The intuition that fertility intentions and behaviours entail an element of orientation towards the future can be traced back to seminal discussions of declining birth rates during the period when Europe and North America were undergoing the demographic transition. Frank Notestein famously stated that “the reduction of fertility requires a shift in social goals”, brought about “in response to drastic changes in the social and economic setting” (Notestein 1945, 40). As Kirk (1996, 364) noted, Notestein drew from Carr-Saunders (1922), who focused on the role of customs and codes of conduct in regulating fertility in pre-transition societies in his book The Population Problem. By contrast, Caldwell (1976, 325) criticised this “Malthusian” emphasis on “attitudes, beliefs, traditions and irrationality”, and instead stressed the role of rational choice in fertility behaviours. Building on Simon (1974), who argued that “fertility is everywhere subject to much rational control”, Caldwell (1976, 355) observed that “fertility is high or low as a result of economic benefit to individuals, couples, or families in it being so”.

According to Caldwell, fertility intentions primarily depend on the benefits of childbearing, which are, in turn, determined by the direction and the magnitude of intergenerational wealth flows, and their net balance, “from parents to children and from children to parents – over the period from when people become parents until they die” (Caldwell 1976, 344). Economic formulations of Caldwell’s intuition model fertility as stemming from the maximisation of a dynastic utility function, based on the consumption of agents at one point in time, and on the consumption of the agents’ descendants from that point in time into the future (Becker and Barro 1986). In such a framework, the decision of a couple to start having children is determined by the employment prospects of children, the monetary and psychological costs of avoiding childbearing (Becker 1960), and the interaction between the quantity and the quality of children (Becker and Lewis 1973). It follows that fertility declines if the relative price of having children increases, if income falls, or if there is a change in the shape of the couple’s utility function for children versus for other goods (McDonald 2001).

Most recently, the evidence that that a number of countries are moving towards having low and lowest-low fertility levels has renewed the debate over whether economic rationality represents a satisfying explanatory principle when seeking to account for these developments (Caldwell and Schindlmayr 2003; Billari et al. 2004; Caldwell and Schindlmayr 2004; Caldwell 2004). Lutz et al. (2006) outlined a sociological driver of change by proposing that the formation of family size ideals is supported through socialisation, and is passed on from one cohort to another. The role played by normative pressures has been also integrated into social psychological models, with extensive references to Ajzen’s (1991) theory of planned behaviour. This theory posits that social norms interact with the economic benefits of childbearing perceived by prospective parents, and with the control that they believe they have over their desired behaviours (see, e.g., Philipov et al. 2006; Billari et al. 2009). Sobotka (2009, 390) has suggested that when faced with eroding social norms against childlessness and having only one child, young adults may “reduce their fertility intentions by becoming more realistic when assessing their fertility goals, taking into account competing lifestyle alternatives and their growing awareness of different obstacles that may unfold later in life”.

A connected but relatively unexplored dimension of fertility intentions is how actual behavioural control might moderate the impact of intentions on behaviours. Actual behavioural control refers to the ability to perform a desired action, and, in the case of childbearing, has been variously defined as determined by internal and external constraints, such as wealth, income, employment, education, housing and health status (Billari et al. 2009, 447; see also Schoen et al. 1999, 791). According to Ajzen (2011), a further moderating factor in the intention-behaviour nexus is the ability of individuals to self-regulate. Departing from McEachan et al. (2011), who reviewed empirical evidence indicating that the association between intentions and behaviours weakens with the temporal distance between the measurement of an intention and the observation of a behaviour, Ajzen (2011, 1115) pointed out that “as time passes, an increasing number of intervening events can change people’s behavioural, normative or control beliefs, modify attitudes, subjective norms or perceptions of control, thus generating revised intentions”.

