September 2016
Estimation of Aggregate Demand and Supply Shocks Using
Commodity Transaction Data*
ABE Naohito† (Hitotsubashi University)
INAKURA Noriko‡ (Osaka Sangyo University)
TONOGI Akiyuki§ (Hitotsubashi University)
Abstract
Using commodity-level transaction data for Japan, we estimate aggregate demand and
supply shocks. When the effects of product turnover are ignored, the estimated aggregate
demand shocks are generally negative, except for periods before the Great Financial
Crisis and after mid-2015, even though Japan increased its consumption tax rate in early
2014. However, when the effects of product turnover are considered, the aggregate
demand shocks become positive from mid-2013, which is more consistent with the
conventional view of stockpiling behavior occurring before a tax rate increase. After 2013,
the supply shocks become negative and remain at a very low level, which suggests that
increases in prices during those periods reflect, at least in part, inward shifts of supply
curves.
Keywords: Demand shocks, Supply shocks, Elasticity, Scanner data
JEL Classifications: D40, D12, D22, E30, E32
This paper is conducted as a part of a project entitled “Decomposition of Economic Fluctuations for
Supply and Demand Shocks: Service Industries” undertaken at the Research Institute of Economy,
Trade, and Industry (RIETI) and a product of a joint research project conducted with the Institute of
Economic Research, Hitotsubashi University, the New Supermarket Association of Japan, and
INTAGE Inc. The authors are grateful for comments by Yoshihiko Nishiyama, Yoko Konishi, Kyoji
Fukao, and seminar participants at RIETI and Hitotsubashi University. This work was supported by a
Japan Society for the Promotion of Science Grant-in-Aid for Scientific Research (A) (No.15H01945)
and (C) ((No.15K03349).
† Corresponding Author. Institute of Economic Research, Hitotsubashi University , Naka, Kunitachi,
Tokyo, 186-8603, JAPAN. E-mail: [email protected].
‡ Osaka Sangyo University
§ Institute of Economic Research, Hitotsubashi University
*
1
1. Introduction
The source of business cycle fluctuations is one of the most important factors to
consider when choosing appropriate macroeconomic policies to address the fluctuations.
If the fluctuations are caused by supply-side factors, such as an increase in the imported
oil price or natural disasters, then countercyclical fiscal or monetary policies, intended to
increase domestic demand, might cause stagflation. On the other hand, if the fluctuations
are caused by demand-side factors, such as the Global Financial Crisis (GFC) that
commenced in September 2008, then fiscal policies, such as increases in government
purchases, might contribute to mitigating such shocks.
Because of the significance of this topic, a number of econometric methods have been
proposed to identify the source of shocks. As far as we are aware, all the papers to date
that identify demand and supply shocks in business cycles utilize macroeconomic data
and vector autoregressive models. 1 This contrasts sharply with the situation in many
other macroeconomic fields of study—such as studies of consumption, unemployment,
and investment—in which an increasing number of researchers employ micro rather than
macro data to identify macroeconomic structures.
In fields of applied microeconomics, such as industrial organization, the identification
of demand and supply curves and their shifts has been the active focus of many studies
for a number of decades. Since Working (1927), studies have proposed a variety of
identification procedures, such as instrumental variables or full structural models of
production and demand, to identify the shapes of the demand and supply curves.2
The purpose of this paper is to bridge the divide between these micro and macro
approaches. More specifically, we estimate aggregate demand and supply shocks using
micro data. Using a large commodity transaction data set from Japan, we are able to
estimate product category-level demand and supply curves. Based on these estimated
demand and supply curves, we then identify demand and supply shocks that shift either
or both curves and, then, we aggregate the demand and supply shocks to obtain the
macroeconomic demand and supply shocks. Our weekly frequency data set covers the
period between 2007 and 2015 in Japan. During this period, a number of significant
shocks took place, including the GFC, the Great East Japan Earthquake (GEJE) in March
2011, and, in April 2014, an increase in Japan’s rate of consumption tax. We quantify the
impact of these shocks on demand and supply.
1
See, e.g., Blanchard and Quah (1989) and Uhlig (2005).
For instance, see Berry et al. (1995) for an instrumental variable approach and Byrne et al. (2015)
for a structural approach.
2
2
When estimating demand and supply curves, we employ two types of estimation
strategies: one based on the commodity code (barcode) level of information; and another
based on producer-level information. We consider these to be complementary approaches.
When using the barcode-level micro data set, we treat commodities with different
commodity codes as different goods. Thus, we do not encounter empirical complexities
such as the quality adjustments that are frequently required when constructing price
indexes. However, because our estimation requires yearly differences in both prices and
quantities to control for seasonal movements, we have to restrict our data set to goods
that are present in both the base and the current periods. This means that we need to
remove around 30% of the goods in the data set that appear for the first time between the
base and current periods. Alternatively, if we use producer-level information, we need to
aggregate different goods produced by the same manufacturers. Because we take into
account more than a billion commodities, based on detailed barcode-level characteristics,
it is practically impossible for us to conduct detailed quality adjustments, such as hedonic
regressions. Therefore, we need to adopt a simplified approach, using unit value prices
and volumes.
More specifically, we first convert the prices and quantities of all goods belonging to
the same commodity category to the same unit value, such as price per gram, and then
conduct aggregations over the same producers and categories. The advantage of this
method is that we can include new goods that appear between the base and current periods.
As Bernard et al. (2010) argue, if product turnover plays a significant role in business
cycles, ignoring it when estimating demand and supply shocks could result in serious
bias.3 We consider that these two estimates of the aggregate demand and supply shocks—
one based on barcode-level data that ignores new goods, and the other based on producerlevel aggregate unit values that assumes all products have the same qualities apart from
their volumes—are extremes. We assume that the actual aggregate demand and supply
shocks lie somewhere between these two extremes.
The differences in the estimated aggregate demand shocks between the two approaches
are large for two of the periods during our sample, namely, the period before the GFC in
2008 and the period preceding the change in Japan’s consumption tax in April 2014. The
demand shocks based on the barcode-level information are negative before the GFC but
3
Feenstra (1994) shows that, in some circumstances, we can capture the effects of product turnover
on economic welfare through the changes in product variety. Although there is an emphasis on variety
expansion in the macroeconomic growth literature, for commodity level data, we suspect a blue
bus/red bus problem, that is, for most people, an addition of new bus with different color might not
increase economic welfare through variety expansion, is serious. Thus, we do not take this approach.
See Nevo (2010) for more details on a blue/red bus problem when using commodity level data.
