Developing a consumption measure, with examples of use for poverty and inequality analysis: a new research product from BLS

Although consumption may be a better measure of economic well-being than income, determining the actual consumption level for an individual or a group of individuals—a person or family that forms a consumer unit (CU), for example—is difficult because it depends on the individual’s or group’s particular circumstances, choices, use of purchases, and time usage.⁶ In addition, data limitations make the problem of valuing consumption even more difficult. For example, combining expenditures data with time-use data has been particularly challenging.⁷ As a result, many researchers have used expenditures as a proxy for consumption.⁸ The U.S. Bureau of Labor Statistics (BLS) Consumer Expenditure Surveys (CE) have been the primary source of expenditures data at the CU level for researchers and government agencies since the late 19th century. And, as such, the CE is the oldest BLS product that collects household or consumer expenditures as a measure of living conditions for the United States.⁹

BLS has long been interested in the creation of a measure of consumption with CE data as the base,¹⁰ as have other researchers.¹¹ For example, Fisher, Johnson, and Smeeding (2015)¹² produced a measure of consumption to study inequality, and Meyer and Sullivan (2012, 2013)¹³ produced a consumption measure to study poverty and inequality. However, recent experiences with the COVID-19 pandemic have highlighted the need for a more comprehensive consumption measure than has previously been produced. Throughout the pandemic, family and household members have played an increasingly important role in the well-being of other members of their households through the provision of services such as childcare, eldercare, more home-cooked meals, and education, all of which are nonmarket transactions and not captured with expenditures alone. For these transactions, a comprehensive consumption measure needs to account for the time household members spend in home production for their own consumption. These shortcomings of expenditures as an economic measure of well-being have motivated researchers at BLS to develop a more comprehensive consumption measure that accounts for a broader set of in-kind benefits than accounted for in previous measures and for the value of home production for own consumption.

Our first attempt at BLS to produce a consumption measure was published in the May 2022 edition of AEA Papers and Proceedings as “Building a consumption poverty measure: initial results following recommendations of a federal interagency working group.”¹⁴ In this article, we build upon our earlier work by expanding what is included in consumption and refining our methods to produce the measure. The primary difference between the earlier consumption measure and the one presented in this article is that we introduce a value of health insurance; this makes our most recent consumption measure more like the ones produced by Meyer and Sullivan (2012, 2013). This enables us to produce three consumption measures: one that excludes health insurance, one that includes health insurance capped as a percentage of consumption, and one that includes health insurance not capped. The measure based on capping health insurance expenditures for poverty measurement is based on a recommendation made by the Interagency Technical Working Group on Evaluating Alternative Measures of Poverty (ITWG).¹⁵ As illustrations of how the measures can be used, we produce simple means for the consumption measures and components and inequality statistics for 2019, 2020, and 2021. The consumption-based poverty and inequality statistics are compared with statistics based on pretax income and CE-defined total expenditures.¹⁶

In our analysis, we find that the value of CU consumption without health insurance averaged about 16 percent less than CE total expenditures over the 3-year period, and that the values of CU consumption with health insurance (capped and uncapped) was about 14 percent more than CE total expenditures in 2019 and 2020, and about 12 percent more in 2021. For all consumption measures, compared with total expenditures and pretax income, poverty rates are lower when based on a relative concept. The absolute concept that we used for this study sets all poverty rates the same in 2019; such an approach allows us to see how poverty rates changed over the period while holding the base poverty rate the same. Our results show a noticeable decrease in the poverty rate from 2019 to 2021 when it is measured by consumption without health insurance. For all 3 years, consumption distributions are more equal than distributions of total expenditures and pretax income. Relative to 2019, we find that consumption poverty fell in 2020, with the onset of the pandemic, and that distributions of consumption became more equal. Compared with 2020, for all the consumption measures, poverty rose in 2021 when we used a relative measure, and inequality rose but not to the same levels as in 2019.

The consumption measures presented in this article represent work in progress. Next steps in the development of the BLS research consumption measure include the addition of the value of time for home production. During the 2021–23 period, BLS sponsored research to impute values of select home-production activities using the BLS American Time Use Survey and other data in combination with the CE. Once the value of home production is included, the goal of BLS to produce a more comprehensive measure of consumption can be realized. In addition, BLS plans to publish a consumption-measure research series based on internal CE data and imputation; this series will be available to the public through published tables of means. The series will not be an official BLS production series, but it can be referred to as an official research series. Depending upon available resources, BLS plans to release an auxiliary public-use data file that includes research-based consumption values.

Background and related literature

In standard economic models, individual utility is a function of consumption. In the life-cycle model, individuals choose the level of consumption in each period to maximize utility subject to lifetime income. The consumption level in a given period can be more or less than the income level in that period because individuals can use savings and borrowing to smooth consumption over their life cycle. The implication of the life-cycle model is that income, consumption, and savings will follow a predictable pattern over an individual’s life. Income is expected to exceed consumption during the prime working years and be less than consumption early in life and in retirement. Individuals can also use savings and borrowing to smooth consumption in response to unexpected fluctuations in income. This view of consumption was developed by Modigliani and Brumberg (1954) and extended by Friedman (1957).¹⁷ As noted by Jappelli and Pistaferri (2017), models of consumption developed in the 1970s through the 2000s further account for intertemporal choice under uncertainty, as well as the tradeoff between leisure and home production.¹⁸

Given its direct relationship with utility, consumption is often considered a better unidimensional measure of well-being compared with income or expenditures. In a framework focused on households, it is assumed that the members of the household collectively choose a level of consumption that maximizes utility for a given budget constraint defined by available resources. Consumption can be viewed as an outcome variable reflecting what has been achieved, with income acting as one of many inputs. This contrasts with economic well-being outcome measures that focus on what could be achieved, such as income, for example.

