Cyclicality, Mortality, and the Value of Time: The Case of Coffee Price Fluctuations and Child Survival in Colombia

Abstract

Recent studies demonstrate procyclical mortality in wealthy countries, but there are reasons to expect a countercyclical relationship in developing nations. We investigate how child survival in Colombia responds to fluctuations in world Arabica coffee prices – and document starkly procyclical child deaths. In studying this result’s behavioral underpinnings, we highlight that: (1) The leading determinants of child health are inexpensive but require considerable time, and (2) As the value of time declines with falling coffee prices, so does the relative price of health. We find a variety of direct evidence consistent with the primacy of time in child health production.

1. Introduction

Prominent economic studies published in the past decade document procyclical mortality in the United States and other wealthy nations (Ruhm 2000, Dehejia and Lleras-Muney 2004, Neumayer 2004, Ruhm 2005, Tapia Granados 2005, Ruhm 2006).¹ In developing countries, however, there are reasons to expect the opposite relationship. Credit constraints and other market imperfections lead to incomplete consumption smoothing, potentially making health vulnerable to economic downturns (Sen 1981, Behrman and Deolalikar 1988, Sen and Drèze 1989). Caloric intake and dietary quality may deteriorate as a consequence, and the use of important health services may fall.

Casual empirical observation in Colombia does not match this expectation. For example, Figures 1a and 1b show a distinctly procyclical relationship between the aggregate unemployment rate and de-trended infant and child mortality rates.² Previous developing country studies are remarkably mixed, reporting both procyclical and countercyclical deaths alike as well as acyclical patterns of health.³ These differences could partly be due to variation in methodologies for disentangling the complex interrelationship between economic conditions and mortality. Importantly, such heterogeneity may also be explained by the variety of different behavioral mechanisms implicated by homogeneously-termed “macroeconomic shocks.”⁴

Figure 1.

To address the endogeneity of economic cycles (and to focus on one key behavioral mechanism), we investigate how infant and child survival in Colombia respond to abrupt changes in world Arabica coffee prices. Colombia is the world’s leading producer of washed Arabica coffee beans (second only to Brazil in total coffee production), and as Colombia’s principal export throughout much of the 20^th century, coffee plays an important macroeconomic role (Palacios 1980, Vinod 1985, Cárdenas 1994, Uribe 2000, and CEDE 2002).⁵ Because of Colombia’s prominence on world markets, we examine three external events that caused abrupt changes in world coffee prices: frosts that destroyed a large share of Brazilian coffee groves in 1975, the Brazilian drought of 1985, and the 1989–90 collapse of the International Coffee Agreement (governing cartel-like behavior among coffee producing nations). Because price shocks occurred in both directions, we are able to study both booms and crashes (Meyer 1995).

Given the shortcomings of developing country vital registries, we also employ a novel measure of survival (or cumulative mortality): cohort size (Jayachandran 2009). Inferring mortality from cohort size offers several distinct advantages over other measures. These include freedom from under-reporting, coverage of fetal deaths, and the ability to capture lagged mortality.⁶

Overall, we find a stark pattern of countercyclical cohort size. For a county with median coffee cultivation, a 25% birth year price increase (decrease) is accompanied by a 0.4% – 2.0% decrease (increase) in cohort size.⁷ Because we measure cohort size by place of birth rather than by place of residence, migration is unlikely to play a role in explaining this result. We also restrict our analyses of each price shock to cohorts conceived before a shock occurred, minimizing the influence of fertility. Attributing our estimates to mortality, implied procyclical changes in deaths under age five are approximately 15%.

We then seek to understand the behavioral foundations underlying these results. In doing so, we highlight that for institutional reasons, coffee price fluctuations in Colombia implicate one primary behavioral factor: the return to working (or the opportunity cost of time). Time is a critical health input, so all else equal, the relative price of health falls with the return to working (Grossman 1972, Gronau 1977, Rosenzweig and Schultz 1983, Mwabu 1988, Vistnes and Hamilton 1995). This is especially true for child health in developing countries: the most important determinants of child health are inexpensive but require large amounts of time (bringing pure water from distant sources, practicing good hygiene, and traveling to distant facilities for free preventive and primary health services, for example). Other behavioral factors present in related studies are largely absent in this context: coffee is not a health input (unlike many agricultural commodities); it accounts for a negligible share of total household consumption; and given Colombia’s institutions, coffee prices are unrelated to the financing of local public programs (including health programs).

Changes in the return to coffee-related work are of course accompanied by competing income and substitution effects. Although we cannot separately identify each, we are able to estimate their combined influence. Because they work in opposite directions, this provides a direct test of the relative importance of time vs. current income in household production of child health. Our results imply the primacy of time. Using intermediate outcomes, we also find additional evidence on the importance of time in health production (and in explaining procyclical mortality). First, coffee price reductions are associated with substantial declines both in the probability that adults work and in hours of work if employed (Dube and Vargas 2008). These declines are twice as large for women (the primary caregivers of children) as for men. Second, time-intensive child health investments are countercyclical. Third, childhood morbidity is procyclical – and when children are ill, parents are more likely to be at home rather than working outside the home if coffee prices are low. Fourth, births are countercyclical with a one year lag, suggesting the value of time to be an important determinant of other vital events as well (as prominently reported by Schultz (1985)).

We conclude by considering alternative explanations for procyclical mortality and conducting a variety of informal validity tests. We also cast our results in the context of broader debate about the wealth-health relationship (McKeown 1976, Pritchett and Summers 1996) and the importance of parental time in the production of child quality (Mayer 1997, Blau 1999, Goldin 1999, Schady 2004, Kruger 2007, Price 2008). Overall, our findings are consistent with growing evidence that the relative price of health is a more powerful determinant of mortality than wealth (Preston 1975, Jamison et al. 2001, Cutler and Miller 2005, Cutler, Deaton, and Lleras-Muney 2006, Deaton 2006).

2. Coffee Cultivation in Colombia

2.1 The Ecology of Coffee

Arabica caturra coffee (the predominant variety grown in Colombia since the early 1970s) is a tropical plant requiring very specific environmental conditions for cultivation: temperatures between 15 – 24° C, annual rainfall between 1500 and 2000 mm (depending on seasonal rainfall patterns and moisture retention of the soil), slopes of certain degrees at high altitudes (over 1700m but below frost lines), and, depending on the circumstances, generous shade (Clifford and Wilson 1985). Tropical, high-altitude regions of Colombia are particularly well-suited for coffee cultivation, especially the states of Antioquia, Caldas, Quindio, and Risaralda. Most coffee-producing nations have a single annual harvest, but Colombia’s unusual ecology and rainfall allows for two harvests each year in some areas – a primary harvest between October and December and a secondary harvest in April-May (or the reverse in some areas). Figure 2 shows the geography of coffee cultivation in Colombia during the 1970s and 1980s at the municipio (hereafter ‘county’) level.

Figure 2.

2.2 Labor and Coffee Cultivation

Coffee cultivation requires considerable non-harvest maintenance (including weeding, pruning, fertilizing, pest control, and renovation), and the use of labor for these purposes has a significant impact on the size of current-year coffee harvests (Clifford and Willson 1985, Ortiz 1999, Bacca 2002, CEDE 2002).⁸ With the introduction of green revolution coffee varieties (caturra in particular) and the arrival of new fungal parasites (“rusts”) in Colombia during the 1970s and 1980s, non-harvest labor as a share of total labor has grown considerably. During the harvest, coffee cherries must be picked immediately upon ripening to maximize their quality. Harvest windows for a given farm last approximately two weeks; during this period, picking cherries at their optimal stage of development can require visiting a single tree as many as eight times. After the harvest, coffee cherries must be processed (generally by the “wet” method in Colombia). This involves the use of pulping machines to soak and remove pulp from ripe cherries the day they are picked, fermenting the beans in tanks for 12–24 hours to loosen the remaining pulp and mucilage, washing to remove fermented residues, and drying either in the sun or using mechanical drying silos (CEDE 2002). This ‘parchment’ coffee is then sold to distributors who mill and bag Colombia’s ‘green’ coffee beans for export and sale to roasters.⁹

Labor on Colombian coffee farms generally falls into one of three categories: small farm owners who supply their own non-harvest labor, day laborers who live nearby and work year-round on the same farm, and seasonal migrant harvest workers. Since Colombia’s agrarian reforms during the 1960s, most coffee is grown on small farms of seven hectares or less (in 1997, average farm size was 1.7 hectares). This reduction in average farm size has made it possible for farm owners and their families to perform much of the non-harvest maintenance themselves. Larger farms also employ additional day-labor for in non-harvest seasons; day-laborers are generally men, while many wives are hired to prepare meals for field workers and provide general support (Ortiz 1999). Finally, picking coffee cherries at their optimal ripeness requires additional short-term seasonal labor supplied by migrant harvest workers. These harvest workers are almost exclusively young unmarried men and comprise a modest (but important) part of the coffee workforce, moving from region to region for several months each year following the harvest (which varies according to altitude, rainfall, soil composition, etc.).