3
virtually zero before the consumption tax increase. Conversely, the estimates of the
aggregate demand shocks based on the producer-level information are positive before the
GFC and positive before the change in the consumption tax. Both estimates are negative
following the GFC. These differences suggest that product turnover plays a significant
role in Japanese business cycles. However, in contrast to the aggregate demand shocks,
the two sets of estimates do not result in large discrepancies in the measurement of the
aggregate supply shocks. Both estimates indicate positive aggregate supply shocks after
the GEJE. The supply shocks also exhibit temporary increases before the change in the
rate of the consumption tax in April 2014. During other periods, the supply shocks are
negative, but the negative supply shocks increased in magnitude after the change in the
consumption tax rate. Along with the positive aggregated demand shocks, the negative
supply shocks contributed to an increase in the price index after 2014.
The remainder of the paper is organized as follows. Section 2 describes the model and
Section 3 details the identification strategy. Section 4 describes the data set used and
Section 5 discusses the estimation results. In Section 6, we discuss time-varying
elasticities. Section 7 concludes the paper by providing a brief summary and identifying
some prospects for future research.
2. The Model
The representative consumer has the following separable utility function at time t:
𝑈𝑡 = 𝑈(𝐶𝑡1 , 𝐶𝑡2 , … , 𝐶𝑡𝐽 ),
𝜎𝑗 −1
𝑗
𝐶𝑡 = (∑𝑖∈Θ𝑗 𝑎𝑡𝑖 𝑥𝑡𝑖
𝜎𝑗
𝜎𝑗
𝜎𝑗 −1
)
𝑡
, 𝑎𝑡𝑖 ≥ 0, , 𝜎𝑗 > 0,
𝑗
where 𝑥𝑡𝑖 is consumption of commodity i at time t, 4 Θ𝑡 is the commodity space of
𝑗
category j at time t, 𝐶𝑡 is the aggregate consumption of category j at time t, 𝑎𝑡𝑖 is the
time-varying weight for commodity i at time t, and 𝜎𝑗 is a constant parameter. 𝑈 is
a twice-continuously differentiable utility function that satisfies standard assumptions on
preferences. Because the utility function is separable across categories and, given a
constant elasticity of substitution (CES) for the category-level utility function, the optimal
4
Note that commodities that have the same bar code (commodity code), b, such as the universal
product code, but that are sold at different stores, are treated as different products in this paper.
4
consumption for commodity i, given the category aggregate, is given by the following
simple compensated demand function:
𝜎𝑗
1−𝜎𝑗
𝑗
𝑥𝑡𝑖 = Ct ( ∑ 𝑎𝑡𝑘
𝜎𝑗
𝑝𝑡𝑘
1−𝜎𝑗
𝜎
𝑎𝑡𝑖 𝑝𝑡𝑖
)
−𝜎𝑗
.
𝑗
𝑘∈Θ𝑡
Denoting,
1
1−𝜎𝑗
𝑗
𝑃𝑡 = ( ∑ 𝑎𝑡𝑘
𝜎𝑗
𝑝𝑡𝑘
1−𝜎𝑗
)
𝑗
𝑘∈Θ𝑡
and taking logged time differences, we obtain:
𝑗
𝑗
∆ln(𝑥𝑡𝑖 ) = ∆ ln(𝐶𝑡 ) − 𝜎𝑗 ∆ ln(𝑝𝑡𝑖 ) + 𝜎𝑗 ∆ ln(𝑃𝑡 ) + 𝜀𝑡𝑖 ,
(1)
where 𝜀𝑡𝑖 = 𝜎𝑗 ∆ ln(𝑎𝑡𝑖 ).
𝑗
The first term on the right-hand side of (1), ∆ ln(𝐶𝑡 ), represents the income effect, the
second and third terms reflect the effects of the relative prices, and the final term, 𝜀𝑡𝑖 ,
captures the commodity-specific demand shocks. Note that the category-specific demand
𝑗
𝑗
shocks, ∆ ln(𝐶𝑡 ) and ∆ ln(𝑃𝑡 ), do not include subscript i, which implies that these
shocks are common to all commodities belonging to the same category. Given that
category-specific demand shocks are likely to be related to supply shocks, we need to
eliminate these components from (1) when estimating the demand elasticity, 𝜎𝑗 . For this
purpose, Feenstra (1994) proposed taking the difference from a reference country to
control for the shocks. Similarly, in our context, it would be possible to take the difference
from a reference good, such as the commodity with the largest market share.
However, we do not take this approach for two main reasons. The first reason is that
the high rate of commodity turnover makes it difficult to identify a reference good. As
described in Section 3, our data set contains a number of processed foods, daily
necessities, cosmetics, and drugs. In these categories of goods, it is very difficult to find
a commodity that is a market leader for the entire subsample period (more than nine years).
The second reason that we avoid this approach is that, in taking the difference from a
reference good, we would remove the observations relating to the reference good itself,
which would result in the loss of a large amount of information. Instead, we take the
5
difference between the quantity of commodity i from the categorical average quantity
without commodity i, but within the same barcode, b. More specifically, we use the
following double difference:5
∆ln(𝑥𝑡𝑖 ) −
1
∑
𝑗,𝑏
#(Θ𝑡 ) − 1
∆ln(𝑥𝑡𝑘 )
𝑗,𝑏
𝑘∈Θ𝑡 ,𝑘≠𝑖
= −𝜎𝑗 (∆ ln(𝑝𝑡𝑖 ) −
𝜀̃𝑡𝑖 = 𝜎𝑗 ∆ ln(𝑎𝑡𝑖 ) −
1
∑
𝑗,𝑏
#(Θ𝑡 ) − 1
1
𝑗,𝑏
#(Θ𝑡 ) − 1
∆ln(𝑝𝑡𝑘 )) + 𝜀̃𝑡𝑖
𝑗,𝑏
𝑘∈Θ𝑡 ,𝑘≠𝑖
∑
𝜎𝑗 ∆ ln(𝑎𝑡𝑘 ),
𝑗,𝑏
𝑘∈Θ𝑡 ,𝑘≠𝑖
𝑗,𝑏
where #(Θ𝑡 ) is the number of products included in the product set for the barcode b,
𝑗,𝑏
Θ𝑡 . For simplification, we denote the following:
𝑥̃𝑡𝑖 = ∆ln(𝑥𝑡𝑖 ) −
𝑝̃𝑡𝑖 = ∆ ln(𝑝𝑡𝑖 ) −
1
𝑗,𝑏
#(Θ𝑡 ) − 1
1
𝑗,𝑏
#(Θ𝑡 )
−1
∑
∆ln(𝑥𝑡𝑘 ),
𝑗,𝑏
𝑘∈Θ𝑡 ,𝑘≠𝑖
∑
∆ln(𝑝𝑡𝑘 ).