Many researchers have attempted to construct measures that reflect consumption. Early examples include Cutler and Katz (1991) and Slesnick (1991), who use a flow-of-services approach to value durables consumption.¹⁹ Later studies that examined household consumption as distinct from expenditures include Johnson, Smeeding, and Torrey (2005); Krueger and Perri (2006); Attanasio, Battistin, and Ichimura (2007); Heathcote, Perri, and Violante (2010); Aguiar and Bils (2011); Coibion, et al (2012), and Attanasio, Hurst, and Pistaferri (2012).²⁰ The studies from the economics literature most related to our attempt to construct a consumption measure at the household level are those by Meyer and Sullivan (2013, 2023) and Fisher, Johnson, and Smeeding (2015).²¹ (The next section provides a more detailed discussion of the similarities and differences between the Meyer-Sullivan and Fisher-Johnson-Smeeding measures and those presented in this article.)

There is a large body of literature produced by international and national organizations that provides guidance on the development of a consumption measure. At the international level, guidance has been provided by the International Labour Organization (ILO), the Organisation for Economic Co-operation and Development (OECD), the United Nations Economic Commission for Europe, and the World Bank.²² At the U.S. national level, the ITWG included in its 2021 report a comprehensive review of the literature on the development of consumption measures and their use in poverty measurement.²³ Then, in 2022, BLS published a report by Curtin et al. called “A conceptual framework for the U.S. Consumer Expenditure Surveys,” which includes a measure of consumption.²⁴

Data

In this section, we discuss the data and methods used to construct our consumption measures. We describe the methods used to produce the earlier consumption measure and those reflecting the addition of health insurance and improvements to our methodology. We also provide a discussion of the scope of our measures and then compare our measures with those produced by other researchers.

Consumer Expenditure Surveys and variables

The CE Interview Survey is the primary data source for the consumption measures presented in this article. The CE uses two survey instruments to collect expenditures at the CU level: the Diary Survey and the Interview Survey. Respondents are selected to participate in only one of the two surveys. The Diary Survey focuses on collecting expenditures for certain frequently purchased goods and services, such as food at home and food away from home, apparel, and other products and services purchased on a regular basis. In contrast, the Interview Survey collects expenditures more comprehensively, capturing approximately 95 percent of total expenditures, which is why we use it for our consumption measure.²⁵ The Interview Survey is conducted throughout the year on a rolling basis and has a 3-month reference period. CUs complete the Interview Survey up to four times at 3-month intervals.

For our analysis, we compare consumption with CE-defined total expenditures and a measure of income. To define income, we start with the CE definition of pretax income available in the microdata files and subtract the value of Supplemental Nutrition Assistance Program (SNAP) benefits.²⁶ This income measure matches the U.S. Census Bureau measure of money income used to produce the official U.S. poverty statistics.²⁷ Note that unemployment benefits are included in this definition, so the impact of changes to benefit generosity are captured in the absolute magnitude of benefits collected in the CE from year to year. However, benefits related to the COVID-19 pandemic, such as the Economic Impact Payments, the expanded Child Tax Credit, and other tax credits are not included in the income measure because they are income-tax adjustments; however, the use of these credits are likely to be reflected in our consumption measures.

Consumption measure

In defining the scope of the consumption measure, we generally follow the ILO and OECD guidelines and the consensus recommendations of the ITWG.²⁸ We deviate from the ILO and OECD guidelines, which include education and health insurance in final consumption, and instead follow the ITWG recommendations by excluding education but including values for health insurance. We also considered the research of others in the development of the consumption measures presented in this article.

In many countries, including the United States, household surveys are used to collect expenditure data that can be used, in part, to create a consumption measure. For most categories, consumption will equal expenditures in a given period. However, not all components of consumption satisfy this equality. For durable goods, the value of consumption is defined as a flow of services over the life of the product, rather than as the expenditures needed to acquire the good. International guidelines recommend that the flow of services from housing and vehicles be represented by, for example, rental equivalence or user cost. For our measures—the earlier measure and those presented in this article—we use rental equivalence for owner-occupied shelter as collected in the CE, and we impute elements of user costs for cars and trucks not collected in the CE (i.e., depreciation and the opportunity costs of capital).²⁹

International guidelines and the ITWG also recommend that the value of in-kind benefits be counted in consumption; however, the value of these benefits is often not collected in national household surveys. In the CE, the value of most in-kind benefits—those from the government (e.g., subsidized school meals and energy assistance) and from employers (e.g., health insurance)—are not collected. One exception is that SNAP benefits are collected in the CE. Because SNAP benefits are administered in the form of electronic benefit transfer (EBT), we assume that they are used like money income and are implicitly included in reported food expenditures. For other government and employer in-kind benefits, imputed values are included in our measure of consumption. Government in-kind benefits for food, energy, and rental assistance were included in the earlier BLS consumption measure; two of the consumption measures presented in this study include government and employer in-kind benefits in the form of health insurance. Although we do not include home production in the current version of the consumption measure, we plan to include it in future versions.

Expenditures in categories such as education and health can be better thought of as investments rather than as providing for current consumption. Following the recommendation of the ITWG,³⁰ in this and in our earlier consumption measure, we exclude education expenditures. We also exclude out-of-pocket spending on medical goods and services because the ITWG did not reach a consensus recommendation for the treatment of those expenditures. But health insurance provides current utility in the form of risk protection, unlike expenditures for medical goods and services, which provide utility indirectly by their effect on health. The inclusion of health insurance in a consumption measure for poverty analysis is controversial. Following the ITWG recommendation, and as presented in this article, we produce consumption measures with and without health insurance. But one concern with including health insurance in the measure for poverty analysis is that the value of government-provided health insurance can make up an outsized share of total consumption and almost singlehandedly push people out of poverty. To prevent overstating the effect of government-provided health insurance, we produce a consumption measure that restricts health insurance to be no more than half of total consumption with health insurance (uncapped) and use this measure in our poverty analysis. For consistency, we also use consumption with health insurance capped for our inequality analysis.

