1553
American Economic Review 2008, 98:4, 1553–1577
http://www.aeaweb.org/articles.php?doi=10.1257/aer.98.4.1553
The “Law of Demand,” which holds that as the price of a good increases, consumers’ demand
for that good should decrease, is one of the bedrock principles of microeconomics. Economists
have long recognized, however, that the axioms of consumer theory do not guarantee that demand
curves must slope downward, and that the Law of Demand, while descriptively valid in many
situations, may not apply to very poor consumers facing subsistence concerns. Alfred Marshall
first publicized this idea in the 1895 edition of his Principles of Economics:
As Mr. Giffen has pointed out, a rise in the price of bread makes so large a drain on
the resources of the poorer labouring families and raises so much the marginal utility of
money to them, that they are forced to curtail their consumption of meat and the more
expensive farinaceous foods: and, bread being still the cheapest food which they can get
and will take, they consume more, and not less of it (208).
Since Marshall’s time, a discussion of “Giffen” behavior has found its way into virtually every
basic economics course, despite a lack of real-world evidence supporting Marshall’s conjec-
ture.
Studies by George J. Stigler (1947) and Roger Koenker (1977) argue that demand for
neither bread nor wheat was upward sloping in Britain during Marshall’s time. The standard
textbook example of a Giffen good, potatoes during the Irish Potato Famine of 1845–1849 (Paul
A. Samuelson 1964), has also been discredited (Sherwin Rosen 1999). Not only are there no data
to support the claim, but at a more basic level it is unlikely that consumption of potatoes could

We use the term “Giffen behavior” rather than “Giffen good” to emphasize that the Giffen property is one that
holds for particular consumers in a particular situation and therefore depends on, among other things, prices and
wealth. Thus, it is not the good that is Giffen, but the consumers’ behavior.
Giffen Behavior and Subsistence Consumption
By Robert T. Jensen and Nolan H. Miller*
This paper provides the first real-world evidence of Giffen behavior, i.e.,
upward sloping demand. Subsidizing the prices of dietary staples for extremely
poor households in two provinces of China, we find strong evidence of Giffen
behavior for rice in Hunan, and weaker evidence for wheat in Gansu. The data
provide new insight into the consumption behavior of the poor, who act as
though maximizing utility subject to subsistence concerns. We find that their
elasticity of demand depends significantly, and nonlinearly, on the severity of
their poverty. Understanding this heterogeneity is important for the effective
design of welfare programs for the poor. (JEL D12, O12)
* Jensen: Watson Institute for International Studies, Brown University, 111 Thayer St., Providence, RI 02912,
and National Bureau of Economic Research (e‑mail: robert_jensen@brown.edu); Miller: John F. Kennedy School of
Government, Harvard University, 79 JFK St., Cambridge, MA 02138 (e-mail: nolan_miller@harvard.edu). We thank
Alberto Abadie, Chris Avery, Sebastian Bauhoff, Amitabh Chandra, Suzanne Cooper, Daniel Hojman, Brian Jacob,
Elizabeth Lacey, Erzo Luttmer, Mai Nguyen, Albert Park, Rodrigo Wagner, Sangui Wang, Richard Zeckhauser, and
two anonymous referees for valuable comments and discussions, and Frank Mou, Dulles Wang, and Fan Zhang for
research assistance. We gratefully acknowledge financial support from the National Institute of Aging, the William F.
Milton Fund at the Harvard Medical School, the Dean’s Research Fund at the John F. Kennedy School of Government,
the Center for International Development at Harvard University, and the Hefner China Fund.
September 20081554 THE AMERICAN ECONOMIC REVIEW
have increased when the price rose during the famine, at least in the aggregate, precisely because
the price rise was caused by a blight that destroyed much of the crop.
While some laboratory
studies have found evidence of Giffen behavior, these experiments have been far from removed
from reality.
In this paper we present data from a field experiment exploring the response of poor households
in China to changes in the prices of staple food items, which provide the first rigorous,
empirical evidence of real-world Giffen behavior. In fact, we find Giffen behavior with respect to
two goods, rice and wheat. Further, these goods, and the populations that exhibit Giffen behavior,
meet some basic but common conditions that suggest this behavior may be widespread in the
developing world. Thus, the absence of previously documented cases most likely results from
inadequate data or empirical strategies rather than from their nonexistence.
Giffen behavior has long played an important, though controversial, role in economic pedagogy,
as well as in the history of economic thought.
Finding convincing evidence of such behavior
is important, however, for economic theory more broadly. The fact that there has, to date, been
no convincing evidence of Giffen behavior stands as a minor embarrassment to economists (John
H. Nachbar 1998), one that is reflected in the discussion of the Giffen phenomenon often being
presented as a paradox of economic theory rather than as a real (or even possible) mode of behavior
(e.g., Stigler 1947). This lack of evidence has prompted a range of reactions among economists.
Some have interpreted it as support for the descriptive validity of the Law of Demand:
Perhaps as persuasive a proof [of the “Law of Demand”] as is readily summarized is this:
if an economist were to demonstrate its failure in a particular market at a particular time,
he would be assured of immortality, professionally speaking, and rapid promotion while
still alive. Since most economists would not dislike either reward, we may assume that the
total absence of exceptions is not from lack of trying to find them.
Stigler (1966, 24).
Others’ reactions to the lack of validation for the Giffen phenomenon have been more extreme,
interpreting it as an indictment of neoclassical consumer theory. Along these lines, Lawrence A.
Boland (1977) points out that not only is the theory unable to rule out Giffen behavior, it is also
unable to explain why it is not observed. Put another way, if the neoclassical model is correct,
then under certain (albeit uncommon) conditions, Giffen behavior should exist. If it has not been
observed, it is either because the appropriate conditions have not been satisfied, the appropriate
data have not been available to measure it, or the theory is incomplete or flawed.

Gerald P. Dwyer Jr. and Cotton M. Lindsay (1984) present a summary of the basic case against the potato version
of the Giffen paradox. See also Terrence McDonough and Joseph Eisenhauer (1995). In both the bread and potato cases,
it is possible that poor individuals exhibited Giffen behavior but the market overall did not. However, the data to test
this hypothesis do not exist.

Raymond C. Battalio, John H. Kagel, and Carl A. Kogut (1991) find evidence of upward sloping demand among
rats given limited “budgets” and the choice between root beer and a quinine solution, and R. J. DeGrandpre et al. (1993)
find in a laboratory setting that human smokers, given the choice between brands of cigarettes and a limited budget of
“puffs,” can exhibit Giffen behavior.

The lack of verified examples has raised numerous concerns about the pedagogical role of the Giffen story: “Since
the Giffen paradox is not useful for understanding the Irish Experience, is it asking too much for future writers of
elementary texts to find another example? Fictions have no place in the teaching of economics” Rosen (1999); “We shall
have to find a new example of the positively sloping demand curve, or push our discussion of it deeper into footnotes”
Stigler (1947).

Others have argued that it is not our understanding of consumers that is flawed, but rather our understanding of
markets. For example, William Dougan (1982) argues that markets with upward sloping demand curves are inherently
unstable, and thus unlikely to be observed, while Nachbar (1998) shows in a general equilibrium framework that
observing the equilibrium price and quantity of a good move in the same direction in response to a supply shock implies
that the commodity is normal, not inferior, and thus not Giffen at all. Thus, economists looking for Giffen behavior at
the level of the market are unlikely to find it. Yoram Barzel and Wing Suen (1992) argue that if consumers can exploit
—
VOL. 98 NO. 4 1555Jensen and Miller: Giffen Behavior and Subsistence Consumption
Beyond documenting the existence of Giffen behavior, our field experiment also provides an
opportunity to study more broadly the consumption behavior of the “extreme poor,” a population
that worldwide includes more than one billion people living below the World Bank’s extreme
poverty line of one dollar per person per day. These households, like Marshall’s “labouring
families” and those in our sample, are often highly dependent on a single staple food for the bulk
of their nutritional needs. Consequently, fluctuations in the prices of these staple foods can have
large effects on real wealth and purchasing power. Anecdotally, such price fluctuations, even
fairly large ones, are increasingly common in developing countries.
And while there is a large
literature examining household vulnerability and responses to income shocks, there is comparably
little evidence with respect to price shocks. Our analysis, by focusing on the extremely poor
and by introducing exogenous price changes for staple foods, is useful for understanding this
vulnerability.
In an earlier study using panel data from the China Health and Nutrition Survey (Jensen and
Miller 2002), we found suggestive evidence that poor households in China exhibited Giffen
behavior with respect to their primary dietary staple (rice in the south; wheat and/or noodles
in the north).
However, because the study relied on possibly endogenous variation in market
prices, we were unable to identify a causal relationship between price changes and consumption.
To address this concern, for the present study we conducted a field experiment in which, for
five months, randomly selected households were given vouchers that subsidized their purchases
of their primary dietary staple. Building on the insights of our earlier analysis, we studied two
provinces of China: Hunan in the south, where rice is the staple good, and Gansu in the north,
where wheat is the staple. Our analysis in these provinces focused on households classified as the
“urban poor,” a population that includes approximately 90 million individuals throughout China
(Martin Ravallion 2007).
Using consumption surveys gathered before, during, and after the subsidy was in place, we
find strong evidence that poor households in Hunan exhibit Giffen behavior with respect to rice.
That is, lowering the price of rice via the experimental subsidy caused households to reduce their
demand for rice, and removing the subsidy had the opposite effect. This finding is robust to a wide
range of empirical specifications. In Gansu, the evidence for Giffen behavior is somewhat weaker,
due to the partial failure of two of the basic conditions under which such behavior is expected;
namely that the staple good has limited substitution possibilities, and that households are not so
poor that they consume only staple foods. Focusing our analysis on those whom the theory identifies
as most likely to exhibit Giffen behavior, we find stronger evidence of its existence.
We also provide important new insights into the consumption behavior of poor households.
In particular, we find the consumption response to an increase in the price of a staple good follows
a previously undocumented inverted-U pattern predicted by consumer theory in the presence
of subsistence concerns, with the very poorest and the least poor of the poor responding by
decreasing demand in response to an increase in the price of the staple in the standard way, while
the group in the middle increases demand (i.e., exhibits Giffen behavior). These results have
intertemporal variations in prices, they will never purchase more of a good when its price increases. The very poor,
however, are likely to be liquidity constrained and thus lack this ability.