2.3 The National Federation of Coffee Growers of Colombia

Until industrialization occurred during the 1970s, coffee was Colombia’s leading export. As a consequence, an unusual institutional arrangement to coordinate and manage coffee production developed in the form of the National Federation of Coffee Growers (NFCG, or Federación Nacional de Cafeteros de Colombia). The Federation was established as a cooperative organization of coffee growers in 1927, but coffee’s importance in the Colombian economy led the government to share in oversight and governance. It seeks both to advance the industrial interests of the Colombian coffee sector and to promote the welfare of coffee growers more broadly in rural regions historically neglected by government programs. It also ensures that Colombia adheres to its coffee export quotas under standing International Coffee Agreements (international treaties governing cartel-like behavior among coffee producing nations) and operates an internal price-support system for domestic growers. This system sets internal prices paid to growers as a function or world prices and partially shields them from world price volatility, paying more than growers would otherwise receive during bad years and less during good years (net of export costs and other mark-ups).¹⁰ These prices are uniform across the country – growers in all regions of Colombia are paid the same price for their coffee at any given point in time. Figures 3a and 3b show how internal prices paid to Colombian coffee growers vary with world coffee prices.

Figure 3.

3. Data and Empirical Strategy

3.1 Data

We obtained average annual coffee prices for years 1970 to 2006 from the National Federation of Coffee Growers and the International Coffee Organization. Although household choices do not influence world prices (and hence the determination of internal prices), we focus our analyses on price shocks known to originate outside of Colombia. As shown in Figures 3a and 3b, these are frosts that destroyed much of Brazil’s coffee harvest in 1975, a drought in Brazil in 1985, and the collapse of the International Coffee Agreement in 1989–1990 that led to the abandonment of export quotas by 1991. All three supply shocks led to dramatic changes in Colombia’s internal coffee prices by as much as 40% of the long-run mean, and they allow us to study both sudden price increases and decreases.

Because prices paid to Colombian coffee growers do not vary by region, our identification strategy relies on the interaction between birth year coffee price and the intensity of birth county coffee production. We construct county-level intensity measures using the NFCG’s decennial coffee censuses. For planning and monitoring purposes, the NFCG conducts decennial enumerations of all coffee farms in Colombia. We use this data from the early 1970s, early 1980s, and 1997 to measure hectares of coffee (or “intensity” of coffee cultivation) in each Colombian county as shown in Figure 2.¹¹ (See the Appendix 1 for a complete description of this data.) The timing of the NFCG’s coffee censuses is convenient because they were generally conducted a few years before each major price shock that we analyze. New coffee cultivation cannot yield a harvest in less than three to four years (the biologically-determined amount of time required for new coffee plants to produce their first fruit), so we can reasonably assume that our coffee intensity measures apply to shock years (Ortiz 1999, NFCG personal communication). Table 1 shows descriptive statistics for all of Colombia’s counties and for counties with and without coffee cultivation.

TABLE 1.

Descriptive Statistics

Panel A: Counties in the 1993 Population Census
	All Counties (N=1,060)		Counties with Coffee (N=589)		Counties without Coffee (N=471)
	Mean	Std Dev	Mean	Std Dev	Mean	Std Dev
Number of Households	6,532	(34,598)	6,675	(23,535)	6,354	(44,767)
Number of Individuals	30,942	(174,651)	30,200	(106,237)	31,869	(233,683)
Share Female	0.49	(0.02)	0.49	(0.02)	0.49	(0.02)
Age	25.31	(2.48)	25.93	(2.30)	24.53	(2.48)
Share Under Age 5	0.13	(0.02)	0.12	(0.02)	0.13	(0.02)
Share Married or in Free Union (over Age 14)	0.53	(0.12)	0.54	(0.07)	0.52	(0.16)
Share Born in Current Municipality	0.66	(0.17)	0.65	(0.16)	0.67	(0.19)
Share Literate (over Age 4)	0.76	(0.18)	0.79	(0.11)	0.71	(0.23)
Share in School (over Age 4)	0.26	(0.05)	0.25	(0.04)	0.27	(0.05)
Years of Education (over Age 4)	3.62	(0.90)	3.65	(0.84)	3.59	(0.97)
Share Employed (over Age 9)	0.41	(0.10)	0.43	(0.06)	0.38	(0.12)
Children Ever Born (Females over Age 14)	3.66	(0.52)	3.65	(0.50)	3.67	(0.55)
Children Alive (Females over Age 14)	4.11	(0.49)	4.07	(0.48)	4.15	(0.49)
Age at Last Birth (Females over Age 14)	29.85	(1.40)	29.91	(1.29)	29.76	(1.53)
Share with Brick or Prefabricated Walls	0.47	(0.24)	0.47	(0.22)	0.48	(0.25)
Share with Adobe or Pressed Dirt Walls	0.20	(0.24)	0.21	(0.24)	0.18	(0.24)
Share with Dirt Floors	0.28	(0.22)	0.23	(0.18)	0.34	(0.24)
Share with Water Access	0.55	(0.26)	0.61	(0.23)	0.48	(0.28)
Share with Sewage Access	0.32	(0.25)	0.40	(0.24)	0.22	(0.23)
Share with Electricity	0.67	(0.27)	0.73	(0.22)	0.60	(0.30)
Number of Household Rooms	3.14	(0.43)	3.22	(0.45)	3.05	(0.39)
Share Owning Home	0.70	(0.17)	0.70	(0.12)	0.69	(0.22)
Share Renting or Leasing Home	0.17	(0.11)	0.18	(0.10)	0.16	(0.11)
Hectares of Coffee (Early 1980s)	979	(1,871)	1,726	(2,209)	0	(0.00)

Panel B: Children in the Pooled Demographic and Health Survey Sample (1986, 1990, 1995, and 2000 Waves)
	All Counties (N=70,695)		Counties with Coffee (N=22,313)		Counties without Coffee (N=48,382)
	Mean	Std Dev	Mean	Std Dev	Mean	Std Dev
Mother’s Age	36.01	(8.16)	36.24	(8.17)	35.91	(8.16)
Mother’s Years of Education	5.28	(3.83)	4.77	(3.55)	5.52	(3.93)
Number of Household Members	6.16	(2.61)	6.38	(2.61)	6.06	(2.60)
Mother’s Total Number of Births	4.40	(2.73)	4.76	(3.00)	4.24	(2.58)
Mother’s Age at First Birth	19.84	(3.92)	19.83	(3.90)	19.84	(3.93)
Mother’s Age at First Marriage	18.97	(4.14)	19.03	(4.20)	18.94	(4.11)
Preceding Birth Interval (Months)	35.89	(27.13)	34.44	(26.45)	36.57	(27.42)
Share Receiving Prenatal Tetanus Toxoid	0.68	(0.46)	0.65	(0.48)	0.70	(0.46)
Share of Mothers Receiving Prenatal Care	0.84	(0.37)	0.81	(0.39)	0.85	(0.36)
Share of Medically-Supervised Births	0.82	(0.39)	0.80	(0.40)	0.83	(0.38)
Months Breastfed	9.66	(8.69)	9.11	(8.69)	9.91	(8.68)
Share Receiving BCG Vaccine	0.92	(0.27)	0.93	(0.25)	0.91	(0.28)
Share Receiving DPT Vaccine	0.92	(0.28)	0.93	(0.26)	0.91	(0.28)
Share Receiving Polio Vaccine	0.93	(0.25)	0.94	(0.24)	0.93	(0.26)
Share Receiving Measles Vaccine	0.64	(0.48)	0.68	(0.46)	0.62	(0.49)
Hectares of Coffee in Municipality (Early 1980s)	1,899	(3,136)	2,974	(3,493)	0	(0.00)

Panel C: Children in the Familias en Acción Survey Sample
	All Counties (N=13,732)		Counties with Coffee (N=5,176)		Counties without Coffee (N=8,556)
	Mean	Std Dev	Mean	Std Dev	Mean	Std Dev
Age	3.50	(1.92)	3.49	(1.91)	3.50	(1.93)
Share Compliant with Growth and Development Monitoring	0.32	(0.47)	0.37	(0.48)	0.29	(0.45)
Mother’s Age	31.15	(7.01)	31.05	(6.82)	31.21	(7.13)
Mothers: Share with No Education	0.15	(0.35)	0.12	(0.32)	0.16	(0.37)
Mothers: Share with Primary Education	0.21	(0.41)	0.20	(0.40)	0.21	(0.41)
Mothers: Share with Secondary Education or More	0.06	(0.24)	0.06	(0.23)	0.06	(0.24)
Mothers: Share Worked in Preceding Week	0.35	(0.48)	0.34	(0.47)	0.35	(0.48)
Mothers: Monthly Hours of Work	36.69	(77.19)	38.30	(78.55)	35.66	(76.30)
Household Head’s Age	40.57	(12.33)	39.97	(12.00)	40.95	(12.53)
Household Head: Share with No Education	0.23	(0.42)	0.19	(0.39)	0.26	(0.44)
Household Head: Share with Primary Education	0.11	(0.31)	0.09	(0.29)	0.11	(0.32)
Household Head: Share with Secondary Education or More	0.04	(0.20)	0.04	(0.19)	0.04	(0.21)
Household Head: Share Worked in Preceding Week	0.91	(0.29)	0.91	(0.29)	0.91	(0.28)
Household Head: Monthly Hours of Work	173.93	(91.92)	175.20	(88.99)	173.17	(93.64)
Number of People in Household	6.60	(2.53)	6.53	(2.51)	6.64	(2.54)
Number of Children in Household (Ages 0–6)	2.09	(1.13)	2.08	(1.12)	2.09	(1.14)
Share of Children Suffering Acute Diarrhea in last 15 days	0.14	(0.35)	0.13	(0.34)	0.15	(0.36)
Share of Children Suffering Acute Respiratory Infection in last 15 days	0.43	(0.50)	0.39	(0.49)	0.46	(0.50)
Share of Children Suffering Any Illness in last 15 days	0.55	(0.50)	0.52	(0.50)	0.57	(0.49)
Share of Children Not Participating in Usual Activities in last 15 days	0.22	(0.41)	0.20	(0.40)	0.23	(0.42)
Travel Time in Minutes to Town Center	56.29	(101.12)	55.04	(64.39)	57.08	(118.54)
Travel Time in Minutes to Nearest Health Care Facility	41.74	(61.31)	44.30	(53.75)	40.13	(65.58)