𝑗,𝑏
𝑘∈Θ𝑡 ,𝑘≠𝑖
Then, the equation becomes:
𝑥̃𝑡𝑖 = −𝜎𝑗 𝑝̃𝑡𝑖 + 𝜀̃𝑡𝑖 .
(2)
Following Feenstra (1994) and Broda and Weinstein (2010), we assume the following
simple supply function:
𝑗
∆ ln(𝑥𝑡𝑖 ) = 𝜔𝑗 ∆ ln(𝑝𝑡𝑖 ) + 𝑆t + 𝛿ti ,
(3)
where 𝜔𝑗 is a constant parameter, 𝛿ti is a supply shock that shifts the supply curve, and
𝑗
𝑆t is a category-specific shock that is common across commodities in category j that
could be correlated with the demand shocks. Taking additional differences within the
same barcode, as for the demand curve, we obtain:
5
Note that by taking the difference from the barcode-level mean, we not only control for categoryspecific shocks, but also for commodity (barcode)-specific shocks.
6
∆ln(𝑥𝑡𝑖 ) −
1
𝑗,𝑏
#(Θ𝑡 )
−1
∑
∆ln(𝑥𝑡𝑘 )
𝑗,𝑏
𝑘∈Θ𝑡 ,𝑘≠𝑖
= 𝜔𝑗 (∆ ln(𝑝𝑡𝑖 ) −
𝛿̃𝑡𝑖 = 𝛿ti −
1
𝑗,𝑏
#(Θ𝑡 ) −
1
𝑗,𝑏
#(Θ𝑡 )−1
1
∑
∆ln(𝑝𝑡𝑘 )) + 𝛿̃𝑡𝑖
𝑗,𝑏
𝑘∈Θ𝑡 ,𝑘≠𝑖
∑𝑘∈Θ𝑗,𝑏 ,𝑘≠𝑖 𝛿tk .
𝑡
Using the same notation as in (2), we obtain the following supply curve:
𝑥̃𝑡𝑖 = 𝜔𝑗 𝑝̃𝑡𝑖 + 𝛿̃𝑡𝑖 .
(4)
3. Identification
Our main identification assumption for estimating the elasticities, 𝜎𝑗 and 𝜔𝑗 , is the
orthogonality between the shocks for supply and demand, 𝛿̃𝑡𝑖 and 𝜀̃𝑡𝑖 . Note that these are
error terms, after controlling for the category-specific time-varying shocks, such as
natural disasters and large macroeconomic shocks, which could shift both the demand
and the supply curves at the same time. As we have two unknown parameters, 𝜎𝑗 and
𝜔𝑗 , we require two or more moment conditions to identify them. In this analysis, we use
the following three moment conditions:6
𝐸[𝛿̃𝑡𝑖 𝜀̃𝑡𝑖 ] = 0,
2
E [𝛿̃𝑡𝑖 𝜀̃𝑡𝑖 ] = 0,
2
E [𝛿̃𝑡𝑖 𝜀̃𝑡𝑖 ] = 0,
Note that the obvious moment conditions, 𝐸[𝛿̃𝑡𝑖 ] = 0 or 𝐸[𝜀̃𝑡𝑖 ] = 0, cannot be used for estimating
the elasticities because, by construction, the sample moments of 𝛿̃𝑡𝑖 and 𝜀̃𝑡𝑖 are always zero for all j
and t, i.e., ∑𝑖∈Θ𝑗,𝑏 𝛿̃𝑡𝑖 = 0 and ∑𝑖∈Θ𝑗,𝑏 𝜀̃𝑡𝑖 = 0.
6
𝑡
𝑡
7
which implies that the model is overidentified. 7 As pointed out by Leamer (1981),
identification based on orthogonality between the residuals cannot yield unique estimates
of the parameters. The reason for this is simple. Suppose that a set of parameters (𝜎̂, 𝜔
̂) =
(𝛼, 𝛽) satisfies the above moment conditions. Then, another set, (𝜎̂, 𝜔
̂) = (𝛽, 𝛼), also
satisfies the moment conditions because the system is symmetric for demand and supply.
In our analysis, we consider an estimate with a negative slope as indicating the elasticity
of the demand curve, whereas one with a positive slope indicates the elasticity of the
supply curve. Therefore, if the estimated pair of elasticities indicates that both slopes have
the same sign, we remove the category.8
After obtaining the estimates of the elasticities, (𝜎̂, 𝜔
̂), we plug them into (1) and (3),
which gives us the following two sets of shocks:
𝜀̂𝑡𝑖 ≡ ∆ln(𝑥𝑡𝑖 ) + 𝜎̂∆ ln(𝑝𝑡𝑖 ),
and
𝛿̂𝑡𝑖 ≡ ∆ ln(𝑥𝑡𝑖 ) − 𝜔
̂∆ ln(𝑝𝑡𝑖 ),
where 𝜀̂𝑡𝑖 and 𝛿̂𝑡𝑖 contain both commodity- and category-specific shocks that shift the
category-level demand and supply curves, respectively.
The final step of the estimation is aggregation over commodities and categories. We
use the Törnqvist weights for sales for both aggregations. It is also possible to adopt a
Sato–Vartia-type aggregation formula, given the estimates of the CES utility function.9
In this paper, we choose to employ the Törnqvist weights for two reasons. First, the
Törnqvist index is a superlative index, considered to be a good approximation for a large
class of expenditure/cost functions.10 Because of these characteristics, we do not need to
7
Feenstra (1994) and others, including Broda and Weinstein (2010) and Soderbery (2015), used only
a single condition, 𝐸[𝛿̃𝑡𝑖 𝜀̃𝑡𝑖 ] = 0, instead of three moment conditions. To identify the two parameters,
2
2
2
Feenstra (1994) used the relationship 𝛿̃𝑡𝑖 𝜀̃𝑡𝑖 = 𝑥̃𝑡𝑖 + (𝜎𝑗 − 𝜔𝑗 )𝑥̃𝑡𝑖 𝑝̃𝑡𝑖 − 𝜔𝑗 𝜎𝑗 𝑝̃𝑡𝑖 and regressed 𝑝̃𝑡𝑖 on
𝑥̃𝑡𝑖
2
and 𝑥̃𝑡𝑖 𝑝̃𝑡𝑖 , with the product indicators as instruments. Although this approach is frequently used
2
in the literature, to implement it, 𝛿̃𝑡𝑖 𝜀̃𝑡𝑖 needs to be uncorrelated with 𝑥̃𝑡𝑖 𝑝̃𝑡𝑖 and 𝑥̃𝑡𝑖 , which requires
additional assumptions in models, such as time-invariant parameters. In Section 6, we consider a
model with time-varying elasticities, which means that we cannot adopt the procedure of Feenstra
(1994) but, instead, we conduct a nonlinear estimation with additional moment conditions.