For example, occasional reports from China note rice prices that double from year to year in some localities
(“Surge in Consumer Prices Puts China on Guard,” China Daily, April 22, 2004). Jed Friedman and James Levinsohn
(2002) note that the mean price of rice increased by almost 200 percent in Indonesia (where the typical household
spends nearly 30 percent of its total household budget on rice) during the 1997/98 financial crisis.

Ours is not the first study to suggest rice as a likely candidate for Giffen behavior. Dwyer and Lindsay (1984) propose
(but do not test) this possibility for Singapore, and John-Ren Chen (1994) finds suggestive evidence of positively
sloped demand for rice in Taiwan.
September 20081556 THE AMERICAN ECONOMIC REVIEW
important implications for the design, targeting, and evaluation of programs aimed at improving
nutrition among the poor.
The paper continues in Section I, where we present a discussion of the consumption behavior
of the poor that motivates Giffen behavior. Section II discusses the field experiment, data, and
estimation strategy. Section III presents the results, and Section IV concludes.
I.  Giffen Behavior and Consumption among the Poor
The conditions under which we would expect Giffen behavior can be demonstrated by elaborating
Marshall’s statement.
Imagine an impoverished consumer near a subsistence level of
nutrition, whose diet consists of only two foods, a “basic” or staple good (in Marshall’s case,
bread) and a “fancy” good (meat). The basic good offers a high level of calories at low cost, while
the fancy good is preferred because of its taste but provides few calories per unit currency. A poor
consumer will therefore eat a lot of bread in order to get enough calories to meet his basic needs
and use whatever money he has left over to purchase meat. Now, if the price of bread increases,
he can no longer afford the original bundle of foods. And if he increases his consumption of
meat, he will fall below his required caloric intake. So, he must instead increase his consumption
of bread (which is still the cheapest source of calories) and cut back on meat.
The Giffen phenomenon illustrates the potential significance of the wealth effects of price
changes for extremely poor households. Although the price increase makes the staple less attractive
in relative terms, the fact that it makes the consumer so much poorer (in real terms) forces
him to consume more bread. Translating this to the language of consumer theory, the conditions
under which Giffen behavior is likely to be observed, therefore, include that the good in question
be strongly inferior and that expenditure on that good comprise a large portion of the consumer’s
budget. As can be seen from the elasticity version of the Slutsky equation, ep 5 ep
h
– bew , where
ep is the observed price elasticity of demand, ep
h
(# 0) is the Hicksian compensated elasticity, ew
is the wealth elasticity, and b is the budget share of the good, only then can the wealth effect of
a price increase be large enough to offset the pure substitution effect.
In light of these observations, we can state a set of conditions under which Giffen behavior is
most likely to be observed:
C1: 	Households are poor enough that they face subsistence nutrition concerns.
C2: 	Households consume a very simple diet, including a basic (staple) and a fancy good.
C3: 	The basic good is the cheapest source of calories available, comprises a large part of the
diet/budget, and has no ready substitute.
When dealing with extreme poverty of the sort exhibited by the urban poor in China, another
requirement becomes important. While consumers who are too wealthy will not exhibit Giffen
behavior, those who are too poor also cannot exhibit Giffen behavior. To take an extreme example,
consider a consumer who is so poor that he consumes only bread. When the price of bread
increases, he has no choice but to consume less bread. Thus, it is critical to the Giffen story that the
consumer be consuming at least some of the fancy goods (e.g., meat) that are more enjoyable but

Much of the theory of Giffen behavior has previously appeared elsewhere. The interested reader should see the
online Appendix to this document (http://www.aeaweb.org/articles.php?doi=10.1257/aer.98.4.1553) for a discussion of
the theory underlying this behavior.

Some of these conditions have been noted before by, for example, Otis W. Gilley and Gordon V. Karels (1991).
VOL. 98 NO. 4 1557Jensen and Miller: Giffen Behavior and Subsistence Consumption
more expensive sources of calories so that he has something to substitute away from when the price
of the staple increases. In light of this, we add the following requirement to the three stated above:
C4:	Households cannot be so impoverished that they consume only the staple good.
The theory thus predicts that only consumers who are poor, but not too poor, will exhibit Giffen
behavior. Panel A of Figure 1 depicts the indifference curves for a typical consumer choosing
how much of the basic and fancy goods to consume. The basic or staple good is relatively high
in calories, while the fancy good offers more “taste,” i.e., the enjoyable but nonnutritive aspects
of food.10
The consumer’s indifference map can be divided into three regions. The outer set of
indifference curves corresponds to the standard case, where the consumer’s calorie intake is well
above subsistence. Over this range the consumer trades off between calories and taste (and thus
between the basic and fancy goods) in an ordinary way. As the consumer’s calorie consumption
decreases, he crosses into a “subsistence zone.” Over this range, caloric intake becomes much
more important to the consumer. Consequently, the consumer’s indifference curves take on the
familiar “elbow” shape associated with Giffen behavior.11
Consumers in this range behave as if
they maximize taste, subject to the constraint that they meet their minimum caloric needs. As
the consumer’s calorie consumption decreases even further, he crosses from the subsistence zone
to the calorie-deprived zone. In this region, the consumer’s calorie intake is below subsistence
levels. Hence, his primary concern is maximizing calories, and the consumer’s indifference
curves are, in effect, iso-calorie curves.
The consumer’s response to an increase in the price of the staple good will differ across the
three regions of his indifference map. When the consumer is relatively wealthy, he will demand a
bundle of goods in the standard zone. In this case, as illustrated in panel B of Figure 1, we expect
the consumer to respond to an increase in the price of the staple good by consuming less of it.
Thus, demand is downward sloping. As wealth decreases, the consumer’s demand moves into the
10
The substitution across goods with varying nutritional and nonnutritional attributes also motivates the literature
concerned with the income elasticity of demand for calories (see John Strauss and Duncan Thomas 1995; Angus
Deaton 1997).
11
See the online Appendix for more discussion of the relationship between the shape of indifference curves needed
to generate Giffen behavior and subsistence concerns.
Figure 1. Zones of Consumer Preferences
September 20081558 THE AMERICAN ECONOMIC REVIEW
subsistence zone, and the consumer focuses more on maintaining caloric intake as his primary
goal. It is over this region that Giffen behavior arises, as the consumer responds to an increase in
the price of the staple good by substituting toward the cheaper source of calories, which is still
the staple good. Over this range, the consumer still trades off calories against taste, although
caloric intake is given much greater importance. A consumer in the subsistence zone behaves, in
effect, as if he maximizes taste subject to the constraint that calories reach a certain minimum
requirement. Finally, as the consumer’s wealth decreases even further, he is unable to afford to
meet his subsistence calorie needs. Calories are maximized by consuming only the staple good,
and so the consumer has no choice but to respond to an increase in the price of the staple good
by consuming less of it.
This set of predictions provides important insight into the search for Giffen behavior and
may also help explain why such behavior has not been previously detected. First, we should not
expect to observe Giffen behavior at the market level; a subset of consumers might exhibit Giffen
behavior with respect to a particular commodity while the overall market exhibits downward
sloping demand. Thus, the search should not be for a “Giffen good,” but for Giffen behavior. In
addition, the nonlinear response suggests the search is even more nuanced than just focusing on
the poorest households. Selecting just the very poorest, or even aggregating over a broader set
of households that includes both those in the calorie-deprived and subsistence zones, may not
be sufficient. Consumers in the intermediate, subsistence range must be isolated in order to find
such behavior.
II.  Empirical Strategy
A. The Experiment
A central problem in documenting Giffen behavior, and indeed in any analysis of demand, is
finding both sufficient and exogenous price variation. As a practical problem, whether data are
cross-sectional, time-series, or panel, there is often not a great deal of variation in prices for the
kinds of goods likely to be candidates for Giffen behavior. This applies especially to cross-sectional
data, as arbitrage should eliminate spatial price differences, especially for easily storable
and nonperishable commodities such as grains. Further, the prices for staple goods might even
be fixed by the government for the poorest households, such as under India’s Public Distribution
System, and any remaining price variation may be due to unobservable quality differences. A
more serious concern is that, even with sufficient price variation, the source of that variation is
often potentially endogenous, since price is the equilibrium of a system of simultaneous equations.
A positive correlation between price and consumption could simply represent shocks to, or
differences in, demand over space or time; higher demand leads to higher prices, which could be
misinterpreted as Giffen behavior. Although instrumental variables could address this problem,
finding instruments that shift supply but do not directly affect demand is difficult.12
12
Most previous studies of Giffen behavior have failed to address this identification concern. A few cases have used
instrumental variables, but with problematic instruments. For example, Anthony E. Bopp (1983) uses refinery utilization
rates and the price of crude oil as instruments for the price of kerosene; however, both instruments likely also affect
the price of substitute fuels, and are likely to be driven by other unobserved factors also affecting fuel demand, such as
weather. Shmuel Baruch and Yakar Kannai (2001) use the lagged prime interest rate as an instrument for the price of a
low-grade Japanese alcohol (shochu), which is likely be a poor predictor of the price of shochu, or, to the extent that it
does predict the price, will likely also affect the prices of substitutes (or income—and thus demand). Rainfall is commonly
suggested as an instrument for price. However, rainfall will be an invalid instrument for the price of a given food
item, since it likely also affects the prices of other foods, as well as wages and income. One exception to the endogeneity
concern in the search for Giffen behavior is the study by David McKenzie (2002), which uses the elimination of tortilla
subsidies and price controls as a natural experiment to test for (and ultimately reject) such behavior in Mexico.
VOL. 98 NO. 4 1559Jensen and Miller: Giffen Behavior and Subsistence Consumption
To overcome these challenges, we conducted a field experiment in which we provided randomly
selected poor households in two Chinese provinces with price subsidies for staple foods.
In Hunan, a southern province, rice is the staple good, and in Gansu, a northwestern province,
wheat is the staple good (consumed primarily as buns, a simple bread called mo, or noodles).
These regional differences in preferences are primarily determined by geography, climate, and
history, with wheat the dominant crop grown in Gansu and rice dominant in Hunan. Accordingly,
we subsidized rice (only) in Hunan and wheat flour (only) in Gansu.