Notes: The data summarized in Panel A is from the complete 1993 Colombian population census matched to the National Federation of Coffee Grower’s early 1980s coffee census; the data summarized in Panel B is from pooled child records from the 1986, 1990, 1995, and 2000 Colombian Demographic and Health Surveys also matched to the National Federation of Coffee Grower’s early 1980s coffee census; the data summarized in Panel C is from the Familias en Acción Survey (2002 first wave characteristics only shown) matched to the National Federation of Coffee Grower’s 1997 coffee census. Standard deviations shown in parentheses.

To estimate how cohort size changes both with birth year coffee prices and in proportion to the intensity of birth county coffee cultivation, we construct birth county × birth year population counts using the 100% 1993 Colombian population census. Building population counts by county of birth rather than by county of residence minimizes the influence of migration (there was very little foreign emigration from Colombia during the years we study (Gaviria 2004, Garay and Rodríguez 2005, CELADE 2006)). To address fertility responses, we also restrict our analysis of each price shock to cohorts born in shock years or earlier (i.e., only those conceived prior to shocks). We therefore interpret our population counts as reflecting survival (or cumulative mortality). This measure is free from many shortcomings of developing country vital registries.¹²

We also use Colombia’s Demographic and Health Surveys (DHS) (and county-level child vaccination records described in Appendix 1) to study more directly how time-intensive child health investments respond to coffee price fluctuations. These surveys contain detailed pregnancy and child health histories for nationally-representative samples of reproductive age women (defined as 15–49) in 1986, 1990, 1995, and 2000. We pool all four waves together to create a sample of child-level records that includes birth dates, maternal characteristics, preceding birth intervals, and detailed child health investment histories. Specific investments include maternal use of prenatal care, prenatal tetanus vaccinations, breastfeeding duration, and EPI child vaccinations (BCG, DPT, polio, and measles).¹³ Women report this information for each of their children – regardless of their survival to the time of the survey – with the exception of child health histories, which are reported only for children born within five years of the survey. Table 1 Panel B shows descriptive statistics for Colombian children and their mothers in the pooled DHS sample.

Additionally, we use data from Colombia’s Familias en Acción evaluation survey to investigate how local labor market conditions and parental time use change with coffee price fluctuations. This survey was first administered to 11,502 households in 122 Colombian counties in 2002; follow-up waves were conducted in 2003 and 2005 (Attanasio and Vera-Hernández 2004). Topical modules broadly cover household composition and demographic characteristics, labor force participation and other labor market conditions, and child health. Although this data was collected after the major coffee price shocks that we study, it contains detailed information not available from any other sources on employment, hours worked, time spent at home, travel time to health care facilities, and prevalence of childhood illness. Table 1 Panel C shows descriptive statistics for children and their parents in the Familias en Acción sample.

Finally, Appendix 1 describes these data sources as well as others used for supplementary analyses in Section 5 in greater detail.

3.2 Empirical Strategy

In a cohort study framework, we first estimate how cohort size varies in a way that is proportionate both to (i.) internal birth year coffee price and (ii.) the economic importance of coffee in one’s birth county (defined as farmland dedicated to coffee). Our specific measure of the impact of coffee price utilizes both sources of variation simultaneously. Constructed at the birth county × birth year level for counties (municipios) m and birth cohorts c, we implement this measure as:

$${(\mathit{Birth}\mathit{county}\mathit{coffee}-\mathit{growing}\mathit{intensity})}_m\times {(\mathit{Internal}\mathit{birth}\mathit{year}\mathit{coffee}\mathit{price})}_c.$$

For simplicity, we write this term as: (g_m×p_c).¹⁴ As noted, this approach exploits the fact that health is considerably more fragile in utero and during the first critical year of life than during the second or subsequent years (Dobbing 1976, Barker 1992).¹⁵ Because adjacent birth cohorts experience nearly identical conditions at every age except for the first year, we associate differences in survival between them with our coffee price measure during this critical developmental period.

We analyze each price shock shown in Figure 3 separately: the 1975 Brazilian frost, the 1985 Brazilian drought, and the 1991 price collapse following the abandonment of International Coffee Agreement export quotas. We begin by restricting our analyses to samples of those in their first two years of life (ages 0–1 and age 1–2) at the time that a coffee price shock occurred. Specifically, we estimate:

$$ln(s_{cm})=\alpha +\lambda (g_m\times p_c)+{\delta }_m+{\delta }_c+{\epsilon }_{cm},$$ (1)

where s is the size of birth cohort c born in county m, δ_m and δ_c represent birth county and birth cohort fixed effects, and the parameter of interest is λ. Negative estimates of λ would imply countercyclical cohort size (which is consistent with procyclical mortality). We also estimate equation 1 using samples of those in their first three years of life when a price shock occurred, both with and without county-specific linear trends.¹⁶ To test for differential reallocation of household resources by children’s gender in response to price shocks, we estimate equation 1 separately for males and females as well.

After establishing the basic cyclical pattern of cohort size, we next consider what behavioral responses might explain these patterns. To study changes in the value of time, we use the Familias en Acción data (collected after the major coffee price shocks that this paper examines) to investigate the relationship between coffee price fluctuations and labor market outcomes underlying time-intensive child health investment decisions. For adult individuals i observed in years y and counties m, we estimate:

$$o_{\mathit{iym}}=\alpha +\lambda (g_m\times p_y)+{\delta }_m+{\delta }_y+{\delta }_m\times y+{\epsilon }_{\mathit{iym}},$$ (2)

where o is a labor market outcome, δ_m, δ_y, and δ_m×y represent county and year fixed effects and county-specific linear time trends (respectively), and the other variables are defined as before. Specific labor market outcomes that we examine separately for mothers and household heads include whether or not one worked the week prior to the survey (estimated using a probit model) and hours of work in the past month (both unconditionally and conditional on working in the past week).

We also use variants of equation 2 to investigate how the prevalence of childhood morbidity changes with price fluctuations. Then, to study how adult time use responds to price shocks when children are sick, we use a probit model to estimate:

$$Pr(t_{\mathit{iym}}=1)=\mathrm{\Phi }\left[\begin{array}{l}\alpha +\theta s_{\mathit{iym}}+\lambda (g_m\times p_y)+\mu (g_m\times s_{\mathit{iym}})+\sigma (p_y\times s_{\mathit{iym}})+\rho (g_m\times p_y\times s_{\mathit{iym}})+{\delta }_m+{\delta }_y+{\delta }_m\times y+{\epsilon }_{\mathit{iym}}\end{array}\right],$$ (3)

where t is a dichotomous indicator for adults’ primary activity the day before the survey (either being at home or working outside the home), Φ[·] is the standard normal cumulative density function, s is a dummy variable for whether or not a child in adult i’s household was sick at the time of the survey, ρ is the parameter of interest, and all other variables are again defined as in previous estimating equations.

Finally, we use the pooled child-level DHS sample to examine specific health behaviors that are important in the first year of life. These analyses are again restricted to children ages 0–2 the year that a price shock occurred. Most health behavior measures are dichotomous, so we use probit models for children i, birth cohorts c, and counties m to estimate:

$$Pr(b_{\mathit{icm}}=1)=\mathrm{\Phi }\left[\alpha +\lambda (g_m\times p_c)+\sum _k{\phi }_k{w}_{ik}+{\delta }_m+{\delta }_c+{\epsilon }_{\mathit{icm}}\right],$$ (4)

where b is a dichotomous health behavior of interest (receipt of a prenatal tetanus toxoid vaccine, prenatal care, medical birth assistance, BCG vaccine, DPT vaccine, polio vaccine, or measles vaccine), w is a vector of maternal characteristics (mother’s age, education, number of household members, number of preceding births, age at first birth, and age at first marriage), and all other variables are defined as before. (In complementary analyses, we also use Colombia’s official county-year vaccination records to estimate variants of equation 1.)

4. Results: The Cyclicality of Child Survival, Time Use, and Child Health Investments

4.1 Cohort Size

Before formally estimating the association between cohort size and coffee prices, we first explore this relationship graphically. For years 1970 to 1993, Figure 4 shows internal coffee prices and the difference in mean residual birth cohort size (net of county fixed effects) between counties with above- and below-median coffee intensity.¹⁷ Although this graph includes all birth cohorts – and therefore does not distinguish between changes in fertility and changes in mortality – it clearly exhibits a pattern of countercyclical cohort size. (Despite the limitations of Colombia’s mortality statistics, they exhibit a similar pattern – Appendix Figure A1 re-constructs this graph using infant deaths rather than cohort size, showing procyclical infant mortality.)¹⁸

Figure A1.

