In School or at Work? Evidence from a Crisis

Abstract

This paper makes use of the income variability generated by the macroeconomic crisis of 2001/2002 to examine schooling outcomes in Argentina. The effect of this macroeconomic swing is examined with a focus on whether the income or substitution effect dominates in the decision-making of young people. It is demonstrated that the probability of being in school was 6.5–10 percentage points higher in May 2002 than in 2001 for 15–18-year-olds. This is probably the largest (positive) effect found in the developing country literature so far and is comparable to the effect of a 10% increase in household income. For 19–25-year-olds, the probability is between 2 and 6 percentage points higher. Results are robust to a wide range of controls and specification checks. Difference-in-difference panel estimation corroborates these findings and shows that the increase in schooling seems to be driven by a decrease in school exits during the crisis.

Notes

The author wishes to thank Francis Teal for very helpful discussions. The paper also benefited from comments received from Adrian Wood, Simon Appleton, Martin Browning, Marta Favara, Julian Aramburu, Oxford University's Gorman and Labour and Applied Microeconometrics, RES/IDB, IEA-2008, NEUDC-2008, LACEA-2008 and IZA Employment and Development 2009 seminars participants. This paper has circulated as “How do crises affect schooling decisions? Evidence from changing labour market opportunities”. This document reflects the opinions of the author and does not represent the opinions of the Inter-American Development Bank or its Board of Directors. The usual disclaimer applies.

¹ Related papers on the positive (negative) impact of crises (economic booms) on health outcomes are Ruhm (2000, 2003, 2005), Dehejia & Lleras-Muney (2004) and Schady & Smitz (2010).

² The results were contradictory for two reasons: first, the age group chosen by these authors was inappropriate; and second, their identification strategy was directed towards analysing different aspects of the crisis (i.e. only via the effect on household income).

³ To the author's knowledge, panels for the full country have not been used in any published paper, as there are some doubts over the consistency of household identifiers over time.

⁴ From these observations, it therefore seems unlikely that differential access to credit or changes in rates of return to education can explain the patters of attendance in Argentina, at least over the recession-crisis period of 1998–2002.

⁵ In terms of the recession years, the first reduction in GDP took place in the last quarter of 1998 following the Russian financial crisis (see INDEC web page, www.indec.gov.ar). However, as I am taking May waves from EPH, the GDP reduction will only be reflected in the May 1999 survey.

⁶ Despite the small “t”' (t = 11) I still show the Zivot–Andrews (1992) unit root test in Figure 1.

⁷ The law sets the minimum age for employment at 14 years, which is why we only show employment rates for those older than 15 years.

⁸ Given that surveys cover only urban areas, most statistics are not significantly affected by seasonality issues.

⁹ Each panel therefore will have 25% of the population of a wave (i.e. as 25% of the sample “leaves” in each wave, then after 2 years only 25% of the original sample is left).

¹⁰ The panels have exactly the same information as the cross-sectional data, the only difference being that they allow us to follow a fraction of the individuals.

¹¹ More precisely, I use: 791 15–18-year-olds interviewed between May 1998 and May 2000; 698 between May 2000 and May 2001; 793 between October 2001 and May 2002; and 739 between May 2002 and May 2003. For the panel exercise, I extend the analysis to 2003, to take into account the fact that school decisions are made in March each year. In other words, the crisis might have had an effect only 1 year later. Note that two rounds of data are used for the first column in the transition matrix in Table 5 as per small sample sizes (see the second note in that table).

¹² According to the Ministry of Education, this group can work only if they have completed compulsory schooling, which normally ends at age 15 (since 1997) and at age 12 (before 1996). Before the 1996 education reform (see Section 4.1), the minimum age required to start high school was 13 years as of 30 June of the entry year. With this reference, an 18-year-old student will be in the last year of high school.

¹³ The EPH data show that those 15–18-year-olds that report themselves as being employed are working an average of 37 hours per week.

