Household food consumption

Our primary source of household consumption data comes from the Nielsen HomeScan household panel for the years 2005–2013. The Nielsen HomeScan data are a nationally representative sample of household food purchasing behavior. Participants in the sample are asked to scan their daily purchases from supermarkets, convenience stores, mass merchandisers, club stores, and drug stores using at-home scanner technology. These data include household food purchasing information for approximately 40,000 households for 2005–2006 and over 60,000 households for 2007–2013. The data are drawn from over 20,000 zip codes across the USA.^{Footnote 2} The panel nature of these data allows us to track the purchases of households over time. The Nielsen data include rich information on demographics such as household structure, income, education level, age, and race. This information is updated on an annual basis so we are able to control for time-varying demographic effects. The most important feature of these data is that they record the purchase information, including quantity, price, and product characteristics at Universal Product Code (UPC) level, which we use to create our measures of diet quality.^{Footnote 3}

While these data provide a number of benefits, they also have limitations. Although Nielsen eliminates households that report only a small fraction of their expenditures, the dataset still contains some households who did not remain in the panel for a reasonable length of time or report inconsistent or unreasonable purchases (Dubé et al. Reference Dubé, Hitsch and Rossi2018). To obtain more reliable data, we apply several filters to the raw data. First, we drop households that are in the data for less than six months; second, we drop households that do not spend at least $125 on food per quarter; third, we only keep households with a ratio of quarterly food consumption to family income greater than 0.1% and less than 200%; and finally, we exclude households with income less than 130% of poverty threshold because some income-support programs, such as Supplemental Nutrition Assistance Program (SNAP), expanded during the Great Recession, which was a time of rapidly falling house prices.

House prices

Our second data set is zip-code-level house price indices from Zillow. Zillow constructs these indices based on estimates from micro-level hedonic models using price data collected from county assessors, real-estate agencies, and self reports. Zillow house price indices have better geographic coverage than other publicly available data, such as Federal Housing Finance Agency or Standard & Poor’s Case-Shiller index, although all of these indices have similar time-series trends. In addition, Zillow computes median sales price indices. Stroebel and Vavra (Reference Stroebel and Vavra2019) indicate that median sales price indices are more accurate compared to standard repeat-sales indices at geographically disaggregated levels such as zip codes, which have limited numbers of housing transactions. Therefore, Zillow house price indices are widely used in the economics and finance literature. We use their quarterly data at the zip-code level from 2005 to 2013.

Variable construction

Outcome variables

To test the hypotheses posited in this article, we construct two main outcome variables – a variable for total food expenditure and a variable for diet quality. Measuring total expenditure is straightforward. The main challenge is how to measure diet quality properly.

Following previous literature, we construct three different measures for diet quality based on household-by-quarter expenditure shares. We begin by grouping 600 broad Nielsen food categories into 23 food groups based on the methods in Volpe et al. (Reference Volpe, Okrent and Leibtag2013). From these 23 groups, we further divide them into healthful and unhealthful food classes based on the USDA Dietary Guidelines for Americans (DGA) and the Quarterly Food-at-Home Price Database (Table 1). Using these groupings, we generate our first two measures of diet quality – expenditure shares on fresh fruits and vegetables and expenditure shares on all healthful food – using raw expenditure share data for each household and quarter.

Table 1. USDA Food Categorization

Note: The share of each food category is the real expenditure share for the average household in the Nielsen sample over our study period. The recommended expenditure share is from a representative family according to the liberal food plan specified by USDA (Volpe et al. Reference Volpe, Okrent and Leibtag2013). The representative family includes one male and one female, age 19–50, one child age 9–11, and one child age 6–8. The Nielsen HomeSacn sample is not representative for the US population. Comparing to the US population, our data include more households with higher age and living in metropolitan areas.