Several previous studies have created measures of consumption. Table 1 provides a comparison between the research consumption measures presented in this article and a selection of other measures. The measures most closely aligned with our measure are those created by Meyer and Sullivan (2012, 2013) and Fisher, Johnson, and Smeeding (2015).³¹ Although these measures are similar to ours, there are some important distinctions.³² With respect to the components of consumption, both Meyer and Sullivan and Fisher, Johnson, and Smeeding used rental equivalence for owner shelter but not for vacation homes.³³ We include rental equivalence for owner shelter and vacation homes in our earlier consumption measure and in the current consumption measure. Our measures and those of Meyer and Sullivan replace reported rent with imputed market rents for renters who report receiving government subsidies and for those who reside in public housing. Meyer and Sullivan assigned the value of rent as the maximum of the reported rent versus the imputed rent.³⁴ In contrast, we assign the imputed rent regardless of its size relative to the reported value. In addition, we impute rents for CUs living in rent-controlled units, those living rent free, and those living in college dormitories. Fisher, Johnson, and Smeeding imputed government rental and public-housing subsidies and added them to reported rents. For vehicles, Meyer and Sullivan estimated varying depreciation rates, but they did not account for opportunity cost.³⁵ Fisher, Johnson, and Smeeding accounted for opportunity costs but used a fixed rate (5 percent); in addition, they used a fixed rate of depreciation (10 percent).³⁶ Finally, Meyer and Sullivan did not include out-of-pocket health expenditures or education in their consumption measure. However, they included an imputed value of health insurance, which is capped at 30 percent of total consumption in their poverty analysis.³⁷ We take a similar approach in this article, but we cap health insurance at 50 percent of consumption.³⁸ In contrast, Fisher, Johnson, and Smeeding included out-of-pocket health and education expenditures in their consumption measure.³⁹

Imputations for consumption

In cases in which in-kind benefits are not collected as part of the CE, a value of in-kind benefits must be imputed. For some types of in-kind benefits, such as government rental assistance and health insurance (both employer provided and government), participation is captured in the CE; for these programs, we only need to assign a value to the benefits. For other in-kind benefits, such as the National School Lunch Program (NSLP), the Women Infants and Children (WIC) program, and the Low Income Home Energy Assistance Program (LIHEAP), the CE does not ask about participation. For these programs, participation and benefits must be imputed. Finally, in terms of the flow of services from housing and durable goods, as noted previously, the CE asks about rental equivalence for owner-occupied housing; however, rental equivalence is not asked for vehicles. Instead, we apply a user-cost approach to impute values of the service flows arising from vehicle (i.e., car and truck) ownership. For the consumption value of consumer units who are not homeowners but are identified as renters, select imputations are also produced. In this section, we provide a summary of the imputation methods. (See the appendix for a more detailed description of imputation methods.)

As previously noted, the CE does not collect program-participation or benefit values for NSLP, WIC, or LIHEAP. To impute benefit values to the CE, we start by modelling participation for these programs and LIHEAP benefits as reported in the Current Population Survey Annual Social and Economic Supplement (CPS ASEC). Once participation in these programs is imputed to CUs in the CE, a value of the in-kind benefits is assigned. This value comes from administrative sources for NSLP and WIC. For LIHEAP, the reported benefit amounts from the CPS ASEC are used to impute a value of the benefits. Imputations for NSLP and LIHEAP for the measures presented in this article are the same as those included in the earlier measure; however, improvements in methodology to produce WIC imputations are introduced.

In addition to owner-occupied housing, in certain instances it is expected that reported rents do not reflect the full consumption value of rented shelter. Regarding renters living in units for which we consider their reported rents are less than market values, imputations for the value of shelter consumption are needed. We consider renters who are paying additional rental-related expenses—for example, those for maintenance and repairs—as not reporting rents that represent their consumption. In addition, the CE includes questions about whether the renter receives government assistance in paying rent, lives in public housing, lives in rent-controlled units, or lives rent free. However, not asked is the value of the difference in what renters pay for shelter and the market value of similar units. We use regression models for those who pay full market value to calculate the implied market value for those who pay less than the full value. The imputed market values represent the consumption value of rental shelter for these renters. Improvements in the rent imputations are introduced with the current measure, and imputed rents for renters living rent-free were added.

The CE asks about different types of health insurance coverage and collects data on out-of-pocket insurance premiums associated with each type of insurance. The CE also collects out-of-pocket expenditures on medical goods and services. As noted previously, for one of the measures presented in this article, the consumption value of health insurance is included. To derive this value, we first impute the full value of health insurance on the basis of insurance type. For private health insurance, the full value is based on the market price. This includes the out-of-pocket premiums that are captured by the CE and the employer contribution for employer-provided plans, or the value of any subsidy received for individual plans. For public insurance programs, the full value is based on the cost to the government (including administrative costs).

Finally, for vehicles, we begin with the CE definition for all transportation expenditures, which includes the expenditures for the purchase of all new and used vehicles, vehicle finance charges, gasoline and motor oil, maintenance and repairs, vehicle insurance, public transportation, and vehicle rental licenses and other charges. We remove expenditures for the purchase of all vehicles and vehicle finance charges that were made during the 3-month reference period. These expenditures are replaced with a flow-of-services value for cars and trucks that is imputed using a user-cost approach. The user cost is defined as the depreciation plus the opportunity cost of capital (current estimated value of the car multiplied by an interest rate) plus maintenance and repair costs. To derive a flow of services from the ownership of vehicles, which we limit to cars and trucks, we include imputed values of vehicle depreciation and opportunity costs of capital. Other components of user costs are already accounted for in the transportation expenditures that we keep. We estimate vehicle depreciation using vehicle purchase information in the CE by comparing similar vehicles purchased at different ages. A consumption flow-of-service value is imputed to CUs for the stock of cars and trucks owned for nonbusiness use. Although the flow of services from cars and trucks was included in the earlier BLS consumption measure, improvements in the methodology have been introduced and are reflected in the current measure. Note that our user costs will undervalue the flow of services from the stock of vehicles that are not cars or trucks; these other vehicles include, for example, planes, boats with motors, and motorized campers. 

