The Effects of Income Fluctuations on Rural Health and Nutrition

Our working paper, âThe Effects of Income Fluctuations on Rural Health and Nutrition,â provides causal evidence on how the income fluctuations poor households confront across the globe influence health and nutrition outcomes across the life cycle. We use individual-level data from a 13-year, nationally-representative rotating panel survey of Kyrgyzstan to estimate the effects of fluctuations in the incomes of agriculture-dependent households on the heights and weights of young children (age 0â5) and on the incidence of overweight and obesity among children and adults. Our focus on departures of income from trend is distinct from analysis of the effects of long-term changes in income. It offers insight into how health responds to income fluctuations that are ubiquitous in developing countries rather than the impacts of global shifts in a countryâs prosperity. We address the endogeneity of income to health and consumption using an instrumental variables approach; we instrument for income with predicted income, obtained using the householdâs initial period share of income from six different revenue sources, agricultural production costs from two different sources (crop and livestock), and aggregate growth rates of each of these eight revenues and costs over time. We find that young children (age 0-5) exposed to reductions in income experienced reductions in height. At the same time, older children and adults saw decreases in BMI andâfor adultsâdecreases in the incidence of overweight.


Introduction
How do income fluctuations affect health and nutrition outcomes, and how do these effects vary by gender and across the life cycle? Studies looking both across and within countries reveal strong correlations between income and health (Cutler et al. Cutler et al., 2006Adda et al. Adda et al., 2009 raise healthcare expenditures, reduce productivity, and lead to premature mortality (Withrow andAlter Withrow andAlter, 2011 2011;Dee et al. Dee et al., 2014. Childhood and adolescent obesity also negatively affect social and emotional well-being and academic performance (Sahoo et al. Sahoo et al., 20152015. Nonetheless, public health campaigns have largely failed to curb over-nutrition (Swinburn et al. Swinburn et al., 20112011Ng et al. Ng et al., 2014. 3 3 These trends have been attributed to technological change which has lowered the cost of calories-especially that of refined grains, added sugars, and added fats, which are now among the lowest-cost sources of dietary energy (Drewnowski andDarmon Drewnowski andDarmon, 2005 2005)-while making work more sedentary and raising the cost of physical activity (Cutler et al. Cutler et al., 2003Lakdawalla et al. Lakdawalla et al., 2005World Health Organization World Health Organization, 2009Strombotne Finkelstein andStrombotne, 2010 2010). Low-income individuals in some settings are also more likely to be overweight or obese (Drewnowski andDarmon Drewnowski andDarmon, 2005 2005), and some research has linked economic insecurity to the consumption of foods that cause weight gain (Gundersen andKreider Gundersen andKreider, 2009 2009;Smith Smith, 20112011. Nonetheless, other research finds no relationship (Alaimo et al. Alaimo et al., 20012001Bhargava et al. Bhargava et al., 20082008 or a negative relationship (Jimenez-Cruz et al. Jimenez-Cruz et al., 2003Rose andBodor Rose andBodor, 2006 2006) between economic insecurity and obesity. And empirical evidence from Sweden shows that winning the lottery increases children's health care utilization and may reduce obesity risk (Cesarini et al. Cesarini et al., 20162016. The net impact of changes in income on the incidence of overweight and obesity is somewhat ambiguous. In this paper, we use individual-level data from a 13-year, nationally-representative rotating panel survey of Kyrgyzstan to estimate the effects of fluctuations in the incomes of agriculture-dependent households on the heights and weights of young children (age 0-5) and on the incidence of overweight and obesity among children and adults. Our focus on departures of income from trend is distinct from analysis of the effects of long-term changes in income. It offers insights into how health responds to income fluctuations that are ubiquitous in developing countries rather than the impacts of global shifts in a country's prosper-ity. Our focus on a low-income, developing country is motivated by research showing that poor households tend to under-insure against reductions in income (Townsend Townsend, 19941994, 19951995 Jalan and Ravallion Jalan andRavallion, 1999 1999;Dercon Dercon, 2002Yang Yang, 20082008, making them more vulnerable to such fluctuations. The poor also face a higher arrival rate of health shocks (Currie and Stabile Currie andStabile, 2003 2003), and their negative health impacts accumulate over time (Case et al. Case et al., 2002. As poorer households' inability to smooth their consumption over time has been shown to disproportionately affect women (Dercon andKrishnan Dercon andKrishnan, 2000 2000), we estimate health impacts by gender.
We address the endogeneity of income to health and consumption using an instrumental variables approach; we instrument for income with predicted income, obtained using the household's initial period share of income from six different revenue sources, and agricultural production costs from two different sources (crop and livestock), and aggregate growth rates of each of these eight revenues and costs over time.
Our paper extends existing literature in several ways. First, we focus on the impacts of recurring fluctuations in income-in our case due to price shifts-as opposed to either extreme shocks that form natural experiments or targeted cash transfer programs. This helps isolate the health and nutrition consequences of income fluctuations from the trauma of extreme shocks, and increases the external validity of the findings beyond individuals targeted by transfer programs. Second, by exploiting not only time but also spatial variation in exposure to macroeconomic shocks to income and shifts in the costs of agricultural inputs, we are able to separate the effects of reductions in household income from other countrywide or region-wide shocks. These include everything from the quality of public services to the relative prices of different foods and essential nutrients. Third, we address identification challenges related to the simultaneity of income and health using an instrumental variables approach following Bartik Bartik (19911991. Showing causal relationships is broadly challenging. Some efforts to achieve identification have included instrumenting for income with five-year changes in terms of trade (Pritchett andSummers Pritchett andSummers, 1996 1996) and past rainfall (Bengtsson Bengtsson, 2010; we build on these efforts. Fourth, we add to a scant literature on the impacts of fluctuations in income on the incidence of overweight and obesity-two phenomena currently experiencing a rapid (Popkin et al. Popkin et al., 2012 and costly (Thorpe et al. Thorpe et al., 2004Thorpe et al., 2004 rise worldwide. 4 4 Fifth, relatively little attention has been paid to the differential impacts of fluctuations in income across the life cycle. The focus has generally been on young children-often disaggregated by gender-while ignoring potentially important impacts on older children and adult women and men. When studies do consider adults, they often consider the impacts of negative shocks in utero or during early childhood on subsequent adult health, rather than the impacts of contemporaneous fluctuations in income. Finally, we explore potential causal channels related to consumption, dietary diversity, expenditure on healthcare, and fertility. We find that young children (age 0-5) exposed to reductions in household income experienced reductions in height and height-for-age z-scores that were largest for girls and those under age two-groups that additionally experienced increases in stunting. Both girls and boys experienced reductions in weight, weight-for-age z-scores, and weight-for-height z-scores. Reduced consumption of healthy foods, reduced dietary diversity, and less expenditure on healthcare may help explain the results. A channel possibly offsetting negative impacts is a decrease in fertility. At the same time, older children and adults saw decreases in BMI following reductions in income and-for adults-decreases in the incidence of overweight. The effects of reductions in income on the BMIs of older children and youth appear to be mostly driven by men; we find no evidence that fluctuations in income affect the BMI or incidence of overweight in older female children (age 5-18) or female youth (age 18-35).
In contrast, reductions in income among older adults (age 35 and older) lower BMIs and reduce the incidence of overweight among both men and women.
This paper adds to a growing literature employing shift-share instruments, following Bartik Bartik (19911991. For example, Card Card (20012001) studies the effects of immigrant inflows on labor market outcomes by interacting initial immigrant composition of a place with immigration flows from origin countries. Acemoglu and Linn Linn (2004 2004) investigate the effect of potential mar-4 Thorpe et al. Thorpe et al. (2004Thorpe et al. ( 2004 find that 27 percent of the rise in per capita health care spending in the United States between 1987 and 2001 was due to increased spending on obese people. ket size on entry of new drugs and pharmaceutical innovations by constructing an instrumental variable that is the interaction of age-group spending patterns with demographic changes. Dube andVargas Dube andVargas (2013 2013) consider how income shocks affect armed conflict by instrumenting for the product of hectares of coffee grown in a base year and internal coffee prices in the current year with the product of hectares grown in the base year and current year international prices; they do a similar calculation involving oil production. Acemoglu et al. Acemoglu et al. (2013Acemoglu et al. ( 2013 examine the effects of income on health expenditures by instrumenting for local area income The paper is organized as follows. Section 2 2 provides background on Kyrgyzstan's economy, healthcare system, and levels of health. Section 3 3 describes our data, empirical strategy, and method for identifying the causal effects of fluctuations in household income. Section 4 4 presents results, while Section 5 5 explores potential causal mechanisms driving them. Section 6 6 shows the robustness of our findings to several alternative specifications and provides evidence to support the assumption of parallel trends. Finally, Section 7 7 concludes.

