3.1. Data

We use four types of data: human capital and labor market data from the CLHNS, weather data from the National Climatic Data Center (NCDC) and Water Resources Center of the University of San Carlos (WRC), historical emissions data from the RETRO (REanalysis of the TROpospheric chemical composition over the past 40 years) database, and polluter source location and industry/sector data collected specifically for this study from various governmental and non-governmental agencies in Cebu. The CLHNS provides birth outcomes including birth weight, measures of length/height throughout life, test scores measuring cognitive development, and the labor market outcomes of hours worked and hourly earnings. Weather data from the NCDC and WRC identify extremes and deviations from seasonal patterns in temperature and precipitation. Combining the RETRO database describing historical emissions in Metro Cebu with specific information on the locations and industries of pollution sources generates temporal and spatial variation in emissions. Emissions are translated into exposures using detailed residential and exposure window information for individuals in the CLHNS birth cohort.

The years of life documented and the human capital and labor market outcomes measured in the CLHNS provides the unique opportunity to assess the effects of the early-life environment in a developing economy context. The CLHNS randomly sampled 33 barangay (17 urban and 16 rural) in Metro Cebu in order to form a cohort of pregnant women (see figure 1). Barangays are the smallest administrative district in the Philippines. The 33 sampled barangays contained in total roughly 28,000 households in 1982, all of which were canvassed in search of pregnant women. Women from the selected barangays who gave birth between 1 May 1983, and 30 April 1984 are included in the baseline sample that took place during the 6th or 7th month of pregnancy. In total, 3,327 women were surveyed at baseline and 3,122 were resurveyed at childbirth. Following the child's birth, the mother-child pair was resurveyed every two months for the first two years of the child's life, and then in 1991, 1994, 1998, 2002 and 2005 (followed by limited tracking surveys).

Table 1 provides summary statistics of mothers, fathers, household characteristics, residence, migration, attrition, and children's health and human capital at birth, adolescence, and in adulthood. Anthropometrics such as height and weight were collected for women beginning with the baseline survey and for children beginning at birth. Parents in the sample are young (26–28 years old on average), and not highly educated (approximately 50 per cent with primary or less education). Fifty-three per cent of children born in the sample were male and the majority of children were born in the lengthier Amihan season between September/October and May/June. Mothers are relatively small (151 cm tall) and are, on average, 26 years old at the time of the birth of the CLHNS cohort member. Just under 50 per cent of the sample reside in Cebu City at birth and over three-quarters of the sample reside in urban barangays. Furthermore, 24 per cent of the sample permanently attrit from the three-decade long survey, while 17 per cent temporarily attrit for a period of time and are observed in later periods. Thirteen per cent of children are low birth weight (less than 2,500 g). Height and length until adulthood are expressed in z-scores; the z-score system expresses the anthropometric value as a number of standard deviations or z-scores below or above the international age and sex specific reference means. Between birth and ages 1 and 2, mean child height z-scores go from $-$.32 to $-$1.43 and $-$2.11. Miscarriages, stillbirths and deaths within a week of birth represent only 2 per cent of pregnancies in the sample, and infant mortality within the first year is 4 per cent.Footnote ²

Hourly observations of wind direction and wind speed from the NCDC provide high frequency observations of temperature, wind speed and humidity. WRC data provides daily precipitation measures from five observation stations throughout Metro Cebu. Table 1 provides summary statistics of early-life exposures to extreme temperature and precipitation. Generally, the cohort experienced greater exposure to extreme temperatures than to extreme precipitation during their exposure window between conception and age 2. Weather patterns in Cebu are influenced by the seasons: the dry season, or Amihan, from September/October to May/June, and the wet season, or Habagat. Online appendix figure A1 shows trends and variation in daily precipitation and maximum and minimum temperatures across the seasons before, during and after the early-life period of the CLHNS cohort (1978 to 1987).

