Assessing and improving the accuracy of the CPI

The Consumer Price Index (CPI) is a measure of the average change over time in the prices paid by urban consumers for a market basket of consumer goods and services. Americans use the CPI extensively, including to adjust historical data, to escalate federal payments and tax brackets, and to adjust rents and wages. The CPI directly affects lives, so it must be as accurate as possible. For example, the CPI measured the 12-month change in prices from December 2021 to December 2022 as 6.5 percent. How confident can we be in that estimate?

This article looks at some ways the U.S. Bureau of Labor Statistics (BLS) has improved the accuracy and precision of its indexes and how we responded to questions about the accuracy and precision of the CPI. The first section gives an overview of the CPI production process, from the geographic sample to the samples of prices. Next, the article looks at the sampling error of the CPI and issues of data quality. We then discuss potential biases in the CPI. Finally, we examine recent and ongoing improvements to the CPI involving increased use of alternative data and more frequent expenditure weight updates.

Producing the CPI

The CPI is produced by using several interrelated surveys. These include a geographic sample, a survey of consumer expenditures, and two separate samples of prices (one for shelter and one for other commodities and services).

The CPI geographic sample

Because it would be prohibitively costly to collect price data in all geographic locations in the United States, the CPI program scientifically selects a sample of geographic areas to estimate price change in the nation. Specifically, the CPI program selects a set of core-based statistical areas (CBSAs) as defined by the Office of Management and Budget. In the current sample design, areas were first classified into one of nine census divisions: New England, Middle Atlantic, East North Central, West North Central, South Atlantic, East South Central, West South Central, Mountain, and Pacific.¹ (See chart 1.) Each area was also classified into one of two population-size classes: self-representing or nonself-representing. Areas with more than 2.5 million residents are defined as self-representing and their weight in the CPI corresponds to their population relative to the U.S. urban population. Areas with fewer than 2.5 million residents represent not only themselves but other, similar areas in their census region and size class as well.

The current CPI geographic sample, implemented in 2018, consists of 75 CBSAs, which are called primary sampling units (PSUs). In total, there are 23 self-representing PSUs. These include 21 areas whose population is greater than 2.5 million and two additional areas: Anchorage, Alaska, and Honolulu, Hawaii.² These 23 self-representing PSUs are combined with 52 nonself-representing PSUs to form the 75 total PSUs. For index calculation, the 75 PSUs are consolidated into 32 index areas. (See chart 2) The current area design yields 7,776 basic indexes (243 item strata for each of the 32 index areas) for the U.S. all-items CPI.

CPI market basket and outlet sample

BLS uses data from the Consumer Expenditure Surveys (CE) to create its market basket and outlet frame for the CPI.³ The CE program is an ongoing set of surveys that collects information from samples of consumers on what they buy and where they buy it. Information from the CE is used to create the expenditure weights for the categories of goods and services in the CPI. Starting in 2023, these weights are updated annually. Expenditure weights in 2023 are based on CE data from 2021. (Prior to 2023, weights were updated once every 2 years and based on 2 years of CE data). Additionally, since 2019, the CE surveys provide data on the retail outlets—department stores, supermarkets, hospitals, and other types of stores and service establishments—from which metropolitan and micropolitan households purchase commodities and services that are priced in the CPI. This use of the CE surveys for the outlet frame represents a change from the prior method of using a separate, dedicated telephone survey (a telephone point of purchase survey, or TPOPS) as the basis for the outlet sample. The transition to the CE surveys, completed in 2019, eliminated redundancies and inefficiencies in the operations of the surveys and resulted in lower household respondent burden.⁴

CPI pricing surveys

A fundamental part of creating the CPI is collecting price data. BLS conducts two separate, ongoing consumer pricing surveys, one for commodities and services and one for rent. Traditionally, BLS has collected prices for the CPI by personal visit, with a small amount of data collection by telephone. Recently, collecting prices online has increased in prevalence. Moreover, in some cases BLS has moved to collecting data from nontraditional sources. This collection will be explored later in this article.

Commodities and services price survey

BLS collects prices for approximately 100,000 goods and services. Prices are collected each month in 75 urban areas across the country from approximately 21,500 retail establishments.⁵ All taxes directly associated with the purchase and use of items are included in the index.