Background
Kyrgyzstan declared its independence from the Soviet Union in August 1991, and joined prior to that, it was pooled at the oblast level (Ibraimova et al. Ibraimova et al., 20112011. As a result of these reforms, the country experienced a significant increase in access to healthcare during the 2000s and a reduction in reports of unofficial, informal payments for healthcare. Nonetheless, informal payments for some services (e.g., anesthesia) reportedly persist in some places, making affording healthcare still problematic for the poorest (Falkingham et al. Falkingham et al., 2010

Empirical Strategy
We hypothesize that fluctuations in household income will influence households' consumption and various other decisions-with implications for health and nutrition outcomes. To test this, we estimate the following fixed effects model: where i indexes individuals, j indexes households, k indexes the oblast (i.e. region) in which the household resides, t indexes years, and s is the chosen lag structure in years (either 1 or 2). O ijkt is a health or nutrition outcome in our main analysis, which is measured in the first quarter of the year (January-March). We consider several other outcomes, described in Section 3.4 3.4 and analyzed in Section 5 5, when we explore the mechanisms driving our main results; these relate to consumption, dietary diversity, healthcare expenditure, and fertility.
H jkt is total annual household income; it includes non-agricultural income (paid employment, self employment, one-time work, pensions and other benefits, and capital income) and five types of agricultural incomes (crop production, livestock sales, meat production, hunting/ gathering, and production of processed food)-all net of the costs of agricultural production (crop production and livestock rearing). We log the total annual household income, but also present robustness checks showing similar results when using the level of income, described in Section 6.3 6.3. X j,k,t=0 is a vector of household-level controls, taken from the first year the household j entered the sample, and Y ijkt is a vector of individual-level controls, both described in Section 3.5 3.5. α k are oblast fixed effects while µ t are year fixed effects. If all variation in H jkt were random, β 1 would provide the causal effect of an increase in total household income on health and nutrition-related outcomes. In Section 3.2 3.2, we explain how we account for the likely endogeneity of this variable.
We consider both specifications in which we use lagged household income, H jk,t−1 as well as specifications in which we instead consider a two year lagged value, H jk,t−2 . The former examines impacts in the first quarter (January-March) following the calendar year over which annual income is measured. The latter allows an additional year for impacts to materialize. The speed with which impacts materialize may depend on the particular outcome in question and age group, motivating analysis of both. We do not consider longer lags given the short number of years (four at the median) that households appear in our sample; the sample and it selection are described in detail in Sub-section 3.1 3.1.

Data
Our data source is the Kyrgyzstan Integrated Household Survey (KIHS), a nationallyrepresentative, rotating panel household survey carried out quarterly starting in 2003. These data were collected by the National Statistical Committee (NSC) of Kyrgyzstan, with financial and technical support from the UK Department for International Development (DFID) and Oxford Policy Management (in round 1); they are described in detail by Esenaliev et al. Esenaliev et al. (20112011. The panel gradually rotated new households in and old households out-though it has not been officially detailed how this was done, motivating our analysis of attrition in Sub-