¹⁴ As both these categories are much more numerous in Argentina than in other Latin American countries.

¹⁵ The multinomial logit suffers from a shortcoming when analysing this population, because it assumes the hypothesis of independence of irrelevant alternatives (IIA hypothesis). In this context, this would imply that labour market choices are independent from education. Luckily, Hausman & McFadden (1984) proposed a Hausman-type test of the IIA property based on the comparison of two estimators of the same parameters. One estimator is consistent and efficient if the null hypothesis is true, while the second estimator is consistent but inefficient. Moreover, Hausman & McFadden (1984) suggest that if a subset of the choices is truly irrelevant with respect to the other alternatives, omitting it from the model will not lead to inconsistent estimates.

¹⁶ The student category includes those who combine school with work. For simplification I pool both full-and part-time work, but I split them as a robustness check (not reported). I also pool all types of employment (i.e. self, salaried formal or informal, public or private sector).

¹⁷ It is worth clarifying that by including year dummies, I am controlling for aggregate effects in a particular year.

¹⁸ I have also included per capita household income in the regressions and replicated Rucci's strategy, and the results remain unaltered. I have also used predicted income because of the potential endogeneity of actual income in a child employment regression, and the results also remain unaltered. See Table 3.

¹⁹ There is also a literature, less directly related to this paper, on commodity price changes (Edmonds & Pavcnik, 2005; Kruger, 2007) and production shocks (Jensen, 2000; Beegle et al., 2005) in a variety of contexts.

²⁰ Sosa & Marchionni (1998) found that drop-out rates were higher for boys than for girls aged 13–19 in the GBA. Rucci (2004) also found that males were more likely to drop out of school during the 1998–2002 Argentine crises. The types of job available for 13–18-year-olds in Argentina seem to be better suited to boys than to girls. Unfortunately, if girls are working at home, this will not recorded in the household surveys.

²¹ One might have expected larger coefficients for boys if the opportunity cost hypothesis is valid given different attitudes on the part of parents towards boys versus girls in terms of employment.

²² Running a regression on the 1995 dummy shows that this crisis had no significant effect on attendance. Results are available on request.

²³ It is worth mentioning that although I am reporting only the 2002 dummy, the recession-year dummies (May 1999–May 2001) are all significant and positive.

²⁴ To avoid endogeneity of income with household head's education, regressions that transform years of schooling of the household head into a categorical variable by levels (primary, secondary and university completed) and interact this variable with the crisis dummy were also analysed. The interaction coefficients show that households with heads with primary education (a very small number in the sample) have sought to protect their investments in schooling less than their more educated counterparts. On the other hand, there were no significant differences between the coefficients of the interactions between secondary-and university-educated heads.

²⁵ I have distinguished between short distance commuter (for instance, in the case of GBA this would be either within the City of Buenos Aires or within Conurbano) and long distance commuter (i.e. to places outside the GBA).

²⁶ It is important to note that while McKenzie (2003a) uses wave-to-wave variation, I am unable to do so for two reasons: the smaller sample sizes for the age group I look at; and, most importantly, schooling decisions are yearly decisions, made in March every year in Argentina, whereas labour decisions (i.e. the ones he studies) are more volatile and could be captured in a few months.

²⁷ The May 2000–May 2002 waves of the EPH are used to form overlapping panels of two observations per household or individual, giving four periods of changes. With that I run a probit, which gives results similar to those in the text (available on request). I have also tried to estimate the schooling– employment decisions using a multilevel latent variable model (Generalized Linear Latent And Mixed Models, GLLAMM) in order to follow the MNL (Multinomial Logit) model. Unfortunately, STATA experienced convergence problems, possibly due to the small sample size.

²⁸ Assuming the income effects are larger than substitution effects during growth periods, which seems to be the case in most of the articles reviewed for this paper. This relationship seems to be different during contractions, as shown in this paper.