The RETRO database of historical emissions describes the temporal variation in CO and $\textrm {O}_{3}$ precursors for Metro Cebu during the first two years of the lives of CLHNS birth cohort members. RETRO contains global monthly emissions by sectors for the years 1960–2000 at a $0.5 \times 0.5$ latitude-longitude degree of spatial resolution.Footnote ³ RETRO combines five global-scale numerical models of atmospheric transport and chemistry to achieve statistically robust and temporally consistent estimates of emissions (Schultz et al., Reference Schultz, Heil, Hoelzemann, Spessa, Thonicke, Goldammer, Held, Pereira and van Het Bolscher2007). Figure A2 in the online appendix shows examples of the RETRO database for CO and $\textrm {O}_{3}$ precursor emissions for a randomly selected month and an example sector (residential/commercial). We focus solely on the single $0.5 \times 0.5$ grid located at the 10.25 degrees North and 123.75 degrees East latitude-longitude which covers Metro Cebu. This $0.5 \times 0.5$ degree, 55 km $\times$ 55 km grid is shown as the outer border in the bottom panel of figure 1. The RETRO database is particularly relevant for the current study because the $0.5 \times 0.5$ grid covering Metro Cebu does not cover any other land mass or population hub that could contribute to emissions.

Available monitored emissions as well as descriptions of economic activity (energy and solvent usage, and biomass burning), technology, behavior (legislation, economic and industrial policies), population, and meteorology form inputs to the atmospheric transport and chemistry models yielding monthly emissions by pollutant and sector for each $0.5 \times 0.5$ latitude-longitude grid (Schultz et al., Reference Schultz, Heil, Hoelzemann, Spessa, Thonicke, Goldammer, Held, Pereira and van Het Bolscher2007). Comparisons to regional observational data have demonstrated the validity of the estimated RETRO emissions database and shown accurate representation of temporal atmospheric variability and chemical state (Schultz et al., Reference Schultz, Heil, Hoelzemann, Spessa, Thonicke, Goldammer, Held, Pereira and van Het Bolscher2007).Footnote ⁴ Where deviations from existing databases exist, the RETRO estimated emissions are generally conservative (Schultz et al., Reference Schultz, Heil, Hoelzemann, Spessa, Thonicke, Goldammer, Held, Pereira and van Het Bolscher2007).

Table 1 also provides summary statistics of early-life monthly average exposures.Footnote ⁵ Encompassing various industries and activities, the sectors represented in the RETRO database are: industrial combustion, power generation, manufacturing with solvents/chemicals, fossil fuel extraction and distribution, agriculture, residential and commercial, shipping, road and other land transportation, and waste disposal. Among point sources, monthly emissions of $\textrm {O}_{3}$ precursors exceed those of CO and the majority of emissions come from the power generation sector, which is entirely composed of two coal-fired power plants in Metro Cebu. CO is the most commonly emitted pollutant by non-point sources, the majority emitted by residential and commercial sources.

While the RETRO emissions database provides temporal variation in total emissions of CO and $\textrm {O}_{3}$ precursors for Metro Cebu, the database contains no spatial variation to determine individual CLHNS birth cohort member exposure. Spatial variation in exposure is generated by linking RETRO to the locations of point and non-point sources of pollution in Metro Cebu. Because Metro Cebu is the only area of economic and population density on the island and within the RETRO $0.5 \times 0.5$ latitude-longitude grid, the sources of pollution within Metro Cebu are assumed to characterize the complete set of sources that contribute to the emissions described in RETRO.

Point sources are single, identifiable sources including immobile structures like power and manufacturing plants, while non-point sources emit from more diffuse areas like agricultural land or roads. Online appendix figure A3 shows the locations of point sources by industry and sector (top row) and non-point sources of pollution (middle and bottom rows). Telephone directories from the Directories Philippines Corporation locate and describe the point sources that existed during the years 1982–1986. Telephone directories provide information regarding the existence, location and industry of Metro Cebu firms. Sources that belong to industries that required pollution permits during the 1999–2012 period are assumed to require them during the 1982–1986 period and are included as point sources for the years 1982–1986. The remaining point sources included in the data come from the Provincial Mining Office (PMO) and the Provincial Planning and Development Office (PPDO) and describe large and small scale mines. Only mines that existed between 1982–1986 are included (per PMO and PPDO data). In total, 21 large and small scale mines existed during the early 1980s in Metro Cebu; 12 of the 21 mines were copper mines, five were coal mines, and others were clay, gold and silver mines. To describe non-point sources within Metro Cebu during the years 1982–1986, we use maps describing land use (from the PPDO), zoning (each municipality planning and development offices), and the road network (PPDO) generated from data collected from the PPDO. Land use and zoning maps describe agricultural and commercial/residential pollution sources. The road network is limited to roads in existence during the 1982–1986 period. Traffic flows are estimated using a standard gravity model incorporating relative populations of barangays from the 1980s census collected from the National Statistics Office and supplemented with 1980–1985 zoning information from the Information Services Offices of the various municipalities of Metro Cebu in order to describe commuting flows between each barangay (Fernandez and Santos, Reference Fernandez and Santos2014).