In this paper, I draw from advances in the field of behavioural economics to offer a simple analytical operationalisation of behavioural control. In particular, I focus on the role played by time discounting, or the preference for gratification sooner rather than later (Hariri et al. 2006). Ubiquitous in human decision-making, discounting may lead individuals to opt for rewards that happen sooner but are smaller, rather than for rewards that are larger but delayed. For example, a person might favour a $100 reward today over a $110 reward tomorrow.¹ Empirically, rates of discounting are found to decline over time. Following on the previous example, individuals might indeed prefer getting $100 today to receiving $110 tomorrow, but may be willing to wait 31 days to get $110 if the alternative is getting $100 30 days from now. In other words, willingness to wait increases as a function of the expected delay in getting the reward (Loewenstein and Thaler 1989).

Crucially, this discount structure, which is also called ‘hyperbolic discounting’, creates “a conflict between today’s preferences, and the preferences that will be held in the future” (Laibson 1997, 445). According to Loewenstein and Thaler (1989, 186), “if the discount rate declines over time, people will always consume more in the present than called for by their previous plans”.² Thus, time discounting can be seen as a realistic integration of the classical assumption of rationality of economic agents, or as way of incorporating the role of ‘impatience’ into analyses of human decision-making (Thaler 1981; 2018). While discounting has been investigated in a wide variety of settings, from saving to health behaviours (Strotz 1955; Laibson 1997; Barro 1999; Harris and Laibson 2001; Gruber and Kőszegi 2002; Diamond and Kőszegi 2003; Barlow et al. 2016), and particularly with regard to the implications for households in developing economies (Banerjee and Mullainathan 2010), little attention has been paid to the role discounting might play in childbearing decisions.

To the best of my knowledge, only three previous studies have investigated the childbearing implications of discounting (Wrede 2011; Wigniolle 2013; Robson and Szentes 2014). Among these studies, Wrede (2011) and Wigniolle (2013) derived analytical solutions from a standard quantity-quality trade-off à la Becker, augmented by adding (quasi-)hyperbolic discounting to the utility function of a unitary household. Quasi-hyperbolic discounting penalises future rewards by a coefficient βkt, composed by a constant factor, β<1, which captures the presence of a negative preference for delays; and by an exponential factor, kt, for any t>0, which grows at a constant rate as the time from the reward increases. By setting βk=1, this utility formulation is equivalent to that of a rational agent, while a fully hyperbolic formulation of the discount factor, such as (11+kt)β/α, with β and α greater than zero, can also be used to capture decreasing rates of discounting.

To analyse the childbearing implications of discounting, I build on Wigniolle (2013, 72), and depart from a unitary household model with three periods. The number of children (quantity) is set in the first period, while the quality is set in the second period. Because parents may be assumed to want children, each child enters positively the utility of the household in period 1, but at an opportunity cost in terms of consumption due to the resources devoted to childrearing. In the second period, the household selects the level of ‘quality’, which also entails a consumption trade-off because investing in children is costly. In the third and final period, the quantity and the quality jointly enter the household’s utility. This may reflect the altruism of parents who value both the number and the quality of their children, or the fact that children may represent an investment good, as they can provide support to their ageing parents (Wrede 2011, 1056–1057).³

In this toy model, the representative household chooses the optimal number of children by maximising the instantaneous utility obtained across all three periods. In this simple set-up, quasi-hyperbolic discounting has straightforward implications: by penalising rewards that come later rather than sooner, it implies that the benefits of quantity, which enter utility in period 1, bear more weight in the household’s maximisation problem than the benefits of quality, which enter utility in period 2. Thus, ceteris paribus, a ‘hyperbolic discounter’ household will invest relatively more in quantity than in quality. The main hypothesis is as follows: countries characterised by a greater bias towards the present (less future-oriented) will also have higher total fertility rates (TFR) than countries that are less present-biased (more future-oriented) (Hypothesis 1).

Given that in Becker’s framework, quantity and quality are connected through the intertemporal budget constraint, setting a higher number of children in period 1 may lead the household to resort to a dynamic inconsistent solution in terms of quality – i.e., a lower-than-desired investment – in period 2, which is a feature of hyperbolic discounting. At the same time, because hyperbolic discounting implies that time preferences decrease over time, a higher level of future orientation – i.e., a greater willingness to wait – may lead the household to appreciate the prospect of receiving more benefits at a later point in time. This can be thought of as a shift in the relative weight of period 3. Thus, as the bias towards the present decreases (future orientation increases), the relative weight of future periods for the household’s utility also increases, leading households to appreciate the benefits of old-age support (or of altruism) more. Because in this model the benefits of old-age support (altruism) are determined by both quality and quantity, a highly future-oriented household might have more children than a moderately future-oriented household. I therefore expect that at high levels of future orientation, the negative relationship between future orientation and fertility reverses (Hypothesis 1a).