Each outlet is assigned a number of specific goods or services, defined in the CPI as entry-level items (ELI) for price collection. BLS staff then visit each selected outlet and use a multistage probability selection technique to select specific items among all the items the outlet sells that fall within the entry-level item definitions.⁶ CPI sampling procedures are designed to create a sample that includes a wide variety of specific items and that corresponds roughly to consumer behavior; more popular items have a greater chance of being selected for the sample.⁷

Housing prices survey

In addition to the broad survey of commodities and services prices, the CPI uses a survey of rental prices to estimate price changes for both rent and owners’ equivalent rent.⁸ The housing survey follows the rents of a sample of renter-occupied housing units selected to represent both renter- and owner-occupied housing units in the urban United States. The sample is divided into six panels, with approximately 8,000 units in each panel. Each panel is priced once every 6 months. The estimate of monthly price change is the sixth root of the 6-month price change in the panel being priced.

The sample is updated regularly, with one-sixth of the sample replaced every year on the basis of the latest available U.S. Census Bureau data. Rent data in the sample are collected by personal visit or phone. Unlike many other measures of rent price change, which survey only rents available to new tenants, this CPI survey includes units where the same tenant is remaining in the unit. This method aligns with the measurement goal of the CPI, which is to measure the change in shelter costs for all consumers, not just consumers who are changing housing units.

Data quality in the CPI

As a measure of price change, the CPI can be evaluated both in terms of its precision (the sampling error of the estimate) and possible biases it might have (whether and how much it is consistently above or below the true value). In this section, we look at the precision of the CPI, examine some procedures related to data quality, and discuss several possible biases that CPI users should understand and consider.

Sampling error in the CPI

If the CPI sample were comprehensive and included every price within its scope, the CPI program’s estimates of price change would be precise; the standard error would be zero. However, although the CPI program collects over a million prices per year, this is only a tiny fraction of all the prices in the economy. So, like other surveys that make estimates from samples of data, the CPI surveys are subject to sampling error. In the case of the CPI program, this error is the difference between the CPI program's estimate and what the estimate would be if the CPI program were able to collect all prices. Sampling error is a metric of uncertainty, one measured by a statistic known as standard error.

The CPI sample design accommodates error estimation by making two or more selections (replications) of items and outlets within an index area. Therefore, two or more samples of quotes in each area are available. With this structure, which reflects all stages of the sample design, variance estimation techniques using replicated samples can be used.

BLS constantly tries to improve the precision of the CPI. Variance and sampling error are reduced by using the largest possible samples of retail prices, given resource constraints. To optimize its allocation of resources, BLS developed a model that indicates the number of prices that should be observed in each geographic area and each item category in order to minimize the variance of the U.S.-city-average all-items index.⁹ Because the CPI uses a probability sampling method, the sampling error of survey estimates can be computed directly from the sample data. The CPI program publishes measurements of sampling error for all its indexes.

The 2022 sampling error is quite small for the CPI for All Urban Consumers (CPI-U) U.S.-city-average, all-items index, which is the broadest measure of inflation. In 2022, the median standard error for 1-month price changes was 0.04 percent. For example, if the all-items index increases 0.50 percent in a month, 95 percent of the estimates of the actual rate of inflation would be between 0.42 and 0.58 percent (that is, 0.50 plus or minus 2 times the standard error).

The sampling error for 12-month changes in the all-items CPI was also small, with a median standard error of 0.12 percent. So, for example, if the all-items CPI rises 6.50 percent over 12 months, 95 percent of the estimates of the actual rate of inflation would be between 6.26 and 6.74 percent.

However, it is important to note that sampling errors are typically larger for smaller geographic regions and smaller CPI item categories. For example, the 2022 12-month median standard error for the Northeast census region all-items CPI was 0.23 percent, almost twice as large as the 0.12-percent standard error for the United States as a whole. Standard errors for the census divisions, such as the Middle Atlantic or the Mountain, and local metropolitan areas, such as Boston or Philadelphia, would be even larger.

Similarly, CPI item categories usually have larger standard errors than the all-items index. For example, the 12-month standard error for the housing index was 0.22 percent—almost twice as high as that for the all-items index. For some index series, the standard errors are significantly higher. For example, the 12-month standard error for apparel was 0.80 percent, meaning that a 12-month increase of 2.9 percent would have a 95-percent confidence interval of between 1.3 and 4.5 percent. For this reason, BLS encourages users to employ broader indexes when using the CPI for purposes of escalation. The broadest index, the CPI-U U.S.-city-average index for all items, is often used even when more specific indexes might be considered.