Table 2 then shows estimates of λ (the coefficient for the interaction between birth year coffee price and birth county coffee intensity) obtained by estimating equation 1. Each panel reports results for different price shocks (1975, 1985, and 1991), and each column corresponds to a different sample or specification (those ages 0–2 at the time of a price shock, those ages 0–3, and those ages 0–3 controlling for county-specific linear trends). Because the dependent variable is in logarithmic form, coefficient estimates can roughly be interpreted as percent changes in cohort size. To aid in interpretation, implied changes in cohort size are also shown for median coffee-growing intensity and a 500 peso price change.

TABLE 2.

Coffee Price Shocks and ln (Cohort Size)

	Sample/Specification
	Ages 0–2	Ages 0–3	Ages 0–3 with Trends
Panel A: 1975 Brazilian Frost	−0.17*** (0.03)	−0.03*** (0.01)	−0.08*** (0.02)
County Fixed Effects	Yes	Yes	Yes
County-Specific Linear Trends	No	No	Yes
Implied Change	−2.16%	−0.40%	−0.99%
N	2215	3319	3319
R2	0.99	0.99	0.99
Panel B: 1985 Brazilian Drought	−0.16*** (0.04)	−0.14*** (0.04)	−0.23*** (0.09)
County Fixed Effects	Yes	Yes	Yes
County-Specific Linear Trends	No	No	Yes
Implied Change	−2.04%	−1.69%	−2.89%
N	2208	3310	3310
R2	0.99	0.99	0.99
Panel C: 1990 ICA Collapse	−0.10*** (0.03)	−0.05*** (0.01)	0.08 (0.06)
County Fixed Effects	Yes	Yes	Yes
County-Specific Linear Trends	No	No	Yes
Implied Change	−1.22%	−0.58%	---
N	2203	3305	3305
R2	0.99	0.99	0.99

Notes: County-year cohort size data from the complete 1993 Colombian population census; coffee cultivation data from the National Federation of Coffee Grower’s early 1970s and early 1980s coffee censuses; annual internal coffee price data from the National Federation of Coffee Growers. Estimates and standard errors (in parentheses, clustered by county) shown for the interaction between coffee growing intensity and coffee price in the first year of life obtained by estimating equation 1 (controlling for county and year fixed effects and county-specific linear trends as shown above). Coffee area is in thousands of hectares and coffee prices are in thousands of pesos per kilogram. Implied changes are calculated for 250 hectares of coffee and a 500 peso per kilogram price change.

^*p<0.1,

^**p<0.05,

^***p<0.01.

Overall, Table 2 presents evidence of countercyclical cohort size for all three price shocks.¹⁹ For a county with median coffee cultivation, implied countercyclical changes in cohort size range from about −0.4% to −2.0% (with implied elasticities of −0.01 to −0.05). Interpreting cohort size as cumulative survival, this is consistent with procyclical mortality. Because the results in Table 2 use population counts in 1993, it also suggests that deaths linked to coffee price fluctuations generally occur at young ages (the 1985 shock estimates are not smaller than the 1975 shock estimates, but the 1991 estimates are smaller than the 1985 estimates). If we assume that excess mortality related to price shocks occurs by age five, our estimates imply that the largest price fluctuations since 1970 are associated with changes in child survival of approximately 15%.²⁰

Table 3 then shows estimates of λ obtained by estimating equation 1 separately for males and females. In contrast with other studies of consumption smoothing, intrahousehold resource allocation, and child mortality (Rose 1999), we find no evidence of statistically meaningful differences in cohort size (or survival) between boys and girls. This equivalence is consistent with other suggestions of little gender bias in intrahousehold resource allocation in Colombia (PROFAMILIA 2005, Hincapié 2006).

TABLE 3.

Coffee Price Shocks and ln(Cohort Size) by Gender

Male-Only Sample				Female-Only Sample
Sample/Specification				Sample/Specification
	Ages 0–2	Ages 0–3	Ages 0–3 with Trends		Ages 0–2	Ages 0–3	Ages 0–3 with Trends
Panel A: 1975 Brazilian Frost	−0.17*** (0.04)	−0.04** (0.02)	−0.08*** (0.02)	Panel D: 1975 Brazilian Frost	−0.17*** (0.03)	−0.03** (0.01)	−0.08*** (0.02)
County Fixed Effects	Yes	Yes	Yes	County Fixed Effects	Yes	Yes	Yes
County-Specific Linear Trends	No	No	Yes	County-Specific Linear Trends	No	No	Yes
Implied Change	−2.09%	−0.48%	−1.01%	Implied Change	−2.15%	−0.38%	−0.97%
N	2207	3305	3305	N	2205	3307	3307
R2	0.99	0.99	0.99	R2	0.99	0.99	0.99
Panel B: 1985 Brazilian Drought	−0.16*** (0.05)	−0.15*** (0.05)	−0.20*** (0.11)	Panel E: 1985 Brazilian Drought	−0.17*** (0.06)	−0.13*** (0.05)	−0.25** (0.11)
County Fixed Effects	Yes	Yes	Yes	County Fixed Effects	Yes	Yes	Yes
County-Specific Linear Trends	No	No	Yes	County-Specific Linear Trends	No	No	Yes
Implied Change	−2.04%	−1.81%	−2.49%	Implied Change	−2.13%	−1.63%	−3.13%
N	2199	3299	3299	N	2205	3305	3305
R2	0.99	0.99	1.00	R2	0.99	0.99	0.99
Panel C: 1990 ICA Collapse	−0.09** (0.04)	−0.05** (0.01)	0.03 (0.08)	Panel F: 1990 ICA Collapse	−0.10*** (0.04)	−0.04*** (0.01)	0.11 (0.08)
County Fixed Effects	Yes	Yes	Yes	County Fixed Effects	Yes	Yes	Yes
County-Specific Linear Trends	No	No	Yes	County-Specific Linear Trends	No	No	Yes
Implied Change	−1.13%	−0.67%	---	Implied Change	−1.23%	−0.46%	---
N	2199	3300	3300	N	2198	3298	3298
R2	0.99	0.99	0.99	R2	0.99	0.99	0.99

Notes: County-year cohort size data from the complete 1993 Colombian population census; coffee cultivation data from the National Federation of Coffee Grower’s early 1970s and early 1980s coffee censuses; annual internal coffee price data from the National Federation of Coffee Growers. Estimates and standard errors (in parentheses, clustered by county) shown for the interaction between coffee growing intensity and coffee price in the first year of life obtained by estimating equation 1 (controlling for county and year fixed effects and county-specific linear trends as shown above). Coffee area is in thousands of hectares and coffee prices are in thousands of pesos per kilogram. Implied changes are calculated for 250 hectares of coffee and a 500 peso per kilogram price change.

^*p<0.1,

^**p<0.05,

^***p<0.01.

4.2 Local Labor Markets, Infant/Child Morbidity, and Adult Time Use

After establishing how cohort size co-varies with coffee prices and the intensity of coffee cultivation, we then seek to understand the underlying behavior that explains these results. To focus more directly on the value of time, we first investigate how local labor market outcomes and adult time use respond to coffee price fluctuations. Microdata on labor market outcomes during the 1970s and 1980s are unavailable, so Table 4 reports estimates of λ obtained by estimating equation 2 with the Familias en Acción survey data.

TABLE 4.

Coffee Price Shocks and Local Labor Markets

	Estimate	Standard Error	Implied Change	N	R2/Pseudo R2
Mother - Worked last Week?	0.02*	(0.01)	0.05	11,110	0.04
Household head - Worked Last Week?	0.02***	(0.01)	0.05	18,093	0.05
Mother - Hours of Monthly Work	4.81*	(2.77)	9.81	8,757	0.10
Household Head - Hours of Monthly Work	5.85***	(1.96)	11.93	16,737	0.05
Mother - Hours of Monthly Work (Conditional on Worked Last Week)	6.84	(4.88)	---	3,369	0.08
Household Head - Hours of Monthly Work (Conditional on Worked Last Week)	2.06	(1.72)	---	15,236	0.04

Notes: Individual-level labor market participation data from the Familias en Acción panel survey (2002, 2003, and 2005 waves); coffee cultivation data from the National Federation of Coffee Grower’s 1997 coffee censuses; annual internal coffee price data from the National Federation of Coffee Growers. Estimates and standard errors (in parentheses, clustered by county) shown for the interaction between coffee growing intensity and coffee price obtained by estimating equation 2 (controlling for county and year fixed effects and county-specific linear trends as shown above). For dichotemous dependent variables (the first two rows), a probit model was used to estimate equation 2, and marginal probabilities are reported. Coffee area is in hundreds of hectares and coffee prices are in hundreds of pesos per kilogram. Implied changes are calculated for 250 hectares of coffee and a 500 peso per kilogram price change.

^*p<0.1,

^**p<0.05,

^***p<0.01.