Going from expenditures to consumption

To help readers understand the relationship between consumption and total expenditures as defined by BLS and presented in its published tables, we present a step-by-step description. Because we are interested in the CU’s consumption level, and not the consumption among people who do not live in that CU, we first remove expenditures for goods and services purchased to be given to someone outside the CU; these are identified as “gifts.” Next, we remove expenditures related to home ownership and the purchase of vehicles because these values will be captured instead by the flow-of-services value for shelter and vehicles, respectively. And, as noted previously, we also remove education and noninsurance medical expenditures because we view these as being more investment than consumption. For the measure of consumption with health insurance (uncapped and capped), we add the imputed value of health insurance. Finally, we do not include certain expenditures that are included in CE total expenditures that can be better thought of as financing future consumption (e.g., retirement contributions, life insurance purchases, etc.).

Methods for analysis

In this section, we present the methods used to examine the impact of moving to a consumption measure from CE total expenditures. Then, we describe the methods used to examine the impact of using consumption as opposed to total expenditures or income for poverty and inequality analysis.

Means

We analyze five measures of expenditures and consumption: total expenditures as defined by the CE, total expenditures as defined by the CE excluding gift expenditures, consumption without health insurance, consumption with the value of health insurance capped, and consumption with health insurance uncapped. To calculate the averages, we pool four quarters of data starting with the second quarter of the current calendar year through the first quarter of the following year. Quarters refer to the calendar period when the CE data were collected; for example, quarter one includes expenditures collected during interviews that took place from January to March; the reference period for January interviews is October through December of the previous year, while March interviews reference expenditures made in December of the previous year through February of the current year. Thus, when we refer to “quarterly” expenditures, these are actually 3-month values. For example, to calculate the quarterly averages for 2019, we use data from the second quarter of 2019 through the first quarter of 2020. We present CU-level quarterly averages that are weighted using one-fourth of the CU weight included in the CE data file; CE data are weighted quarterly and thus scaling by one-fourth accounts for the fact that four quarters of data are pooled together. Annual means can be produced by taking the quarterly mean and multiplying by four, but these values will differ from the means reported in the published CE tables.⁴⁰

Poverty and inequality analysis

We present results for three consumption measures: one that does not include health insurance, a second that includes health insurance capped, and a third that includes health insurance uncapped. For comparison, we also present results for CE total expenditures and pretax income. The income measure that we use differs from the income measure that appears in BLS published tables with CE expenditures data.⁴¹ Unlike the pretax income measure in the published CE tables, the one in this study does not include SNAP benefits or food and rent as pay. For each measure, we create equivalized values using a three-parameter equivalence scale.⁴² Poverty and inequality statistics are based on equivalized values.

For studying poverty, we produce two sets of poverty rates. The first set is based on purely relative thresholds. These thresholds are defined to be 60 percent of the median equivalized value for each measure for each year. The second set of poverty rates is based on the implicit thresholds that result when all poverty rates are anchored to the same rate for a single measure; this is a type of absolute measure. We anchored the poverty rates for all the measures in 2019 to be the same as the 2019 relative poverty rate for consumption with health insurance capped. For 2020 and 2021, the implicit 2019 thresholds (based on equivalized-measure values) resulting from this anchoring are updated to account for inflation. To approximate inflation, we create an annual index as the average of the monthly Chained Consumer Price Index for All Urban Consumers for all items and all urban areas.⁴³ When the respective measure for the CU is below the threshold, the CU is considered poor, and all members of that CU are considered poor. For example, if the CU’s equivalized consumption without health insurance is below the poverty threshold for consumption without health insurance, then the CU and everyone in it are considered poor. The poverty rates that we show refer to the percentage of people in the United States below the thresholds.

For studying inequality, we produce Gini indexes and Lorenz curves for each measure for each year. CUs are ranked on the basis of their equivalized value of each measure, and population weights are used to produce the distributions. For each measure, the Lorenz curve is a plot of the cumulative share of the population plotted against cumulative share of each measure. For example, 60 percent of the population accounts for 40 percent of overall consumption without health insurance.

Results

The results are calculated for 2019, 2020, and 2021 and are broken into three parts. The first section presents a discussion of the means. The second and third sections present the results of the poverty and inequality analysis, respectively. Results for 2021 are labeled as “preliminary” because some of the underlying non-BLS data used to produce the consumption measures are not finalized. Specifically, select 2021 health insurance values are subject to revision. In addition, WIC data are not finalized until up to a year after initial release by the U.S. Department of Agriculture (USDA). When updates to these non-BLS data are released, we will produce revised estimates for 2021.

Means

Quarterly CU averages are presented in tables 2 and 3. Table 2 includes results for 2019, 2020, and 2021 for the three consumption measures and two expenditures measures. Table 3 includes means for the consumption and expenditure subcomponents for 2020. (Detailed results for all 3 years are presented in table A-1 of the appendix for comparison at the subcomponent level.)

As shown in table 2, the trend from 2019 to 2021 is similar across all five measures, although the levels are slightly different. Means are lowest for 2020 relative to 2019 and 2021; this pattern is expected because of changes in consumption and expenditure patterns during the first year of the COVID-19 pandemic.⁴⁴ Quarterly means are lowest for consumption without health insurance (from $12,158 to $13,562). The next highest means are for total expenditures that do not include those for gifts, followed by total expenditures that include those for gifts. Quarterly means for both expenditure measures are about $2,000 higher than those for consumption without health insurance. The quarterly means for consumption with health insurance capped and uncapped are more than $4,000 higher than the means for all years for consumption without health insurance.