Response rates in the CPI

Response rates measure the proportion of completed surveys to those issued and help measure the accuracy of the CPI. Lower response rates can reduce the precision of the CPI by reducing the amount of data on which CPI estimates are based. Additionally, low response rates can cause bias in the CPI if the nonresponse is correlated with something that is related to price change. For example, in our sample of physicians’ services, there was a lower response rate for private insurance quotes compared with self-pay quotes, creating a possible bias in the index if left untreated.¹⁰

BLS publishes response rates annually to assist the data user in judging the accuracy of BLS products.¹¹ The published tables show response rates for all data included in the CPI-U, U.S.-city average, by major group.¹² At the U.S. level, the housing (excluding shelter) category includes items like household fuels and utilities, and furnishings and operations; data for shelter are shown separately. BLS also publishes rates for commodities and services and for shelter in selected areas.

Response rates are calculated for the CPI-U at the data collection and data estimation phases. The response rate at the data collection phase is the number of responding sample units divided by the number of eligible sample units. A sample unit is eligible if it belongs to the defined target population and if it was selected to provide information for one or more items. The percentage of the sample used at estimation is defined as the number of sample units used in estimation divided by the number of eligible sample units.

In recent years, like many other statistical surveys, BLS surveys have seen a decline in response rates.¹³ This decline has created additional urgency to explore more use of nontraditionally acquired data. Some examples of this exploration are discussed later in the article.

Despite the increase in their number, missing or incomplete responses have always been a part of surveys, and the CPI program uses different methods to account for nonresponse. In order to use all the available data, the CPI program fills in missing values in a systematic way, usually through a process called “imputation.”

Imputation

Nonresponse necessitates the use of imputation in the CPI. Price change has to be estimated in some way for quotes that are not collected. Frequently, this is done by imputation.

Imputation is a statistical procedure for handling missing information. The CPI program uses imputation for several cases. Some of these cases involve items in the sample being temporarily unavailable, such as for respondent refusals or items that are out of season. Additionally, when items are replaced in the sample, imputation may be used when quality change cannot be satisfactorily estimated. Replacement items that can be neither directly compared nor quality adjusted are deemed to be noncomparable. For noncomparable replacements, an estimate of constant-quality price change is made by imputation. There are two main imputation methods used in the CPI: cell-relative imputation and class-mean imputation.

Cell-relative imputation

If there is no reason to believe that the price change for an item is different from the price change observed for the other items in its basic index, the CPI program uses cell-relative imputation. This method is used for missing values because no information is available about the observation in such cases. It is also used when a substitution has been made to a new item, but the items are noncomparable (this is common for food and service items). The price change between the original item and the noncomparable replacement item is assumed to be the same as the average price change of all similar items in 1 month for the same geographic area.¹⁴ When there is a new version of the item that is not comparable to the previous version, the price of the new version is not used in calculations for period t, but it will be used in the subsequent period, t + 1, as it will then be the previous price.

Class-mean imputation

For some commodities and services item strata, such as those for durable goods and for apparel, the CPI program uses a class-mean imputation for noncomparable replacements. In this case, “class” refers to all the comparable and quality-adjusted replacement observations in the same ELI and PSU. The logic behind the class-mean procedure is that price change is closely associated with the annual or periodic introduction of new lines or models for many items. For example, at the introduction of new model-year vehicles, there are often price increases, while later in the model year price decreases are common. The CPI program uses the quality-adjustment method for approximately 40 ELIs to handle item replacements that occur when product lines are updated.¹⁵

Bias in the CPI

Sampling error influences the precision of the CPI, but bias is perhaps more a serious issue concerning the accuracy of the indexes. Bias is defined as the difference between the expected value of an estimator and the true value being estimated. In general, sampling error tends to even out in the long run, but if the CPI is persistently understating or overstating inflation because of a bias, a growing gap between true price change and the CPI measure will occur. Three sources of potential bias in the CPI have received the most attention in recent decades: substitution bias, quality-change bias, and new-goods bias.

Substitution bias

Substitution is when consumers change their purchasing behavior and buy one type of good instead of another. There is an expectation that, other things being equal, an increase in a good’s price will cause consumers to reduce their purchases of that good and instead purchase a substitute with a relatively lower price. However, it is not always price that causes consumers to substitute; it could be new features or new goods, or just changes in tastes.

