Professional Documents
Culture Documents
Internet Rising, Prices Falling Measuring Inflation in A World of E-Commerce
Internet Rising, Prices Falling Measuring Inflation in A World of E-Commerce
https://doi.org/10.1257/pandp.20181038
The e-commerce share of retail spending in 2003; Goolsbee and Klenow 2006; Brynjolfsson
the United States has almost tripled in the last and Oh 2012; and Varian 2013).
10 years and stands at close to 10 percent over- We document lower inflation online than in
all and more than 50 percent in several major the CPI—1.3 percentage points lower inflation
categories according to the US Census Bureau per year—for the same categories. The data also
(2018). If online pricing or inflation is funda- shows that the entry of new products and the exit
mentally different than traditional retail, it could of old products is extremely important for most
have a rising impact on the overall consumer categories of goods. In a constant elasticity of
price index (CPI) or bias in it.1 substitution (CES) model with a standard model
We use Adobe Analytics data on online trans- elasticity, the net entry of new goods during the
actions for millions of products in many differ- sample implies that m atched-model price indi-
ent categories from 2014 to 2017 to shed light on ces overstate true inflation by an additional 1.5
how online inflation compares to overall infla- to 2.5 percentage points per year.
tion, and to gauge the magnitude of new prod-
uct bias online. The Adobe data is similar to the I. Adobe Data
Billion Prices Project of Cavallo and Rigobon
(2016), which scrapes list prices from the web, Adobe Analytics provides a variety of services
except the Adobe data contains actual transac- to e-commerce merchants who share their trans-
tion prices and includes quantities purchased. action data for Adobe to analyze. Adobe clients
We follow two literatures. One uses detailed include 20 of the 30 largest employers in the
scanner data from grocery stores to analyze nation and 80 percent of Fortune 500 retailers.
new product introductions, such as Broda and The underlying data are quantities and reve-
Weinstein (2010). Another studies consumer nue from individual transactions (not including
surplus from the internet and e-commerce in taxes or shipping costs). The product codes are
particular (e.g., Brynjolfsson, Hu, and Smith merchant-specific so our definition of a good
will be the good-merchant combination. We
* Goolsbee: Booth School of Business, University of
use Adobe’s data aggregated up to the monthly
Chicago, Chicago, IL 60637, and NBER (email: goolsbee@ level: total quantity and average transaction
chicagobooth.edu); Klenow: Department of Economics, price for each product for a given month. Adobe
Stanford University, Stanford, CA 94305, and NBER anonymizes the data so we cannot identify any
(email: klenow@stanford.edu). Mathias Jimenez and retailers or customers.
Matteo Leombroni provided superb research assistance;
Alberto Cavallo, Charles Hulten, and Leonard Nakamura We use a subset of the products and mer-
helpful comments; and Luis Maykot and Siddharth Kulkarni chants from the full Adobe dataset. This subset
assistance in accessing and understanding the Adobe data. covers categories making up about 19 percent of
†
Go to https://doi.org/10.1257/pandp.20181038 to visit the CPI relative importance weights in Bureau
the article page for additional materials and author disclo-
sure statement(s).
of Labor Statistics (2018), and the revenue
1
Economists have long known about the potential for amounts to about 15 percent of all e-commerce
new products to bias upward the inflation measured in the in US Census Bureau (2018). The sample goes
CPI. See Boskin et al. (1996) and, more recently, Groshen from January 2014 through September 2017.
et al. (2017). Recent business press articles have argued that Table 1 shows the number of products in the
online commerce may be leading to growing problems in
measuring inflation and complicating monetary policy deci- Adobe data overall and by CPI major group,
sions at the Federal Reserve, as in Cohan (2017), Torrey and averaged over the January 2014 to September
Stevens (2017), and Gross (2017). 2017 period. The dataset contains over 2 million
488
VOL. 108 INTERNET RISING, PRICES FALLING 489
Index
products
96 96
Headline 19 2.1M DPI
94 94
Food and beverages 49 1M
Housing 7 90 K 92 92
Apparel 100 128 K
Education and communication 9 400 K
7
−1
−1
−1
−1
−1
−1
−1
−1
Medical care 9 23 K
l
Ju
Ju
Ju
Ju
Transportation 3 125 K
Ja
Ja
Ja
Ja
Recreation 32 200 K
Other goods and services 42 90 K Figure 1. Cumulative Inflation, DPI versus CPI
Notes: Headline is all covered categories. The next rows are Notes: For the 65 ELIs covered by the Adobe digital price
CPI major groups. The last column gives the average num- index (DPI). CPI relative importance weights for each ELI.
ber of products from 2014 through 2017. The second to last Source: Authors’ calculations using Adobe Analytics and
column gives the monthly average percentage of ELI cover- BLS data.
age for the respective category.