Focusing on 2020, table 3 presents a complete breakdown, by subcomponent, of the quarterly means for total expenditures with and without gifts, consumption without health insurance, and consumption with health insurance uncapped. The quarterly mean for total expenditures is $14,555. The quarterly mean drops, as expected, to $14,386, when we remove expenditures for goods and services purchased to be given to people who live outside the CU (i.e., identified as gifts by BLS). When we move to consumption without health insurance, the quarterly mean drops further, to $12,158. As revealed by the subcomponent means, the lower mean for total consumption without health insurance relative to the means for both measures of total expenditures is largely due to the removal of health insurance and the deduction of out-of-pocket expenditures, as well as switching from the net purchase price of vehicles to a flow of services. However, the decline was mitigated by increases in the consumption values of owned dwelling, rented dwelling, and other lodging; these reflect the movement from expenditures to rental equivalence for owned housing and imputed rents for renter shelter. Including health insurance increases the consumption measure to $16,767. Capping health insurance to 50-percent of consumption only slightly reduces the mean, to $16,675.

Poverty

Chart 1 presents the poverty rates defined using relative thresholds (set at 60 percent of each equivalized measure value). For exposition purposes, we focus on consumption with health insurance capped, consumption without health insurance, total expenditures, and pretax income. Relative consumption poverty rates with or without health insurance (about 11 and 14 percent, respectively) or health insurance capped (about 11 percent) are lower than relative poverty rates based on total expenditures (averaging 20 percent) and those based on pretax income (averaging 28 percent).⁴⁵ Relative poverty rates for all measures fell in 2020 and increased slightly in 2021. Focusing on 2019 relative to 2021, we see that there was little to no change in relative consumption poverty.⁴⁶ For consumption without health insurance, total expenditures (that include gifts) and pretax income, there were declines in relative poverty.

Using the same relative thresholds as for the total population, chart 2 shows the poverty rates for the population who are less than 18 years of age. The rates for this younger population relative to the rates for the total population are higher for all measures. Unlike for the total population, for which poverty increased from 2020 to 2021, for people less than 18 years of age, there was a continued decline in consumption poverty and in total expenditures poverty in 2021.⁴⁷ Income poverty increased in 2021, returning to its 2019 level. Note that the income measure we use does not include the expanded Child Tax Credit or the Economic Impact Payments issued during the pandemic. However, receipt of these subsidies could be reflected in the consumption estimates.⁴⁸

Chart 3 presents the poverty rates based on an absolute concept of poverty, rather than a relative concept. Thresholds are absolute in that they are “fixed” in 2019 and updated only for inflation. The thresholds for each measure are derived in such a way that the 2019 poverty rates for all the measures equal the 2019 relative poverty rate based on consumption with health insurance capped. Thus, the starting poverty rates for all measures are set to equal 11.2 percent. Using this measure, we find that poverty fell from 2019 to 2021 for each of the consumption measures, with the largest drop occurring for consumption without health insurance. Poverty rates for total expenditures also dropped, while pretax income poverty rose from 2020 to 2021.

Chart 4 presents the poverty rates for children (under age 18) that are based on the same absolute thresholds as those used for the total population (with the poverty rate set at 11.2 percent). Child poverty rates are not anchored. As with the relative thresholds, poverty rates for children are higher than those for the total population. However, by 2021, consumption-based child poverty rates had fallen more (from 2.4 to 4.0 percentage points) than did the poverty rates for the total population (from 1.0 to 2.3 percentage points). In contrast, the pretax-income child poverty rate for 2020 to 2021 increased by more (2.0 percentage points) than did the total population income poverty rate (0.9 percentage point).

Inequality

Table 4 presents the Gini indexes for the different measures. Gini indexes are a summary measure of inequality. They range from 0 to 1, with higher values corresponding to less equal distributions. Consumption is distributed more equally than total expenditures, which are distributed more equally than pretax income. Adding health insurance to the consumption measure makes the distribution more equal. The greatest inequality across the 3 years occurred in 2019 for all measures except total expenditures. The distributions of all the measures became more equal in 2020; but inequality increased in 2021.

Chart 5 presents the Lorenz curves for 2020. Lorenz curves are a visual representation of inequality and show the cumulative share of the population (ranked by the value of each measure from lowest to highest value) on the x-axis at or below the cumulative share of the measure as represented on the y-axis. Perfect equality is represented by the 45-degree line; the closer the Lorenz curve is to the line for perfect equality, the more equal the distribution. If a measure is equally distributed across the population, an equal share of the population would account for an equal share of the measure—for example, 50 percent of the population would account for 50 percent of consumption. An example of an unequal distribution is shown in chart 5 by the measure for consumption without health insurance: 50 percent of the population accounts for 30 percent of consumption. As with the Gini index results, pretax income is the least equally distributed of the measures, and consumption with health insurance capped is the most equally distributed. The Lorenz curves for consumption with health insurance and with health insurance capped are indistinguishable, and thus only one of the curves is presented. Lorenz curves for 2019 and 2021 (based on preliminary data) exhibit similar patterns.

Conclusion

In this article, we construct an initial version of a comprehensive consumption measure for the United States and use it to study poverty and inequality from 2019 to 2021. Our results show that the average value of consumption without including health insurance is lower than the average value of total expenditures. However, when we include health insurance, capped or uncapped, average consumption exceeds average total expenditures. We also find that consumption poverty rates defined using relative thresholds are lower than expenditure- or income-based poverty rates, regardless of whether health insurance is included. Consumption is also more equally distributed than expenditures or income. Poverty rates across all our measures fell in 2020 relative to 2019, and the distributions of the measures became more equal. This general finding is likely an effect of the COVID-19 pandemic, but verifying that claim is beyond the scope of this article.