Through their digital behaviour, internet users leave traces that can be leveraged to measure their relative interest in the present, the past or the future. Preis et al. (2012) argued that the levels of relative interest in the future or in the past can be measured by comparing the volume of searches for specific ‘years’. Exploiting the fact that Arabic numerals are widely used across the world, they built a ‘future orientation index’ as the ratio of Google search volumes for future (t+1) and past (t-1) ‘number-years’ (e.g., interest in 2021 compared to interest in 2019, with reference t=2020). In Figure 1, I show how interest in different ‘number-years’ varies across time by plotting monthly worldwide search volumes for the numbers ‘2013’, ‘2014, ‘2015’ and ‘2016’, from 1 January 2012 to 31 December 2017.⁴ I obtained these volumes through Google Trends, which draws from a subsample of all the queries submitted to Google.com. Search volumes are divided by the total number of searches, and then rescaled from zero to 100; thus, they represent a measure of the relative interest in specific keywords in a specific time period and geography.

Looking at Figure 1, it is possible to observe that queries for the ‘number’ capturing a given year start rising during the immediately preceding months, with a peak corresponding to December of the previous year. The relative interest in a given ‘number-year’ remains relatively stable throughout the following months, with peaks related to specific events. In 2014, for example, the male 2014 FIFA World Cup caused a surge in searches for ‘2014’ during the summer months of that same year. However, as a new year approaches, searches for the current ‘number-year’ start declining quickly, while searches for the next ‘number-year’ rise. After each year has passed, the respective ‘number-year’ searches stabilise at low levels with long right tails, which indicates continuous interest in past events. This pattern is quite regular across the six years of data plotted in Figure 1. Throughout the same period, it can also be observed that searches for the present (current ‘number-year’) systematically outnumber searches for both the past and the future.

In this paper, I leverage cross-country variations in relative interest in the future rather than in the present, and use it as a proxy of future orientation. In particular, I propose a country-year measure of future orientation leveraging Google search data from 2010 to 2016.⁵ This Future Orientation Index is computed as the ratio of the aggregated yearly volume of Google search queries for the ‘next year’ to Google search queries for the ‘current year’, corrected for differentials in internet adoption rates across countries. Formally, this can be written as:

where FOIjt is the Future Orientation Index for a given year (t) and country (j). Equation (1) is composed of two terms. The first term captures the relative interest in the immediate ‘future’ compared to in the ‘present’. In 2016, for example, future orientation is captured as the ratio of the aggregated country-level volume of searches for the keyword ‘2017’ to the aggregated country-level volume of searches for the keyword ‘2016’. Relative to the measure devised by Preis et al. (2012), comparing searches for the present (t) and the future (t+1), as proposed in Equation (1), has the advantage of yielding a measure that is directly interpretable in terms of a more theoretically relevant ‘bias towards the future’.

In Equation (1), the second term corrects raw ‘future orientation’ for cross-country differentials in internet penetration, as these differences might introduce biases into the representativeness of the data. The share of the population with internet access varies considerably across counties: in 2016, this figure ranged from 1.17% for Eritrea to 98.24% in Iceland. Wide differentials in internet access raise two issues: a) an instability bias, as short-run variations in internet access might make within-country year-to-year comparisons less meaningful; and b) a selective adopter bias, as internet usage and access are not orthogonal to individual-level characteristics that are likely to be correlated with future orientation, such as income and education. Therefore, especially in countries with low internet penetration, the observed volumes might be skewed towards specific demographics, such as towards individuals who are younger, more educated and more urban, and who might therefore display higher levels of future orientation compared to the baseline for the country.