Identification
Reverse causality and omitted variables may bias ordinary least squares (OLS) estimates of the health and nutrition effects of income fluctuations. 7 7 That improved health and nutrition are likely to mechanically raise incomes should bias upward OLS estimates of the effects of increases in income. However, a host of factors is likely to simultaneously affect incomes and health, and it is difficult to sign the bias that these may generate. For example, poor weather (e.g., droughts or flooding) may lower incomes from agriculture. It is also likely to reduce food production and thus food consumption-possibly leading to a deterioration in health. At the same time, poor weather may provide adults with more leisure time and thus ability to invest in child health. It may also spur contraceptive use and thus allow parents to invest more in existing children's health as opposed to a child in utero. These effects would tend to downward bias OLS estimates. Similarly, reductions in sexism and prescribed gender roles may grow incomes by spurring women's greater involvement in the labor force, and this may increase health through increased food consumption-thus leading to upward biased OLS estimates. But they may also lead to reductions in investments in children by increasing the opportunity cost of such investments by women-accordingly downward biasing OLS estimates. This makes it difficult to sign the bias in OLS estimates. It also makes it important to empirically examine the mechanisms driving our results.
We follow a large literature based on Bartik Bartik (19911991 to identify the causal effects of fluctuations in income. Specifically, we predict logged total household income in year t = n by taking the baseline (year t = 0) values of six sources of household revenue (income from the non-agricultural sector, crop production, livestock sales, meat production, hunting/ gathering, and production of processed food) and two costs (crop production costs and livestock others who happen to be born on the same year, month, and day as another household member, this problem is likely to be minimal. In our raw data, for example, only 0.45 % observations (where an observation is an individual-year) are duplicates in terms of year, household identifier, gender, and exact birth date-on par with the 2015 twin birth rate in the U.S. (Martin et al. Martin et al., 2017. 7 Despite the difficulties of signing the direction of bias for OLS estimates, Christian and Barrett Christian and Barrett (2018 2018) make compelling arguments for discussing its likely direction clearly in the context of shift-share and other interacted instruments. production costs) that jointly sum to total income, and multiplying each by the oblast × area type (rural or urban) aggregate growth rate in this revenue (or cost) source between t and t + n. 8 8 Formally: where g j,k,r,t and g j,k,c,t are the average growth rates of revenue source r and cost source c, respectively, between year t = 0 and year t in the oblast × area type. We then use this predicted (i.e. "projected") household income variable as an instrument for actual total household income.
Using this instrumental variables strategy, our first and second stage equations are: 9 9 where γ k and δ k (η t and σ t ) are oblast (year) fixed effects in the second and first stages, respectively. Our year fixed effects absorb the impacts of nation-wide movements in revenues from different sectors over time, while our oblast fixed effects capture regional differences in the composition of income. Importantly, identification in no way comes from endogenous household decisions to change the household's relative reliance on different sectors (e.g., the non-agricultural sector vs. crop agriculture) over time. We further control for the logged value of initial (year t = 0) revenue, revenue j,r,t=0 that the household earned from each of the six revenue sources, r = {1, . . . 6}, and logged value of the initial year costs, cost j,c,t=0 that the household incurred from each of the two cost sources, c = {1, . . . 2}. And we also control for the interaction of each logged income or cost with a linear time trend, to allow households to be on different trends according to their initial reliance on different sources of revenue, or exposure to different costs. Finally, in case households are also on different secular trends according to their initial year total income, H j,k,t=0 , we additionally control for logged initial total income and its interaction with a linear time trend.
Our IV strategy thus exploits that part of household income that is due to exogenous changes in the profitability of different forms of earning income, and the costs of different ways of earning income. Our key identifying assumption is that predicted (i.e. projected) total income only affects health-related outcomes through its effects on household income.
As for all IV estimates, our estimates reflect average effects for observations that comply with the instrument-that is, we estimate a local average treatment effect Angrist Imbens andAngrist, 1994 1994). Compliers are observations that experience higher total household incomes following increases in the average earnings in sectors on which they are especially reliant and/or that experience higher total household incomes following decreases in the costs of types of production on which they are heavily reliant. In other words, our IV estimates are not driven by the effect of having higher total household income for households and individuals whose total household income is unaffected by changes in the average regional profitability of sectors in which they are heavily engaged.
It is useful to consider the size of typical fluctuations in annual income that sample households experience from one year to the next. In Figure 2  As Table 2 2 shows, total household income is indeed strongly correlated with predicted income from agriculture. Whether or not we use our full control set, and regardless of which sample we use-young children, older children, youths, older adults, or a household-level sample-the coefficient on predicted income is between 0.62 and 0.73, and our first stage F statistic is always above 360-far from suggesting any problems of weak instruments.

Attrition
As we lack data on the precise procedure employed for dropping households from the rolling panel each year, we are unable to distinguish households that exit the sample due to planned exit versus more problematic attrition-such as that due to household refusal to participate or inability to locate the household or its members. To assess whether attrition is likely to be non-random as opposed to random (i.e. planned attrition due to the rolling nature of the panel), we used our household-level sample to analyze the extent to which household income, or lagged income, predict the household leaving the sample in the following year.
In Table A3 A3, we find little evidence that household income significantly influences household attrition. While higher income in the current year predicts lower attrition in column 1 (which includes only our basic control set), the coefficient is incredibly small in magnitude: a 36 percent reduction in income (corresponding to the sample mean year-over-year change in income, as shown in Figure 2 2) leads to a 0.024×ln(1.36) = 0.007 percentage point increase in the likelihood of attriting from the sample in the following year. Thus, we effectively obtain a precisely estimated zero attrition rate. Additionally, that significance of the coefficient on income disappears when we include our full control set (column 2). Further, we find no evidence that income lagged one year, or lagged two years, has any impact on attrition-either with our limited control set (columns 3 and 5) or our full control set (columns 4 and 6). We conclude that non-random attrition is unlikely to impact our interpretation of the results.