Effective with January 1999 indexes, BLS changed the way it calculated the CPI for many of the basic item-area indexes, moving from a Laspeyres formula to a geometric-means formula.¹⁶ The geometric formula effectively presumes modest consumer substitution within item categories, mitigating possible bias from consumers substituting between goods within CPI item strata (lower level substitution bias). That is, the formula assumes that consumers will substitute away from one brand or type of item as that brand or type becomes relatively more expensive compared with other brands or types of that product. Examples might include switching from one brand of car to another, or one cut of steak to another. The geometric-means formula is only used within CPI categories and so does not correct for substitution across item categories (upper level substitution bias). For example, pork and beef are two separate CPI item categories. If the price of pork increases while the price of beef does not, consumers might shift away from pork to beef. However, because any substitution from pork to beef would be across item categories, the use of a geometric-means method has no effect. In any case, the use of geometric means for most categories has had the effect of lowering the CPI-U in the United States by 0.2 or 0.3 percent per year, on average.¹⁷ The Chained Consumer Price Index for All Urban Consumers (C-CPI-U), introduced in 2002, uses updated expenditure weights; rather than make any assumptions about substitution, it derives its weights from expenditure measures both before and after a price change. It is thus free of upper level substitution bias. As would be expected, it tends to run slightly lower than the CPI-U. Therefore, those who believe that upper level substitution bias is important can focus on this measure.¹⁸ However, an important limitation of the C-CPI-U is that it is revised and not made final until 10 to 12 months after initial publication.

With the geometric-means formula in place to account for consumer substitution within item categories, and the C-CPI-U designed to account for consumer substitution between item categories, any remaining substitution bias would be quite small.

Quality-change bias

The CPI program strives to adjust for quality change within the CPI market basket. Because the CPI seeks to approximate a cost-of-living index (COLI), the CPI is, conceptually, a constant-quality index. However, quality change can be difficult to quantify, and to the extent the CPI is unable to accurately adjust for changes in quality, a bias in the index can arise. This bias can potentially be in either direction.

The CPI is calculated using prices for a fixed basket of goods and services through time. Although the basket is periodically revised to reflect changing consumer expenditures in the marketplace, some items that were selected to be priced in the sample come and go from the marketplace for a variety of reasons, which adds complexity to collecting these prices from month to month. When an item is no longer available in the marketplace, a similar replacement item is selected. In some cases, there are no similar items from which to choose. As a result, a less comparable item is selected, potentially introducing quality change and an associated price differential into the index. Thus, when the quality of goods and services in the market basket changes, BLS must make some estimate of the value of such changes in order to remove it. Economists continue to debate whether the CPI appropriately adjusts for quality or whether there is an upward or downward bias.¹⁹

To understand these issues, it is helpful to look at how BLS responds to changes in the goods it prices. Operationally, the CPI measures quality change in several different ways. To price a given item, the BLS economist in the field determines if the item has changed in any way (that is, if the item has been replaced with a new version). If there is a new version and the two are essentially the same, a commodity expert may deem them directly comparable and use the price comparison as if no quality change had occurred. If the versions are substantially different, then a quality adjustment must be implemented—either by imputation procedures or by direct quality adjustments. Although there are different types of imputation procedures, such procedures essentially mean that the price is assumed to change at the same rate as other similar items that month.

The simplest type of direct quality adjustment can be made when the difference is easily quantifiable, such as a size decrease. Data collection procedures vary for different products and services; therefore, the impact of product size change is handled differently based on the item. An effective price per standard size, usually a price per ounce, is calculated for items in which size is reported. The effective price per ounce is the collected price divided by size. For example, if a half-gallon (64 ounces) of Brand A milk is priced in December 2022 at $3.85, the effective price per ounce is $0.060 per ounce ($3.85 divided by 64 ounces). If, in January 2023, the same Brand A milk is reduced in size to 60 ounces, but the price is still $3.85, the effective price per ounce would be $0.064 per ounce. This is a 6.7-percent increase in the price per ounce of the milk, and the CPI would include this price increase.²⁰