BLS will continue to develop a comprehensive consumption measure for the United States with the CE providing the core data. The next main area of improvement is to incorporate the value of home production. Because home production is not a market transaction and relies instead on household members spending time performing tasks that add value to the household, we must impute time use to CUs participating in the CE. Home production includes the provision of childcare and eldercare. When purchased, both services can represent heavy financial burdens for families, some of whom will be priced out of the market. In such cases, children and elders in these families often receive care from other family members or friends through nonmarket transactions. We expect that the inclusion of home production in the CE will help us better understand consumption dynamics during the COVID-19 pandemic, when many childcare centers were closed.

The consumption measures presented in this article represent research in progress. However, BLS plans to release to the public a research series of consumption values in tables similar to the currently published data tables for CE total expenditures. These tables will include average consumption values. In addition, and dependent upon available resources, BLS is considering making available an auxiliary public-use microlevel data file that researchers could use to reproduce these consumption measures or create new ones of their own.⁴⁹ Finally, BLS expects its ongoing development work to result in methodological improvements that will be reflected in the CE data and in the consumption measure particularly.

Appendix: Imputation methods

In this appendix, we present the details of the imputation methods for NSLP, WIC, and LIHEAP benefits, market rents for which reported rents are expected to undervalue consumption, health insurance, and vehicle flow of services that are used in the construction of our consumption measures.

NSLP, WIC, and LIHEAP

The basic approach for imputing these in-kind government benefits is based on that of Garner and Gudrais (2018) in their research on producing Supplemental Poverty Measure (SPM) thresholds that account for in-kind benefits.⁵⁰ For the BLS consumption measure, we draw upon more recent methods developed by BLS (described below) and those developed by the Census Bureau to assign NSLP benefits to CPS-ASEC households for the SPM resource measure during the COVID-19 pandemic period.⁵¹

To produce the new consumption measure, the first step is to restrict, when necessary, the CPS-ASEC sample to households that are potentially eligible for in-kind programs. The full sample is asked the LIHEAP questions, and thus there are no sample restrictions for the imputation of LIHEAP benefits. However, only subsets of CPS-ASEC households are asked NSLP and WIC questions. The CPS ASEC only asks the NSLP-participation question of survey respondents in households with school-aged children (ages 5 to 18) who usually ate a hot lunch at school. During CPS-ASEC data collection, if a Census Bureau field representative was asked what is meant by “usually,” the response would be more than 50 percent of the time when the reference period was 2019; for the 2020 and 2021 reference periods, the response to what is meant by “usually” would have been when students were in school prior to the pandemic or when schools remained open during the pandemic.⁵² The WIC question is only asked of households with women ages 15 to 45 with no children (to include women who could potentially be pregnant) and women ages 15 and older with children under 5 years of age. These same sample restrictions for the NSLP and WIC benefit program are applied to the CE, with one exception. In the CPS ASEC, the NSLP questions are only asked of households with children who usually ate a hot meal at school; this information is not collected in the CE, which means we cannot restrict the CE sample along this dimension. The CE and CPS-ASEC data are then pooled. In the pooled data, NSLP and WIC participation and LIHEAP benefits are not missing in the CPS ASEC, but they are missing in the CE by design. A logistic regression model is used to impute participation when participation values are missing (in this case, the CE). For LIHEAP, a logistic regression model is used to first impute participation and then for those with imputed receipt, an ordinary least squares regression model is used to impute LIHEAP-benefit values from the CPS ASEC to the CE. For the statistical models, the explanatory variables are demographic variables that are defined the same for the CPS ASEC and the CE. For the imputed participation in these programs, the estimated model coefficients are used to generate participation probabilities for CE respondents, and then, using these predicted probabilities, participation (yes or no) is assigned randomly.⁵³

From the CPS-ASEC data, there are three possible outcomes for NSLP: (1) receive free or reduced school lunch, (2) pay full price for school meals, and (3) does not consume school meals. All school lunches provided to children are subsidized. Thus, children in the second group are NSLP participants, but we refer to them as “paid” in that they paid fully for their school lunches. For “free” and “reduced,” the CPS-ASEC question for NSLP participation does not distinguish between these two levels of meal support; thus, we assign children to “free” and “reduced” by using a method similar to that used by the Census Bureau for assigning benefits to participating households when calculating SPM resources. NSLP-program benefit values for the three categories—free, reduced, and paid—are based on data from the USDA.⁵⁴

Two different methods were used to assign free and reduced benefits, one for 2019 and a different one for 2020 and 2021. The method used for 2019 is based on CU pretax income (not including SNAP benefits) and a random assignment. If this pretax income measure is less than 150 percent of the official poverty threshold, children in the CU are all assumed to receive free meals. CUs with an NSLP-participation assignment of free or reduced and income equal to or above this threshold are randomly assigned to either being free or reduced NSLP participants.⁵⁵ For 2019, the number of school days for which these benefits are assigned is based on the state average number of school days in an academic year.⁵⁶ In contrast, for SPM resources, the Census Bureau assigns NSLP benefits using a national average of 179 school days.