Outcomes
Our primary outcomes consist of several measures of the nutritional status of young children (ages 1-5), older children and adolescents (ages 5-18), youth (ages 18-35), and older adults (ages 35 and over). These outcomes are measured in the first quarter of the year (i.e. between January and March), during the initial visit with the household in which roster data (including on gender and exact age) were also collected; the median date of data collection is February 15. Fluctuations in income naturally will take time to influence measures of long-term health and nutrition. As such, we are interested in the impacts of lagged income.
When we use a single lag of income (H t−1 ), we are considering income over the calendar year that began (ended) approximately 13.5 months (1.5 months) prior to measurement of our health and nutrition outcomes. And when we use a two year lag of income (H t−2 ), we are considering the calendar year that began (ended) approximately 25.5 months (13.5 months) prior. This is depicted visually in Figure 3 3.
We omit children under 12 months from our main analysis as none of them would have been in utero by the start of the year over which H t−2 is measured, and over 40 percent would not yet have been in utero by the start of the year over which we measure H t−1 . Table   A4 A4 illustrates the ages of children (e.g., A months pre-pregnancy, B months in utero, or C months old) at the start and end of each year over which we measure income. Excluded children (ages 0-11 months) are shown in grey shade. For example, the table shows that if we had included children aged 4 months old, they would be between 0.5 months pre-pregnancy (i.e. not yet conceived) and 2.5 months old during the year over which we measure income when we lag income by one year, and they would be 12.5 months pre-pregnancy to 0.5 months pre-pregnancy during the year over which we measure income when we lag income two years. We show that results are robust to including all children.
Nutritional status in young children is generally assessed using height-for-age Z-scores (HAZs), weight-for-age Z-scores (WAZs), and weight-for-height Z-scores (WHZs). In addition to considering height and weight themselves as outcomes, we thus construct each of these measures, which also utilize information on child gender, exact age (in years, months, and days), and global child growth standards from the World Health Organization World Health Organization (2006 2006).
Z-scores measure deviation from the WHO (2006) reference population's mean; a Z-score of 0 means that the individual has the mean score. Having a HAZ < −2 is known as stunting.
For older children and adolescents (ages 5-18) as well as adults, we additionally compute the individual's body mass index (BMI) and consider whether they are overweight (BMI ≥ 25 kg/m 2 ) or obese (BMI ≥ 30 kg/m 2 ). We complement these objective measures of health with household head reports about the subjective well-being of members-specifically, we code a dummy for whether each member is in good health. 10 10 Subjective impressions of the main respondent are likely to be noisy, so we interpret findings with this caveat in mind. impacts of fluctuations in household income on health and nutrition outcomes, we consider three additional sets of outcomes: those related to household consumption, dietary diversity, healthcare expenditure, and fertility. These broadly capture whether nutrition and health impacts are due to changing diets and selection into child bearing. 12 12 10 Exact question wording was: "In the opinion of [NAME], what is the state of his/her health?" Response choices included: "very good," "good," "satisfactory," "not bad, not good," "Poor," and "very poor". 11 This rate of stunting is similar to that found in World Bank World Bank (2019b 2019b), which shows average rates of stunting among children under age 5 of 36 percent in 1997, 23 percent in 2009, and 13 percent in 2014. 12 Since the seminal work of Becker Becker (19601960, it has been recognized that fertility decisions are affected by household wealth; we assess whether they are also affected by fluctuations in household income. We employ several measures of household consumption and dietary diversity, all summarized in Table 1 1. Consumption data come from a two week recall and do not indicate which household members consumed the food. However, they were collected at quarterly intervals during the calendar year (i.e. we have four observations per household in a given year) as opposed to once, helping us achieve greater between-household variation in consumption. 13 13 We code dummies for the household consuming each of 11 mutually-exclusive categories of food (cereals, eggs, fruits, meat and poultry, pulses/ legumes/ nuts, roots and tubers, fresh vegetables, fish/ seafood, dairy products, oils, and sugar); they take on a one only if the food category was consumed during all four quarters. These comprise our measures of the extensive margin of consumption. We additionally computed the logged average amount consumed, across all four quarters, of each food category-capturing the intensive margin of consumption. Each food category is measured either in liters or in kilos (kg), as appropriate. 14 14 To better understand the dairy products category, we further sub-divide it into milk products (such as milk, cream, or kefir, all measured in liters) and cheese products (such as cheese, curds, butter, sour cream, or yogurt, all measured in kg) and code both a dummy and the logged amount consumed of each. There is a twelfth, "other" category for which we can code a dummy for consumption but cannot compute the logged amount given it contains a diverse mix of foods varyingly measured in liters or kilos. 15 15 Beyond total amounts of food consumed, existing research shows that there is a strong association between child dietary diversity and nutritional status, and that dietary diversity reflects diet quality (Arimond and Ruel Ruel, 2004 2004;Moursi et al. Moursi et al., 20082008Rah et al. Rah et al., 2010. We thus code two dietary diversity indices. The first is the household dietary diversity score (HDDS); to construct it, we count the total number of our 12 categories of food for which the 13 24 hour or 7 day recalls are more commonly used in studies of dietary diversity (Arimond and Ruel Ruel, 2004 2004). This was an additional motivation to use our four observations per household per year in this way.
14 A number of studies similarly measure consumption in terms of quantity (e.g., liters or kilos) consumed (Ali and Tsou Tsou, 1997 1997;Ives Ives, 2002Suryanarayana andSilva Suryanarayana andSilva, 2007 2007). 15 While we use this as a food category in computing our dietary diversity index, we do not analyze the dummy or the logged amount of its consumption independently.) household head reported its consumption during each of the four visits. 16 16 Our choice of these 12 categories and our method of combining them into a HDDS follows Swindale and Bilinsky Swindale andBilinsky (2006 2006). 17 17 The second is a "healthy" HDDS, which we code similarly but considering only a subset of four relatively healthy food categories: fruits, pulses/ legumes/ nuts, vegetables, and fish/ seafood. This is similar to an index created by Imamura et al. Imamura et al. (2015Imamura et al. ( 2015. 18 18 We also measure whether or not women are pregnant as well as two associated measures: whether or not the woman normally practices contraception (this outcome is missing for pregnant women), and whether or not she wants more children. We include in this analysis all women of reproductive age (15-49) who have had their first period and are married or otherwise sexually active. In these regressions, we control for the number of children a woman already has (a continuous variable); on average, sample women already have 2.8 children.