Focusing on labor force participation, the first and second rows of Table 4 show marginal probabilities (calculated using probit estimates and independent variable means) implying that a 500 peso price decline in a county with median coffee cultivation is associated with a 5 percentage point decrease in the probability that either a mother or a household head worked during the preceding week. Because mothers generally work less than household heads, this reduction is a 14% decline for mothers vs. a 5% decline for household heads. The third and fourth rows then imply that the same price shock is associated with working 10 and 12 fewer hours per month, respectively.²¹ These reductions are again larger in relative terms for mothers (27%) than for household heads (7%). Conditional on working the week prior to the survey, the fifth and sixth rows show that reductions in hours worked are statistically indistinguishable from the unconditional estimates but are also not statistically different from zero at conventional levels.

Next, we use the Familias en Acción data to investigate how coffee price fluctuations are related to the incidence of childhood illness – and to adults’ choices to stay at home or work outside the home when children are sick. Using dichotomous indicators for various childhood illnesses within the past fifteen days (acute diarrhea, acute respiratory infection, any illness, or missing usual activities due to illness) as dependent variables, Table 5 Panel A reports marginal probabilities calculated from probit estimates for the interaction between coffee cultivation and coffee price. Childhood morbidity is procyclical for all illness measures; implied changes range from 10% to 30% for a 500 peso price change at median coffee cultivation.

TABLE 5.

Coffee Price Fluctuations, Infant/Child Morbidity, and Adult Time Use

	Estimate	Standard Error	Implied Change	N	R2/Pseudo R2
Panel A: Child Morbidity within the Preceding 15 Days
Acute Diarrhea	0.009**	(0.004)	0.12	29,786	0.03
Acute Respiratory Infection	0.022***	(0.006)	0.28	29,771	0.04
Any Childhood Illness	0.022***	(0.006)	0.28	29,775	0.03
Child Absent from Usual Activities Due to Illness	0.025***	(0.005)	0.32	29,707	0.02
Panel B: Adult Time Use the Preceding Day Conditional on Child Illness
Working outside of the home for money	0.007	(0.014)	---	20,554	0.04
Working in self-employed occupation	0.011*	(0.006)	0.13	20,557	0.05
At home doing household activities	−0.027*	(0.016)	−0.34	18,779	0.06

Notes: Individual-level child morbidity symptom data and adult time use data from the Familias en Acción panel survey (2002, 2003, and 2005 waves); coffee cultivation data from the National Federation of Coffee Grower’s 1997 coffee censuses; annual internal coffee price data from the National Federation of Coffee Growers. Panel A shows marginal probabilities (and corresponding standard errors in parentheses, clustered by county) for the interaction between coffee growing intensity and coffee price and were calculated using estimates from equation 2 fit using probit models (controlling for county and year fixed effects and county-specific linear trends). Panel B shows marginal probabilities (and corresponding standard errors in parentheses, clustered by county) for the three-way interaction among coffee growing intensity, coffee price, and a child in the household having missed usual activities due to illness in the preceding 15 days and were calculated using estimates from equation 3 fit using probit models (controlling for the childhood illness measure, all lower-order interactions, day of the week, county and year fixed effects, and county-specific linear trends). Coffee area is in hundreds of hectares and coffee prices are in hundreds of pesos per kilogram. Implied changes are calculated for 250 hectares of coffee and a 500 peso per kilogram price change.

^*p<0.1,

^**p<0.05,

^***p<0.01.

To investigate how adult time use varies with coffee prices when children are ill, we then estimate equation 3 using several dichotomous measures of adult time use the day before the survey. These include working outside the home for money, self-employed working at home, and not working but at home doing household chores – we interpret the last as including child care. Our measure of childhood illness is having missed usual activities due to illness in the past fifteen days. Table 5 Panel B shows marginal probabilities for the three-way interaction of coffee cultivation, coffee price, and childhood illness. When a child is sick and coffee prices are high, adults in areas with greater coffee cultivation are more likely to be working (in self-employed occupations in particular) and less likely to be doing household chores (that presumably include child care).

Taken together, these patterns of labor market activity, childhood illness, and adult time use when children are ill suggest that parental time may play an important role in explaining procyclical infant and child mortality.

4.3 The Value of Time and Infant/Child Health Investments

We next study the cyclicality of time-intensive child health investments. Many health investments that we observe are linked to the use of primary and preventive health services. Importantly, these services are generally inexpensive or free in Colombia but require considerable travel and long waiting times. Other important time-intensive child health investments that we do not observe (trips to distant clean wells and good household hygiene, for example) can reasonably be assumed to vary in the same cyclical manner.

Using the pooled DHS sample to estimate equation 4, Table 6 shows marginal probabilities corresponding to estimates of λ for preventive health service use in the first year of life.²² All estimates that are distinguishable from zero are negative, implying countercyclical health investments. As Section 5.1 explains, local coffee revenue and local public health budgets are not linked, so these relationships are presumably demand-driven.

TABLE 6.

Coffee Price Shocks and Infant Health Investments

	Estimate	Standard Error	Implied Change	N	R2/Pseudo R2
1985 Brazilian Drought
Prenatal Tetanus Toxoid	−1.08***	(0.37)	−0.135	1082	0.19
Prenatal Care	−0.36	(0.31)	---	1055	0.17
Birth Assistance	−0.17	(0.32)	---	1034	0.25
Months Breastfed	−0.53	(4.93)	---	1218	0.20
BCG Vaccine	0.29	(0.37)	---	490	0.15
DPT Vaccine	0.00	(0.00)	---	531	0.13
Polio Vaccine	−0.01***	(0.01)	−0.002	540	0.12
Measles Vaccine	0.33	(0.48)	---	574	0.12
1990 ICA Collapse
Prenatal Tetanus Toxoid	0.25	(0.51)	---	1790	0.17
Prenatal Care	−0.67*	(0.42)	−0.084	1815	0.13
Birth Assistance	0.87	(0.42)	---	1668	0.18
Months Breastfed	12.60	(10.00)	---	1939	0.17
BCG Vaccine	−0.04	(0.28)	---	1397	0.15
DPT Vaccine	−0.04	(0.25)	---	1631	0.25
Polio Vaccine	0.02	(0.16)	---	1743	0.25
Measles Vaccine	−0.06	(0.17)	---	1096	0.25

Notes: Individual-level infant health input data from the pooled 1986, 1990, 1995, and 2000 Colombian Demographic and Health Survey sample; coffee cultivation data from the National Federation of Coffee Grower’s early 1980s coffee census; annual internal coffee price data from the National Federation of Coffee Growers. Marginal probabilities (and corresponding standard errors in parentheses, clustered by county) are shown for the interaction between coffee growing intensity and coffee price and were calculated using estimates obtained from equation 4 (controlling for county and year fixed effects and county-specific linear trends as well as mother’s age, education, number of household members, number of preceding births, age at first birth, and age at first marriage) with the exception of “Months Breastfed,” for which an OLS estimate is reported. Breastfeeding results are generally invariant to methods of addressing censoring in the distribution of months breastfed. Coffee area is in thousands of hectares and coffee prices are in thousands of pesos per kilogram. Implied changes are calculated for 250 hectares of coffee and a 500 peso per kilogram price change.

^*p<0.1,

^**p<0.05,

^***p<0.01.

We then use Colombia’s official county-year EPI (WHO Expanded Program on Immunization) vaccination records from years 1998–2007 to estimate equation 1 (Appendix 1 describes this data). These vaccines include polio, DPT (diphtheria, pertussis, and tetanus), BCG (against tuberculosis), hepatitis B, haemophilus influenzae type b (against Hib disease manifestations like pneumonia and meningitis), and MMR (measles, mumps, and rubella). Table 7 shows estimates for the interaction parameter λ – all estimates are negative and statistically distinguishable from zero, suggesting countercyclical fluctuations in vaccination use.

TABLE 7.

Coffee Price Fluctuations and WHO EPI Bundle Vaccinations

Vaccination Type	Estimate	Standard Error	Implied Change	N	R2
ln(Polio Vaccionations)	−0.44***	(0.11)	−5.54%	7,600	0.94
ln(DPT Vaccinations) (Diptheria, Pertussis, and Tetanus)	−0.38***	(0.12)	−4.78%	7,603	0.92
ln(BCG Vaccinations) ( Bacille Calmette-Guerin)	−0.30***	(0.12)	−3.76%	7,573	0.94
ln(Hepatitis B Vaccinations)	−0.60***	(0.12)	−7.51%	7,604	0.93
ln(Haemophilus Influenzae Type b Vaccinations)	−0.58***	(0.15)	−7.25%	7,018	0.90
ln (MMR Vaccinations) (Measles, Mumps, and Rubella)	−0.53***	(0.12)	−6.68%	7,611	0.93

Notes: County-year EPI (WHO Expanded Program on Immunization) vaccination data for years 1998–2007 from Fundación Santa Fe de Bogotá ; coffee cultivation data from the National Federation of Coffee Grower’s 1997 coffee census; annual internal coffee price data from the National Federation of Coffee Growers. Estimates and standard errors (in parentheses, clustered by county) shown for the interaction between coffee growing intensity and coffee price obtained by estimating equation 1 (controlling for county and year fixed effects and county-specific linear trends). Coffee area is in thousands of hectares and coffee prices are in thousands of pesos per kilogram. Implied changes are calculated for 250 hectares of coffee and a 500 peso per kilogram price change.

^*p<0.1,

^**p<0.05,

^***p<0.01.