8
When estimating the elasticities, we used STATA command for the generalized method of moments
(GMM) with a diagonal weight matrix.
9 Sato (1976) provides details.
10 See Diewert (1976).
8
specify the functional form of the cost function as the micro foundation of the supply
curve. Second, aggregation with the Törnqvist weights creates very similar results to
those found using other formulas, such as the Sato–Vartia formulation.
One drawback of the above approach is the restriction imposed on the product space.
To take the first differences in price and quantity for each product, we need information
on price and quantity in two different periods, i.e., the current period and the base period.
In other words, we are unable to calculate first differences for new products that do not
exist in any base period. Therefore, we are obliged to ignore these observations in our
estimations. As long as the appearance of new products is uncorrelated with either
demand or supply shocks, exclusion of new (and even withdrawn) goods might not be a
serious problem when estimating demand and supply shocks.
However, if producers experiencing large supply shocks, or households experiencing
large demand shocks, alter their product choices more often than those not experiencing
shocks, an estimation based only on continuing goods could lead to a biased estimation
of the magnitude of shocks. We could consider the effects of product turnover on the cost
of living index by including variety-expansion effects (and their corresponding terms) in
(1) and (3), as proposed by Feenstra (1993). 11 However, if the effects of variety
expansion or contraction are common to all of the commodities in the same category, then
they are already included in our model as the category-specific effects. That is, our
estimates of the aggregate demand and supply shocks include the effects through changes
in the variety expansion as long as the variety effects shift the demand or supply curves
to the same extent within each category.
To consider the effects of new products that are not category-level common shocks,
we need to include new products when estimating the demand and supply elasticities.
Consideration of product turnover and new products has been a central problem in the
construction of price indexes. Although overlap and hedonic methods are widely used
procedures, unfortunately they are virtually impossible to use with our large-scale scanner
data. One practical method is to use unit value prices and quantities so that all the products,
including new and discontinued products, are evaluated at a price per gram or liter. The
obvious drawback with this approach is that we are unable to control for quality
differences between products. To mitigate this drawback, in our paper, we assume that
the quality of products is heterogeneous, but that products made by the same producers
have the same quality. That is, the differences in quality between products are assumed to
be captured by the differences between producers. More specifically, we first construct a
total volume and unit value price by dividing total sales by total volume (i.e., price per
11
See Feenstra (1994) for more detail.
9
gram) for each category and producer within a store. Then, after taking the first difference,
we take additional differences from the average quantity for each producer within the
same category, after excluding the same producer to control for category-specific demand
and supply shocks. With these variables, we conduct an estimation of the demand and
supply elasticities and shocks. We refer to these estimation results based on producerlevel unit values as producer-level estimates. That is, we adopt two extreme strategies,
one employing barcode-level information that disregards product turnover, and the other
employing producer-level unit values that disregards the heterogeneity of commodities
produced by the same producer. We assume that the reality lies somewhere between these
two results.
4. Data
We use Japanese store-level scanner data, known as SRI,12 collected by INTAGE Inc.
The data set include sales records of processed foods, daily necessities, cosmetics, and
drugs, each with a Japanese Article Number (JAN) code. 13 , 14 The sample period is
between January 2007 and February 2016. The data set covers about 3,000 stores, located
all over Japan, that can be classified into four different types: general merchandise stores
(GMS), supermarkets (SMT), drug stores (DRG), and convenience stores (CVS).
One of the noteworthy characteristics of the data set is its detailed commodity
classification, with commodities classified into 1,057 different subcategories. This is
about seven times more than the number of classifications adopted by the Japanese
official statistics for similar types of goods. Table 1 presents the basic statistics for the
data set and Figure 1 illustrates some of the key variables from the table.
SRI is the abbreviation for “Syakaichosa-kenkyujo Retail Index,” translated from Japanese as the
Retail Index by the Institute of Social Research. The original frequency of the data set is weekly. In
this paper, we transform the data set to monthly data to mitigate the effects of bargain sales, which
might give us extremely large estimates of elasticities.
13 The data set provides pretax price information.
14 Fresh foods are not included in the data set because of a lack of suitable commodity codes.
12
10
Table 1: Summary of Monthly Transaction Data
Mean
GMS
Total
Sales (million yen)
Continuing Goods
New Goods
New Product Ratio
Number of Stores
Average Number of
Categories within a Store
Producers within a Store
Items within a Store
Foods
Sales (million yen)
Continuing Goods
New Goods
New Product Ratio
Number of Stores
Average Number of
Categories within a Store
Producers within a Store
Items within a Store
Daily Commodities
Sales (million yen)
Continuing Goods
New Goods
New Product Ratio
Number of Stores
Average Number of
Categories within a Store
Producers within a Store
Items within a Store
SMT
DRG
Std. Dev.
CVS
GMS
SMT
DRG
CVS
12,851
9,025
3,826
0.30
204.40
19,713
14,271
5,442
0.28
989.46
6,309
4,272
2,037
0.32
1010.32
2,432
1,252
1,180
0.48
743.35
748
803
552
0.04
8.69
1,115
1,111
895
0.04
23.28
457
445
203
0.04
18.30
269
133
166
0.03
30.84
685.01
2422.77
7584.47
574.32
1675.84
4524.23
340.66
1116.62
3687.51
156.31
397.96
989.70
13.25
109.56
678.68
6.20
49.90
268.06
28.68
93.27
435.07
6.57
9.91
35.43
11,147
7,910
3,237
0.29
204.40
18,372
13,389
4,983
0.27
989.46
2,543
1,779
764
0.30
1004.20
2,374
1,214
1,160
0.49
743.35
654
698
492
0.04
8.69
1,122
1,090
840
0.04
23.28
324
277
104
0.04
18.37
265
132
164
0.03
30.84
565.15
1895.46
5353.00
490.67
1430.42
3687.88
286.41
745.38
2047.21
124.80
345.10
896.10
11.05
88.66
454.54
6.37
50.67
246.65
24.20
77.49
282.60
5.33
9.38
37.79
1,705
1,115
590
0.35
204.40
1,341
882
459
0.34
985.66
3,766
2,493
1,273
0.34
1010.03
58
38
20
0.34
743.29
148
132
83
0.05
8.69
88
71
80
0.05
23.96
211
217
128
0.03
18.28
11
6
6
0.05
30.82
121.86
548.19
2349.56
90.55
280.54
1001.87
111.05
525.08
2051.13
32.90
55.48
99.55
2.91
27.08
265.37
1.06
5.10
45.39
2.82
17.56
172.19
2.10
3.82
9.06
Note: Monthly point of sale data for GMS, STM, DRG, and CVS from January 2007 to February,
2016. Tobacco is not included. “Foods” include processed food such as milk, bottled water, canned or
frozen foods, pasta, sugar, and salt. “Daily Commodities” include various items available in the above
stores, such as detergent, shampoo, and toilet papers.