Controls
We present results with and without our full set of controls. In general, results are not sensitive to inclusion of controls and our first stage is always strong. We include in all specifications geographic and time fixed effects, a quadratic in age, and a male dummy. We further include in all specifications controls for the initial period income for the household from each of the six income sources, the initial period costs faced by the household from each of the two cost sources, as well as the initial period value of total household incomeall logged-plus a linear time trend interacted with each of these nine variables.
Our full set of controls additionally includes individual-level dummies for relationship with the household head, being married, and having a general secondary degree or higher. 19 19 It also includes several household-level controls, summarized in Table 1 1. These include a dummy for residing in an urban area, logged land area farmed, 20 20 the number of unique agricultural goods the household produces annually, dummies for household size, a quadratic in age for the household head, and dummies for the head having a general secondary degree or higher, being married, and being male. All household-level controls are taken from the year in which the household enters the sample.
A few features of our sample are noteworthy. As Table 1  these are growing through 2009, then declining, and then experiencing another period of growth before declining again during 2013-2016. However, through this period, there is stable income from meat production. Panels B, C, and D of Table 1 1 summarize additional household-level outcomes. On average, households consumed only 1.9 of the four relatively healthy food categories used to construct our "healthy" HDDS (fruits, pulses/ legumes/ nuts, vegetables, and fish/ seafood) and 8.7 of the 12 food groups used to construct our HDDS.
20 Zero land was imputed to 0.1 square meters of land. Table 3 3 presents regression results where our outcomes are the height (columns 1-2), HAZ (columns 3-4), and a dummy for stunting (columns 5-6) of young children (aged 1-5). Panels A and B show ordinary least squares (OLS) results when total household income is measured during the calendar year immediately preceding the quarter 1 (Q1) measurement of health and nutrition outcomes (t − 1), and when it is measured two calendar years prior (t − 2), respectively. Panels C and D present the instrumental variables (IV) second stage analogues of panels A and B, respectively. We present specifications with both our basic (odd-numbered columns) and full (even-numbered columns) control sets. Focusing on our preferred, full controls specifications, a reduction in household income predicts significantly lower child height and HAZ regardless of lag structure employed and method of estimation (OLS or IV)-though IV coefficients are modestly larger. While OLS results suggest that reductions in income increase stunting in young children, this finding does not hold in the IV results,

Young child height
where coefficients are smaller in magnitude and statistically insignificant. The negative effects of income on height and HAZ in the specifications in which we measure income in year t−1 become larger and generally more significant in the specifications measuring income in year t − 2, despite the lower sample size (and thus statistical power) of this further lagged specification; this is consistent with the impacts taking time to materialize.
Our IV specification with the full set of controls reveals that a 36 percent reduction in income (corresponding to the sample mean year-over-year change in income, as shown in Figure 2 2) leads to a 0.187 × ln(1.36) = 0.057 standard deviation reduction in HAZs for children aged 1-5 after one year, which grows to a 0.252 × ln(1.36) = 0.077 standard deviation reduction in HAZs after two years. While we excluded 0-1 year olds as many were not yet conceived when income is lagged by two years, effects are similar if we instead use children aged 0-5, as shown in Table A5 A5. Statistical significance is largely unchanged, and point estimates are simply slightly smaller-consistent with the very youngest children in the 0-5 year old age group being unexposed (or much less directly exposed) to fluctuations in income because they pre-date their conception. In comparison, Minoiu  As Table A6 A6 reveals, these impacts are driven by girls. For neither lag structure, and for neither height nor HAZ, do we find impacts on boys. Compared to the full sample, the same 36 percent reduction in income reduces girls' HAZ scores a year later by an even higher 0.086 standard deviations, which grows to 0.093 standard deviations after two years. Girls further experience a statistically significant increase in stunting not seen among boys in response to reductions in income; the magnitude is non-trivial, contributing to a 0.113×ln(1.36) = 0.035 percentage point increase in stunting after one year, and a 0.123×ln(1.36) = 0.038 percentage point increase after two. Boys may be protected from reductions in households income in ways that girls are not-with long-term impacts on their health and nutrition.
Effects of income on height, HAZ, and stunting are concentrated on children under age 2 (Table A7 A7). Compared to the overall effects, effects on 2-5 year olds are smaller in magnitude and statistically insignificant. However, effects are larger and more statistically significant compared to overall effects when we restrict attention to 1-2 year olds, in columns 2, 5, and 8. Table A8 A8 similarly shows that statistically significant effects of income on height, HAZ, and stunting of 0-1 year olds and 0-2 year olds are in all cases larger than comparable effects on all 0-5 year olds. This is consistent with very early childhood being a critical period.

Young child weight
In addition to impacts on height, Table 4 4 shows that we also find statistically significant reductions in the weight, weight-for-age (WAZ) Z-scores, and weight-for-height Z-scores (WHZ) of children aged 1-5 in response to reductions in income. In both OLS and IV specifications, regardless of the control set employed, young children exposed to reductions in income experiences lower weight and WAZ. OLS results suggest no impacts on WHZ, but our preferred IV specifications show significant reductions in WHZ that appear to be larger when we lag income by one year than by two years. Considering our preferred, fully controlled specifica-  (Table A8 A8).
In contrast to our findings for young child height, we find no evidence that the weight, WAZ, or WHZ scores of girls and boys are differently affected by reductions in income (Table A9 A9). Nor do we find that the weight, HAZ, and WHZ of 1-2 year olds are generally more heavily affected by reductions in income than are those of 2-5 year olds (Table A10 A10).
Negative effects of reductions in income on weight and related measures for young children appear to matter across genders and ages.