Within each sample cluster (described below), households were randomly assigned to either a
control group or one of three treatment groups. Households in the treatment groups were given
printed vouchers entitling them to a price reduction of 0.10, 0.20, or 0.30 yuan (Rmb; 1 Rmb
≈$0.13) off the price of each jin (1 jin 5 500g) of the staple good (the subsidy level stayed fixed
for each household over the course of the study). These subsidies represented substantial price
changes, since the average preintervention price of rice in Hunan was 1.2 yuan/jin, and the average
for wheat flour in Gansu was 1.04 yuan/jin. The vouchers were printed in quantities of 1, 5,
and 10 jin, and the month’s supply of vouchers was distributed at the start of each month, with
each household receiving vouchers for 750g per person per day (about twice the average per
capita consumption). All vouchers remained valid until the end of the intervention. Households
were told in advance they would receive vouchers for five months and that any unredeemed
vouchers would not be honored afterward.
The vouchers could be redeemed at local grain shops. The merchants in these shops agreed
to honor the vouchers in exchange for reimbursement and a payment for their participation.
Households and merchants were told they were not permitted to exchange the vouchers for anything
but the staple good, that there would be periodic auditing and accounting to make sure they
were in compliance with the rules, and that any violations would result in their being removed
from the study without any additional compensation. Households and merchants were explicitly
told that selling the vouchers for cash or reselling rice or wheat bought with the vouchers would
result in dismissal from the program.
There are several points about the intervention worth noting. First, all foods in China are
sold in free markets, at market determined prices. A 1993 reform of the grain distribution system
largely put an end to price controls, state food stores, or free rations. Second, the number
of subsidized households in each sample site is trivial relative to the size of the population (all
sites were county seats, most with populations over one million), so the intervention could not
have affected market prices.13
Third, the experiment is predicated on the assumption that either
households are limited in their ability to borrow and save, or they have short planning horizons;
otherwise, the wealth effect of the five-month subsidy would be trivial, making Giffen behavior
unlikely. To the extent the wealth effect of the price change can be smoothed over the lifetime,
this will bias us against finding Giffen behavior. Fourth, limiting the quantity of vouchers to
750g/person/day might limit the potential demand response for the staple good (although the
amount is still quite generous), but it should not induce Giffen behavior, as might be the case if
we limited the vouchers to a quantity smaller than what they would prefer to consume.14
Finally,
while staple foods such as rice can be found in varying qualities or varieties with different prices,
because the households in our sample are extremely poor, our data show that they consume
almost exclusively only the lowest-cost variety. Therefore, quality substitution in response to the
13
Similarly, because the samples were drawn from lists of the poor spread throughout large cities, we believe it is
unlikely the various study participants knew each other or the benefits others were receiving.
14
One concern is that by limiting the potential increase in consumption in response to the price decline, we might
skew the average consumption change toward a decline (i.e., Giffen behavior). However, in practice, almost no households
even approached the voucher limit, most likely due to their extremely low incomes and a lack of access to credit,
so this is unlikely to be a major concern.
September 20081560 THE AMERICAN ECONOMIC REVIEW
price subsidy is not a concern for our analysis. Two final concerns with the experiment, namely
whether there was cheating (in the form of cashing out or reselling) despite our rules against
doing so or whether the vouchers might create a “salience” or signaling effect, are discussed with
the results in Section IIIE.
B. Data
The survey and intervention were conducted by employees of the provincial-level agencies of
the Chinese National Bureau of Statistics. The sample consisted of 100–150 households in each
of 11 county seats spread throughout Hunan and Gansu provinces (Anren, Baoqing, Longshan,
Pingjiang, Shimen, and Taojiang in Hunan, and Anding, Ganzhou, Kongdong, Qingzhou, and
Yuzhong in Gansu), for a total of 1,300 households (650 in each province), with 3,661 individuals.
Within each county, households were chosen at random from lists of the “urban poor” maintained
by the local offices of the Ministry of Civil Affairs.15
Households on this list fall below
a locally defined poverty threshold (the Di Bao line), typically between 100 and 200 yuan per
person per month, or $0.41−$0.82 per person per day, which is below even the World Bank’s
extreme poverty line of one dollar per person per day. It is estimated that about 90 million individuals
throughout China live below the Di Bao threshold.
The questionnaire consisted of a standard income and expenditure survey, gathering information
on the demographic characteristics of household members, as well as data on employment,
income, asset ownership, and expenditures. A key component of the survey was a 24-hour food
recall diary completed by each household member. Respondents were asked to report everything
they ate and drank the previous day, whether inside or outside the home, by specifically listing
the components of all foods eaten.16
These foods were recorded in detail in order to match
with the 636 detailed food items listed in the 1991 Food Composition Tables constructed by the
Institute of Nutrition and Food Hygiene at the Chinese Academy of Preventative Medicine. As
we will see below, because households are very poor, most diets are very simple and consist of
a small number of basic (not processed, prepared, or packaged) foods like rice, bean curd, or
stir-fried cabbage, so concerns about coding the specific quantities of the various ingredients in
a complex dish or meal are not significant.17
Data were gathered in three waves, conducted in April, September, and December of 2006.
After completing the first survey, treatment households were told they would receive the subsidies
for five months, from June through October. Thus, the initial interviews occurred before
treatment households knew of or received the subsidies, the second occurred after the subsidy
had been in place for slightly more than three months, and the final interviews were conducted
one to two months after the subsidy had ended, by which time treatment households would likely
have exhausted any stocks of rice or wheat flour they may have purchased with the subsidy, and
would therefore again be purchasing at the full market price. Sample attrition was extremely low,
since the three rounds occurred in a relatively short span. Only 11 of 1,300 households (less than
1 percent) in the first round did not appear in the second round. All households in the second
15
We chose urban areas because in smaller towns or rural areas many of the poorest households grew rather than
purchased their staple food, and lower population density meant fewer households living in extreme poverty, which
would have both required a greater number of sample clusters and prevented varying the treatment within clusters.
16
While it may seem difficult to recall or estimate how many grams of, say, rice were eaten with a meal, for the
extreme poor who are on a very limited budget, food is often apportioned and accounted for much more carefully.
Further, diets for these extremely poor households often vary little or not at all from day to day, except on special occasions,
so recalling the quantity of specific food items is not as difficult.
17
Similarly, because households were so poor, almost all food (98 percent) was at-home consumption, so respondents
were aware of the exact ingredients and quantities used.
VOL. 98 NO. 4 1561Jensen and Miller: Giffen Behavior and Subsistence Consumption
round were interviewed in the third round. Means and standard deviations for key variables are
presented in Table 1.18
Table 2 shows the basic consumption patterns for households in the two provinces. The dominance
of, and difference in, staple goods in the two regions is evident. In Hunan, the average
per capita consumption of rice per day is 330g, comprising 64 percent of daily caloric intake.
18
While there are some differences in variables across control and treatment groups, these arise largely due to random
variation, given the relatively small sample size. Randomization was done blindly by the authors, rather than the
field teams, so any differences should not be systematic. Further, any differences in variables across households based
on treatment assignment will be eliminated because our analysis uses household fixed effects.
Table 1—Means and Standard Deviations of Key Variables
	 Control	 0.1 yuan/jin subsidy	 0.2 yuan/jin subsidy	 0.3 yuan/jin subsidy
Panel A: Hunan
Family size	 2.8	 2.9	 3.0	 2.7
	 [1.3]	 [1.3]	 [1.4]	 [1.1]
No. of children	 0.46	 0.46	 0.44	 0.38
	 [0.68]	 [0.6883]	 [0.6687]	 [0.61]
Female head	 0.34	 0.37	 0.37	 0.40
	 [0.47]	 [0.4844]	 [0.4844]	 [0.49]
Income per capita	 604	 557	 703	 751
	 [1227]	 [797]	 [959]	 [2451]
Expenditure per capita	 316	 330	 299	 361
	 [252]	 [316]	 [290]	 [483]
Calories per capita	 1767	 1783	 1817	 1851
	 [628]	 [588]	 [549]	 [601]
Rice per capita	 317	 325	 340	 338
	 [122]	 [129]	 [128]	 [120]
Meat per capita	 50.4	 42.4	 40.7	 52.8
	 [81.6]	 [61.0]	 [59.2]	 [70.3]
Rice calorie share	 0.639	 0.636	 0.645	 0.642
	 [0.188]	 [0.186]	 [0.158]	 [0.152]
Observations	 161	 162	 162	 159
Panel B: Gansu
Family size	 2.9	 2.7	 2.7	 2.7
	 [1.1]	 [1.1]	 [0.95]	 [1.1]
No. of children	 0.56	 0.55	 0.54	 0.54
	 [0.64]	 [0.69]	 [0.66]	 [0.60]
Female head	 0.44	 0.40	 0.44	 0.44
	 [0.50]	 [0.49]	 [0.50]	 [0.50]
Income per capita	 694	 694	 724	 726
	 [663]	 [652]	 [800]	 [697]
Expenditure per capita	 202	 228	 198	 216
	 [247]	 [214]	 [231]	 [201]
Calories per capita	 1737	 1732	 1716	 1655
	 [496]	 [553]	 [500]	 [520]
Wheat per capita	 353	 353	 341	 329
	 [132]	 [147]	 [136]	 [120]
Meat per capita	 13.9	 9.7	 13.5	 13.6
	 [30.9]	 [23.8]	 [33.7]	 [31.1]
Wheat calorie share	 0.691	 0.691	 0.678	 0.680
	 [0.176]	 [0.172]	 [0.181]	 [0.165]
Observations	 163	 162	 162	 162
Notes: Standard deviations in brackets. All consumption figures are in grams per capita. Calorie share
is the percent of total calories attributable to the particular food category. Income and expenditure per
capita are in 2006 yuan (Rmb). 1 jin 5 500g. The only pair-wise difference that is statistically significant
(at the 10 percent level) is meat per capita consumption in Hunan for the 0.3yuan/jin versus 0.2yuan/
jin groups.
September 20081562 THE AMERICAN ECONOMIC REVIEW
The consumption of wheat is much lower, with only 42g of daily consumption per person on
average, comprising just 8 percent of total caloric intake. By contrast, Gansu features almost the
exact reverse pattern; wheat-based foods are the dominant staple, with 344g of consumption per
person per day, comprising 69 percent of total calories, whereas rice consumption is only 35g.