Data on the point and non-point sources of pollution in Metro Cebu enables the overall pollution levels described by RETRO to be disaggregated spatially. In the next section we will describe how this disaggregation is performed in order to generate more refined measures of exposure. Additionally, the counts of point sources and the area of non-point sources will be interacted with the weather instruments as part of the set of many, invalid instrumental variables. These instruments and their use will be further described in the following section.

3.2. Econometric specification

A standard linear functional form is adopted for the human capital production function, simplifying interpretation and avoiding specification pitfalls, and human capital outcomes are examined at the ends of periods $p$ : birth ($p = 0$), age 1 ($p = 1$), age 2 ($p = 2$), ages 10–12 ($p = 3$) and ages 21–23 ($p = 4$). Let $Y_{ibp}$ denote the human capital outcomes, namely height, cognition and labor market outcomes including labor sector, hours and earnings, of individual $i$ residing in barangay $b$. The outcome is observed at the end of period $p$. Equation (1) shows the relationship between the human capital outcomes and early-life environmental factors:

(1)\begin{equation} Y_{ibp} = \alpha + \sum_{l}\beta _{l}^{W}\boldsymbol{W}_{iblt} + \sum_{j}\beta _{j}^{\sigma }\sigma _{ibjt} + \delta \boldsymbol{X}_{ibt} + \mu _{g} + \epsilon _{ibp}. \end{equation}

Here $\boldsymbol {W}_{iblt}$ denotes the exposure of individual $i$ to extreme weather and $b$ denotes individual $i$'s barangay of residence. Let $l$ denote the type of extreme weather exposure and let $t$ denote the time period of exposure. Let $j$ denote the pollutant type such that $\sigma _{ibjt} = \{\sigma _{i,b,CO,t},\sigma _{i,b,O_{3},t}\}$ describes the exposure of individual $i$ to pollutant $j$ during time $t$. For early-life outcomes observed at birth, age 1, or age 2, $t$ and $p$ are equal (i.e., for birth outcomes the period of exposure is from the estimated date of conception until birth, and at age 1 the period of exposure is from estimated conception until age 1). For later life outcomes, $t$ and $p$ are not equal. For these outcomes, $t$ describes the early-life time of exposure (prenatal up to age 2), while $p$ describes the period when the outcome is observed (age 10–12 or 21–23). Furthermore, let $\boldsymbol {X}_{ibt}$ denote a rich set of observable control variables. The set of observable control variables includes child gender and season of birth,Footnote ⁶ indicators for household environmental quality (solid fuel use, sanitary conditions and access to piped water), as well as maternal, paternal and household risk factors including per capita income, mother's education, mother's smoking, mother's alcohol consumption, mother's consumption of prenatal vitamins, the number of the mother's previous pregnancies, mother's height, mother's age, father's presence in the household, father's education and father's age. For test score outcomes, month of year when the tests were taken is also included in order to control for contemporaneous exposures. Finally, let $\mu _{g}$ capture urban/rural fixed effects within the broader urban/rural geographic category denoted $g$.