Anthropometric outcomes for older children
We next consider the impacts of income fluctuations on the height, weight, and BMI of children and adolescents (ages 5-18) in Table 5 5. We do not consider as outcomes dummies for these individuals being overweight or obese as these rates are incredibly low in this population ( These are incredibly small but precisely estimated impacts. By comparison, the standard deviation of the BMI for this group is 2.30 points (Table A2 A2). Compared to very young children, older children's health is scarcely affected by reductions in income.  (Table A2 A2). We interpret this as effectively a precisely estimated zero effect. More notably, however, youth also experience a decrease in the incidence of overweight after one

Anthropometric outcomes for youth and older adults
year that is sustained when we instead measure income with a two year lag. Specifically, a 36 percent reduction in income leads to a 2.1 percentage point reduction in the incidence of overweight after one year (0.067×ln(1.36)), which declines to 1.5 percentage points after two years (0.049 × ln(1.36)). These compare to a mean (standard deviation) of the overweight dummy among youth of about 0.18 (0.39). We find no impacts of fluctuations in income on the incidence of obesity in youths, possibly due to the small mean of this variable in the youth population (only 0.02) (Table A2 A2).
Among those over age 35, we obtain similar results (Table 7 7), though coefficients are larger in magnitude and statistical significance compared to the youth sample. We also estimate statistically significant impacts on the weights of older adults that are not found in the youth sample. A 36 percent reduction in income leads to a 0.28 unit reduction in older adults' BMIs after one year that is roughly sustained after two years (0.911 × ln(1.36))relative to a standard deviation of the older adult BMI of 3.9. Also, older adults subjected to an average-sized decline in income of 36 percent tend to experience a 0.141×ln(1.36) = 0.043 percentage point decline in the incidence of overweight after one year that is roughly sustained for a second year. We interpret the null effects of income on adult heights as a useful placebo analysis. It suggests that the results do not reflect some form of systematic measurement error as opposed to real reductions in height following negative economic shocks.
Among young children, girls' anthropometric outcomes were most susceptible to reductions in income. Among adults, however, we find distinct patterns, shown in Table A13 A13; for youth, it is men whose BMI and incidence of overweight are most sensitive to the level of income; we find no significant impacts on women for either outcome. Among older adults (age 35 and older), in contrast, income affects both men's and women's BMIs and incidence of overweight-with impacts generally larger in magnitude for women compared to men.
Overall, the results suggest that economic downturns may be good on some metrics for the health of youth and older adults. To the extent that high BMIs and overweight pose public health concerns, downturns in household income may reduce them. Importantly, though, we identify very modestly-sized impacts. While our results suggest that high BMI and the incidence of overweight are likely to become growing problems as income situations improve in the Kyrgyz Republic, they do not reflect a looming public health epidemic.

Subjective well-being
We next turn to subjective impressions of health-specifically, a dummy for being in good health (Table 8 8). These findings come with the caveat that they are reports from the main respondent rather than the individual whose health status we measure, but we argue that this should if anything introduce noise that makes it more difficult to identify statistically significant effects. Despite reductions in objective measures of health following reductions in household income, we only find robust evidence of deteriorations in subjective impressions of health among 5-18 year olds following reductions in income. OLS results suggest positive impacts on all groups, but these are not robust to IV, where we find the coefficients generally lower in magnitude and statistically insignificant. This may reflect that adults' subjective health is not vulnerable to such shocks, and that respondents almost uniformly reported 1-5 year olds to be in good health (97.6 percent are reported to be in good health). Among 5-18 year olds, effects are very small; a 36 percent reduction in income yields about a 1 percentage point improvement in subjective impressions of health in the specification with a one year lag that is sustained and slightly larger after two years. This provides some evidence that survey respondents-most typically household heads-do not feel that household members' health suffers drastically following a reduction in household income.

Mechanisms
Overall, we have observed that reductions in household income have statistically significant and meaningful effects on not only young children but also on older children as well as adults. These findings hold even after accounting for the endogeneity of income to health outcomes using our instrumental variables strategy. Young children face significant reductions in height, HAZs, and WAZs-critical measures of long-term health and nutrition. They also experience reductions in WHZ, where a low value indicates acute malnutrition. Older children and adolescents experience reductions in weight though not in height. However, youth and older adults benefit from decreases in the incidence of overweight.
To better understand the potential causal mechanisms driving these results, we consider three sets of outcomes: those related to household consumption and dietary diversity, healthcare expenditure, and fertility. These help capture whether nutrition and health impacts are due to changing diets, other investments in health, and/or selection into child bearing. As we noted earlier, while the measures we employ have important caveats, the results can at least provide suggestive evidence on the mechanisms at work.

Consumption and Dietary Diversity
We first explore the food security implications of reductions in household income by considering whether they influence what food groups households consume, the quantities consumed, and dietary diversity. Here, we consider the effects of contemporaneous income and income lagged one year rather than considering longer lags as we expect income to have immediate impacts on consumption-in contrast to its effect on health and nutrition outcomes, which are further down the causal chain. The effects of income on our consumption and dietary diversity variables indeed decline when we go from measuring income in the current period to measuring it with a one year lag ( relative to its standard deviation of 0.616, this is a 0.07 standard deviation decline. Examining consumption of specific foods, Panel B provides evidence on the extensive margin (outcomes are dummy variables for consumption) while Panel C shows evidence on the intensive margin (outcomes are logged quantities consumed). From both panels, we see that following a decrease in household income, individuals are less likely to consume fruits, fresh vegetables, roots and tubers, and dairy products-possibly explaining declines in young children's health. However, consumption of sugar also goes down with declines in incomepossibly helping explain declines in BMI and the incidence of overweight among adults.
Consumption of meat and poultry is interestingly counter-cyclical-possibly indicating that households are more likely to slaughter and consume their animals in times of economic downturn as a coping strategy. However, effects of income on meat and poultry consumption are found only on the extensive margin (whether meat and poultry are consumed) and not on the intensive margin (logged amount of consumption).

Healthcare expenditure
In

Fertility
Next, we consider whether reductions in household income influence decisions on fertility in women aged 15-49; these results are reported in Table 11 11, and importantly reflect selfreports. Column 1 reveals that declines in income predict a greater likelihood of practicing contraception 21 21 -though these results are not statistically significant at conventional levels.
As women's discomfort with talking about contraception with an enumerator may tend to introduce noise, we consult a related outcome measure that is slightly less personal: whether a woman wants to bear additional children (column 3). When we consider a one year lag of income, we find that a typical, 36 percent reduction in income decrease the reported desire to have additional children by 0.09 × ln(1.36) = 0.028 percentage points-a sizeable reduction compared to the variable's mean of 0.43. This is largely sustained a year later, when it declines only slightly to a 0.078 × ln(1.36) = 0.024 percentage point reduction.
Next, we find a small but precisely estimated decrease in the incidence of pregnancy following a reduction in income. We estimate a 0.023 × ln(1.36) = 0.007 percentage point 21 The contraception variable is missing for women who are already pregnant.
reduction in the probability of being pregnant when using income lagged by one year, which grows to a larger 0.040 × ln(1.36) = 0.012 percentage points when lagging by two years.
As the mean (standard deviation) of our pregnancy dummy is only 0.065 (0.247), this is a sizeable decrease. The results suggest that women avoid pregnancy in response to bad economic conditions. With fewer children and without the demands of a pregnancy, parents can conceivably invest more in the children they have. Thus, changes in fertility patterns may help blunt the negative effects of reductions in household income on very young children.