In both provinces, the relevant staple good is a dominant source of calories for most households,
with 80 percent of households in Hunan relying on rice for at least half their calories and 75
percent of households in Gansu similarly relying on wheat.19
The reliance on these basic foods for nutrition is underscored even more by the fact that the
average total calorie share from all cereals or grains is 72 percent in Hunan and 77 percent
in Gansu. Further, in both provinces, on average 13 percent of calories come from edible oils
(mostly vegetable oil), which are primarily used in cooking, and are generally not a substitute for
other forms of consumption or nutrition. Thus, the consumption of all other foods combined on
average contributes only 10 percent of calories in Gansu, and 15 percent in Hunan. In Hunan, the
greatest remaining share comes from meat, with 42 grams of consumption per person per day on
average, comprising 7 percent of average caloric intake. By contrast, in Gansu meat consumption
is much lower, averaging only 13 grams per person per day and contributing less than 1 percent
of total caloric intake. Consumption of pulses is in fact greater than consumption of meat in
Gansu. This is likely due to the lower income levels in this province; pulses are often referred to
as “poor man’s meat” because they are a cheaper source of protein (when combined with other
foods typically eaten as staples). Therefore, while the consumption patterns in Hunan match up
well with the basic conditions under which we predict Giffen behavior, in Gansu the patterns do
not fit quite as well due to relatively low consumption of meat (the fancy good in our setup).
19
These goods also fit the basic Giffen conditions in that they are the cheapest source of calories in each province:
rice in Hunan yields 1,399 calories/yuan, while wheat in Gansu yields 1,655 calories/yuan. By contrast, the calories per
yuan for other common foods are: wheat (1,221), millet (537), pork (331), bean curd (239), and cabbage (141) in Hunan,
and millet (1,105), rice (980), pork (340), bean curd (224), and cabbage (173) in Gansu.
Table 2—Daily Consumption Per Capita and Calorie Shares for Food Categories
	 Hunan	 Gansu
	 Consumption (g)	 Calorie share	 Consumption (g)	 Calorie share
Rice	 330	 0.64	 35	 0.07
	 [125.4]	 [0.17]	 [69.5]	 [0.13]
Wheat	 42	 0.08	 344	 0.69
	 [60.2]	 [0.12]	 [134.3]	 [0.17]
Other cereals	 1.5	 0.00	 4.2	 0.01
	 [21.3]	 [0.022]	 [24.2]	 [0.050]
Vegetables and fruit	 341	 0.05	 232	 0.07
	 [194.6]	 [0.044]	 [141.6]	 [0.045]
Meat (including eggs)	 47	 0.07	 13	 0.01
	 [68.6]	 [0.11]	 [30.1]	 [0.037]
Pulses	 62	 0.02	 36	 0.02
	 [102.3]	 [0.043]	 [68.1]	 [0.056]
Dairy	 1	 0.00	 19	 0.01
	 [7.4]	 [0.0031]	 [56.6]	 [0.039]
Fats	 26	 0.13	 23	 0.13
	 [20.4]	 [0.095]	 [16.3]	 [0.090]
Calories	 1,805	 —	 1,710	 —
	 [591.7]		 [517.4]	
Observations	 644	 644	 649	 649
Notes: Standard deviations in brackets. All consumption figures are in grams per capita. Calorie share is the percent of
total calories attributable to the particular food category.
VOL. 98 NO. 4 1563Jensen and Miller: Giffen Behavior and Subsistence Consumption
C. Estimation Strategy
Given the random assignment of the price change and the panel nature of our survey, our basic
strategy is to simply compare the household-level changes in dietary intake20
of the staple good
for treatment and control groups. Since assignment to treatment and control groups was randomized
within sample counties, we add county*time fixed effects, so that we are in effect comparing
the changes for households with different subsidy levels within the same community. This
strategy controls for any county-level factors that change over time, such as the price of foods,
labor market conditions, or the value of government transfer programs.
We regress the percent change in intake of the staple good for household i in period t on the
change in the subsidy (in percent). The percent change formulation normalizes for factors such
as household size, composition, and activity level and allows us to interpret the coefficients as
elasticities. For each household, we observe two changes: the change between periods 2 and 1
(t 5 2), capturing the effect of imposing the subsidy; and the change between periods 3 and 2
(t 5 3), capturing the effect of removing the subsidy. Thus, we estimate
(1) 	 %Dstaplei, t 5 a 1 b%Dpi, t 1 gg%DZi, t 1 gdCounty*Timei, t 1 Dei, t  ,
where %Dstaplei, t is the percent change in household i ’s consumption of the staple good, %Δpi, t is
the percent change in the price of the staple due to the subsidy (negative for t 5 2 and positive for
t 5 3), %DZ is a vector of percent changes in other control variables including income (split into
earned and unearned (government payments, pensions, remittances, rent, and interest from assets)
sources) and household size, and County*Time denotes a set of county3time dummy variables.
We compute all changes as arc-percent-changes (i.e., 10031xt – xt212/11xt 1 xt212/22).21
The percent
change in the subsidy is computed as 100 times the change in the subsidy divided by the average
(net of subsidy) price of the staple good in the two corresponding rounds. The results are robust to
a wide range of alternative specifications, some of which we discuss in Section IIIB.
D. Refining the Test for Giffen Behavior
The discussion in Section I highlighted a nuanced prediction of the standard consumer model
in light of subsistence concerns. The poorest of the poor should have a negative price elasticity,
the poor-but-not-too-poor may have a positive price elasticity, i.e., exhibit Giffen behavior, and
the relatively wealthy should once again have a negative price elasticity. Although our primary
concern in this paper is documenting the presence of Giffen behavior, we are also interested in
testing this broader prediction of the theory.
Unfortunately, classifying households or individuals directly by consumption zone is not possible.
Not only is there no consensus on what constitutes a subsistence level of calories, but any
such threshold would certainly vary widely by age, sex, height, weight, body fat, muscle composition,
level of physical activity, health status, and a range of other factors. As a result, although
we can compute caloric intake for each individual, identifying whether specific individuals are
below, near, or above their subsistence level of caloric requirements is not possible. For the
20
While we also gathered data on food purchases and expenditures, actual daily intake is likely to be a better measure
of consumption or demand. This is due to the fact that food is storable, purchases are lumpy or infrequent, and
households’ recall of what they ate the day before the survey is likely to be more accurate than recall of purchases over
the past month.
21
We prefer the arc-percent-change specification over the simple percent change because the subsidies represent
large changes and because the arc formulation has the desirable property of being symmetric over time. The results are
largely unchanged, however, if we use the simple percent change instead.
September 20081564 THE AMERICAN ECONOMIC REVIEW
same reason, it is not possible to define these regions based on income or expenditure; individuals
with different characteristics will require different amounts of expenditures or income to
achieve nutritional sufficiency. Any such cutoffs would be imperfect, including some people
who, because of high weight or activity levels, are unable to achieve maintenance nutrition with
the specified income, and excluding others who have lower than expected nutritional (and thus
income) needs because of small stature or low activity levels.
The method for classifying households we employ is based in the theory. Those who are so
poor that they cannot achieve subsistence nutrition will consume a very high proportion of their
food in the form of the staple good, regardless of size and activity level. Thus, splitting the data
by the pre-intervention or initial share of caloric intake from consumption of the staple (Initial
Staple Calorie Share, or ISCS) provides a more direct measure of whether a consumer or household
is well off enough that they could, potentially, exhibit Giffen behavior. Consider panel B of
Figure 1. Along the x-axis, 100 percent of calories come from the staple, while along the y-axis,
the share is zero. In between these extremes, the share of calories from the staple is constant along
rays from the origin, with the share decreasing monotonically as the slope of the ray increases.
This provides a method of identifying whether a consumer is in the calorie-deprived, subsistence,
or standard zone based on the share of calories received from the staple good. Consumers
in the calorie-deprived zone will have a high ISCS, consumers in the standard zone will have a
low ISCS, and consumers in the subsistence zone will have an intermediate ISCS.
While just using ISCS does not overcome the problem of identifying the exact threshold cutoff
for moving from the calorie-deprived to subsistence zones, the advantage of this measure is
that it is more “need neutral,” in that it normalizes for individual differences in caloric requirements.
The measure also captures the simpler idea that if a household is so poor that it does not
consume any of the fancy good, it cannot respond to a price increase by consuming less of it.
While the ISCS may not be a perfect indicator of whether a household is near the subsistence
zone (because of unobserved taste variation, for example), we believe it to be superior to other
available measures.22
Exploratory calculations using a simplified version of a minimum-cost diet problem (see, for
example, the discussion in Robert Dorfman, Samuelson, and Robert M. Solow 1958) for China
suggest that the ISCS associated with a minimum-cost, nutritionally sound diet (designed to
ensure adequate consumption of calories and protein, and consisting of rice or wheat and bean
curd) is much less variable than either required calories or required expenditure (details provided
in the online Appendix). We compute the minimum cost diet for a range of weight/age/
gender/activity level combinations, and find that the ISCS associated with the minimum-cost,
nutritionally sound diet ranges between only 0.79 and 0.86 in Hunan and 0.78 to 0.85 in Gansu.
Consumers or households wealthy enough to be consuming a diet with a lower ISCS would
seem to be those who could, in principle, exhibit Giffen behavior. In light of this, our baseline
specification splits households based on whether their ISCS is less than 0.80 (this corresponds
22
The broad conclusions of our analysis hold if we instead use, say, staple budget shares to parse the data (see
the online Appendix). We believe, however, that the ISCS is a more appropriate measure. First, expenditure data are
notoriously noisy, especially due to large but infrequent purchases such as durable goods that can skew budget shares.
By contrast, the diets of households in our sample rarely contain more than three or four items and typically do not
vary from day to day. Second, as shown, ISCS is not very need dependent, whereas budget share thresholds will vary
considerably by household, due to differing housing, health care, education, and nutritional needs. Further, while some
expenditures such as entertainment are highly discretionary, others categories such as housing, health care, and utilities
are much less so. Thus, unlike the fairly precise ISCS cutoffs derived, clean cutoffs based on budget shares are difficult
to derive, since the amount of truly discretionary income is difficult to measure. These thresholds would therefore
contain many classification errors. For example, households with school-aged children would appear “richer” by the
budget share measure because of expenditures on school fees; however, these expenditures are not really discretionary,
and in fact make the household less “wealthy” than an identical household not facing these fees.