For our final diet quality measure, we use USDAScore, which is an index measure based on raw household expenditure outcomes and USDA dietary recommendations. It was introduced to the literature by Volpe et al. (Reference Volpe, Okrent and Leibtag2013) and has been widely used in recent studies (Chen et al. Reference Chen, Jaenicke and Volpe2016; Freedman and Kuhns Reference Freedman and Kuhns2018; Allcott et al. Reference Allcott, Diamond, Dubé, Handbury, Rahkovsky and Schnell2019). USDAScore measures the extent to which a household’s expenditure on a set of broad food categories deviates from recommendations from USDA’s Center for Nutrition Policy and Promotion (CNPP).^{Footnote 4} Specifically, the USDAScore for household h in time t is calculated as follows:

$$\eqalign{USDAscor{e_{ht}} = \left[ {\sum\limits_{j \in {J_{Healthful}}} {{{(S{h_{jht}} - Sh_{jh}^{CNPP})}^2}} |S{h_{jht}} \lt Sh_{jh}^{CNPP}} \right. \cr {\left. { + \sum\limits_{j \in {J_{Unhealthful}}} {{{(S{h_{jht}} - Sh_{jh}^{CNPP})}^2}} |S{h_{jht}} \gt Sh_{jh}^{CNPP}} \right]^{ - 1}}, \cr} $$

where j represents the CNPP food categories, $S{h_{jht}}$ denotes the percent of household h’s food expenditures in quarter t on products in category j, and $Sh_{jh}^{CNPP}$ is the expenditure share of category j that the CNPP recommends for household h.

To construct the recommended expenditure shares for each household, we need to convert the individual-level recommended shares to household-level shares. As the total expenditure for adults and children are different, we cannot treat them equally when we combine the recommended expenditures. Following Allcott et al. (Reference Allcott, Diamond, Dubé, Handbury, Rahkovsky and Schnell2019), we assign a larger weight to adults and a smaller weight to children using the OECD equivalence scale. The weight for an adult is ${w_{adult}} = {{{{1 + ({n_{adult}} - 1)*0.5} \over {{n_{adult}}}}} \over {1 + ({n_{adult}} - 1)*0.5 + {n_{children}}*0.3}}$ and for a child it is ${w_{child}} = {1 \over {({n_{adult}} - 1)*0.5 + {n_{children}}*0.3}}$ , where ${n_{adult}}$ is the number of adults and ${n_{adult}}$ is the number of children in a given household. The recommended family expenditure shares are $Sh_{jh}^{CNPP} = \sum {w_i}Sh_{ijh}^{CNPP}$ , where i represents a member of household h. Each household has a recommended expenditure share of food group j based on the demographic structure. This measure penalizes households for purchasing food which is different from the guidelines, while it does not penalize them for purchasing more healthful food. It also values food diversity as the diet quality decreases if households consume only from a select group of healthful food products. Therefore, USDAScore emphasizes the food structure and provides a more complete picture for diet quality in comparison to raw expenditure share measures.

All of our measures of diet quality are based on expenditure shares from a select group of food categories. This has both advantages and disadvantages. First, as Nielsen HomeScan data record the exact purchased products and their prices, we have accurate calculations for expenditure shares for different food groups. Second, as the expenditure shares of food groups, such as fruits and vegetables, are widely used it is easy to compare our results with other studies. One of the major weaknesses, however, is that our measures may be affected by prices. For example, our measures would suggest that diet quality increases if the prices for fruits and vegetables increase and food purchases stay the same. We deal with this issue later in the paper.

Summary statistics

Table 1 lists the 23 food groups used in constructing our household-level diet variables. In this table, we show the expenditure share of average households in our sample and the recommended share for a representative household.^{Footnote 5} As is shown, the expenditure shares in the healthful food groups – dark vegetables, fruits, canned and dry beans, lentils, peas, and white meat – are consistently below the amounts recommended by USDA, while households consume significantly more than is recommended from the unhealthful categories – soft drinks, sodas, fruit drinks, sugars, sweets, candies, and frozen or refrigerated dinners. Overall, the total expenditure on healthful food is 35%, while the recommended share is 82%.

Table 2 presents summary statistics for the variables used in our econometric analysis. The sample we use in this study comprises 887,183 household-quarter observations. Panel A shows the descriptive statistics for our outcome variables; Panel B provides summary statistics for our control variables; and Panel C provides summary information for our instrumental variables.

Table 2. Summary Statistics

Note: The above summary is based on the quarterly data we organized from Nielsen HomeScan household panel for the years 2005–2013. The Nielsen HomeSacn data are not representative for the US population. Comparing to US population, our data include more households with higher age and living in metropolitan areas.