In 2023, BLS published an article that examined product upsizing and downsizing and explored how these concepts affect measurements in the CPI.²¹ Downsizing, commonly called “shrinkflation,” is common across food and household commodities, such as potato chips, paper towels, cereal, cleaning supplies, and candy. Manufacturers change sizes because market research indicates that consumers are more sensitive to price change than size change.²² Downsizing impacts the amount of a good a customer receives; therefore, goods that are sold by a specific unit of weight or volume do not experience downsizing. For example, gasoline or steaks generally do not experience downsizing because they are sold per gallon or per pound.²³

In other cases, direct quality adjustment may be used when the change is not simply a matter of a different size or quantity. In some cases, a technique called hedonic quality adjustment is used, which involves using regression techniques to estimate the value of specific bundles of characteristics, such as the sleeve length and fabric design of a shirt or the capacity and number of cycles of a dryer.²⁴ Hedonic adjustment has generated a fair amount of attention and is sometimes criticized as being intentionally designed to lower the CPI.²⁵ However, it is used on a narrowly defined part of the total index, and research suggests that the effect of hedonic techniques on the all-items index is very small. Hedonic adjustments may capture both higher and lower quality adjustments and result in faster price increases in some categories and slower increases in others, with the net effect close to zero.²⁶ The hedonic quality-adjustment method removes any price differential attributed to a change in quality by adding or subtracting the estimated value of that change from the price of the old item. For example, the hedonic quality adjustments for rent and owners’ equivalent rent are used primarily to adjust for the age of a rental unit and for utility adjustments.

Although it is widely assumed that hedonic adjustments are downward, resulting in a lower CPI, adjustments can be in either direction. BLS has used hedonic models in the CPI shelter and apparel components for almost three decades, and on average, hedonic adjustments usually increase the rate of change of those indexes. Hedonic models have been introduced in several other components, mostly consumer durables such as personal computers and televisions, but these newer areas only have a combined weight of about 1 percent in the CPI. An article by BLS economists estimated that the hedonic models currently used in the CPI outside of the shelter and apparel areas have increased the annual rate of change of the all-items CPI, but by only about 0.005 percent per year.²⁷

New-goods bias

Additionally, the CPI market basket does not perfectly reflect what is being purchased in real time, and new goods might not be fully represented in the CPI market basket for some time after their introduction into the market. New goods often decrease in price after they are introduced (examples include cell phones, computers, and smart watches). If new goods are not quickly incorporated into the market basket of the index, the initial downward price trend might be missed and there could be an upward bias in the index.

Currently, BLS updates the CPI market basket with a new survey of consumer expenditures every year and rotates most of its sample of items every 4 years. These procedures, which began in 2023, are designed to keep the CPI market basket as up to date as possible. New goods and services are introduced into the CPI in several steps. Often, respondents in the CE report purchases of goods and services not already included in the CPI, and other times economists identify a new product and adjust the survey documents used in the CPI to accommodate its eventual selection into the sample. Once identified, the new CE item must be mapped to an existing CPI ELI, or a new ELI might need to be created. When a new product is added to the CPI sample, BLS creates a checklist for it or updates an existing checklist to accommodate the new item. For example, BLS made changes to accommodate subscription meal kits in 2017, drones in 2018, and bike and scooter rentals in 2019.²⁸

The process from when a new good or service first shows up in the expenditure data to actually pricing it in the sample varies depending on the item. If the new good or service is sold in entirely new outlets that are not already in our sample, such as a new clothing subscription service sold through its own website, it can take up to 2 years to get the new outlet into our sample. But if the new good or service is available in outlets that sell other similar items already in our sample, the new item can be incorporated during regular 6-month initiation cycles. All of this means that the CPI does not always include new goods in its sample in real time, but it also means that price change experienced by early adopters of new items—typically introduced at a high level and then dropping—may not be included in the CPI either. Nevertheless, as part of its commitment to measuring price change as accurately as possible, the CPI program proactively monitors new goods entering the economy and strives to get major new goods into its sample as quickly and appropriately as possible.

Beyond the commodities and services and housing surveys

When it comes to data that are not collected through the traditional commodities and services and housing surveys, the CPI program identifies two distinct sources: corporate data and secondary source data.

Corporate data are price or transaction information that a survey respondent provides, typically using data directly from corporate headquarters. Generally, the information included is what the respondent is willing to provide, and BLS must adapt the data elements and format to fit its processing systems. Discussions with respondents often involve finding a level of aggregation that addresses any of their confidentiality concerns while also matching the needs of BLS estimation objectives.