VOL. 98 NO. 4 1565Jensen and Miller: Giffen Behavior and Subsistence Consumption
approximately to the eightieth percentile of the staple calorie share distribution). However, we
also explore the robustness of the results to different thresholds.
While the theory suggests we should also exclude the wealthier households in the standard
zone of consumption, unlike the threshold for segregating households that are too poor, it is
unfortunately not possible to estimate the threshold for this region. Further, because our sample
is drawn from the poorest households, there is no guarantee we even have any households in
this zone. Therefore, we begin by taking the conservative approach of using only the threshold
excluding the poorest. If our theory is correct, if anything, keeping the lower tail of the staple
calorie share distribution will make it less likely we find Giffen behavior, since we are potentially
including households with downward sloping demand among our potential Giffen consumers
(we explore this possibility in Section IIIC).
III.  Results
A. Hunan
The estimation results for equation (1) for Hunan are presented in Table 3. For all regressions,
we present standard errors clustered at the household level. Starting with the full sample of
households and excluding all other controls, in column 1, a 1 percent increase in the price of rice
causes a 0.22 percent increase in rice consumption (i.e., consumption declines when the subsidy
is added, and increases when it is removed).23
While the estimate of the elasticity is positive, the
coefficient is not statistically significant at conventional levels (the p-value is 0.14). Column 2
23
Although our intervention caused a price decrease between rounds 1 and 2 and a corresponding increase between
rounds 2 and 3, for ease of exposition and interpretation we will typically refer to the effects of a price increase, the
more traditional and intuitive way of describing Giffen behavior.
Table 3—Consumption Response to the Price Subsidy: Hunan
	
Dependent variable: Rice Dependent variable: Meat	
							 ISCS		 Initial intake
	 Full sample	 Full sample	 ISCS #0.80	 ISCS #0.80	 ISCS .0.80	 ISCS .0.80	 0.60−0.80	 Full sample	 .50g
	 (1)	 (2)	 (3)	 (4)	 (5)	 (6)	 (7)	 (8)	 (9)
%∆Price(rice)	 0.224	  0.235*	 0.451***	 0.466***	 20.61**	 20.585**	    0.640***	 20.325	 21.125*
	 (0.149)	 (0.140)	 (0.170)  	 (0.159)   	 (0.296)	 (0.262)	 (0.192)	    (0.472)	   (0.625)
%∆ Earned		    0.043***		 0.047***		 0.024	 0.030	   0.028	  0.105
		 (0.014)		 (0.016)   		 (0.023)	 (0.019)	   (0.050)	   (0.069)
%∆Unearned		 20.044*  		 20.038     		 20.058  	 20.053* 	   0.061	   0.084
		 (0.025)		 (0.030)   		 (0.049)	 (0.030)	   (0.079)	  (0.104)
%∆People		   0.89***		 0.83*** 		   1.16***	   0.79***	 20.08    	 0.03
		 (0.08) 		 (0.09)    		 (0.15) 	 (0.14) 	  (0.27)	 (0.36)
Constant		  4.1***		 5.7***  		 21.8     	 0.8  	 212.3***	 249.0*** 
		 (1.0)   		 (1.1)     		 (1.7)  	 (1.3)  	 (3.1)	 (3.7)  
Observations	 1,258	 1,258	 997	 997    	 261	 261	 513	 997	 452
R2
	 0.08	 0.19    	 0.09	 0.20    	 0.15	 0.33	 0.24	 0.09	 0.28
Notes: Regressions include County*Time fixed effects. The dependent variable in columns 1−7 is the arc percent
change in household rice consumption, and in columns 8−9 it is the arc percent change in household meat consumption.
Standard errors clustered at the household level. %∆Price(rice) is the change in the subsidy, measured as a percentage of
the average price of rice; %∆Earned is the arc percent change in the household earnings from work; %∆Unearned is the
arc percent change in the household income from unearned sources (government payments, pensions, remittances, rent,
and interest from assets); %∆People is the arc percent change in the number of people living in the household. ISCS
(Initial Staple Calorie Share) refers to the share of calories consumed as rice in the preintervention period. *Significant
at 10 percent level. **Significant at 5 percent level. ***Significant at 1 percent level.
September 20081566 THE AMERICAN ECONOMIC REVIEW
adds changes in income (earned and unearned) and household composition. Controlling for these
other variables will help absorb any residual variation, and isolate the “pure” price effect of the
intervention, as opposed to any behavioral effects the intervention may have on household size
or either source of income (though in regressions for both provinces and for all population subgroups,
the effect of the subsidy on these other variables is small and not statistically significant,
suggesting the treatment had no such behavioral effects). Adding these other control variables
changes the results only very slightly, increasing the coefficient and improving precision. While
this coefficient is statistically significant only at the 10 percent level, it provides our first suggestive
evidence of Giffen behavior in Hunan. As would be expected for households exhibiting
Giffen behavior, the income effect is negative for unearned income, confirming that rice is an
inferior good. The point estimate of the elasticity of unearned income is small, though there is
likely to be significant measurement error in this variable, biasing the coefficient toward zero.24
As we have emphasized, however, Giffen behavior is likely to be exhibited only by a specific
subset of the poor. Therefore, in columns 3 through 6 we refine the test by parsing the data
according to the theory, separating households by whether their preintervention staple calorie
share suggests they are likely to be too poor to purchase something other than rice. For the group
consuming at least some substantial share (20 percent) of calories from sources other than rice
(column 3), i.e., the poor-but-not-too-poor, we find very strong evidence of Giffen behavior. A 1
percent price increase causes a 0.45 percent increase in consumption, and the effect is statistically
significant at the 1 percent level (and little changed by adding in the other control variables).
Thus, as theorized by Marshall and others, when faced with an increase in the price of the staple
good, these households do, indeed, “consume more, and not less, of it” (Marshall 1895).
By contrast, but again consistent with the theory, the group consuming more than 80 percent
of their total calories from rice (i.e., those still largely unable to consume meat), respond in
the opposite direction (columns 5 and 6), with a large decline in rice consumption. Since these
households consume essentially only rice, they have no choice but to respond to an increase in
the price of rice by reducing demand. Thus, beyond finding evidence of Giffen behavior, the
results also provide initial support for the subsistence model underlying such behavior. We find
Giffen behavior where the theory predicts it, and downward sloping demand elsewhere. We
explore the subsistence model further in Section IIIC below.
B. Robustness
The finding of Giffen behavior is robust to a wide range of alternate specifications, shown in
Table 4. Since including the change in household size or either source of income rarely makes
more than a marginal difference on our estimates of the price elasticity, for conciseness of presentation
we show only the results with these additional control variables included. Columns 1
to 3 present results from a log-log specification, regressing the change in the log of household
rice consumption on the change in the log of the net-of-subsidy price of rice and changes in
the logs of the other control variables. The results again reveal Giffen behavior for households
consuming less than 80 percent of their calories from rice, and downward sloping demand for
those above this threshold. The point estimates of the elasticities are much greater here than for
the arc percent changes in Table 3. However, this difference is largely attributable to the greater
weight given to very low values with a log specification; for example, if we trim just the ­lowest
24
The coefficient on earned income is positive (though also small); however, since greater caloric intake may
improve productivity and earnings (Thomas and Strauss 1997), especially among those with very low nutritional status,
this coefficient may be biased due to endogeneity. Unfortunately, we lack convincing instruments for changes in earned
income. Dropping this variable does not change the results.
VOL. 98 NO. 4 1567Jensen and Miller: Giffen Behavior and Subsistence Consumption
1 percent of rice consumers in Hunan, the price elasticity coefficients are almost identical to
those in Table 3 (0.229 (0.183), 0.461 (0.218), and 20.558 (0.250) for the full sample and the
less than and greater than 80 percent staple calorie share groups, respectively). Returning to our
main specification for the independent variables (equation (1)) but using the level change in rice
consumption per capita (rather than total household consumption)25
as the dependent variable
(columns 4–6) or the percent change in consumption using individual-level data (adults only;
columns 7–9) again reveals Giffen behavior for the group with less than 80 percent calorie share
(though the results for those with greater than 80 percent, while negative, are no longer statistically
significant).
To explore the robustness of the conclusions to an alternative way of classifying households
into consumption zones, columns 10–13 return to equation (1) but split households by preintervention
expenditure per capita.26
As described earlier, due to variations in individual and household
characteristics, we believe expenditure to be an inferior method of classifying consumers
into different consumption zones. Nevertheless, doing so provides a useful robustness check.
Lacking in this case a threshold based on a cost minimization problem, we simply stratify households
based on whether they are above or below the fifteenth or twenty-fifth percentile of the
expenditure distribution. We again see evidence of Giffen behavior among the poor-but-not-toopoor.
Those above the bottom quartile (column 10) respond to a 1 percent increase in the price of
rice by increasing rice consumption by 0.29 percent, though the effect is statistically significant
at only the 10 percent level. And unlike the case of stratifying by staple calorie share, the poor
group in this case does not decrease consumption in response to a price increase; this is likely
due to the relative imprecision of relying on the expenditure-based threshold. Using the fifteenth
25
Using the percent change in consumption per capita yields nearly identical results to those in Table 4.
26
Ideally, we would use the data from each particular round to assess living standards rather than using only the preintervention
data, since Giffen behavior depends on a consumer’s budget at the time he makes his decisions. However,
expenditure in the round with the subsidy is obviously endogenous with respect to the subsidy; income would encounter
enodgeneity as well (the increased consumption afforded by the subsidy might affect earnings).