Several methods have been proposed to address bias in non-representative internet samples (see Zagheni and Weber 2015 for a discussion). The correction proposed in Equation (1) equals one in the country-year with the highest proportion of internet users, and moves towards zero as the percentage of internet users declines. Therefore, the magnitude of the change in the raw future orientation measure for each country is inversely proportional to the rate of internet adoption: for high values of adoption, FOIjt will not differ substantially from the raw future orientation measure. This follows from the assumption that the representativeness of the data increases as the internet adoption rate rises. Reflecting the further assumption of a negative monotonic relationship between individual-level future orientation and the likelihood of having internet access, the raw future orientation measure is reduced proportionately as the percentage of internet users declines. I repeat the procedure set out in Equation (1) for 200 countries and territories across the world in any given year from 2010 and 2016, and report summary statistics of this FOI in Table 2.

Because my weighting approach always returns low FOI values for those countries with low internet penetration rates, social and economic development factors that are related to modernisation, and are thus associated with both fertility and internet penetration rates, might lead to omitted variable bias when testing for the correlation between the future orientation measure and fertility at the country level. To test the robustness of my findings, I employ in this paper two sets of alternative future orientation measures. The first is the uncorrected FOI based on raw Google Search volumes. The second set of measures is obtained by multiplying my raw FOI by the correction factor proposed by Zagheni and Weber (2012). This correction factor (CF) can be written as: CF=pj(e-k-1)(e-kpj-1) where pj is the internet penetration rate by country (j) and k is a parameter that captures the intensity of the selective adopter bias.⁶ According to Zagheni and Weber (2012, 351), “the parameter k determines how fast the selection bias increases with a decrease in Internet penetration rates. Low values of k imply that the selection bias is small, even at very low Internet penetration rates. Conversely, high values of k mean that the selection bias increases substantially when Internet penetration rates become lower”. Thus, varying the value of k allows me to account for varying degrees of (unobserved) self-selection in the sample. Following Zagheni and Weber (2012), I experiment with several values of the k parameter (k=0.005;k=0.05;k=0.5;k=5;k=50).⁷ Finally, I also use the naïve procedure proposed by Preis et al. (2012), who excluded countries with less than five million internet users from their sample. The correlation coefficients between these corrected FOI and the raw measure given by FOIjt=GoogleSearchfort+1jtGoogleSearchfortjt are reported in Table 1.

To further stress the appropriateness of my measure for capturing future orientation, I validate my index of country-level FOI against ‘ground truth’ data. In Figure 2, I plot the country-year FOI thus obtained with the Long-Term Orientation Index proposed by Hofstede (1991), and measured by Galor and Özak (2016) based on Wave 1 to 5 (1981–2009) of the World Value Survey (WVS), for 87 countries across the world.⁸ This Long-Term Orientation Index (LTO) was constructed based on the following WVS item: “Here is a list of qualities that children can be encouraged to learn at home. Which, if any, do you consider to be especially important?” As explained in Galor and Özak (2016, S93), “An individual is considered to have Long-Term Orientation if she answered ‘Thrift, saving money and things’ as an especially important quality children should learn at home”. The country-level LTO also spans from zero to 100, and has been shown to correlate with national saving rates and other theoretically relevant macro-level outcomes (Hofstede et al. 2010, 38). The Pearson’s coefficient for the correlation between FOI and LTO is 0.43.

As can be seen from Figure 2, I also validate my FOI against a measure of time preference obtained from the Global Preference Survey (GPS), a large international survey covering 80,000 respondents in 76 countries that represent approximately 90% of the world’s population (Falk et al. 2016, 2018).⁹ In the GPS, the measure of time preference, or ‘patience’, is derived from the combination of responses to a quantitative and a qualitative survey: in the quantitative survey, respondents were confronted with a series of five interdependent binary choice tasks in which they had to decide between receiving a payment today or larger payments in 12 months. In the qualitative survey, respondents were asked to self-assess their ‘willingness to wait’ on an 11-point Likert scale, based on the following question: “How willing are you to give up something that is beneficial for you today in order to benefit more from that in the future?” The two sets of responses were then combined with a 71% weight assigned to the quantitative survey (Falk et al. 2018, 1653–1654). I find that the FOI, based on volume data from Google Search and corrected for internet adoption rates, correlates positively and strongly with the GPS’s measure of patience (with a 0.70 Pearson correlation coefficient).