Baseline model

To identify the impact of house price shocks on food expenditure and dietary outcomes, we begin with the following baseline linear specification:

(1) $${Y_{hzq}} = {\beta _0} + {\beta _1}lnH{P_{zq}} + {X_{hq}}\delta + {\lambda _h} + {\mu _q} + {\varepsilon _{hzq}},$$

where ${Y_{hzq}}$ represents our outcome variab le of interest for household $h$ , in zip code $z$ , in quarter $q$ ; $lnH{P_{zq}}$ is the natural log of house prices at the zip code level in quarter $q$ ; ${X_{hq}}$ is a vector of control variables; ${\lambda _h}$ is a household fixed effect; ${\mu _q}$ is quarterly fixed effect; and ${\varepsilon _{hzq}}$ is an error term. The control variable $Xhq$ includes family income, household size, dummies indicating a household head’s education level, household head’s age, a dummy indicating the household head’s marital status, a dummy indicating the presence of children, and dummies for race. To ensure that our results are not driven by households migrating between zip codes, we consider the same household living in different zip codes in different periods as separate observations. Therefore, we track diet changes for the same household in the same zip code using household fixed effects ${\lambda _q}$ . In this way, we are able to avoid selection bias as households who value health may move to more expensive communities with a better food environment. Finally, we control for quarter fixed effects ${\mu _q}$ to capture time-varying macro shocks. ${\beta _1}$ , which is our coefficient of interest, represents the elasticity of dietary outcomes with respect to house prices.

Owners vs. renters model

Theory suggests that changes in house prices may impact expenditure on other goods and services through wealth or collateral channels – changes in asset prices (housing) relax borrowing and budget constraints and change expenditure on other goods as a result of increased income. This result, however, should only hold for asset owners, or homeowners in our case. Thus, we do not expect a positive relationship to exist between house prices and food expenditure for renters. To exploit these differential responses, we estimate the following model:

(2) $${Y_{hzq}} = {\beta _0} + {\beta _1}lnH{P_{zq}}*ow{n_{hz}} + {\beta _2}lnH{P_{zq}} + {X_{hq}}\delta + {\lambda _h} + {\mu _q} + {\varepsilon _{hzq}},$$

where $ow{n_{hz}}$ represents the homeownership status for household $h$ in zip code $z$ , ${\beta _2}$ shows how changes in house prices affect the dietary outcomes of renters, and ${\beta _1}$ captures the differential effect of house prices on homeowners. We expect ${\beta _1}$ to be positive based on the theory discussed above. As our data do not provide direct information on the homeownership status of each household, we use two distinct strategies to identify homeowners. First, we use homeownership rates at the zip-code level as a proxy for homeownership. In this setting, we exploit variation in homeownership rates across zip codes – i.e., we expect house price changes to have a differential effect for households living in areas with different rates of homeownership. Second, we use the house-type variable (defined as single-family or multi-family residences) in the Nielsen data and follow Stroebel and Vavra (Reference Stroebel and Vavra2019) to determine whether a household is a renter or owner. Specifically, we identify households living in one-family and non-condo residences as homeowners and families living in 3+ family residences as renters.

Instrumental variable model

The estimates from Equations (1) and (2) are unbiased if house prices are exogenous. Here, endogeneity may arise for several reasons. First, common factors such as financial liberalization and expectation over future income are likely to drive up house prices and food consumption simultaneously (Campbell and Cocco Reference Campbell and Cocco2007). Conversely, omitted variables on time allocated to shopping are likely to be negatively correlated with house prices but have a positive effect on food spending. To address these issues, we extend our baseline and homeowner-renter specifications and estimate an instrumental variable (IV) model.