Secondary source data are compiled by a third-party data aggregator or regulator and usually include prices for goods or services from multiple establishments. The data aggregator has made some effort to standardize the data elements and structure across business establishments. Often, BLS needs to purchase this secondary source data; however, in limited cases, the data are available free of charge. Although the data are standardized, BLS still needs to adapt the datasets to fit its processing systems and develop methods to match the level of aggregation in the secondary source data to meet BLS needs.

Data from these sources can be obtained via a variety of methods, many of which BLS is researching or already using. Methods could include direct transfer of data files, web-scraping prices, or programmatic access such as through application programming interfaces (APIs).

The reason that BLS is evaluating, and in some cases adopting, different collection methods is to improve both the quality of its data and the processes by which those data are collected. Alternative data can substitute for what is presently being collected from respondents and provide additional information to supplement BLS-collected data. Additionally, alternative data may be used in research to improve CPI estimates, such as to improve quality adjustment or to consider methodological changes. The BLS strategic plan includes the goal of integrating alternative data into its programs.

When BLS is considering alternative data for data validation, adjustment, or replacement, it is critical to assess the representativeness of the source relative to the CPI’s targeted improvements. BLS must also consider factors such as the intended coverage of the alternative data, systematic inclusions or exclusions of various population subgroups, and any additional adjustments made by alternative data vendors. BLS has worked on several applications of alternative data, and some of those uses are discussed below.

Improvements to the new-vehicles index

As discussed above, BLS is consistently updating its collection methods to improve the quality of its data and the process by which those data are collected. BLS achieves both goals in the use of alternative data in the new-vehicles index. This index is now estimated using a transactions dataset that includes observed transaction-level prices and detailed vehicle information. Each observation includes a transaction price as well as a set of 40 variables including rebate values and vehicle characteristics.

Prior to using the J.D. Power transaction data, BLS relied on dealer-estimated prices for a sample of hypothetical vehicle configurations based on consumer purchasing patterns that may have been several years old. The J.D. Power transaction data allow BLS to use observed purchase prices and to reflect real-time market conditions in terms of the quantities and expenditure shares of vehicles sold. In addition to better representing the sales in the current month, the J.D. Power data provide substantially more observations and broader geographic coverage.

Although the transaction data have many advantages over the survey data, incorporating transaction data into the CPI presented several challenges. To address these challenges, BLS developed an alternative methodology that allows it to make better use of the transaction data.²⁹ BLS research found that with the transaction data, monthly, matched-model price indexes declined persistently even as average prices (the unit price index) rose. The matched-model indexes only show the price change for the same model-year vehicle. When a new model-year version of a vehicle is introduced, it is offered at a relatively high price. As the months pass, the price drops as dealers offer steeper concessions and manufacturers begin to offer more rebates and incentives. Linking these price drops together results in a steadily declining price index. The existing CPI methodology also showed these declines but offset them by showing the price change from a vehicle’s end-of-life price to the early life price of its next model-year replacement. This method of offsetting did not perform well when applied to the transaction data because the indexes were very sensitive to how these offsetting price changes were weighted.

Instead of using these offsets, BLS focused on the price change across model years for similarly aged vehicles. This method avoids the complications of weighting within model-year price change with cross model-year offsets. Arguably, this year-over-year price measurement is more consistent with a cost-of-living index (COLI), because a COLI should not show the price change from early adopters buying the newest version of a product and bargain hunters buying at the end of a model year. Instead, the year-over-year index compares the price of a vehicle sold today with the price of the previous model-year vehicle 12 months ago.

The J.D. Power monthly new-vehicle sample size is approximately 250,000 transaction prices per month. BLS removes vehicles specifically referenced as fleet vehicles, as well as vehicles not typically purchased for consumer use. Unlike the CPI program, J.D. Power does not attempt to construct a representative sample. However, our research showed that market shares in the J.D. Power data were similar to the representative sample of vehicles in the CPI.

J.D. Power collects a large dataset of observed transaction-level new car prices and detailed vehicle information for approximately one-third of new U.S. consumer-class automobile sales. Each observation includes a transaction price as well as a set of 40 variables showing size rebates, vehicle characteristics, and, sometimes, the cost of the vehicle to the dealer. In J.D. Power samples, prices are collected without sales tax; the sales tax is applied by BLS.

Once the data are received from J.D. Power, BLS filters them to find specific vehicle combinations that are new to the sample. Economists then review the new combinations in order to make comparability decisions and to apply quality adjustments. BLS economists apply the decisions and adjustments to the data using methodology identical to that used on traditionally collected data.