11
Figure 1: Comparison between Store Types
(1) Volume of Sales (Continuing vs. New Goods) (2) Volume of Sales (Food vs. Daily Commodities)
(million yen)
(million yen)
20,000
20,000
15,000
15,000
10,000
10,000
5,000
5,000
0
0
GMS
SMT
DRG
Continuing Goods
GMS
CVS
Foods
New Goods
(3) New Product Ratio (Food vs. Daily Commodities)
0.49
0.50
0.35
0.40
0.29
0.34
0.27
0.30
0.34
989
1,010
743
800
0.34
CVS
(4) Number of Stores
1000
600
0.20
400
0.10
200
204
0
0.00
GMS
SMT
Foods
DRG
GMS
CVS
SMT
DRG
CVS
Number of Stores
Daily Commodities
(5) Number of Categories within a Store
800
700
600
500
400
300
200
100
0
DRG
Daily Commodities
1200
0.60
0.30
SMT
(6) Number of Producers within a Store
3000
685
2500
574
2,423
2000
341
1,676
1500
1,117
1000
156
398
500
0
GMS
SMT
DRG
GMS
CVS
SMT
DRG
CVS
Producers within A Store
Categories within A Store
Note: When calculating the number of producers within stores, we count the number for each product category, and
aggregate them over the categories. That is, if a retail store sells two products that belong to different categories, but
are produced by the same manufacturer, we double count the producers when aggregating the number of producers
over categories.
12
In this paper, a continuing good is defined as a commodity that exists both in the
current month and in the same month one year before. In other words, continuing goods
are commodities for which we can calculate yearly differences in prices and quantities.
On the other hand, a new good is defined as a commodity without a sales record in the
same month one year before. As shown in Table 1, a significant number of new goods are
present in the data set, representing about 30% of all goods included.
Figure 2 illustrates the changes in the sales of new goods compared with total goods.
A sharp increase before the GFC in 2008 and a sharp decrease after the GFC in 2009 is
evident. The surge in the new product ratio in 2012 reflects the shortage of goods
immediately following the GEJE in March 2011. In early 2014, the new product ratio
began to increase again. As convenience stores have completely different product cycles
compared with the other store formats, we exclude convenience stores in the subsequent
analysis.
Figure 2: New Product Ratio
Foods
Daily Commodities
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
1 6 11 3 8 13 5 10 2 7 12 4 9 1 6 11 3 8 13 5 10 2 7 12
1 6 11 3 8 13 5 10 2 7 12 4 9 1 6 11 3 8 13 5 10 2 7 12
2007 2008 2009 2010 2011 2012 2013 2014 20152016
GMS
SMT
DRG
CVS
2007 2008 2009 2010 2011 2012 2013 2014 20152016
GMS
13
SMT
DRG
CVS
Figure 3 plots the movements of the aggregate first-difference series of prices and
quantities. Figure 4 depicts the movements of price and quantity when we use the
producer-level unit value price and total volume, which includes new goods as well as
incumbent goods. Because we use the unit value, the product space used in Figure 4 is
much larger than that in Figure 3. That is, new goods, for which there are no data one year
prior, are included in Figure 4 but not in Figure 3.19 Figures 5 and 6 compare the changes
in quantities and prices between the barcode-level aggregate and the producer-level
aggregate, based on the unit value. As is clear in Figure 5, the changes in quantities are
generally negative when we use barcode-level information. We suspect that these
negative values reflect the movements of products that are about to exit from the market.
That is, the changes in quantities are subject to a serious downward bias when we use
continuing goods only. If we include new goods, as shown in Figure 5, the changes in
quantities increase significantly.
From Figure 6, we can observe that the changes in prices are very similar between the
barcode-level and producer-level aggregates. However, a notable difference emerges in
mid-2013, when we observe higher price increases in the producer-level aggregate,
suggesting that significant effects arose from the introduction of relatively more
expensive goods during that period.
19
Aggregate changes in price and quantity are the weighted average of the categorical-level first
differences using the Törnqvist sales weights.
14
Figure 3: Barcode-level Price and Quantity (Continuing Goods)
Barcode Level
0.15
0.10
0.05
0.00
-0.05
-0.10
-0.15
.....................................................................................................................
2007
2008
2009
2010
2011
Δln(x)
2012
2013
2014
2015
Δln(p)
Figure 4: Unit Value Price and Volume × Quantity (All Goods)
0.15
Producer Level
0.10
0.05
0.00
-0.05
-0.10
-0.15
.....................................................................................................................
2007
2008
2009
2010
2011
Δln(x)
15
2012
Δln(p)
2013
2014
2015
Figure 5: Comparison of the Change in Quantity between Barcode- and Producerlevel Estimates
Δln(x)
0.15
0.10
0.05
0.00
-0.05
-0.10
-0.15
.....................................................................................................................
2007
2008
2009
2010
2011
2012
Barcode Level
2013
2014
2015
Producer Level
Figure 6: Comparison of the Change in Price between Barcode- and Producer-level
Estimates
Δln(p)
0.05
0.03
0.01
-0.01
-0.03
-0.05
.....................................................................................................................
2007
2008
2009
2010
2011
Barcode Level
2012
2013
2014
Producer Level
16
2015
5. Estimation Results
Table 2 provides the estimation results for the demand and supply curves. The median
estimate of demand elasticity employing barcode-level information is 9.99, whereas
Broda and Weinstein (2010), employing similar information, estimated elasticity at 11.5.
The supply elasticities are generally smaller than those for demand, suggesting a
relatively flatter supply curve, which we consider to be a reasonable response, given that
the changes occur over only one year. The results of the overidentification tests are mixed.
The results based on the producer-level information generally pass the tests more often
than the results based on the barcode-level information. However, it is important to note
that the sample size for the barcode-level information is much larger than that for the
producer-level information, which increases the power of the statistics.
The producer-level elasticities are much smaller than the barcode-level elasticities,
implying a lower level of substitution between producers than between commodities. In
this paper, we use the category- and producer-level aggregate unit value price and volume
for each store, and the store-level and barcode-level data for the barcode-level estimation.
Thus, if one store conducts bargain sales for one commodity, we would observe a large
increase then decrease in quantity, which would make the estimates of the demand
elasticities based on store barcode-level information very large. We expect that the
nationwide category- and producer-level aggregates are largely unaffected by store-level
bargain sales, which contributes to their smaller estimates of elasticities.