Table 4—Robustness of Results to Alternative Specifications: Hunan
	 log-log specification	 Consumption per capita	 Individual-level data	 Expenditure per capita thresholds
	 Full	 #0.80	 .0.80	 Full	 #0.80	 .0.80	 Full	 #0.80	 .0.80	 $25th	 ,25th	 $15th	 ,15th
	 (1)	 (2)	 (3)	 (4)	 (5)	 (6)	 (7)	 (8)	 (9)	 (10)	 (11)	 (12)	 (13)
%∆Price(rice)	 0.399	 0.694**	 20.718**	 0.762*	 1.348***	 21.348	 0.233	 0.384**	 20.223	 0.286*	 0.139	 0.301**	20.132
	 (0.254)	 (0.304)	 (0.294)	 (0.423)	 (0.476)	 (0.842)	 (0.144)	 (0.169)	 (0.225)	 (0.167)	 (0.238)	 (0.153)	 (0.288)
%∆Earned	 0.010**	 0.012	 0.003	 0.091**	 0.103**	 0.041	 0.041***	 0.046***	 0.022	 0.039**	 0.050**	 0.041***	0.054**
	 (0.006)	 (0.007)	 (0.007)	 (0.043)	 (0.048)	 (0.083)	 (0.014)	 (0.016)	 (0.024)	 (0.017)	 (0.023)	 (0.015)	 (0.026)
%∆Unearned	 20.031**	 20.030	 20.038	 20.107	 20.066	 20.225	 20.061**	 20.051	 20.082**	 20.037	 20.068*	 20.033	 20.104**
	 (0.018)	 (0.020)	 (0.025)	 (0.072)	 (0.080)	 (0.174)	 (0.027)	 (0.033)	 (0.041)	 (0.030)	 (0.040)	 (0.028)	 (0.046)
%∆People	 0.93***	 0.85***	 1.27***	 20.28	 20.55	 0.89	 0.01	 20.08	 0.27	 0.89***	 0.86***	 0.87***	 1.15***
	 (0.10)	 (0.10)	 (0.19)	 (0.32)	 (0.35)	 (0.57)	 (0.09)	 (0.10)	 (0.17)	 (0.10)	 (0.15)	 (0.09)	 (0.18)
Constant	 0.04**	 0.05**	 20.003	 11.9***	 16.7***	 25.3	 5.3***	 6.5***	 0.8	 4.3***	 3.4*	 3.9***	 5.4***
	 (0.02)	 (0.02)	 (0.18)	 (3.0)	 (3.2)	 (6.0)	 (1.0)	 (1.2)	 (1.7)	 (1.1)	 (1.7)	 (1.1)	 (2.0)
Observations	 1256	 997	 259	 1258	 997	 261	 2755	 2191	 564	 971	 287	 1083	 175
R2	 0.11	 0.11	 0.31	 0.09	 0.11	 0.18	 0.05	 0.06	 0.10	 0.18	 0.31	 0.19	 0.35
Notes: Regressions include County*Time fixed effects. Standard errors clustered at the household level. For columns 4–13: %∆Price(rice) is
the change in the subsidy, measured as a percentage of the average price of rice; %∆Earned is the arc percent change in the household earnings
from work; %∆Unearned is the arc percent change in the household income from unearned sources (government payments, pensions, remittances,
rent, and interest from assets); and %∆People is the arc percent change in the number of people living in the household. For columns
1–3, these percent changes result from using the log of the relevant variables. In columns 10−13, Expenditure per capita refers to a household’s
percentile in the distribution of expenditure per capita in the preintervention period. *Significant at 10 percent level. **Significant at 5 percent
level. ***Significant at 1 percent level.
September 20081568 THE AMERICAN ECONOMIC REVIEW
percentile cutoff, we see strong evidence of Giffen behavior for the poor-but-not-too-poor, and
now the coefficient for the poorest is negative, though not statistically significant.
As a final robustness check, since the 80 percent threshold for the rice calorie share was
a rough approximation based on a minimum-cost diet calculation, Table 5 shows the original
regressions using alternative thresholds. As the threshold varies from 70 to 90 percent, the point
estimate of the elasticity for those below the threshold varies only from 0.27 to 0.47, with statistically
significant coefficients in all cases. Therefore, the results point convincingly and robustly
to the conclusion of Giffen behavior in Hunan. Additionally, as might be expected from the
subsistence model, the coefficients broadly increase as the staple calorie share threshold declines
from 0.90 to 0.75, as we are in effect excluding more and more of the least well-off, i.e., those
most likely to respond to a price increase by decreasing consumption. The coefficients for each
corresponding group above the threshold staple calorie share are negative for all thresholds up
to 0.70; however, due in part to the smaller sample sizes in some of the cases, the effects are
statistically significant only at the 10 percent level or better for the 75, 80, and 85 percent thresholds.
The increase in the coefficients as the threshold moves from 0.85 to 0.70 is consistent with
increasingly including some of the least poor of the poor who are in the subsistence rather than
the calorie-deprived zone, for whom the response to a price increase is positive.
Thus, overall, across a range of specifications, alternative thresholds, and ways of classifying
households into consumption zones, the results point to robust evidence of Giffen behavior with
respect to rice in Hunan.27
C. Exploring the Subsistence Model and Refining the Giffen Zone
Beyond providing evidence of Giffen behavior, our study aims to document more broadly
the behavior of extremely poor households in order to highlight some key insights relevant for
27
Two additional refinements are worth reporting. First, if we include an interaction between the subsidy and round
variables, in all cases we cannot reject the hypothesis that the effects are equal for adding versus removing the subsidy.
Second, we find Giffen behavior separately for male- and female-headed households, though the threshold at which the
effects are statistically significant is lower for male-headed households.
Table 5—Robustness to Alternative Staple Calorie Share Thresholds: Hunan
	 #0.70	 #0.75	 #0.80	 #0.85	 #0.90	 .0.70	 .0.75	 .0.80	 .0.85	 .0.90
	 (1)	 (2)	 (3)	 (4)	 (5)	 (6)	 (7)	 (8)	 (9)	 (10)
%∆Price(rice)	 0.362**	 0.461***	 0.466***	 0.382***	 0.270*	 0.004	 20.331	 20.585**	 20.934*	 20.617
	 (0.184)	 (0.174)	 (0.159)	 (0.145)	 (0.143)	 (0.203)	 (0.207)	 (0.262)	 (0.471)	 (0.681)
%∆ Earned	 0.052***	 0.049***	 0.047***	 0.044***	 0.043***	 0.028	 0.028*	 0.024	 0.027	 0.094
	 (0.019)	 (0.018)	 (0.016)	 (0.015)	 (0.014)	 (0.018)	 (0.017)	 (0.023)	 (0.035)	 (0.072)
%∆Unearned	 20.007	 20.027	 20.038	 20.041	 20.044*	 20.093**	 20.076*	 20.058	 0.001	 20.036
	 (0.034)	 (0.031)	 (0.030)	 (0.027)	 (0.026)	 (0.037)	 (0.044)	 (0.049)	 (0.085)	 (0.154)
%∆People	 0.77***	 0.79***	 0.83***	 0.87***	 0.87***	 1.13***	 1.17***	 1.16***	 1.03***	 1.35***
	 (0.11)	 (0.10)	 (0.09)	 (0.09)	 (0.09)	 (0.10)	 (0.12)	 (0.15)	 (0.26)	 (0.22)
Constant	 7.4***	 6.1***	 5.7***	 4.8***	 4.3***	 20.9	 20.3	 21.8	 21.3	 2.8
	 (1.3)	 (1.2)	 (1.1)	 (1.1)	 (1.0)	 (1.3)	 (1.4)	 (1.7)	 (2.3)	 (4.0)
Observations	 777	 883	 997	 1116	 1196	 481	 375	 261	 142	 62
R2
	 0.20	 0.19	 0.20	 0.20	 0.19	 0.30	 0.34	 0.33	 0.31	 0.49
Notes: Regressions include County*Time fixed effects. The dependent variable is the arc percent change in household rice consumption.
Standard errors clustered at the household level. %∆Price(rice) is the change in the subsidy, measured as a percentage of the average
price of rice; %∆Earned is the arc percent change in the household earnings from work; %∆ Unearned is the arc percent change in
the household income from unearned sources (government payments, pensions, remittances, rent, and interest from assets); %∆People
is the arc percent change in the number of people living in the household. *Significant at 10 percent level. **Significant at 5 percent
level. ***Significant at 1 percent level.
VOL. 98 NO. 4 1569Jensen and Miller: Giffen Behavior and Subsistence Consumption
academics and policymakers. We have already seen that consumers with very high staple calorie
shares (i.e., the poorest-of-the-poor) do not exhibit Giffen behavior. In addition, the model also
predicts that once consumers are wealthy enough to pass beyond the subsistence zone into the
standard consumption zone, staple demand should once again slope downward; in effect, we
predict an inverted-U shape, with downward sloping demand (negative coefficients) for low and
high values of staple calorie share, and Giffen behavior (positive coefficients) for intermediate
values.28
As stated, unlike the 80 percent calorie share, it is not possible to define a threshold
beyond which households are likely to be in the standard or normal consumption zone, nor are
we even guaranteed that our sample of the urban poor contains any such households. We therefore
take a simple, flexible approach using a series of locally weighted regressions. At each staple
calorie share point from 0.30 to 0.95 (there are few observations below 0.30 or above 0.95), we
estimate equation (1) using a window of staple calorie shares of 0.05 on either side of that point;
within that window we estimate a weighted regression, where observations closest to the central
point receive the most weight (we use a biweight kernel, though the results are robust to alternatives).
Figure 2 plots the resulting coefficients on the arc percent price change variable (i.e., the
price elasticity) at each initial staple calorie share point for Hunan, along with the associated
95 percent confidence interval. The basic inverted-U shape in staple calorie share is clear. The
elasticity is negative for the lowest and highest staple calorie shares, and positive in between.
The Giffen range, where the point estimate of the elasticity is positive, reaches from 0.53 to 0.84
(which includes nearly two-thirds of the Hunan sample) though it is only statistically significant
from 0.63 to 0.75. The peak of the curve reaches an elasticity of 0.85, at a staple calorie share of
0.70. And the threshold at which the elasticity turns negative is 0.80, which matches surprisingly
well our simple minimum cost diet calculation. In general, the precision of these estimates is
lower than those observed in Tables 3 to 5, since here we are restricting each regression to a band
of 60.05 around a particular point, which reduces the sample size.
Not only does this figure support the theory in that Giffen behavior is most likely to be found
among a range of households that are poor (but not too poor or too rich), it also guides us to a
particular range when theory cannot provide a specific set of thresholds, as with the threshold
between the subsistence and normal consumption zones. In particular, this curve suggests we
restrict the range in which we test for Giffen behavior not only to those with a staple calorie
28
Though, if we do not have enough households wealthy enough to fall into the normal consumption zone, we expect
that the coefficients should at least decline as staple calorie share declines.