When the coefficients of interest are $\beta _{l}^{W}$, the estimation of equation (1) should yield unbiased estimates because the unpredictable extremes of temperature and precipitation denoted by $\boldsymbol {W} _{iblt}$ are unlikely to be correlated with the error term $\epsilon _{ibp}$. For $\beta _{j}^{\sigma }$, the situation is more complicated. The first problem is that pollution exposures are non-random. In the context of Metro Cebu, households of higher income and socioeconomic status live in areas of greater economic activity and high emissions (see online appendix table A4). This suggests that unobserved determinants of human capital (denoted as $\boldsymbol {V}_{ibt}$) in the error term could be correlated with $\sigma _{ijp}$ producing positive omitted variable bias. The second problem is that measures of true exposures to (CO, $\textrm {O}_{3}$) are unavailable. To address this, we estimate $\sigma _{ibjt}$ by combining the emissions described in RETRO, the locations of point and non-point sources of pollution in Metro Cebu, and the early-life exposure window of CLHNS birth cohort members. Intuitively, we do this by dividing the total emissions measured by RETRO across Metro Cebu according to the locations of pollution sources, and then relating the timing of the exposures to the early-life exposure windows of the CLHNS birth cohort members. The first step is to divide the $0.5 \times 0.5$ degree RETRO grid into 400 $0.025 \times 0.025$ degree grids.Footnote ⁷ These smaller grids are matched to CLHNS birth cohort members’ barangay of residence. Assuming equal emissions for each pollution source within a particular sector, we estimate the emissions of pollutant $j$ during quarter of the year $q$ for the $0.025 \times 0.025$ grid $m$, denoted as $\hat {E}_{mjq}$.Footnote ⁸. Figure A4 in the online appendix displays $\hat {E}_{mjq}$ for CO and $\textrm {O}_{3}$ in each $0.025 \times 0.025$ grid of Metro Cebu for a selected year. Estimated emissions levels vary spatially according to the locations of sources, and temporally according to aggregate emissions levels described in RETRO. Finally, the estimated exposure of the CLHNS birth cohort member, or $\hat {\sigma }_{ibjt}$, is the average of $\hat {E}_{mjq}$ across the quarters $q$ of the years corresponding to the individual's exposure window $t$. An underlying assumption in the process of generating $\hat {\sigma }_{ibjt}$ is that within sectors the emissions of each pollutant source are equal. This assumption is unlikely to be true. In other words, $\hat {\sigma }_{ibjt}$ may have measurement error (denoted as $u_{ibjt}$) that may be systematically correlated with the error term. The bias caused by measurement error is also likely to be positive because of the residential sorting patterns in Metro Cebu.

Denoting $\nu _{ibp}$ as the combination of the previous error term, $\epsilon _{ibp}$, with the unobserved determinants, $\boldsymbol {V}_{ibt}$, and the measurement error, $u_{ibjt}$, equation (1) becomes:

(2)\begin{equation} Y_{ibp} = \alpha + \sum_{l}\beta _{l}^{W}\boldsymbol{W}_{iblt} + \sum_{j}\beta _{j}^{\sigma }\hat{\sigma _{ibjt}} + \delta \boldsymbol{X}_{ibt} + \mu _{g} + \nu _{ibp}.\end{equation}

Instrumental variables can address the omitted variable and measurement error biases. As discussed, we use weather instruments to estimate the effects of early-life pollution exposure. The obvious problem is that weather, $\boldsymbol {W}_{iblt}$, is included in equation (2) and may directly affect $Y_{ibp}$. Estimating $\beta _{l}^{W}$ using equation (2) provides our first test of whether weather instruments can be used. For the outcomes where the estimates of $\beta _{l}^{W}$ indicate that weather does not directly affect the outcomes, the empirical specification becomes:

(3)\begin{align} \hat{\sigma}_{ibjt} & = \gamma _{1}\boldsymbol{Z}_{ibt} + \gamma _{2}\boldsymbol{X}_{ibt} + \mu _{g} + \xi _{ibjt}, \end{align}

(4)\begin{align} Y_{ibp} & = \alpha + \sum_{j}\beta _{j}^{\sigma }\hat{\sigma}_{ibjt} + \delta \boldsymbol{X} _{ibt} + \mu _{g} + \nu _{ibp}. \end{align}

$\boldsymbol {Z}_{ibt}$ denotes the set of many instrumental variables, including extreme weather, deviations from seasonal averages, and the interactions of each with location-specific concentrations of pollution sources (counts of point pollution sources and area of non-point pollution sources). Equations (3) and (4) are estimated with LIML and MBTSLS. Comparing the LIML and MBTSLS estimates provides the second test of the direct effects of weather. For the human capital outcomes where weather does not demonstrate direct effects either in equation (2) or by comparing the LIML and MBTSLS estimators, we use the weather instruments to estimate $\beta _{j}^{\sigma }$.