Following Mian et al. (Reference Mian, Rao and Sufi2013), Dettling and Kearney (Reference Dettling and Kearney2014), Chetty et al. (Reference Chetty, Sándor and Szeidl2017), Aladangady (Reference Aladangady2017), and Guren et al. (Reference Guren, McKay, Nakamura and Steinsson2021), we use the interaction of MSA-level housing supply elasticities and national house price trends as an instrument for house prices in Equations (1) and (2). The housing supply elasticity is from Saiz (Reference Saiz2010), and the variance of the housing supply elasticity contains two parts. The first part comes from exogenous geographic features, such as water, oceans, mountains, and wetlands, which are used to measure unavailability of the land. The second part is take from the Wharton Residential Urban Land Regulation Index created by Gyourko et al. (Reference Gyourko, Saiz and Summers2008). The index is used to capture the stringency of residential growth controls. As we use fixed-effect panel data, we build a shift-share instrumental variable by interacting the housing supply elasticities with national house prices. The intuition is that when there are macro-level positive demand shocks to housing, high housing supply elasticity areas, such as Texas, Arizona, and Georgia, experience less of a house-price shock compared to low housing supply elasticity areas, such as California and Massachusetts.

This instrumental variable is not without its limitations. Davidoff (Reference Davidoff2015) points out that housing supply elasticities can be related to other city characteristics and therefore are correlated with long-run growth in demand. Thus, it will pose a problem for cross-sectional analysis. However, our study employs the fixed-effect model using panel data and can overcome this issue. In addtion, Guren et al. (Reference Guren, McKay, Nakamura and Steinsson2021) show that the housing supply elasticity from Saiz (Reference Saiz2010) may lose power for the data before 2000. Our study avoids this time period.

We estimate our IV model using two-stage least squares (2SLS). The first stage of our IV model is specified as follows:

(3) $$lnH{P_{zq}} = {\alpha _0} + {\alpha _1}Elasticit{y_{MSA}}*lnH{P_q} + {X_{hq}}{\delta _1} + {\lambda _h} + {\mu _q} + {e_{hzq}},$$

where $Elasticit{y_{MSA}}$ represents the housing supply elasticity at the MSA level and $lnH{P_q}$ is our national house price index in quarter $q$ . After estimating equation (3), we obtain predicted values of our endogenous house-price variable and estimate the following second-stage model:

(4) $${Y_{hzq}} = {\beta _0} + {\beta _1} {lnH{P_{zq}}} + {X_{hq}}\delta + {\lambda _h} + {\mu _q} + {\varepsilon _{hzq}},$$

where ${lnH{P_{zq}}}$ is the predicted value for $lnH{P_{zq}}$ from the first stage.

To handle price endogeneity in Equation (2), we need two instruments – one for house prices and the other one for the interaction between house prices and homeownership. We follow Dettling and Kearney (Reference Dettling and Kearney2014) and add a second IV to our model by using the interaction of housing supply elasticity, national house prices, and our indicator for homeownership. The first stage of our homeowner–renter IV model consists of two linear regressions given by:

(5) $$\matrix{ {lnH{P_{zq}} = {\alpha _0} + {\alpha _1}Elasticit{y_{MSA}}*lnH{P_q} + {\alpha _2}Elasticit{y_{MSA}}*lnH{P_q}*ow{n_{hz}}} \hfill \cr {\quad \quad \quad \quad + {X_{hq}}{\delta _1} + {\lambda _h} + {\mu _q} + {e_{hzq}}} \hfill \cr } $$

and

(6) $$\matrix{ {lnH{P_{zq}}*ow{n_{hz}} = {\gamma _0} + {\gamma _1}Elasticit{y_{MSA}}*lnH{P_q} + {\gamma _2}Elasticit{y_{MSA}}*lnH{P_q}*ow{n_{hz}}} \hfill \cr {\quad \quad \quad \quad \quad \quad \quad + {X_{hq}}{\delta _2} + {\lambda _h} + {\mu _q} + {\upsilon _{hzq}},} \hfill \cr } $$

where $lnH{P_{zq}}$ and $lnH{P_{zq}}*ow{n_{hz}}$ are both endogenous variables with instruments given by $Elasticit{y_{MSA}}*lnH{P_q}$ and $Elasticit{y_{MSA}}*lnH{P_q}*ow{n_{hz}}$ .

The second stage of our homeowner–renter IV model is given by

(7) $$\matrix{ {{Y_{hzq}} = {\beta _0} + {\beta _1}lnH{P_{zq}}*ow{n_{hz}} + {\beta _2}lnH{P_{zq}}} \hfill \cr {\quad \;\; + {X_{hq}}\delta + {\lambda _h} + {\mu _q} + {\varepsilon _{hzq}},} \hfill \cr } $$

where $ {lnH{P_{zq}}*ow{n_{hz}}}$ and $ {lnH{P_{zq}}}$ are predicted values for $lnH{P_{zq}}*ow{n_{hz}}$ and $lnH{P_{zq}}$ from equations (5) and (6).