17
Table 2: Estimation Results for Demand and Supply Elasticities
Barcode Level
(continuing goods only)
Producer Level
(including new goods)
(1)
(2)
(3)
(4)
2007M1－2016M2
2007 only
2007M1－2016M2
2007 only
995
937
1007
931
For Which the Signs of
σ and ω Are Consistent
with the Model
913
837
897
819
That Pass Overidentifying
Restrictions
173
342
351
425
Specifications
Data Period
Number of Categories:
Achieving Convergence
Basic Statistics
for σ and ω :
Mean
Min
Max
Std. Dev.
Skew
Kurt
p10
p20
p30
p40
p50
p60
p70
p80
p90
σ
11.48
0.53
308.07
14.70
15.05
277.27
5.59
7.21
8.37
9.20
9.99
10.69
11.43
12.58
15.69
ω
8.43
0.00
516.36
25.43
14.44
243.03
1.99
3.19
3.91
4.71
5.36
6.07
7.06
8.46
11.61
σ
14.67
0.60
797.57
32.29
18.32
421.96
4.58
6.64
8.15
9.24
10.31
11.65
12.99
15.75
22.10
ω
10.39
0.01
823.76
33.74
18.04
411.48
1.52
2.68
3.59
4.59
5.49
6.60
8.24
10.64
16.69
σ
8.28
0.65
138.29
8.09
8.20
104.88
3.38
4.32
5.37
6.27
6.96
7.70
8.58
9.94
12.52
ω
7.01
0.00
723.62
26.19
23.41
628.86
1.10
2.11
2.81
3.45
4.10
4.82
5.72
7.39
10.88
σ
11.61
0.78
1291.64
48.08
23.60
617.63
3.08
4.40
5.27
6.08
7.02
8.03
9.14
11.21
15.95
ω
6.90
0.00
198.03
13.36
7.54
79.27
0.88
1.65
2.49
3.05
3.86
4.73
5.95
8.23
13.14
Note: The estimation results are based on GMM with three moment conditions over 1,057 categories.
Only estimates with signs that were consistent with theory (a negative slope for demand and a
positive slope for supply) were included in the basic statistics for  and .
18
Figure 7 depicts the movement of the estimated demand and supply shocks based on
the barcode-level information. 20 Supply shocks were negative during 2008, which is
likely to be a reflection of the increase in the prices of materials and oil that occurred
before the GFC. There are no other notable supply shocks for other periods, except for
negative supply shocks after late 2013. Compared with the supply shocks, the demand
shocks are more frequent and exhibit larger fluctuations. Very large positive demand
shocks took place before the GFC in 2008. As shown in Figure 3, during that period,
prices increased but quantities did not, which suggests that there occurred upward shifts
in the demand curve and downward shifts in the supply curve. After the GEJE, there was
another large positive demand shock. However, the demand shock before the change in
the consumption tax rate in Japan that occurred on April 1, 2014, is negligible, which is
somewhat perplexing. As shown in Figure 3, prices declined during that period, whereas
quantities surged. In our model, given the supply shocks, positive demand shocks arise
when the demand curve shifts up, which enables the equilibrium price to increase.
However, before the 2014 consumption tax increase, as shown in Figure 3, prices did not
increase, which results in very small estimates for the demand shocks.
Figure 7: Demand and Supply Shocks Using Barcode-level Data
Barcode Level
0.30
0.20
0.10
0.00
-0.10
-0.20
-0.30
-0.40
.....................................................................................................................
2007
2008
2009
2010
2011
Demand Shock
2012
2013
2014
2015
Supply Shock
Note: The estimation is for the 797 categories that had signs for  and  that were consistent with
the model. The categories included in this figure are the same as in Figures 8 and 9. Tobacco is not
included.
In Figures 7–9, we dropped the upper 1% of the estimates of  and  for specifications (1) and (3)
in Table 2.
20
19
Figure 8 depicts the estimates of demand and supply shocks using the producer-level
unit value prices and quantities for both new and incumbent goods. Although the general
shapes of the estimated demand and supply shocks in Figure 8 are similar to those in
Figure 7, there are several noteworthy differences. First, the negative demand shocks
during the GFC are much larger in Figure 7 than in Figure 8. Second, the demand shocks
in Figure 7 are negative during 2014, whereas they are positive in Figure 8. In Figure 8,
we can also observe positive demand shocks occurring just before the change in the rate
of the consumption tax in April, 2014. This is more consistent with the conventional view
that stockpiling behavior precedes an increase in a tax rate than were the estimates
obtained using the barcode-level data on continuing goods in Figure 7.
As is clear from Figure 9, compared with the demand shocks, the supply shocks do not
exhibit large differences between Figures 7 and 8. In both figures, we can observe
negative shocks before the GFC, after the GEJE, and after 2013. Large negative supply
shocks arose after the change in the consumption tax rate, which reflects the increases in
prices observed in Figure 6. After mid-2013, the supply shocks remained at a very low
level, except for the temporary increases and decreases resulting from the change in the
consumption tax.
Figure 8: Demand and Supply Shocks Based on All Commodities
0.30
Producer Level
0.20
0.10
0.00
-0.10
-0.20
-0.30
-0.40
.....................................................................................................................
2007
2008
2009
2010
2011
Demand Shock
2012
2013
2014
2015
Supply Shock
Note: The categories included in this figure are the same as those in Figures 7 and 9.
20
Figure 9: Comparisons of the Aggregate Demand and Supply Shocks
0.30
Demand Shock
0.20
0.10
0.00
-0.10
-0.20
-0.30
-0.40
.....................................................................................................................
2007
2008
2009
2010
2011
Barcode Level
2012
2013
2014
2015
Producer Level
0.30
Supply Shock
0.20
0.10
0.00
-0.10
-0.20
-0.30
-0.40
.....................................................................................................................
2007
2008
2009
2010
Barcode Level
2011
2012
2013
2014
2015
Producer Level
Note: The categories included in this figure are the same as those in Figures 7 and 8.
21
6. Time-varying Elasticities
In the previous section, we assumed that the elasticities for supply and demand were
constant, with the macroeconomic shocks shifting the two curves through the changes in
intercepts. Given our detailed category classification, this assumption is clearly restrictive.
In this section, we allow for changes in the demand and supply elasticities when
estimating the macroeconomic shocks. That is, we relax the assumption for demand
elasticities in equation (1) such that they may be time variant. Then, we have:21
ln(𝑥𝑡𝑖 ) = ln(𝐶𝑡𝑖 ) + 𝜎𝑡 ln(𝑎𝑡𝑖 ) − 𝜎𝑡 (ln(𝑝𝑡𝑖 ) − ln(𝑃𝑡𝑖 )).
(1)
We assume that 𝜎𝑡 is a random walk as follows:
𝜎𝑡 = 𝜎𝑡−1 + 𝜐𝑡 , where 𝜐𝑡 is i.i.d.