Robustness
We carry out a number of robustness checks to assess the validity and sensitivity of those results which were statistically significant for at least one lag structure. These include a check for pre-trends, tightening the requirements to be considered an agriculture-dependent household (thus reducing the sample size), and omitting various observations or trimming variables. In this section, we describe each of these in turn.

Check for pre-trends
A within-sample check for pre-trends, shown in right-hand-side income variable), and logged actual total income. 23 23 In these cross-sectional regressions, we control for the full set of controls in our 2SLS regressions. 24 24 We find few statistically significant coefficients-no more than we would expect by random chance. We conclude that our results are not likely due to pre-trends.

Sample inclusion thresholds
In order to appear in our sample, we require only that households have at least some of their total household income coming from agriculture. Our aim is to capture households that are at least somewhat reliant on the rural sector-and who likely rely on a diverse array of income sources. As this is an arbitrary definition of dependence on agriculture,

Sensitivity analysis
Next, Table A12 A12 considers the sensitivity of our results to four separate specifications: a) omitting 2004-2008 data, b) omitting Bishkek and Osh, which contain the two largest cities, c) trimming the top 1 percent of observations of both income and our outcome variables, and d) using the level (instead of log) form of income. All of these analyses, except d), yield smaller datasets and thus less statistical power to detect effects. However, to the extent that our results hold up, they can increase our confidence in the validity of our findings. Omitting 23 The IV is constructed using as its initial year the first year of the second half period-following the same method as our original instrument. In other words, we pretend data in the first half do not exist. Our stage two right-hand-side variable has also been predicted using the new instrument.