Source: Authors’ calculations using Adobe Analytics and the CPI and the DPI. In this way we can rule out
Bureau of Labor Statistics (BLS) data.
that differences between the two indexes arise
from categories which are not covered by the
products in the average month, versus about DPI or the weighting of categories covered.
140,000 per month in the entire CPI. There are We plot the two indices together in Figure 1.
211 CPI categories known as entry level items The DPI exhibits notably more deflation over
(ELIs), and the Adobe data covers 65 of them.2 the period than the CPI for the same categories.
Table 2 shows the average annual inflation rates
II. DPI versus CPI Inflation from 2014–2017. Overall (headline) DPI infla-
tion is more than 1 percentage point per year
We call the matched model price index we lower than CPI inflation. Breaking out by major
construct from the Adobe data the digital price groups, inflation is lower in the DPI than in the
index (DPI) to distinguish it from the CPI. We CPI in every category other than medicine and
start with price changes for overlapping prod- medical supplies.
ucts in months t − 1and t. These are products Now, excess deflation in high frequency,
selling positive quantities in both months. We chain-weighted price indices can result from
take log first difference of average unit prices. To oscillating prices due to recurring discounts.
aggregate price changes across products within This phenomenon is known as “chain drift.”
an ELI, we use Tornqvist weights. These are the Even if the prices and quantities revert to their
average spending share of the product in the ELI starting levels, a chained price index may not
in months t − 1 and t. The spending shares are revert to 1. This has been documented in grocery
based on Adobe data for overlapping products.3 store scanner data by, for example, de Haan and
To facilitate comparison with the CPI, we van der Grient (2011).
aggregate the Adobe ELI inflation rates using To gauge chain drift in the Adobe data, in each
Laspeyres weights. In our base case, we use CPI year we tried adding an artificial “13th month”
relative importance weights for each E LI-month. with the first month’s, prices and quantities and
We use the same set of ELIs to construct both asked whether the price index returns to 1. We
found no evidence of such chain drift. In fact,
the bias went the other way—with the index
2
The Adobe dataset covers home furnishings (appli-
ances, furniture, etc.) but not rent or owner’s equivalent rent,
moving above 1 in the 13th month. Thus, chain
hence we use the label “Household Goods” in our analy- drift does not seem to be a source of deflationary
sis. Similarly, we refer to “Information Technology” as the bias in the DPI relative to the CPI.
goods within Education and Communication that the Adobe
data covers. III. Product Entry and Exit
3
Like the BLS, we do something special for apparel. We
construct a simple index of average unit prices. This is to
avoid extreme deflation from fashion and seasonal cycles for Because the Adobe data include quantities as
clothing. well as prices, we are able to look at spending
490 AEA PAPERS AND PROCEEDINGS MAY 2018
Table 2—Average Annual Inflation Table 3—Adobe Entry and Exit Rates
also report a markedly lower exit rate for food items than
other items.
4 6
We weight by the average monthly sales of a product This should capture improvements in product quality in
during the calendar year across the months the product was addition to brand new types of products, because both are
available. associated with new product ID codes in the Adobe data.
VOL. 108 INTERNET RISING, PRICES FALLING 491
Internet.” American Economic Review 96 (2): Bureau of Economic Research Working Paper
108–13. 22331.
Groshen, Erica L., Brian C. Moyer, Ana M. Torrey, Harriet, and Laura Stevens. 2017. “As
Aizcorbe, Ralph Bradley, and David M. Fried- the Fed Deliberates, Amazon is Making its Job
man. 2017. “How Government Statistics Adjust Difficult.” Wall Street Journal, December 11.
for Potential Biases from Quality Change and https://www.wsj.com/articles/u-s-shoppers-
New Goods in an Age of Digital Technologies: wield-smartphones-to-keep-a-lid-on-
A View from the Trenches.” Journal of Eco- consumer-prices-1513036486.
nomic Perspectives 31 (2): 187–210. US Bureau of Labor Statistics. 2018. Archived
Gross, Daniel. 2017. “How Amazon Fuels Defla- Consumer Price Index Supplemental Files.
tion.” Strategy + Business, August 30. https:// Washington, DC: U.S. Bureau of Labor Sta-
www.strategy-business.com/blog/How- tistics.
Amazon-Fuels-Deflation?gko=1dfcc. US Census Bureau. 2018. Monthly and Annual
Hottman, Colin J., Stephen J. Redding, and David Retail Trade. Washington, DC: U.S. Census
E. Weinstein. 2016. “Quantifying the Sources Bureau.
of Firm Heterogeneity.” Quarterly Journal of Varian, Hal. 2013. “The Value of the Internet
Economics 131 (3): 1291–364. Now and in the Future.” Economist, March 10.
Kaplan, Greg, and Sam Schulhofer-Wohl. 2016. https://www.economist.com/blogs/
“Inflation at the Household Level.” National freeexchange/2013/03/technology-1.