The dataset purchased from J.D. Power also helps to address lower response rates and higher collection costs. In this case, the purchased data allows a large sample to be created at much less expense than manual data collection. By performing extensive research and statistical analysis on the J.D. Power dataset, BLS concluded that this alternative dataset was a viable source for the CPI’s new-vehicle price data. BLS was able to expand CPI’s sample and coverage across the country and use these data in the official calculation of the new-vehicles index. By using this dataset from J.D. Power, BLS addressed the problem of declining response rates and improved the accuracy of CPI measures.

Improvements to the gasoline index

In June 2021, BLS replaced the data collected for the CPI in the gasoline index with data from a secondary source.³⁰ The dataset for gasoline includes the daily average price per gallon observed for a given fuel type at each outlet, as well as the postal code and state of the outlet. Using the same dataset, every month BLS also calculates and publishes special relative series and an average price for each type of gasoline. For the gasoline index, all observations from the secondary source dataset that are within the CPI geography are eligible and used for index-relative calculation.

Given its importance as a consumer good and its frequent sharp price changes, gasoline often is a major factor in overall price change; in a given month, gasoline often accounts for a large fraction of the overall monthly change in the CPI. Fortunately, gasoline is one of the few consumer goods for which there are many sources of price data. In fact, the ease of price collection makes it feasible for other government agencies and even private sources to create reliable measures. On the government side, the Energy Information Administration (EIA) publishes extensive gasoline price data. Among private sources are the American Automobile Association, the Oil Price Information Service, and the Lundberg Survey.³¹

Although these other sources may appear to show different fuel price movements from the CPI, the apparent differences are primarily due to timing. For example, the EIA data are released each week and correspond to prices on a particular day. The CPI gasoline index corresponds to average prices over a calendar month. BLS research shows that when timing differences are considered, the CPI and EIA data are consistent in reflecting price movements.³²

Most instances of alternative data collection in the CPI were implemented to address complications in traditional data collection or poor response rates. This is not the case for gasoline, which is relatively straightforward to price via traditional data collection. However, the alternative secondary-sourced data enabled BLS to base its estimates on a larger and more extensive sample.

Improvements to the medical care index

Medical care is an important category in the CPI, accounting for over 8 percent of the weight as of January 2023. It is also one of the more difficult to collect items, with respondents being hard to reach, respondents’ reluctance to participate given concerns about patient data confidentiality (BLS is not interested in specific patients), and the variety and complexity of medical care.³³

Several of the improvements to the medical care index since 2016 have come from alternative data sources. For example, BLS now uses Medical Expenditure Panel Survey data to obtain a more accurate sample of quotes by payer type for the physicians’ services index. BLS also uses corporately provided transaction data from a retail pharmacy to obtain a more accurate sample of quotes by payer type for the prescription drugs index, which also improves the efficiency of incorporating generic drugs into the prescription drugs index. BLS also improved its traditionally collected data methodology by shifting to longer rotation cycles for hospital services, incorporating mixed medical samples for the physicians’ services and hospital services indexes (as a way to increase the pool of potential respondents), and preselecting pricing specifications for the hospital services index (in order to reduce respondent burden).

Nontraditional data acquisition has the potential to address many of the problems faced in recent years including lower response rates and higher collection costs. The medical care indexes are extremely difficult to price via traditional data collection, but with the potential of nontraditional data acquisition, BLS will be able to expand its sample and coverage across the country and to ultimately use this data in the official medical CPIs and improve data accuracy.

In general, nontraditional data collection has the potential to improve accuracy in the CPI by alleviating response rate issues and other data quality problems. However, the issues that arise from fitting alternative data into the CPI in a way that aligns with CPI measurement goals are often challenging. Continuing to appropriately expand the use of alternatively collected data to improve the CPI is both a challenge and an opportunity for BLS in the years ahead.

Weight updates

The CPI measures the change in the cost of living from one period to the next. Household spending weights are used to average the changes in component goods and services into the all-items index.

The overall goal of BLS is to use consumer spending from as recent a period as possible, and hold the set (or more precisely, the quantity mix) of goods and services purchased fixed over time (until new spending weights are introduced). In general, estimates of current-period inflation calculated with outdated spending weights tend to be higher than inflation estimates calculated with more current spending weights. This overestimation is because consumers change, or substitute, what they buy over time, often shifting purchases away from items that are becoming relatively more expensive to alternatives whose prices are not rising as fast.

