Model M.3 in Table S.3 reports the results obtained from model (Eq. S.6) using as dependent variable the adjusted total fertily rate when such data is available (25 countries for which 505 of the 705 TFR observations are tempo-adjusted). The model excludes the 10 countries for which no time series of tempoadjusted TFR were available. This estimation yields a statistically significant posive value of ˆβ post = 2.85 (Model M.3 in Table S.3), which, albe being of smaller magnude than the coefficient obtained whout tempo-adjustments, supports our hypothesis of a posive HDI TFR relationship between development and fertily at advanced HDI levels. The estimate for HDI for levels below 0.86 is ˆβ pre is (Model M.3), and while not being statistically significant (p = 0.447) as a result of the limed data availabily and the exclusion of countries wh low HDI, this finding is consistent wh our hypothesis of a reversal in the HDI fertily relationship. To further investigate the robustness of our finding wh respect to possible tempo effects, we also estimated a differences-in-differences model, similar to our preferred model in Eq. (S.3), wh an addional control for changes in the timing of fertily: TFR = α B post + β pre HDI pre + β post HDI post + δ Ā + γ t + υ, (S.6) where Ā is change in the mean age of mothers at first birth, measuring directly the postponement of childbearing, that is, the behavioural process that generates tempo effects. Including a direct measure of the change in the timing of fertily as explanatory variable is an alternative using the tempo-adjusted TFR as the dependent variable, and allows us to control for tempo-effects wh a slightly larger sample than in Model M.3. The estimates of Eq. (S.6), shown in Model M.4 in Table S.3, are consistent wh the ones obtained whout adjustment for tempo-effects: The coefficient ˆβ pre, while not being significant (p = 0.447) possibly due to a limed sample size, is negative and of the same magnude as the coefficient obtained whout adjustments for changes in the timing of childbearing (1.57 vs. 1.59). The coefficient ˆβ post is posive, statistically significant (p < 0.01), and only slightly smaller 13

14 than the coefficient obtained whout adjustments for changes in the timing of childbearing (3.06 vs. 4.07). Further direct adjustments in Eq. (S.6) for changes in mean age of mothers at second and higher order births did not change the results. In summary, the addional analyses in Models M.3 M.4 confirm that, even after controlling for changes in the timing of childbearing, the coefficient ˆβ post, which measures the effect of increases in HDI on the adjusted total fertily rate once a country s HDI level is above the crical level 0.86, is posive and statistically significant. Adjusting for tempo effects, eher by including the adjusted TFR or by including changes in the mean age at childbearing in the regression model, therefore, does not change our main substantive conclusion that a posive HDI fertily relationship emerges at advanced stages of development wh HDI levels above Therefore, while our analyses suggest that changes in the timing of childbearing contribute to the reversal of the HDI fertily relationship from negative to posive at advanced development stages, the reversal of the HDI fertily relationship is not driven by these changes in the timing of childbearing. In contrast, the posive HDI fertily relationship at HDI 0.86 is also present in analyses that adjust for tempo effects (Models M.3 M.4 in Table S.3), suggesting that increases in HDI at advanced development stages lead to a higher quantum (or long-term level) of fertily. 14

15 1. Bongaarts, J. & Feeney, G. On the quantum and tempo of fertily. Population and Development Review 24, (1998). 2. Balter, M. The baby defic. Science 312, (2006). 3. Kohler, H.-P., Billari, F. C. & Ortega, J. A. The emergence of lowest-low fertily in Europe during the 1990s. Population and Development Review 28, (2002). 4. Sobotka, T. Is lowest-low fertily in Europe explained by the postponement of childbearing. Population and Development Review 30, (2004). 5. UNDP. Statistics of the Human Development Report (UNDP Human Development Report Office, New York, 2008). Accessed online on 29 September 2008 at 6. World Bank. World Development Indicators Online Database (The World Bank Group, Washington, D.C., 2008). [Online] Available at Accessed on 26 September Fargues, P. Protracted national conflict and fertily change: Palestinians and Israelis in the twentieth century. Population and Development Review 26, (2000). 8. Wooldridge, J. M. Econometric Analyses of Cross Section and Panel Data (MIT Press, Cambridge, MA, 2002). 9. Baltagi, B. H. & Wu, P. X. Unequally spaced panel data regressions wh AR(1) disturbances. Econometric Theory 15, (1999). 10. Lee, R. D. The demographic transion: Three centuries of fundamental change. Journal of Economic Perspectives 17, (2003). 11. Bongaarts, J. & Watkins, S. C. Social interactions and contemporary fertily transions. Population and Development Review 22, (1996). 12. Bryant, J. Theories of fertily decline and the evidence from development indicators. Population and Development Review 33, (2007). 15