The method to assign NSLP benefits for 2020 and 2021 for the BLS consumption measure is an adaption of the method developed by Census Bureau staff when accounting for NSLP benefits in SPM resources.⁵⁷ The method accounts for both school closures and receipt of EBTs to assign school lunch values. It is based on a combination of school operating status, SNAP receipt, and an additional question that was added to the CPS ASEC for 2020 and 2021: “Did your children continue receiving free/reduced price meals through your school or school district if schools were closed during the pandemic?” For 2020 and 2021, free lunch NSLP benefits are assigned to consumers reporting SNAP and with imputed NSLP participation. For 2020 only, reduced lunch NSLP benefits were assigned to CUs with imputed NSLP participation but no SNAP benefits. There were no reduced benefits assigned for 2021. For 2020 and 2021, we only assign NSLP benefits to CUs with school-age children when they are expected to be in school, in person. Because many schools were closed during the pandemic in 2020 and 2021, the USDA used EBTs to administer NSLP benefits. For our measure of consumption, when EBTs were received, we assign a zero value to the NSLP benefit because we assume expenditures based on the use of the EBTs are already reflected in reported food expenditures. This is the same assumption that underlies the treatment of SNAP in the CE because SNAP benefits are administered via EBTs and thus are considered “like cash”; adding SNAP values to reported food expenditures would be double counting.⁵⁸

For 2019 to 2021, if children in CUs were not assigned as receiving free or reduced lunch and were in school, in person, they were assumed to eat a hot lunch and to have received the NSLP-benefit values for paid meals. For 2019, the new BLS consumption measure used the average number of school days in an academic year, by state, to produce benefits for those who paid for their meals. For 2020 and 2021, we employed the method developed by the Census Bureau that accounted for school closings and changes in the NSLP when imputing NSLP benefits for those who paid for their meals.

For WIC, once participation receipt is imputed to the CE, WIC benefits are assigned. Average quarterly WIC-benefit values imputed to the CE are based on the benefits, infant-formula rebates, and infant participation. For those CE-respondent CUs who are imputed to have participants, the values for the in-kind benefits are assigned on the basis of state level per-beneficiary averages for WIC food benefits and infant-formula rebates, as well as timing of the adoption of WIC being electronically administered via EBT.⁵⁹ For consumer units living in states that have fully implemented the distribution of WIC food benefits by EBT, only the infant-formula rebate is added to consumption. To produce quarterly WIC benefits, average monthly values are produced and then multiplied by 3. For example, to compute the average monthly value for 2019 (represented by the CE data for 2019 quarter two to 2020 quarter one), three-quarters of 2019 fiscal-year monthly averages are pooled with one quarter of the 2020 fiscal-year monthly average. In contrast, for our earlier measure, fiscal year 2019 monthly averages were used. Average WIC food benefits are assigned to women and children. In addition, average infant-formula rebates are added to WIC food benefits. The infant-formula rebate benefit is assigned to the fraction of all infants who are assumed to be only formula fed.⁶⁰ WIC benefits and participation rates are based on administrative data published by the USDA.⁶¹

Consumption of rental shelter

For the consumption measure, there are several cases for which we expect the CE-reported out-of-pocket shelter expenditures not to reflect the full consumption value of this shelter. These include CUs receiving government rent subsidies, those living in public housing or those living in rent-controlled units, and those assumed to be paying less than market rents for other reasons. The CE asks about participation in the government programs, but not about the value of the difference in what renters pay for shelter and the market value of similar units, or about the market value of their units. Others who may pay less than the full consumption value of shelter include CUs consisting of people who are living rent free and those who are paying less rent because they are providing services to the landlord in lieu of paying the full rent. In addition, it is unlikely that the rents reported by college students living in dormitories reflect their consumption of shelter.

To impute rents for those receiving government rent subsidies, living in public housing, or living in rent-controlled units, we use the CE-participation variable for each program. In addition, we impute rents for renters who we assume are not paying full market rents: specifically, those reporting expenditures for maintenance and repairs in addition to reported rents. We also use the fact that some CUs consist of people living in dormitories to assign shelter consumption values.

We estimate a model of the full market rental value as a function of demographic, geographic, and housing-unit characteristics using a censored normal regression model, where the observed rental payment is censored from above if the CU received rental assistance (i.e., received government rental assistance, lived in public housing, or lived in a rent-controlled unit), or paid additional rent-related expenses. The estimation sample is restricted to traditional renters (i.e., respondents who report renting and do not report participating in a government rental assistance program, and who do not report out-of-pocket expenditures beyond rent alone). The estimation sample also excludes CUs consisting of people living in college or university dormitories. The rent used in the estimation model as the dependent variable is the logarithm of rent paid for the last month in the reference period, rather than the quarterly reported rent. This last month’s rent is used so that the period more closely matches that of the period covered in owners’ reports of rental equivalence: how much owners think their homes would rent for currently (at the time of the interview) without furnishings and without utilities.⁶² The imputed market rent multiplied by 3 is used in place of the CE-recorded rent for the last 3 months (the reference period) for all respondents who are identified as participating in a government rental assistance program, occupying their residence without payment of rent, and those with out-of-pocket maintenance and repair expenditures in addition to their reported rents. To derive shelter consumption for renters, out-of-pocket expenditures for tenant’s insurance are added to the rents reported by CUs assumed to be paying full market rents, and to the imputed rents for all other renters.

The consumption shelter value for CUs that consist of people living in college or university dormitories is imputed differently. Because dormitories are distinctly different from other rental units, we cannot apply our market-rent imputation model to these cases. Instead, we assign the national average dormitory cost, which we obtained from the National Center for Education Statistics.⁶³ We use this value to calculate an average monthly cost, which is then adjusted to reflect the 3-month cost for the reference period. Because we do not expect dormitories to be rented year round, to get the 3-month reference-period cost, the monthly value is scaled by the number of months in the reference period that overlap with the August–May school year. For example, a CU interviewed in July has a reference period of April, May, and June. This reference period has 2 months that overlap with the August–May school year, so the consumption value for dormitories during this reference period is twice the average monthly cost for a dormitory. The same adjustment is made for the cost of board while living at college.