Figure 2. Coefficient Plots
September 20081570 THE AMERICAN ECONOMIC REVIEW
share less than 0.80, but also to those with at least, say, 0.60. In column 7 of Table 3, doing so
increases the point estimate of the elasticity dramatically, from 0.47 to 0.64, as we are in effect
removing the wealthiest households.29
And even with the smaller sample, the effect is statistically
significant at the 1 percent level, again strongly supporting the conclusion of Giffen behavior
in Hunan.
A second prediction of the subsistence model we can explore is that, in response to an increase
in the price of the staple good, consumers facing a subsistence constraint will consume not only
more of that good, but also less of the fancy good, which we identified here as meat. Column 8 of
Table 3 shows regressions like (1) above, but using the arc percent change in meat consumption
as the dependent variable (we focus on the sample of households with less than 80 percent rice
calorie share, though the results are robust to other thresholds). We find that the point estimate
of the elasticity of meat consumption with respect to the price of rice is negative as predicted,
though it is not statistically significant. However, one limitation of this analysis is that in Hunan,
only about 45 percent of households reported meat consumption.30
Therefore, in column 9 we
focus on households that consume at least 50g of meat per person in round 1, which is still a
very modest amount.31
Here, the results are more evident; a 1 percent increase in the price of rice
leads to a large (1.13 percent), statistically significant decrease in meat consumption, as predicted
by the model.
Thus, again, while our primary goal was to document the existence of Giffen behavior, these
two results (the inverted-U shape of the response of rice consumption to a change in price and the
decline in meat consumption in response to a change in the price of rice) support the subsistence
model of consumption with a staple good and a taste-preferred but more expensive source of
calories (such as meat) outlined above.
D. Gansu
As shown in Table 2, wheat-based foods (primarily buns, the simple bread mo, and noodles),
are the staple good in Gansu. However, not all wheat-based foods are made at home from flour;
most notably, noodles are often either consumed at restaurants or road-side food stalls, or purchased
from shops as a prepared or packaged food. Since the subsidy we provided applied only to
the purchase of wheat flour, for our analysis we use only the consumption of wheat foods typically
produced at home from flour.32
And, as suggested by the calculations in the online Appendix,
because there is some consumption of these other forms of wheat, our threshold staple calorie
share for Giffen behavior based on wheat flour alone is closer to 0.70.33
Table 6 presents the main
results. In contrast to the case of Hunan, the coefficient is negative for the full sample in column 1.
Even focusing on those below the staple calorie share threshold of 70 percent (column 2), while the
coefficient is positive, it is extremely small and not statistically ­significantly different from zero.
In addition, there is no evidence that wheat is even an inferior good in these cases.
29
This coefficient differs slightly than the peak coefficient in the figure since the latter arises from a weighted
regression, with more weight assigned to points closer to the peak of the curve.
30
Although we condition on the staple calorie share in our regressions, the residual is not simply calories from
meat.
31
While it may seem natural to have run all the specifications above stratifying based on meat consumption rather
than staple calorie share, the latter is more general and does not rely on our ability to specifically identify meat as the
(only) fancy good.
32
Over 90 percent of the consumption of wheat-based foods in Gansu was reported as “wheat flour,” with most of
the remainder reported as noodles. However, we cannot rule out that some noodles were made at home from flour but
recorded as noodles, or that some consumers mistakenly reported purchased bread as wheat flour.
33
Alternatively, we could use a staple calorie share of 0.80 based on consumption of all wheat foods, rather than just
those produced at home from flour.
VOL. 98 NO. 4 1571Jensen and Miller: Giffen Behavior and Subsistence Consumption
Looking across alternative thresholds in columns 4 through 10, we do find that the coefficients
increase and ultimately turn positive as the staple calorie share decreases toward 60 percent,
consistent with excluding more and more households that are likely to be below the subsistence
consumption zone; however, the coefficient then abruptly declines when the share is lowered to
55 percent, and in none of the cases are the coefficients statistically significant.
As the model suggested and the analysis of Hunan revealed, focusing only on those below a
certain staple calorie share threshold risks including those who may be too wealthy to be Giffen
consumers. While in Hunan we were able to detect Giffen behavior even under the more conservative
approach (i.e., without appropriately parsing the data), it may be that we are simply unable
to in Gansu. Returning to Figure 2, as in Hunan the coefficients from the weighted regressions
for Gansu reveal an inverted-U response of wheat consumption to an own-price change over the
range of initial staple calorie share, though no coefficient is statistically significantly different
from zero at the 5 percent level over this interval. The range of positive point estimates is both
lower and narrower than in Hunan, ranging only from approximately 0.40 to 0.60; correspondingly,
column 11 of Table 6 shows that if we examine households in this range, there is evidence
of Giffen behavior, with a large elasticity (1.07), statistically significant at the 10 percent level.
While we are of course concerned about the inherent biases in searching over many intervals for
a result, both the theory outlined above and the pattern observed in Figure 2 point to the need to
examine only those who are poor, while excluding those who are too poor and not poor enough,
in testing for Giffen behavior. If not as compelling as the evidence in Hunan, the results are at
least suggestive of Giffen behavior in Gansu.
Without discounting this last result, we turn now to consider possible explanations for why
the evidence of Giffen behavior in Gansu is less immediately evident and precisely estimated
than in Hunan. The earlier discussion suggested that Giffen behavior is most likely to be found
among consumers whose diet consists primarily of a single staple good, with relatively few substitutes,
and a fancy good, which is taste-preferred but a more expensive source of nutrition. We
­consider two potential failures of these conditions in Gansu. First, in our sample there is very
Table 6—Consumption Response to the Subsidy: Gansu
												 	 ,50g
	 Full											 .50g	 Substitute
	 sample	 #0.70	 .0.70	 #0.55	 #0.60	 #0.65	 #0.75	 #0.80	 #0.85	 #0.90	 0.40−0.60	 meat	 wheat
	 (1)	 (2)	 (3)	 (4)	 (5)	 (6)	 (7)	 (8)	 (9)	 (10)	 (11)	 (12)	 (13)
%∆Price(wheat)	20.353	 0.024	 20.825**	 20.245	 0.309	 0.128	 0.009	 20.280	 20.321	 20.356	 1.065*	 1.327*	 1.106*
	 (0.258)	 (0.366)	 (0.357)	 (0.453)	 (0.452)	 (0.414)	 (0.326)	 (0.302)	 (0.283)	 (0.268)	 (0.557)	 (0.701)	 (0.566)
%∆ Earned	 0.079**	 0.098*	 0.041	 20.048	 0.023	 0.064	 0.124***	 0.107**	 0.100**	 0.103***	 0.063	 0.139*	 0.156*
	 (0.036)	 (0.052)	 (0.049)	 (0.065)	 (0.062)	 (0.057)	 (0.045)	 (0.042)	 (0.040)	 (0.038)	 (0.074)	 (0.076)	 (0.080)
%∆Unearned	 20.017	 20.048	 0.035	 0.023	 0.045	 20.007	 0.005	 0.063	 0.034	 0.009	 0.189	 0.059	 20.056
	 (0.092)	 (0.129)	 (0.127)	 (0.189)	 (0.173)	 (0.141)	 (0.112)	 (0.105)	 (0.102)	 (0.093)	 (0.181)	 (0.147)	 (0.172)
%∆People	 0.58***	 0.34	 0.80***	 0.18	 0.25	 0.24	 0.40	 0.42*	 0.42*	 0.53**	 0.11	 1.70***	 0.45
	 (0.22)	 (0.30)	 (0.25)	 (0.41)	 (0.34)	 (0.32)	 (0.27)	 (0.25)	 (0.23)	 (0.22)	 (0.32)	 (0.23)	 (0.29)
Constant	 226.1***	 220.8***	 232.8***	 218.7***	 219.5***	 220.3***	 222.9***	 223.3***	 225.8***	 225.7***	 231.6***	 0.82	 226.8***
	 (2.3)	 (3.3)	 (2.9)	 (4.5)	 (4.1)	 (3.7)	 (3.0)	 (2.7)	 (2.6)	 (2.4)	 (4.4)	 (5.1)	 (5.5)
Observations	 1269	 687	 582	 406	 478	 563	 843	 995	 1107	 1199	 266	 107	 247
R2
	 0.08	 0.11	 0.09	 0.17	 0.14	 0.12	 0.09	 0.10	 0.08	 0.08	 0.24	 0.33	 0.22
Notes: Regressions include County*Time fixed effects. The dependent variable is the arc percent change in household wheat consumption.
Standard errors clustered at the household level. %∆Price(wheat) is the change in the subsidy, measured as a percentage of
the average price of wheat: %∆Earned is the arc percent change in the household earnings from work; %∆Unearned is the arc percent
change in the household income from unearned sources (government payments, pensions, remittances, rent, and interest from assets);
%∆People is the arc percent change in the number of people living in the household. Substitute wheat refers to consumption of wheatbased
foods such as noodles or bread that are purchased in a prepared form, rather than made at home from wheat flour. *Significant at
10 percent level. **Significant at 5 percent level. ***Significant at 1 percent level.
September 20081572 THE AMERICAN ECONOMIC REVIEW
­little ­consumption of the fancy good, meat.34
As shown in Table 2, households in Gansu receive
on average only 1 percent of their calories from meat, which is even less than the 7 percent
observed in Hunan; further, only one-quarter of households reported any meat consumption in
our first period consumption diary. The bulk of nonstaple calories comes largely from vegetables
(especially potatoes, which themselves may potentially be a staple food) and oil, neither of which
is likely to be considered a fancy good. With little consumption of the fancy good, it is perhaps
not surprising that most households do not behave like Giffen consumers in Gansu. There is
simply no way for them to finance additional purchases of rice by reducing meat, since they are
consuming almost no meat to begin with.35
This also suggests that the best place to find Giffen
behavior is among those consuming a nontrivial amount of meat. Therefore, in the next-to-last
column of Table 6, we consider only households that consume at least 50 grams of meat per
person in the initial period. Though the sample shrinks considerably because meat consumption
is so uncommon, we do find evidence of Giffen behavior among this group, with a 1 percent
increase in the price of wheat causing a 1.3 percent increase in wheat consumption.