Overall, Tables 6 and 7 establish a general countercyclical pattern of health behaviors requiring substantial parental time. Prominent estimates linking child survival with time-intensive health investments that we do not observe suggest that plausible changes in water fetching (and water quality), household hygiene, and primary health care use alone could explain the mortality changes implied by our cohort size results.²³ Online Appendix 1 presents additional evidence on the value of time in determining the use of other health inputs and vital events as well.

5. Alternative Explanations for Procyclical Infant/Child Mortality

This section considers alternative explanations for our findings, including those proposed by wealthy country studies documenting procyclical mortality (Ruhm 2000, 2006). Specifically, we examine cyclical patterns of local public spending, NFCG ‘social investments,’ alcohol and tobacco use, traffic fatalities, and other violent deaths. We then summarize several informal tests of the validity of our empirical approach (presented fully in Appendix 3).

5.1 Alternative Mechanisms

5.1.a. Local Public Spending

Coffee cultivation and production in Colombia are not taxed (Cárdenas 1994, Cárdenas and Yanovich 1997, Cárdenas and Partow 1999, Cuellar 2004), and central government transfers to local governments following rigid budget cycles that do not adjust flexibly to local economic circumstances. In exchange for the lack of formal taxation, the NFCG has a longstanding agreement with the Colombian government to finance development projects in coffee growing regions on its own (we analyze this NFCG financing in the next subsection). The result is that coffee price fluctuations do not have an important immediate impact on local public revenue or local public programs.²⁴ To provide econometric support for this assertion, we use a county-year panel of local public finance data for years 1984 through 1993 collected by the Contraloría General de la Nación (the accounting branch of the Colombian government – see Appendix 1 for a description of this data) to regress county-year tax revenue and public spending on interactions between coffee prices and coffee cultivation (following equation 1). Appendix Table A1 shows estimates for the interaction between coffee price coffee cultivation. The absence of statistically meaningful estimates is consistent with our view that coffee price shocks do not act primarily through the financing of public programs.

APPENDIX TABLE A1.

Coffee Price Fluctuations and County-Level Public Finance

	Estimate	Standard Error	N	R2
Total Local Government Income	0.002	(0.024)	8,020	0.95
Transfer Income	−0.005	(0.006)	8,020	0.90
Capital Income	0.000	(0.001)	8,020	0.71
Total Direct Taxes	−0.001	(0.008)	8,020	0.94
Total Indirect Taxes	0.008	(0.012)	8,020	0.96
Industry and Commerce Taxes	0.005	(0.007)	8,020	0.96
Total Spending	−0.005	(0.039)	8,020	0.92
Investment Spending	−0.013	(0.019)	8,020	0.90
Operational Spending	0.003	(0.017)	8,020	0.94
Water Spending	−0.001	(0.002)	8,020	0.39
Infrastructure Spending	−0.008	(0.006)	8,020	0.68
Housing Spending	−0.001	(0.005)	8,020	0.53
Education Spending	−0.001	(0.003)	8,020	0.77
Health Spending	0.001	(0.002)	8,020	0.75

Notes: County-year public finance data for years 1984–1993 from Contraloria General de la Nación; coffee cultivation data from the National Federation of Coffee Grower’s early 1980s coffee census; annual internal coffee price data from the National Federation of Coffee Growers. Estimates and standard errors (in parentheses, clustered by county) shown for the interaction between coffee growing intensity and coffee price obtained by estimating equation 1 (controlling for county and year fixed effects and county-specific linear trends). Coffee area is in hectares and coffee prices are in pesos per kilogram.

^*p<0.1,

^**p<0.05,

^***p<0.01.

5.1.b. Social Contribution Spending by the National Federation of Coffee Growers

Because the National Federation of Coffee Growers devotes some of its revenue to development projects in coffee growing regions (termed NFCG ‘social contribution’ spending), we also investigate whether or not cyclical changes in social contributions could explain our findings. Prima facie evidence against a meaningful role of NFCG social spending is the fact that it generally focuses on infrastructure projects, responding very slowly to revenue fluctuations.²⁵ We re-estimate equation 1 using a department-level panel of NFCG social contribution spending for years 1979 through 2004 (with department rather than county fixed effects and linear trends; see Appendix 1 for a description of this data). We find no statistically meaningful relationship between coffee prices and department-level NFCG social spending.²⁶

5.1.c. Traffic Fatalities and Violent Deaths

Wealthy country studies that report procyclical mortality suggest that traffic fatalities may play an important role (Ruhm 2000). Given high levels of violence in Colombia, cyclical changes in violent deaths may also be quantitatively meaningful.²⁷ To consider these possibilities, we obtained county-year data on traffic fatalities and violent deaths for years 1979–2001 from the Colombian national statistical agency (Appendix 1 describes this data). Appendix Table A2 Panel A shows coefficient estimates for coffee price × coffee cultivation interactions when using these measures as dependent variables in equation 1 – neither estimate is statistically distinguishable from zero.

APPENDIX TABLE A2.

Coffee Price Fluctuations, Traffic Fatalities, and Violent Deaths

	Estimate	Standard Error	N	R2
Panel A: County-Level Traffic Fatalities and Violent Deaths
ln(Motor and Other Land Transportation Deaths)	0.015	(0.024)	12,104	0.81
ln(Violent Deaths)	0.012	(0.007)	11,722	0.85
Panel B: Department-Level Alcohol and Tobacco Tax Revenue
ln(Beer Sales)	−0.006	(0.007)	396	0.99
ln(Liquor Sales)	−0.014	(0.023)	384	0.97
ln(Tobacco Sales)	0.009	(0.011)	394	0.94

Notes: County-year numbers of traffic fatalities and violent deaths for years 1979–2001 from the Departamento Administrativo Nacional de Estadística; department-year beer, liquor, and tobacco sales tax data from Contraloria General de la Nación ; coffee cultivation data from the National Federation of Coffee Grower’s early 1980s coffee censuses; annual internal coffee price data from the National Federation of Coffee Growers. Estimates and standard errors (in parentheses, clustered by county for Panel A and department for Panel B) shown for the interaction between coffee growing intensity and coffee price obtained by estimating equation 1 (controlling for county and year fixed effects and county-specific linear trends in Panel A, controlling for department and year fixed effects and department-specific time trends in Panel B). Coffee area is in thousands of hectares and coffee prices are in thousands of pesos per kilogram.

^*p<0.1,

^**p<0.05,

^***p<0.01.

5.1.d. Alcohol and Tobacco Consumption

Procyclical consumption of harmful normal goods (alcohol and tobacco) is also a potential contributor to procyclical mortality (although the consequences for infants and children may be less than for adults) (Ruhm 2000). Because disaggregated Colombian alcohol and tobacco sales data are unavailable, we obtained annual department-level tax revenue data from beer, hard liquor, and tobacco sales for years 1990–2001 (Appendix 1 describes this data). Although we lack explicit information about tax rates and pre-tax prices over time, Colombian alcohol and tobacco tax rates have been stable in recent decades (with the exception of legislated national changes in 1995 and 2006). Considering these tax revenue measures to be crude proxies for consumption, we use them as dependent variables to re-estimate equation 1. Appendix Table A2 Panel B suggests that none change meaningfully with coffee price fluctuations.

5.2 Validity Tests

We also consider threats to the internal validity of our analyses – Appendix 3 presents informal validity tests which we summarize here. Given our empirical framework, any confounding influence must have varied both over time in the same erratic manner as world coffee prices and across counties in proportion to coffee cultivation intensity. The most natural concerns are (1) that we mistake changes in the composition of births or women giving birth for mortality responses and (2) that interacting plausibly exogenous price shocks with endogenous measures of coffee cultivation causes us to mistake selection effects for true behavioral responses.²⁸ Appendix 3 presents evidence suggesting that neither concern seems important in practice.

6. Conclusion

This paper uses large world coffee price fluctuations to understand the cyclical nature of infant and child mortality in Colombia. Despite the presence of credit constraints and other market imperfects leading to incomplete consumption smoothing, we document a striking pattern of procyclical mortality. We also show that a key behavioral underpinning of procyclical mortality appears to be changes in the value of time associated with world coffee price shocks. Time is a critical health input – many of the leading determinants of child survival in developing countries are inexpensive but require large amounts of time – so all else equal, the relative price of health falls with the return to working.

Our results are also consistent with growing evidence on the primacy of time in the production of overall child quality. Studies of school enrollment in American history and in contemporary developing countries report patterns of procyclical child labor and countercyclical school enrollment – and similarly link these patterns to the opportunity cost of time (Goldin 1999, Schady 2004, Kruger 2007). Other studies of child development have cast doubt over the importance of material resources per se – and transitory income in particular – for the cognitive, social, and emotional development of children (Mayer 1997, Blau 1999, Price 2008).

Finally, our findings directly relate to broader debate about the wealth-health relationship. Although wealth has prominently been proposed as an important determinant of mortality (McKeown 1976, Pritchett and Summers 1996), there is growing suggestion that reductions in the relative price of health (due to technological progress in public health, for example) are more important in explaining observed declines in mortality across time and space (Preston 1975, Jamison et al. 2001, Cutler and Miller 2005, Cutler, Deaton, and Lleras-Muney 2006, Deaton 2006). Our results are consistent with this view.