In the next section, I test for the correlation between future orientation and fertility by fitting a standard OLS model with country and year fixed effects, where country-year TFR¹⁰ is regressed on the FOI obtained from Equation (1). To account for confounders, I include in the model a range of country-level covariates: i.e., with the aim of controlling for a country’s population structure, I include population size¹¹ and the age-dependency ratio.¹² I also control for a country’s level of socio-economic development by adding a measure of gross domestic product (GDP), ¹³ expressed in constant 2010 United States dollars, as well as the widely used Human Development Index (HDI).¹⁴ The HDI has been employed in demographic research to describe the transition from high to low fertility (Bongaarts and Watkins 1996), as well as to outline a recent reversal and the non-linearity of the HDI-fertility relationship (Myrskylä et al. 2009). Consistent with this literature, I allow for a non-linear effect of the HDI on fertility through the inclusion of a squared HDI variable. Furthermore, the inclusion of country-level fixed effects allows me to control for geographic characteristics that are stable across time, while the inclusion of year fixed effects allows me to account for the existence of idiosyncratic time factors that are related to events that shift interest globally.

Despite using information on internet adoption to correct the raw FOI obtained from the Google Trends tool, I also include the levels of internet adoption in the country as an additional explanatory variable for fertility (Billari et al. 2019). Unfortunately, not all of these data are available for all of the countries for which the FOI can be computed. In the next section, I present results using all data for all countries in the sample, including for those countries for which some values are missing across the set of controls. Thus, in order to retain the highest statistical power available, the models are tested on marginally varying samples. However, the results are robust when listwise deletion is employed, and the models are estimated only on the subset of countries for which data for all controls and all years are available (not shown). Table 2 presents summary statistics for the variables employed in the following analysis. All analyses are run in STATA v.16/SE.

I re-express the FOI obtained from Equation (1) on a zero (maximum present bias) to 100 (maximum future orientation) scale, and plot on a world map the country averages for the period 2010–2016. Overall, the mean FOI value is 24.92. Even though in all years the searches for the current ‘number-year’ exceed the searches for the future ‘number-year’ by 9.49 times on average, the variation in the FOI is fairly large, with a standard deviation of 20.90. The country scoring the highest on the future orientation measure is Lichtenstein (85.79), while the country scoring the lowest is Somalia (0.79). Just as countries vary greatly in their overall economic development and in their stage in the demographic transition, they differ in their future orientation levels as well. The average FOI value for countries classified as ‘high-income’ by the World Bank is 48.97, whereas the average FOI for ‘low-income’ countries is only 4.54. Looking at Figure 3, it is also possible to qualitatively observe that the countries with the highest levels of future orientation (Germany, Japan, South Korea) are also characterised by lowest-low fertility, while the countries with high TFRs, such as Sub-Saharan African countries, also tend to have lower levels of future orientation.

In Figure 4, I present some descriptive evidence on the association between the TFR and the FOI from 2010 to 2016 for 200 countries and territories across the world (black circles). The evidence indicates that the TFR quickly declines as the FOI increases. However, this correlation becomes flat when moving towards higher levels of future orientation (lower levels of present bias). In Figure 4, I also separately plot the two elements that constitute the FOI; i.e., the raw search ratio and the standardised levels of internet penetration. When looking at these two factors separately, it appears that the shape of the association between the FOI and the TFR is imposed by the distribution of internet penetration levels, while the raw search ratio shows no clear association with the TFR.