Income

In this part, we examine the effect of house prices on dietary outcomes for different income groups. Kozlova (Reference Kozlova2016) splits the sample into three different income groups: below 130%, above 130% and below 200%, and above 200% of the poverty threshold. These cutoffs have important policy implications. For example, households qualify for the Supplemental Nutrition Assistance Program (SNAP) if their income is below 130% of poverty threshold. Households economically struggle once their income lies below 200% of the poverty threshold. In this study, we divide the sample into income groups below and above 200% of the poverty threshold as we have excluded the observations with income below 130%. As lower income households are more likely to be budget constrained, we expect that house prices will have a larger impact on them compared to households with higher incomes.

Panels A and B of Table 12 present estimates for households below and above 200% of the poverty threshold, respectively. Column (1) reports the effects of house prices on food expenditure. From Panel A, we find that house prices have a significant effect on food expenditure for homeowners and insignificant effect on renters. The food expenditure elasticity is around 0.135 for lower income homeowners. The estimates from Panel B suggest that the food expenditure elasticity of house prices is around 0.126 for homeowners whose incomes lies above 200% of the poverty line. The effects for renters are insignificant for both income groups, which is consistent with previous results. Column (2)–(4) of the table show the effects of house price shocks on diet quality. The effects are not significant in most regressions, although we do find a positive effect for lower income households for the expenditures on of fruits and vegetables.

Table 12. Results for Income Heterogeneity Model

Note: This table presents results for our IV homeowner–renter model for different income groups. We identify homeowners based on the house type. Panel A shows estimates for the sample under the 200% poverty threshold and panel B shows results for households greater than 200% of the poverty line. We identify homeowners based on the house type. Robust standard errors are in parentheses. *Significant at 10% level. **Significant at 5% level. ***Significant at 1% level.

Age

In this part, we divide the sample into two groups according to the age of household heads: above and below 50. ^{Footnote 8} Differential consumption reactions to house-price shocks between younger and older households would suggest the major mechanism by which house prices affect consumption. For example, if the effect is larger for older homeowners, it would suggest that the wealth effect may play a larger role than the collateral effect. This is because older homeowners have strong incentives to sell their large houses and move to small ones with the family size decreasing (Campbell and Cocco Reference Campbell and Cocco2007). If younger homeowners are more affected by house price changes, then the collateral effect is probably the main driver since younger households are more likely to face borrowing constraints. Thus, relaxed borrowing constraints, driven by increasing house prices, would stimulate more spending for younger homeowners as compared to older homeowners.

Table 13 reports our results for different age groups. Column (1) shows that house prices have a positive and significant effect for both younger and older homeowners, while the effects for renters are not significant. The food expenditure elasticity of house prices is 0.135 for homeowners with the household head under 50, and the elasticity is 0.097 for homeowners with the household head over 50. Again, we do not find any evidence that house prices have a significant impact on diet quality in any age group.

Table 13. Results for Age Heterogeneity Model

Note: This table presents results for our IV homeowner-renter model for different age groups based on the household head. We identify homeowners based on the house type. Panel A is for the younger households defined as a household head with age below 50 and Panel B is for the older households with household head age above 50. We identify homeowners based on the house type. Robust standard errors are in parentheses and they are clustered at the zip code level. *Significant at 10% level. **Significant at 5% level. ***Significant at 1% level.

Taken together, we find that the effect of house prices on food expenditure is larger in lower income and younger groups. This suggests that house prices affect food expenditure mainly through the collateral effect rather than the wealth effect as lower income and younger households are more likely to be in borrowing constraints. Aladangady (Reference Aladangady2017) and Berger et al. (Reference Berger, Guerrieri, Lorenzoni and Vavra2018) also find similar results. In addition, we show that house prices do not affect diet quality even in those groups whose food expenditures experience the largest change. This further confirms our conclusion that relaxing budget constraints is not effective in improving diet quality.