Appendix 1: Data

National Federation of Coffee Growers and the International Coffee Organization: Coffee Prices

We initially obtained average annual internal coffee prices paid to Colombian coffee growers for years 1970 to 2006 from two sources: the London-based International Coffee Organization (ICO) and the National Federation of Coffee Growers (NFCG). Internal prices paid to Colombian coffee growers at a given point in time do not vary within the country. The ICO’s price data is obtained directly from the NFCG, so we generally use the latter. (The single exception is our analyses using Familias en Acción survey data. Because the last wave of the survey was conducted in 2005/2006 – but the NFCG only provided price data through 2004 – we use real ICO price data in US dollars). We then converted the NFCG time series price data (obtained in Colombian pesos per kilogram of “green” coffee) to real 1998 terms using the official consumer price index constructed and published by the Colombian Central Bank (Banco de la República). This price index is available on-line at: http://www.banrep.gov.co/estad/dsbb/srea_012.xls.

National Federation of Coffee Growers: Coffee Cultivation

Approximately once per decade, the National Federation of Coffee Growers conducts a complete enumeration of all coffee growers in Colombia for planning and monitoring purposes. The 1970 coffee census combined information collected directly from coffee growers with land use data gathered through aerial photography. Drawing on the experiences of the 1970 census, the 1980–81 coffee census was conducted primarily using aerial photography with field verification for purposes of quality control. The 1997 coffee census was based on a complete enumeration (on the ground) of all coffee growers between 1993 and 1997. In all of our analyses, we use the immediately preceding coffee census conducted prior to a given price shock. (Because new coffee plants require four years to produce their first mature harvest, area dedicated to coffee cultivation cannot respond quickly to changes in world coffee markets.)

The 1970 and 1980 coffee censuses are available only in hard-copy format from the NFCG. With special permission from the NFCG, we digitized county-level indicators of coffee cultivation from each census using these printed volumes. The principle measure relevant to our analyses is hectares of land dedicated to coffee cultivation. The 1997 coffee census is available in electronic format.

100% 1993 Colombian Population Census: Birth Cohort Size

We constructed birth cohort size counts at the county-birth year level using the complete (100%) 1993 Colombian population census obtained from the Colombian National Statistical Agency (Departamento Administrativo Nacional de Estadística, or DANE). These birth cohort counts were generated using detailed geographic identifiers that allow all counties to be recognized according to each individual’s county of birth, not county of residence in 1993. There were 32,451,229 non-institutionalized individuals in 1060 counties recorded in the 1993 population census. These counties account for all of Colombia in a mutually exclusive and collectively exhaustive manner. Birth cohort counts were then matched to (i.) prevailing real internal coffee prices in each cohort’s year of birth and (ii.) the most recent county-level coffee cultivation measures prior to that year in each cohort’s county of birth.

Demographic and Health Surveys: Health Investments and Maternal Socio-Economic Status

Our primary measures of health investments and mothers’ socio-economic status are obtained from four waves of Colombia’s Demographic and Health Surveys (DHS). These are nationally-representative surveys of fertile age women (defined as 15–49) in the year a survey is conducted. We pool the four DHS waves together using variables reported in a comparable manner over time. (The first wave in 1986 was conducted by the Corporación Centro Regional de Población; the 1990, 1995, and 2000 waves were conducted by the Asociación Pro-Bienestar de la Familia Colombiana, or PROFAMILIA.) Public-use DHS data is available for download by registering at: http://www.measuredhs.com/. Using the child recode files matched to maternal characteristics (a pooled sample of 70,695 children), we then match each child to (i.) the prevailing real internal coffee price in his/her year of birth and (ii.) the most recent county-level coffee cultivation measures prior to the child’s birth year according to county of residence at the time of the survey (county of birth is not recorded in the DHS data). Individual counties are not identified in the publicly-available Colombian DHS data, but PROFAMILIA and Macro International (the US-based DHS partner) provided keys that match sampling clusters to individual counties.

Available measures of health investments reported consistently across the four waves include: maternal use of prenatal care, prenatal tetanus vaccinations, birth assistance, breastfeeding duration, and a variety of child vaccinations (BCG, polio, DPT, and measles). This mother-reported information can be divided into two categories: birth histories and child health histories. The birth histories are reported for every live birth (regardless of child survival to the survey date) and include prenatal care, prenatal tetanus vaccinations, birth assistance, and breastfeeding duration. The child health histories are reported for all children born within sixty months of the survey date (regardless of child survival to the survey date) up to a maximum of six children per woman and include BCG, polio, DPT, and measles vaccinations. Maternal socio-economic characteristics for each child that are available in all four waves include: age, educational attainment in years, number of preceding births, preceding birth interval, age at first birth, age at first marriage, and number of household members.

Familias en Acción Evaluation Survey: Local Labor Markets, Child Morbidity, and Time Use

Our data on rural labor markets, childhood morbidity, adult time use, and travel time to health care facilities is drawn from three waves of a household panel survey conducted to evaluate the Familias en Acción program in Colombia. Familias en Acción is a non-randomized conditional cash transfer program similar to Mexico’s Oportunidades program (formerly known as Progressa). The survey was first conducted in 2002 with follow-up surveys repeated in 2003 and 2005/2006. In 2002, the baseline survey was administered to 11,502 households in 122 Colombian counties. Attrition rates for the two follow-up surveys relative to the first wave were 6.3% and 17.1%. Topical survey modules broadly covered household demographic characteristics and composition, consumption, income, school attendance and educational attainment, and labor force participation. In addition, detailed health questions – including adult-reported symptoms of child morbidity and child preventive health care service use – were asked for all children 6 years of age and younger. Although the Familias en Acción survey data does not extend back to the three major world coffee price shocks studied by this paper, it collected complementary data relevant to our analyses but not available in any other data source on employment, wages, hours worked, adult time use, and travel time to health care facilities.

WHO Expanded Program on Immunization (EPI) Colombian Vaccination Records

To investigate how infant/child vaccinations move with coffee prices, we obtained county-year vaccination records covering years 1998–2007 for the following vaccines: polio, DPT (diphtheria, pertussis or whooping cough, and tetanus), BCG (Bacille Calmette-Guerin, a preventive tuberculosis vaccine), hepatitis B, haemophilus influenzae type b (against the variety of Hib disease manifestations, including meningitis and pneumonia), and MMR (measles, mumps, and rubella). The Colombian Ministry of Health is formally responsible for reporting aggregate national vaccination coverage data to the World Health Organization. However, it does not store and maintain disaggregated vaccination data for preceding years. To facilitate the study of vaccination coverage, a private organization (Fundación Santa Fe de Bogotá) has therefore conducted a project in conjunction with the Ministry of Health to collect disaggregated vaccination data from each Colombian county. To date, this project has collected retrospective county-level vaccination data for years 1998–2007. The Fundación has generously provided this data to us, and we have merged these vaccination records with the 1997 coffee census and annual internal coffee prices.

Contraloría General de la Nación: Colombian Local Public Finance Records

We obtained county-year level date on local public revenue and spending collected by Contraloría General de la Nación (the senior accounting branch of the Colombian government). Contraloría obtains this data from each county’s official balance sheets. This data is available for years 1984–1993 and includes: current income, transfers, capital income, tax income, direct taxes, indirect taxes, industry and commerce taxes, total spending, investment spending, operational spending, water spending, infrastructure spending, housing spending, education spending, and health spending. We merged this data with annual coffee prices and county-level coffee cultivation in 1980.

Because data on alcohol and tobacco tax revenue is also relevant to our analyses but is generally not reported at the county level, we also obtained department-year level alcohol and tobacco tax revenue data for years 1990–2001 as well. We then merged this department-year data with department-level coffee cultivation measures from the 1980s and annual internal coffee prices.

National Federation of Coffee Growers: NFCG Social Contribution Spending

At our request, the NFCG has assembled historical information on its annual social contribution for 16 departments (or states) from 1979 to 2006. This data represents the first time that the NFCG has assembled historical information on social contributions from its archival records. Instead of paying taxes, the NFCG makes these social investments to improve quality of life in rural coffee-growing regions of Colombia. These funds are generally spent on public goods for coffee growers, including agricultural research, business development, and publicity as well as four major areas of infrastructure investment (electrification, school construction, water, and roads).

Total social contribution spending is available at the department by year level with the exception of two instances in which NFCG records report combined information for several of Colombia’s most sparsely populated departments (because of the organization of NFCG cooperatives in the most remote areas). These two multi-departmental units are (1) Caquetá, Casanare, Meta and Chocó and (2) Cesar and La Guajira. Our results are insensitive to conducting aggregate regional-level analyses and to excluding these departments. Departments for which the NFCG reported no social spending are assumed to be missing (rather than reflecting no spending) and are omitted from our analyses. We aggregated our county-level coffee cultivation data to the department level and merged the department-year social contribution data with department-level coffee cultivation (using the immediately preceding coffee census) and annual coffee prices.

Colombian Vital Registry Records: Traffic Fatalities and Violent Deaths

Electronic death records at the individual level are available for years 1979–2002 from the Colombian National Statistical Agency (DANE). These records include deaths by age, sex, cause (ICD classification), place of occurrence, place of residence, month and year, marital status, and certification by a medical professional. We provide graphical evidence of infant mortality over time by degree of coffee cultivation but do not generally otherwise make use of Colombia’s mortality statistics because of concerns about data quality and under-reporting (particulary given that the degree of under-reporting thought to be correlated with economic conditions) (Flórez and Méndez 1997, Medina and Martínez 1999, Hill 2003). Appendix 2 provides indirect estimates of under-reporting in Colombia’s vital registry data that range between 30% and 45%. The exception is that we use DANE’s county-year records of traffic fatalities and violent deaths (given that no alternative measure of these deaths is available) in investigating potential alternative explanations for the patterns of procyclical mortality that we observe.