In order to verify whether the association between the FOI and the TFR is an artefact created by the way in which the FOI is computed, I plot in Figure 5 the raw search ratio (GoogleSearchfort+1jtGoogleSearchfortjt) against the TFRs for each country and year by levels of internet penetration (≤20%, >20-≤40%, >60-≤70%, and >80-100%). As was discussed in Section 1, my correction approach assumes that the representativeness of the data increases with the level of internet adoption. Indeed, at low levels of internet penetration (Internet<20%) the raw Future Orientation Index has a higher degree of variability (min=0; max=0.41; sd=0.54), and the association between this search measure and the TFR has a positive slope.¹⁵ However, at the highest level of internet penetration, a significant negative association seems to appear (r=-0.2048, p=0.0069). In the regression analysis, I experiment with different correction approaches to relax the assumptions driving the correction of Equation (1).

In Table 3, I show results for the effects of future orientation on total fertility rates. Column (1) regresses the TFR on the FOI only: in line with Hypothesis 1, the coefficient of this relationship is found to be negative and significant at the 0.001 level. The size of the coefficient implies that a one-unit increase in the FOI leads to a 0.04 decrease in the TFR at the country-year level – or that a standard deviation increase in the FOI causes the TFR to decline by 0.8 points on average, ceteris paribus. This result is robust to the inclusion of year- and country-fixed effects. However, the inclusion of these controls shrinks the size of the coefficient, which becomes only marginally significant (Table 3 Column 3). The introduction of country-year levels of internet penetration (βinternet=0.00, p=9.65) in the regression equation (Table 3 Column 4) decreases the size of the beta coefficient, which also becomes better defined in statistical terms (statistical significance at the 0.001 level).

This result implies that failing to include internet penetration repressed the effect of future orientation on fertility, as the level of internet penetration appears to be negatively correlated with the FOI (implicitly confirming the selectivity of internet users at low levels of internet penetration) and positively correlated with the TFR. In Table 3 Column 5, I control for the population structure by introducing both the population size and the age-dependency ratio. The inclusion of these additional variables (βpopulation=-0.00, p=-1.06; βdependency=0.02, p=18.81) does not alter the significance of the correlation between the FOI and the TFR, while marginally shrinking the size of the correlation coefficient. In Column 6, I control for socio-economic development by introducing variables capturing the GDP (βGDP=0.00, p=2.11) and the HDI (βHDI=-9.54, p=-10.68); while in Column 7, I allow for a non-linear relationship between the TFR and the HDI (βHDIsq=7.23, p=10.70). The correlation between the FOI and the TFR is robust to the inclusion of these controls. Based on the model of Column 6, I conclude that a one-standard-deviation increase in the FOI decreases the TFR by 0.04 points on average.

I also run a set of regressions fitting a non-linear model of future orientation by allowing a FOI squared (FOI2) independent variable to enter the TFR regression (Table 3 Pane b). Even in these models, the FOI is negatively correlated with the TFR. However, the size of the estimated coefficient is noticeably larger across all regressions. Additionally, and consistent with the descriptive evidence presented in Figure 4, I find that at higher levels of future orientation, the correlation with the TFR becomes positive. This result is consistent and significant across all models, even when allowing for non-linear effects of socio-economic development through the inclusion in the model of the HDI and the squared HDI, as is done in Column 7 (significant at the 0.05 level). These results are robust to the inclusion of the FOI with a one-year lag as an explanatory variable (not shown). Overall, these results appear to confirm that future orientation is negatively associated with total fertility, but that this relationship ceases to be negative at high FOI levels.

In Table 4, I replicate the model of Equation (1) using alternative future orientation measures. I first experiment using the raw Google Search ratio of ‘next year’ to ‘current year’ volumes. In the linear model, the correlation between the raw Google search ratio and the TFT is not significant (Column 1 pane a). However, when allowing for non-linear effects in the relationship between the TFR and future orientation, I observe a negative coefficient significant at the 0.10 level for the FOI, and a positive coefficient significant at the 0.05 level for the FOI2 (Column 1 pane b). These results do not change when the sample is restricted to those country-years with internet user numbers above five million, as proposed by Preis et al. (2012). On the other hand, the results from four out of the five correction factors proposed by Zagheni and Weber (2012) confirm that there is a significant and negative correlation between the FOI and the TFR in the linear model (Columns 4 to 7). Moreover, the sign and the significance of the coefficients are robust in the non-linear model as well – with the exception of k=0.05.