Health insurance

The CE Interview asks about different types of insurance coverage and measures the out-of-pocket insurance premium associated with each type of insurance. The CE also measures out-of-pocket expenditures on medical goods and services. For people with insurance, out-of-pocket expenditures on medical goods and services will reflect copayments and coinsurance. One limitation of the CE data on health insurance coverage is that for types of insurance other than Medicare, the only questions asked are whether anyone in the CU is covered and the number of people who are covered. So, there is no way to identify the specific individuals covered.

For the consumption measure with health insurance, we first impute the full value of health insurance on the basis of insurance type. For private health insurance, the full value is based on the market price. This includes the out-of-pocket premiums that are captured by the CE and the employer contribution for employer-provided plans, or the value of any subsidy received for individual plans. For public insurance programs, the full value is based on the cost to the government (including administrative costs).

The methods and data sources used for the imputation of the full value of health insurance are based on the work of Garner, et al. (2022).⁶⁴ For that project, imputed values were added to CE health insurance expenditures to match the scope of the U.S. Bureau of Economic Analysis personal consumption expenditures (PCE) data, which, by definition, include employer contributions to employer-provided health insurance and government insurance programs. The data sources used to impute the values for the CE-PCE project are the same as those used to produce the consumption measures with health insurance (capped and uncapped) presented in this article. However, the values for health insurance used in the production of these consumption measures and those used for the CE-PCE project are not directly comparable because the two measures differ in scope.⁶⁵

For employer-provided health insurance, we use data from the Medical Expenditure Panel Survey Insurance Component (MEPS-IC).⁶⁶ MEPS-IC has data on the average employer contribution and average employer share. For employer-provided coverage, we impute different amounts on the basis of the type of plan: single coverage, plus one, and family. We assign individuals in the CE to different plan types on the basis of the number of people covered by the plan. So, individuals are assumed to have single coverage if the plan covers one CU member, plus one if the plan covers two CU members, and family coverage if the plan covers more than two CU members.

People who purchase individual health insurance plans can receive subsidies in the form of a tax credit. The CE asks individuals who purchase individual plans whether they receive a subsidy. We assign the average subsidy among those who receive a subsidy to CUs in the CE with individual plans who report receiving a subsidy. The data on the average subsidy amounts come from the Centers for Medicare & Medicaid Services (CMS).⁶⁷

For public programs, the imputed amount is the per-enrollee cost to the government, including administrative costs. For Medicare, the average cost (including administrative costs) per beneficiary is calculated from the Trustee’s report for each of the Medicare programs: traditional Medicare (includes Parts A and B), Medicare Part C, and Medicare Part D. These programs also have out-of-pocket premiums. There are two ways to account for out-of-pocket premiums. One way is to add the average cost less premiums to the out-of-pocket premium amount reported in the CE. The other approach is to assign the average cost and ignore the out-of-pocket amounts. Which approach is preferable depends on whether the variation in out-of-pocket premiums for these programs is due to variation in the amount of coverage purchased or to variation in the amount the premium subsidized. If the coverage is the same, then adding the average cost less average premium amount will lead to a lower value for individuals who have received subsidies for their premiums. This is the case for traditional Medicare, so we ignore the CE-reported premiums and impute a value on the basis of the average cost. We treat Part D the same; although there is variation in Part D plans, the variation in out-of-pocket premiums due to subsidies is likely much larger. Part C can be thought of as a bundle of traditional Medicare and supplemental coverage. For the traditional Medicare component, we impute an average cost and ignore the CE-reported Part B premiums. Then, we add any additional premiums paid to reflect the supplemental coverage.

Other public programs are more straightforward because there is less variation in the types of coverage offered. So, we assign a single average-cost value to participants in other public programs and ignore any out-of-pocket premiums. For Medicaid and the Children’s Health Insurance Program, the per-enrollee average cost comes from the CMS National Health Expenditures tables.⁶⁸ Per-enrollee costs for other public programs (e.g., TRICARE, Veterans Affairs benefits, and Indian Health Services) are calculated from the budgets for each of the programs.⁶⁹

Depreciation and opportunity costs of owning vehicles

For vehicles, we employ a user-cost approach to assign a consumption value that is based on the flow of services. The user cost is defined as the depreciation plus the opportunity cost of capital (value of the car times the interest rate), plus maintenance and repair costs. We estimate the depreciation and opportunity costs; the other components of user costs are already included in consumption as nonvehicle-purchase-related expenditures. Depreciation and opportunity costs are imputed to CUs on the basis of vehicle ownership. For our research, we restrict vehicles to cars and trucks, with sport utility vehicles being categorized as trucks. For 2021, about 87 percent of vehicles reported in the CE as being owned were classified as cars or trucks. The remaining 13 percent are other vehicles such as planes, boats with motors, and motorized campers. Because the sample size for other vehicles is quite small and the CE does not collect data on characteristics for these vehicles, depreciation and opportunity costs for “other vehicles” are not calculated and thus are not included in the consumption measure.

We estimate the vehicle depreciation rate using data on vehicles owned as reported in the CE. For all vehicles owned, the CE collects data on the vehicle characteristics, including make, model, and year. The CE also asks when the vehicle was acquired and the purchase price of the vehicle. We estimate a depreciation rate by comparing the purchase price for vehicles purchased at different ages while controlling for vehicle characteristics. We estimate an age-specific depreciation rate for vehicles 10 years old or less. Vehicles over 10 years old are assumed to depreciate at a constant rate, as there are too few transactions involving older vehicles to estimate age-specific depreciation rates. The imputed depreciation value is calculated as the current market value of the vehicle times the age-specific depreciation rate. The current market value is calculated as the estimated new purchase price of the vehicle minus the estimated depreciation for prior years. The opportunity cost of capital is the product of an interest rate and the estimated current market value of each car and truck owned by the CU. The interest rates are the year-specific Treasury Long-Term Average (Over 10 Years), Inflation-Indexed annual rates, not seasonally adjusted, published by the Federal Reserve Bank of St. Louis.⁷⁰