Appendix 2: Indirect Mortality Estimation

To assess the extent of under-reporting in Colombia’s mortality statistics, Appendix Table A3 shows indirect estimates of Colombia’s infant mortality rate (deaths under age one per 1,000 live births) over time. These calculations, taken from Urdinola (2004), are conducted using Colombia’s Demographic and Health Surveys (DHS) and the Brass-Trussell method (United Nations 1990). Specifically, these estimates suggest that under-reporting rates were 31% in 1986, 46% in 1990, 47% in 1995, and 44% in 2000. Similar calculations (not shown) using the Palloni-Helligman (1985) variant of the Brass method yields infant mortality estimates that are roughly equivalent. Appendix Table A3 suggests that the quality of Colombia’s vital registration system may have been deteriorating over time. This is consistent with observations made by others assessing the quality of Colombia’s vital registration system (Medina and Martínez 1999).

APPENDIX TABLE A3.

Infant Mortality Under-reporting in Colombia: Indirect Brass-Trussell Estimates Vs. Official Figures

Source	Year	Estimated Infant Mortality Rate	Vital Registry Infant Mortality Rate
DHS 1986	1986	40.6	27.9
DHS 1990	1990	37.5	20.0
DHS 1995	1995	33.7	18.0
DHS 2000	2000	30.6	17.1

Notes: Vital registry infant mortality data from the Departamento Administrativo Nacional de Estadística; Indirect Brass-Trussell infant mortality rate estimation procedures using the Demographic and Health Survey waves (1986, 1990, 1995, and 2000) described in Appendix 3. Rates shown are deaths under age one per 1,000 live births.

Indirect Estimation of Infant Mortality Rates: the Brass-Trussell Method²⁹

The Brass method of indirect mortality estimation is a standard tool used by demographers to calculate the probability that a child has died (q(x)) by age x in cross-sectional data (Brass, 1974). These probabilities q(x) can therefore be interpreted as age-specific mortality rates commonly found in standard life tables. At minimum, this method requires information on the proportion of infants and children who have died as a share of children ever born to women at each age. Widespread reliance on the Brass Method in producing indirect estimates of age-specific mortality rates has led the United Nations to place the number of children ever born and the number of surviving children on its list of recommended items for national population censuses (United Nations 1990).

The Brass method essentially utilizes differences in child survival rates across age cohorts of mothers to recover information about age-specific child mortality. It exploits the fact that all else equal, children (both alive and dead) born to older women are observed at older ages. Women at varying ages are therefore assumed to provide information about the experiences of all women in the population at each age. An important limitation of this approach is therefore its assumption that cumulative mortality rates depend on age alone. Nevertheless, a large literature in demography demonstrates its usefulness as an approximation in assessing the extent of under-reporting in vital registries.

Following standard notation used in demography, the number of children dead as a share of children ever born (denoted as D_i) among women in reproductive age groups i (i=1 for women 15–19, i=2 for women 20–24, …, i = 7 for women 45–49) are transformed into probabilities of dying (q(x)) between birth and exact age x. For the infant mortality calculations shown in Appendix Table A3, the age x of interest is one. The Brass method’s basic equation is:

$$q(x)=k_i{D}_i,$$ (3)

where k_i is a vector of multipliers derived from fertility measures among women in the population of interest. Through simulations, Brass generated the proportions of children dead, the probabilities of dying, and parity ratios (P₁/P₂, P₂/P₃, etc) linking them. Thus, an estimate of the probability of dying by age 1, q(1), can be derived from the proportion of children dead reported by women aged 15–19, D(1); the probability of dying by age 2, q(2), can be obtained from the proportion of children dead for women aged 20–24, D(2), and so on.

The original Brass method also assumes that mortality rates are constant over time, making cohort and period mortality probabilities identical. This assumption has subsequently been relaxed – if the rate of change over time as assumed to be constant, the reference date of each q(x) can be estimated by making allowances for the age pattern of fertility by means of the parity ratios.

More flexible variants of the Brass method have also been developed. One of the best known is the Trussell (1975) variant, which estimates the multipliers k_i differently. Specifically, the fertility schedule used to produce the ratios P₁/P₂, etc. is taken from the Coale-Demeny (1966) model life tables.³⁰ The Trussell variant also differs in assuming that both infant mortality and fertility patterns remained constant in preceding periods (specifically, the preceding 15 years). Finally, the Palloni-Heligman variation includes a correction using more precise information on birth timing and employs United Nations model life tables for developing countries.³¹ The Coale-Demeny model life table that best fits Colombian vital patterns is the West life table, and the Palloni-Heligman table that best fits Colombian vital patterns is the Latin American variant.

Appendix 3: Informal Validity Tests

In this Appendix we consider threats the internal validity of this paper’s main analyses. Any confounding influence must have varied in a very specific way – both over time in the same abrupt manner as birth year coffee prices and across counties in proportion to the intensity of birth county coffee cultivation. The most natural concerns are: (1) that we mistake changes in the composition of births or women giving birth for mortality responses, or (2) that selection into areas with varying coffee-growing intensity biases our main results. This section presents validity tests that investigate – but fail to corroborate – these concerns.

We first consider how coffee price shocks might alter the composition of births or the types of women giving birth. Economic models of fertility predict that changes in economic circumstances should differentially influence the fertility of women with varying opportunity costs of time (as measured by socioeconomic status, for example) (Becker and Lewis 1973, Ben-Porath 1973, Becker 1981, Butz and Ward 1979, Perry 2003, Dehejia and Lleras-Muney 2004). As already described, we address potential changes in the composition of women giving birth by restricting our analyses to children already conceived at the time a price shock occurred (children born in shock years or earlier). However, because we only know children’s year of birth, we also investigate the possibility of confounding compositional changes directly. Exploiting detailed information on maternal characteristics available in the Demographic and Health Surveys, we use the pooled DHS sample of children ages 0–2 in price shock years to regress measures of maternal socioeconomic status on the interaction between birth year coffee price and birth county coffee cultivation as in equation 2.

Appendix Table A4 shows coefficient estimates for the interaction between coffee price and intensity. There is no evidence of any change in the composition of mothers’ age, education, age at first birth, age at first marriage, preceding number of births, or number of household members. Similarly, estimates for preceding birth intervals – a direct measure of fertility – are statistically indistinguishable from zero. Other confounding compositional changes should be detectable in these analyses as well – including differential migration of pregnant women induced by price shocks and selective mortality among women giving birth – but we find little evidence of them.

APPENDIX TABLE A4.

Coffee Price Shocks, Maternal Characteristics, and Birth Timing

	1975 Brazilian Frost	1985 Brazilian Drought	1990 ICA Collapse
Mother’s Age	0.020 (0.016)	5.616 (17.600)	−0.004 (0.011)
Maternal Education	−0.009 (0.008)	0.001 (0.009)	0.004 (0.009)
Number of Household Members	−0.005 (0.006)	−0.003 (0.010)	−0.009 (0.009)
Mother’s Preceeding Number of Births	0.003 (0.008)	−0.011* (0.006)	0.001 (0.006)
Mother’s Age at First Birth	0.009 (0.007)	−0.009 (0.010)	−0.013 (0.016)
Mother’s Age at First Marriage	0.007 (0.010)	−0.005 (0.011)	−0.004 (0.034)
Preceding Birth Interval	−0.026 (0.074)	0.097 (0.180)	−0.230 (0.204)

Notes: Individual-level maternal characteristics from the pooled 1986, 1990, 1995, and 2000 Colombian Demographic and Health Survey child sample; coffee cultivation data from the National Federation of Coffee Grower’s early 1980s coffee census; annual internal coffee price data from the National Federation of Coffee Growers. Estimates and standard errors (in parentheses, clustered by county) shown for the interaction between coffee growing intensity and coffee price in the year that a woman gave birth obtained by estimating equation 2 (controlling for county and year fixed effects and county-specific linear trends). Coffee area is in hundreds of hectares and coffee prices are in hundreds of pesos per kilogram.

^*p<0.1,

^**p<0.05,

^***p<0.01.

We then explore how selection into counties with varying coffee-growing intensity prior to price shocks might bias our cohort size results. Although we are able to condition our estimates on both fixed and time varying differences across counties, Colombians might sort themselves into counties with varying coffee-growing intensity according to unobserved latent characteristics related to price responsiveness and child survival that become manifest in the presence of price shocks. A testable implication of this concern is that if people in counties with varying coffee-growing intensity were subjected to the same price shock (i.e., one whose impact should not vary with coffee-growing intensity), they would respond differently in ways related to infant and child survival.

To test this concern, we replace internal coffee prices with the Colombian CPI and re-estimate equation 1 using stable coffee price years (1968–69, 1982–83, and 1988–89). During these years, Colombian consumer prices changed by 7%, 20%, and 26% (respectively).³² Appendix Table A5 presents coefficient estimates for the interaction between birth year CPI and coffee-growing intensity. None are statistically distinguishable from zero.