Taking time differences into account, we obtain:
𝑖
∆ln(𝑥𝑡𝑖 ) = Δln(𝐶𝑡𝑖 ) + (𝜎𝑡−1 + 𝜐𝑡 )ln(𝑎𝑡𝑖 ) − 𝜎𝑡−1 ln(𝑎𝑡−1
)
𝑖
𝑖
−(𝜎𝑡−1 + 𝜐𝑡 ) (ln(𝑝𝑡𝑖 ) − ln(𝑃𝑡𝑖 )) + 𝜎𝑡−1 (ln(𝑝𝑡−1
) − ln(𝑃𝑡−1
)),
∆ln𝑥𝑡𝑖 = Δln𝐶𝑡𝑖 − 𝜎𝑡−1 (Δln𝑝𝑡𝑖 − Δln𝑃𝑡𝑖 ) − 𝜐𝑡 (ln𝑝𝑡𝑖 − ln𝑃𝑡𝑖 ) + 𝜀𝑡𝑖 ,
(5)
𝑖
where 𝜀𝑡𝑖 = (𝜎𝑡−1 + 𝜐𝑡 )ln𝑎𝑡𝑖 − 𝜎𝑡−1 ln𝑎𝑡−1
.
Furthermore, taking the difference from the average of the same producer leads to:22
∆ln𝑥𝑡𝑖 − ̅̅̅̅̅̅̅
∆ln𝑥𝑡𝑖 = 𝜀𝑡𝑖 − 𝜀̅𝑡𝑖 − 𝜎𝑡−1 (Δln𝑝𝑡𝑖 − ̅̅̅̅̅̅̅
Δln𝑝𝑡𝑖 ) − 𝜈𝑡 (ln𝑝𝑡𝑖 − ̅̅̅̅̅
ln𝑝𝑡𝑖 ).
(6)
The last term on the right-hand side of Equation (6) takes the difference of the price levels
among different stores. Although we do not include convenience stores, a component of
the price differential between stores may reflect differences in the locations or quality
standards of stores. To control for the store-level effects on the price level, we construct
store effects, defined as follows:
𝑚,𝑠
𝐾𝑦𝑒𝑎𝑟
=
1
#(Θ𝑚,𝑠
𝑦𝑒𝑎𝑟 )
∑𝑖∈𝑚,𝑠,𝑦𝑒𝑎𝑟(ln𝑝𝑡𝑖 − ̅̅̅̅̅
ln𝑝𝑡𝑖 ),
where #(Θ𝑚,𝑠
𝑦𝑒𝑎𝑟 ) is the number of products included for producer m, and for the store s,
in each year. That is, if store s sells all the products made by producer m at prices that are
𝑚,𝑠
3% higher than those charged at other stores, then 𝐾𝑦𝑒𝑎𝑟
becomes 0.03.23
21
22
23
To keep the notation simple, we omit the subscript j that denotes the category for elasticity.
The average is calculated, except in the case of the product itself, as explained in Section 1.
𝑚,𝑠
We drop the upper 1% and the lower 99% from our sample when constructing 𝐾𝑦𝑒𝑎𝑟
.
22
In conducting our estimation, we assume that the elasticities of supply vary less
frequently than those of demand because time is required to increase or decrease
production equipment. Therefore, we use the estimated results on a year-by-year basis in
specification (4) in Table 2 as the time-variant elasticities of the supply function. As 𝜎𝑡−1
in equation (6) is known at time 𝑡, 𝜈𝑡 is the only parameter to be estimated. More
specifically, we employ the following three moment conditions:
𝑚,𝑠
E[(𝜀𝑡𝑖 − 𝜀̅𝑡𝑖 ){(∆ln𝑥𝑡𝑖 − ̅̅̅̅̅̅̅
∆ln𝑥𝑡𝑖 ) − 𝜔
̂𝑦 (Δln𝑝𝑡𝑖 − ̅̅̅̅̅̅̅
Δln𝑝𝑡𝑖 − 𝐾𝑦𝑒𝑎𝑟
)}] = 0,
𝑚,𝑠
E[(𝜀𝑡𝑖 − 𝜀̅𝑡𝑖 )2 {(∆ln𝑥𝑡𝑖 − ̅̅̅̅̅̅̅
∆ln𝑥𝑡𝑖 ) − 𝜔
̂𝑦 (Δln𝑝𝑡𝑖 − ̅̅̅̅̅̅̅
Δln𝑝𝑡𝑖 − 𝐾𝑦𝑒𝑎𝑟
)}] = 0,
2
𝑚,𝑠
E [(𝜀𝑡𝑖 − 𝜀̅𝑡𝑖 ){(∆ln𝑥𝑡𝑖 − ̅̅̅̅̅̅̅
∆ln𝑥𝑡𝑖 ) − 𝜔
̂𝑦 (Δln𝑝𝑡𝑖 − ̅̅̅̅̅̅̅
Δln𝑝𝑡𝑖 − 𝐾𝑦𝑒𝑎𝑟
)} ] = 0.
The estimation procedure is as follows.
(1) Assuming constant elasticity, using the sample for 2007, we obtain the estimate of 𝜎0 .
(2) We set the initial value of 𝜎0 .
(3) Given 𝜎𝑡−1, for each month, we update the demand elasticity using 𝜎𝑡 = 𝜎𝑡−1 + 𝜐𝑡 ,
as long as the GMM results satisfy the following two conditions: (a) the estimate of
𝜈𝑡 is statistically significant at the 10% level; and (b) the overidentification test is
passed at the 10% level. We calculate the entire squared sum of residuals for the entire
sample period. If the two conditions are not satisfied, we retain the previous estimates,
that is, 𝜎𝑡 = 𝜎𝑡−1 .
We change the initial value 𝜎0 a little and repeat (2) to find the initial value that
minimizes the squared sum of residuals.24
(4) For the supply elasticity, 𝜔𝑡 , we use the yearly estimates, assuming that 𝜔𝑡 is
constant within each year.25 Then, Equation (3) with yearly estimates becomes:
𝑗
∆ ln(𝑥𝑡𝑖 ) = 𝜔𝑡 ∆ ln(𝑝𝑡𝑖 ) + 𝑆t + 𝛿ti .
(3)
After obtaining the estimates of the elasticities,(𝜎̂𝑡 , 𝜔
̂),
we plug them into (1) and
𝑡
(3), which gives us the following two sets of shocks:
24
When estimating the time-varying demand elasticities, we restrict our analysis to categories of
goods that contain 50 or more observations each month. For categories with fewer than 50
observations, as long as the overidentification tests are passed, we use the estimates with constant
elasticities in each year. We exclude categories with a small number of observations that are unable to
pass the overidentification tests.