In the first 80 years of producing the CPI, BLS updated the spending weights roughly every 10 to 15 years based on spending information collected in periodic household surveys.³⁴ In the 1980s the CE became continuous, and BLS began updating the CPI spending weights every 2 years, starting in 2002.³⁵ As such, the CE sample was increased to support more frequent weight updates.

Because of the collection and processing times of the CE data, the spending-weight-update schedule resulted in CPI spending weights lagging, on average, 36 months behind the date of the index. For example, consumer purchases made in 2017 and 2018 were used by BLS as the spending weights in January 2020 through December 2021 for the CPI-U, CPI for Urban Wage Earners and Clerical Workers (CPI-W), and CPI for Americans 62 years of age and older (R-CPI-E).

Beginning with the January 2023 indexes, BLS updates the CPI spending weights annually, reflecting spending from 2 years prior. For example, consumer purchases made in 2021 are used by BLS as the spending weights for the January through December 2023 CPI-U, CPI-W, and R-CPI-E indexes. The revised, annual weight-update schedule results in spending weights lagging, on average, 24 months behind the date of the index.

Additionally, BLS updates weights annually for the CPI on the basis of a single calendar year of data, using consumer expenditure data from 2021. This schedule reflects a change from prior practice of updating weights every 2 years using 2 years of expenditure data.

Table 1 compares the index values for January 2023 using the “new” weights (the official indexes based on the 2021 expenditure data) with the “old” weights (what the index values and percentages would be had the previous weights from 2019–20 remained in place). Table 2 shows the same comparison, but for 1-month nonseasonally adjusted percent changes.

A comparison of the index values from January 2023 using the “new” weights (the official indexes based on the 2021 expenditure data) versus the “old” weights (what the index values and percentages would be had the previous weights from 2019–20 remained in place) suggests the effect is relatively small.³⁶

There is one additional caveat to the change to annual weights. The R-CPI-E index uses a subset of the CE surveys sample consisting of urban households with a reference person or spouse age 62 or older. However, since the CE surveys’ sample design does not specifically target those age 62 and over, the number of households used for estimating aggregation weights for use in the research index is relatively small—about one-fifth of the urban CE surveys’ sample. This narrowing means the sample of households used as the basis for expenditure weights in the R-CPI-E is relatively small, leading to a higher sampling error than those used for the larger, official population. Regardless, BLS considers the benefit of the improvement in relevance to outweigh the potential increase in error for the R-CPI-E.

Conclusion

The marketplace, and the data available for characterizing it, has changed dramatically since the year the CPI was created in 1913. What consumers buy, how they buy it, and from where are almost unrecognizable when compared with prevailing norms from 100 years ago. Among the most striking differences are that many more products exist, outlet structures are much more diverse, product turnover (including introduction of new goods and services) has become increasingly rapid, and a higher proportion of the market basket consists of information goods and services.

Measurement of price change in a large economy is sufficiently complex that the accuracy of any estimate is difficult to gauge and is likely to be debated. The CPI cannot claim to be a completely precise measure of inflation. Over the years, potential sources of bias have been identified in the CPI, and subsequently addressed, though there continues to be debate over the extent and direction of any bias that may still exist and the ways in which BLS can continue to increase accuracy.

A key motivation driving CPI data modernization is the potential to improve timeliness, relative to survey alternatives, in detecting what consumers are buying and from where. At no time has this need for updated data collection methods been exposed more starkly than during the COVID-19 pandemic and the ensuing recovery period.

BLS will monitor changes in consumer buying habits and shifts in population and continue to make necessary changes to its methodology. By implementing annual CE surveys, BLS has the flexibility to monitor changing buying habits in a more timely and cost-efficient manner. In addition, the Census, conducted every 10 years by the U.S. Census Bureau, provides information that enables us to reselect a new geographic sample that accurately reflects population distribution and other demographic factors. BLS is continually researching improved statistical methods, so even between major revisions, improvements are made to the CPI.³⁷

The Bureau’s ability to use new data sources and methods for the CPI carries the potential to increase accuracy, detail, and timeliness. These new data sources present new methodological opportunities that will only enhance the data and ensure BLS continues providing gold-standard data now and into the future.