Conclusion
This study provides causal evidence from agriculture-dependent households in Kyrgyzstan that fluctuations in household income can have modest but statistically significant effects on children's long-term health and nutrition status and the BMIs and incidence of overweight in adults. It also provides evidence of several channels possibly explaining these impacts.
Our evidence comes from a study of agriculture-dependent households in a nationally representative, 13 year rolling panel dataset spanning 2004-2016. We address the endogeneity of income to child health, consumption, and other decisions using an instrumental variables approach; specifically, we instrument for household income with predicted income, obtained using the household's initial period share of income from different sources and aggregate growth rates over time in each source. We find that young children (age 0-5) exposed to negative income shocks experienced reductions in height and height-for-age Z-scores that were largest for girls and children under age two. These groups additionally experienced increases in stunting. Reduced consumption of healthy foods, reduced dietary diversity, and less expenditure on health may help explain the results. A channel possibly offsetting negative impacts is a decrease in fertility. At the same time, older children and adults saw decreases in BMI and-for adults-decreases in the incidence of overweight.
Our consumption data were household-level and thus mask important intra-household decision-making regarding how to respond to reductions in household income. More analysis is needed that uses better quality health and nutrition data, such as that captured by 24 hour food diaries, to understand changes in consumption patterns within the home following reductions in income. While our analysis of sex-and age-disaggregated data on anthropometric outcomes and subjective health reports are helpful in assessing the impacts of this process of intra-household decision-making on different members, such data would be helpful to better understand the precise changes in consumption taking place within the household.
Another channel worth further exploring is how these fluctuations in income influence migration patterns for women and men, as well as the household structure itself. Departure of some members and shifts in the time and labor burdens of other members may themselves have profound impacts on child health, and may also be spurred by income fluctuations.
Our findings provide both good news and bad news for the double burden of malnutrition.
While reductions in income, which are ubiquitous in developing country settings and against which households generally under-insure, contribute to under-nutrition in young children, they also reduced over-nutrition in older children and adults. They do so both by decreasing the diversity of diets, leading households to consume less of healthy foods, and reducing overall food consumption. While overall reductions in consumption may be helpful for the problem of over-nutrition in older children and adults, poorer-quality diets combined with lower consumption appear to be contributing to under-nutrition in young children. This suggests the need for public health officials and practitioners in development to respond to fluctuations in household income with tailored solutions that can reduce under-nutrition without simultaneously increasing over-nutrition.         We include in our sample only those households earning at least some income from agriculture. For individuals, we show the first stage for the height outcome. For households, we show the first stage for the household dietary diversity score outcome. The dependent variable is logged total income. Income is measured in 2010 Som. Our instrumental variable is logged predicted total income. Predicted total income is constructed from eight sources of income and cost by multiplying the household's initial period level with the growth rate of the average level in an oblast-rural/ urban combo for each income/ cost component, excluding one's own household. The basic control set includes year, oblast, and urban fixed effects, initial period level of income or cost (logged), the initial period value of household income (logged), and interactions of each of the latter two with a linear time trend. They also include several very basic individual-level controls (present in all regression specifications that include individual-level outcomes): a quadratic in age and a dummy for being male. Our full control set additionally includes several individual level controls: dummies for relationship with the household head, being married, and having a general secondary degree or higher (these latter two controls are omitted when we consider children age 1 to 5). It also includes several household-level controls: the number of unique agricultural goods produced, logged land farmed (with zero land imputed to 0.1 square meters of land), dummies for household size, a quadratic in age for the household head, and dummies for whether the head has a general secondary degree or higher, is married, and is male. All household-level controls are taken from the year in which the household enters the sample. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10.  191 7,191 7,191 7,191 7,191 7,191 191 7,191 7,191 7,191 7,191 7,191 Source: Authors' calculations based on KIHS 2004-2016. Notes: We include in our sample only those households earning at least some income from agriculture. Income (annual) is measured in 2010 Som. We use the logged form of income and its instrument in the regressions. HAZ is the child's height-for-age Z-score computed using WHO 2006standards (World Health Organization World Health Organization, 2006. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10.  191 7,191 7,191 7,191 7,191 7,191 191 7,191 7,191 7,191 7,191 7,191 Source: Authors' calculations based on KIHS 2004-2016. Notes: We include in our sample only those households earning at least some income from agriculture. Income (annual) is measured in 2010 Som. We use the logged form of income and its instrument in the regressions. WAZ and WHZ are the child's weight-for-age and weight-for-height Z-scores respectively, computed using WHO 2006standards (World Health Organization World Health Organization, 2006. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10.   We include in our sample only those households earning at least some income from agriculture. Overweight is defined as having a BMI of at least 25. Obesity is defined as having a BMI of at least 30. Income (annual) is measured in 2010 Som. We use the logged form of income and its instrument in the regressions. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10. We include in our sample only those households earning at least some income from agriculture. Measures are self-reported subjective well-being. The respondent was given 5 choices to evaluate the "state of his/her health:" very poor, poor, satisfactory, good, and very good. Income (annual) is measured in 2010 Som. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10. We include in our sample only those households earning at least some income from agriculture. Data in our regressions are annual, but based on quarterly observations. Consumption dummies (Panel B) takes on a value of 1 only if the food category is consumed in all four quarters. Logged consumption (Panel C) is averaged over the four quarters to get annual outcomes. The household dietary diversity score (HDDS) is constructed by counting the number of the 12 total food categories that have been consumed in the last 2 weeks during each of the four quarterly visits. A "healthy" HDDS is constructed similarly by counting the number of categories a household consumes from: fruits, pulses/ legumes/ nuts, vegetables, and fish/ seafood. In Panel B, the category of dairy products is further sub-divided into milk products (such as milk, cream, or kefir, measured in liters) and cheese products (such as cheese, curds, butter, sour cream, or yogurt, measured in kg) for the purposes of understanding the dairy category better, but these two subcategories are not considered when constructing the HDDS. Panel A and B outcomes are missing in 2007 because 2007 data were annual while other years were quarterly. This does not affect the consumption level used in the panel C outcomes, but it does make it impossible to accurately compare 2007 with other years for the outcomes of panels A and B. Income (annual) is measured in 2010 Som. We use the logged form of income and its instrument in the regressions. The instrumental variable is described in Table 2 2. We include all control variables described in Table 2 2 except for individual characteristics. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10. We include in our sample only those households earning at least some income from agriculture. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10. We include in our sample only those households earning at least some income from agriculture. The universe for all of these outcomes is restricted to women 15-49 who are either married or otherwise sexually active. The universe for the contraception outcome additionally excludes women who are pregnant. The instrumental variable and the full set of controls are described in Table 2 2. Additionally, all regressions control for the number of children a woman already had. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10. We divide all years an individual/a household in our sample into two halves of roughly equal length. We calculate average level of health-related variables in the first period . We take the first year of the second half as the initial year and re-construct the instrumental variable and predict the income used as the second stage right-hand-side income variable. We compute the average income using three variants of income measures, income instrument, predicted income (second stage right-hand-side variable), and the endogenous income in Panels A, B, and C. All regressions in pre-trends analysis include the full set of controls in individual-level regressions in Tables  2 2 except     We include in our sample only those households earning at least some income from agriculture. Income (annual) is measured in 2010 Som. We use the logged form of income and its instrument in the regressions. Dummy -leaving the sample in the next year is a household-level variable coded 1 in the last year that the household stayed in the sample for analysis. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10. Age at end of year t − 1 0.001 16.5 months pre-pregnancy 4.5 months pre-pregnancy 7.5 months in utero 1 15.5 months pre-pregnancy 3.5 months pre-pregnancy 8.5 months in utero 2 14.5 months pre-pregnancy 2.5 months pre-pregnancy 0.5 months old 3 13.5 months pre-pregnancy 1.5 months pre-pregnancy 1.5 months old 4 12.5 months pre-pregnancy 0.5 months pre-pregnancy 2.5 months old 5 11.5 months pre-pregnancy 0.5 months in utero 3.5 months old 6 10.5 months pre-pregnancy 1.5 months in utero 4.5 months old 7 9.5 months pre-pregnancy 2.5 months in utero 5.5 months old 8 8.5 months pre-pregnancy 3.5 months in utero 6.5 months old 9 7.5 months pre-pregnancy 4.5 months in utero 7.5 months old 10 6.5 months pre-pregnancy 5.5 months in utero 8.5 months old 11 5.5 months pre-pregnancy 6.5 months in utero 9.5 months old 12 4.5 months pre-pregnancy 7.5 months in utero 10.5 months old 13 3.5 months pre-pregnancy 8.5 months in utero 11.5 months old 14 2.5 months pre-pregnancy 0.5 months old 12.5 months old 15 1.5 months pre-pregnancy  We include in our sample only those households earning at least some income from agriculture. Income (annual) is measured in 2010 Som. We use the logged form of income and its instrument in the regressions. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10.  We include in our sample only those households earning at least some income from agriculture. Income (annual) is measured in 2010 Som. We use the logged form of income and its instrument in the regressions. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10.  We include in our sample only those households earning at least some income from agriculture. Income (annual) is measured in 2010 Som. We use the logged form of income and its instrument in the regressions. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10.   Notes: Agricultural income as a share of total income is calculated in the year that the household enters the sample. Income is measured in 2010 Som. We use the logged form of income and its instrument in the regressions. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10.  Notes: Agricultural income as a share of total income is calculated in the year that the household enters the sample. Income is measured in 2010 Som. In most regressions, we use the logged form of income and its instrument in the regressions. Note that in (5) we use the level form of the income variables in 100,000 2010 Soms for better presentation of the coefficients. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10. We include in our sample only those households earning at least some income from agriculture. Income (annual) is measured in 2010 Som. We use the logged form of income and its instrument in the regressions. The instrumental variable and the full set of controls are described in Table 2 2. Standard errors are in parentheses and clustered at the household level. *** indicates p<0.01; ** indicates p<0.05; and * indicates p<0.10.