25 Because of space limitations, we have provided the estimation results for 2007 only in specification
(4) of Table 2.
23
𝜀̂𝑡𝑖 ≡ ∆ln(𝑥𝑡𝑖 ) + 𝜎̂𝑡 ∆ ln(𝑝𝑡𝑖 ),
and
𝑖
𝛿̂𝑡𝑖 ≡ ∆ ln(𝑥𝑡𝑖 ) − 𝜔
̂∆
𝑡 ln(𝑝𝑡 ),
where 𝜀̂𝑡𝑖 and 𝛿̂𝑡𝑖 contain both commodity- and category-specific shocks that shift the
category-level demand and supply curves, respectively. The final step of the estimation is
aggregation over commodities and categories. We use the Törnqvist weights of sales for
both aggregations.
Figure 10 shows our estimates of time-varying demand elasticities for the producerlevel estimation. Both median and mean elasticities have an increasing trend. It is worth
noting that the figure does not show any countercyclical or procyclical movements,
suggesting that markups are not related to business cycles in Japan. Figure 11 reports our
estimates of demand shocks with time-variant elasticities on (a) a year-by-year basis and
(b) a month-by-month basis. At first glance, the tilting demand curve does not cause
significant changes in the estimates of the aggregate demand shocks, with the exception
of the year 2015. Based on month-by-month changes, the aggregate demand shocks
indicate a large increase in 2015, which does not occur for the estimates based on yearby-year changes. The increase is mainly driven by milk, for which large increases
occurred both in terms of prices and demand elasticities in 2015.26 By construction, an
increase in both the demand elasticity and the price of milk increases demand shocks.
From Figures 10 and 11, the general movements of our estimates of the aggregate demand
and supply shocks do not differ greatly from our previous estimates using constant
elasticities. Since 2013, the aggregate supply shocks have been negative under several
different specifications, suggesting that the increases in prices during certain periods are,
at least partially, driven by negative supply shocks.
26
In 2015, the price of milk increased by about 5% and the demand elasticity of milk surged from 14
to 29. If milk is excluded, the aggregate demand shock in 2015 becomes much smaller.
24
Figure 10: Changes in Estimates of Demand Elasticities
13
12
11
10
9
8
7
6
5
4
....................................................................................................................
2007
2008
2009
2010
2011
Median
2012
2013
2014
2015
Weighted Average
Note: 1) These estimates are for 336 categories, which have signs consistent with the model for 
and  and which pass the overidentification restrictions. 2) The weighted average is calculated on
the basis of sales volumes for each category.
Figure 11: Demand and Supply Shocks with Time-varying Elasticity
0.40
Producer-Level
0.30
0.20
0.10
0.00
-0.10
-0.20
-0.30
-0.40
....................................................................................................................
2007
2008
2009
2010
2011
2012
2013
2014
2015
(a) Demand Shock: year-by-year basis
(b) Demand Shock: month-by-month basis
(a) Supply Shock: year-by-year basis
(c) Supply Shock: pool (time invariant)
Note: These estimates are for 336 categories, which have signs consistent with the model for  and
 and which pass the overidentification restrictions.
7. Conclusions
25
In this paper, we used barcode-level transaction data to estimate the aggregate demand
and supply shocks in Japan. Both estimated demand and supply shocks exhibit large
fluctuations. When we use the barcode-level information for continuing goods only, there
is a negligible demand shock preceding Japan’s change in its consumption tax rate in
2014, which contradicts the conventional view on stockpiling behavior. This somewhat
perplexing result arises because of the preceding decline in the price index. When we
consider goods that did not exist in the previous year, the estimated demand shock
becomes positive, which suggests that significant stockpiling behavior was associated
with the purchase of new goods. The supply shocks are negative from the end of 2013,
suggesting that increases in prices from that period at least partly reflect inward shifts of
the supply curves. These findings remain unchanged even if we allow for changes in the
demand and supply elasticities over time.
There are several remaining tasks for future research. First, we have not considered
the source(s) of the supply and demand shocks. Macroeconomic shocks, such as currency
depreciations or changes in the prices of oil, grains, and materials, might affect supply,
whereas changes in income taxes, weather, monetary and fiscal policies, and stock prices
could affect demand. To decompose the shocks into their several sources, we would need
to use information from an input–output table and category-level shocks. Second,
investigation of regional variations in shocks is another important topic. For example, the
GEJE in March 2011 occurred in the northeast of Japan. Although it affected the entire
eastern part of Japan, the western part was virtually untouched. Other examples of the
causes of region-specific demand and/or supply shocks include heavy snowfalls in the
northern parts of Japan or typhoons in the south. The investigation of such region-specific
demand and supply shocks will be our next task.
26
References
Bernard, A. B., J. R. Stephen, and P. K. Schott. (2010) “Multiple-product firms and
product switching,” American Economic Review, 100(1), 70–97.
Berry, S., J. Levinsohn, and A. Pakes. (1995) “Automobile prices in market equilibrium,”
Econometrica, 63(4), 841–890.
Blanchard, O. J. and D. Quah. (1989) “The dynamic effects of aggregate demand and
supply disturbances,” American Economic Review, 79(4), 655–673.
Broda, C. and D. E. Weinstein. (2010) “Product creation and destruction: evidence and
price implications,” American Economic Review, 100(3), 691–723.
Byrne, D. P., S. Imai, V. Sarafidis, and M. Hirukawa. (2015) “Instrument-free
identification and estimation of differentiated products models,” Department of
Economics Working Paper No. 1336, Queen’s University, Ontario, Canada.
Diewert, W. E. (1976) “Exact and superlative index numbers,” Journal of Econometrics,
4(2), 114–145.
Feenstra, R. C. (1994) “New product varieties and the measurement of international
prices,” American Economic Review, 84(1), 157–177.
Leamer, E. E. (1981) “Is it a demand curve, or is it a supply curve? Partial identification
through inequality constraints,” The Review of Economics and Statistics, 63(3), 319–
327.
Nevo, A. (2010) “Empirical models of consumer behavior,” Annual Review of
Economics, 3, 51–75.
Sato, K. (1976) “The ideal log-change index number,” Review of Economics and
Statistics, 58(2), 223–228.
Soderbery, A. (2015) “Estimating import supply and demand elasticities: analysis and
implications,” Journal of International Economics, 96(1), 1–17.
27
Uhlig, H. (2005) “What are the effects of monetary policy on output? Results from an
agnostic identification procedure,” Journal of Monetary Economics, 52(2), 381–419.
Working, E. J. (1927) “What do statistical “demand curves” show?,” The Quarterly
Journal of Economics, 41(2), 212–235.
28