Gansu also departs from the ideal conditions for Giffen behavior in that wheat as a staple is
consumed in a number of other forms that may act as substitutes for each other, many of which
are not made directly by consumers at home from wheat flour. Unfortunately, our experimental
design failed to account for this additional complexity.36
In Hunan, the staple good, rice, is
consumed typically only in its basic form. By contrast, in Gansu wheat is consumed as mo and
buns made at home, plus noodles, and other wheat-based, prepared foods like bread, biscuits,
or deep-fried dough purchased from shops or food stalls. While Table 2 showed that average
pretreatment wheat consumption per capita in Gansu was 344g, typically about 34 grams, or 10
percent, of that wheat is from items other than mo or buns. If a household consumes their staple
food in many forms and the price of one increases, they may not need to engage in Giffen behavior
because they can reduce consumption of that one and increase consumption of the other, substitutable
forms of the staple that did not experience the price increase. While this is unlikely to
happen often in reality because the price of all the forms of the staple will be linked to the price
of the raw ingredient (here, wheat), the unique structure of our subsidy did just that, subsidizing
only the form of the staple prepared at home, and not the close substitutes purchased in stores.
This may both explain why we do not find widespread evidence of Giffen behavior in Gansu, and
also suggests we might find such behavior if we focus on those households where consumption
of these other forms of wheat is small or zero in the initial period.37
The final column of Table 6
provides some suggestive evidence of this possibility, focusing on the condition that the household
consumes less than 50g of these alternative forms of wheat. Among this group there is again
statistically significant evidence of Giffen behavior, with a very large elasticity.
Overall, then, while the results for Gansu do not yield as evident, robust evidence of Giffen
behavior as was found for Hunan, we believe this is most likely due to our failure to recognize
ex ante that for the majority of households in our sample, diets do not conform to two of the basic
34
This result was unanticipated, since the northern provinces in our original paper (Jensen and Miller 2002), and
our field test of the survey for the current study, revealed considerably more meat consumption in Gansu.
35
While there is some consumption of pulses and, to a lesser extent, dairy, these goods are also unlikely to be
regarded as fancy goods in the way that meat is, since most households turn to these goods only when they cannot
afford meat. Further, there is no way to cut back consumption of these foods while maintaining protein intake; with
meat, households can reduce consumption but switch to pulses as a less expensive source of protein.
36
Though in selecting sample sites, the authors personally visited only two of the counties in Gansu (Anding and
Yuzhong). These counties, both with significant Muslim populations who traditionally consume primarily the home
made bread mo, fit the pattern better, with 88 percent of all wheat consumption coming from flour, compared to 74
percent in the other three counties. If we limit our analysis to just these two counties, we find a positive coefficient for
all staple calorie share thresholds, though due to the smaller samples, the coefficients are not statistically significant.
37
Some of this variation is geographic or based on religion, as noted above.
VOL. 98 NO. 4 1573Jensen and Miller: Giffen Behavior and Subsistence Consumption
conditions under which we predict Giffen behavior (consumption of a fancy good, and a staple
good for which there are no close substitutes). When we restrict our sample to take these factors
into consideration, we do find evidence of Giffen behavior, though the samples are smaller, precisely
because most households do not conform to the conditions in Gansu. It is possible that if
we sampled a slightly wealthier group of households that consume more of the fancy good, and
perhaps altered our experimental design (e.g., to subsidize all wheat foods, not just wheat flour),
we might find stronger evidence of Giffen behavior.
E. Addressing Potential Alternative Explanations for the Results
The analysis thus far provides robust evidence of upward sloping demand in Hunan, with
somewhat weaker evidence in Gansu. However, two explanations other than Giffen behavior
need to be explored. First, the vouchers themselves may have induced a behavioral effect. For
example, households may have increased demand for the staple because they interpreted being
given vouchers as a signal of its value (e.g., that it had significant health benefits). Alternatively,
the vouchers may have increased the salience of the staple, or households might have felt they
should eat more of it in order to take advantage of the subsidy before it ran out. However, in these
cases we would expect the vouchers to increase consumption, the opposite of what is observed
as Giffen behavior. Alternatively, and perhaps less likely, households may view the vouchers as
providing adverse information about the staple good; for example, they may view the attempt
to sell more rice as an indication that there is something wrong with the current stock, in which
case they might want to consume less of it (though consumers were told the subsidies were being
provided by outside researchers rather than merchants, farmers, or the government). But since
the effects vary by the staple calorie share, to explain our results it would have to be that the
vouchers had a salience or signal effect only for some subset of households based on their calorie
share, which seems unlikely.38
A second concern is the possibility that households cheated,39
for example by swapping vouchers
for cash instead of using them to purchase the staple,40
or by reselling rice or wheat purchased
with the vouchers at a higher price. In the extreme case where a household simply sells all their
vouchers, the program would be a pure wealth transfer, and consumption of inferior goods like
rice or wheat would be expected to decline even though their effective price had not changed. In
less extreme cases where only some vouchers are sold, even at less than face value, the program
would still provide a pure wealth transfer that could in itself decrease consumption of the staple,
which we might again misinterpret as Giffen behavior.41
38
If the consumption of the treatment groups responded both to having received any subsidy at all (i.e., the signal or
salience effect) and to size of the subsidy received, we could eliminate the former and identify the elasticity off of the
size of the subsidy alone by running regressions excluding the control group. Doing so yields similar results, including
continued evidence of Giffen behavior. For example, for Hunan, the price elasticity for the full sample is 0.33(0.22);
for ISCS less than 0.80 it is 0.61(0.25), and for ISCS greater than 0.80 it is 20.75(0.44). Similar results hold for Gansu,
though, as above, the estimates are less precise and in some cases we cannot reject zero even though the point estimates
have the correct sign. These results indicate that the Giffen effect is not driven by some common signaling or salience
effect among the treatment groups. However, it is of course possible that larger subsidies create stronger signaling
effects, so these results do not imply there were no such effects at all.
39
Cheating, where shopkeepers do not provide the full subsidy to consumers (for example, those with poor math
skills), effectively lowers the value of the subsidy, so the Giffen behavior we find would likely have been even stronger
had such cheating not occurred.
40
Most shopkeepers sold only grain, so most households could not have exchanged the vouchers for other foods.
41
If households bought rice at subsidized prices on behalf of (or as a gift to) their friends or relatives but did not
make a profit from doing so, this does not affect the households’ wealth and thus does not bias our experiment. We of
course cannot rule out that treatment households shared some of their rice with others. However, this “leakage” would
tend to diminish the likelihood of finding Giffen behavior, since the effective value of the subsidy is reduced.
September 20081574 THE AMERICAN ECONOMIC REVIEW
Preventing cashing out of the vouchers or resale was one of our primary concerns in designing
the intervention. In doing so, however, we also wanted to ensure that the process of redeeming
the vouchers would be as much like an ordinary market transaction as possible, and to keep the
administrative burden of the intervention manageable. In addition, we wanted to allow for the
fact that a natural reaction to receiving access to discounted rice or wheat would be for households
to build up their stores of these goods, which would look very similar to cashing out (i.e.,
the number of vouchers redeemed exceeds the amount of rice or wheat consumed).
With these concerns in mind, a number of safeguards were built into the experimental design.
First, the consent scripts given to treatment households stated they were explicitly prohibited
from selling the vouchers or the rice or wheat bought with the vouchers. They were also told that
they would be monitored and that violators would be dismissed from the program. Our native
Chinese implementation team, which was very familiar with the population from which our
survey households were drawn, felt that in light of such a rule households would be very unlikely
to cash out the vouchers.
Second, the vouchers were distributed only incrementally (one-fifth each month) over the
course of the intervention (though all vouchers remained valid until the end of the intervention).
As a result, households engaging in early cashing out would be limited in their ability to do so
(since they would have only a small part of the vouchers in hand) and would risk losing the value
of all future vouchers if they were caught. For our purposes, the crucial question is whether there
was cashing out before the second round of the survey, since this is the only round for which the
subsidy was in effect. At the time of the second survey, a significant amount of the benefit of the
program still lay in the future, which reduced the incentive to cash out.
Third, in order to ensure that they were used only by treatment households and not resold, the
vouchers (which were printed in Beijing in multicolor ink and bore a special stamp, making them
difficult to counterfeit in the survey regions) had to be signed by the redeemer at the time they
were used. Shopkeepers had an incentive to ensure the vouchers were used only by the appropriate
households, since they were reimbursed only after the vouchers and signatures had been
validated by our managers.42
Finally, shopkeepers were also given incentives to prevent cashing
out through a lump-sum payment given at the end of the intervention if they were found to have
complied with its guidelines, including preventing resale and cashing out vouchers.
The safeguards discussed above were accompanied by monitoring and auditing to check for
compliance. These audits did not discover any such cheating, and our survey personnel, who visited
the households, did not discover evidence of cashing out. Finally, comparing voucher redemptions
relative to estimated consumption and storage suggests that to the extent it occurred at all,
cashing out could not have been significant or widespread (details in the online Appendix).
IV.  Conclusion
We find strong, clear evidence of Giffen behavior among poor households in Hunan, China,
and somewhat less robust evidence in Gansu. To the best of our knowledge, this is the first rigorous,
real-world empirical evidence of Giffen behavior. It is ironic that despite a long search, in
sometimes unusual settings, we found examples in the most widely consumed foods for the most
populous nation in history. However, the examples were found exactly where theory would predict
they should occur: impoverished (but not too impoverished) consumers, heavily dependent
on a staple good, with limited substitution possibilities. And while our experiment focused on
two areas of China, our framework and analysis suggest that Giffen behavior may not be rare at
42
Our Chinese management team was the residual claimant on the value of unredeemed vouchers, and so they,
themselves, had a strong incentive to enforce the rules of the intervention and prevent cashing out.
VOL. 98 NO. 4 1575Jensen and Miller: Giffen Behavior and Subsistence Consumption
all. To begin, the socioeconomic class we are studying—the urban poor in China—is extremely
large, representing an estimated 90 million individuals. Thus, to the extent that our experiment is
representative of that population, in terms of number of people, Giffen behavior is probably quite
common. Second, our empirical approach differs in important ways from previous attempts to
identify Giffen behavior. Earlier approaches focused on market-level price and quantity data,
which for many reasons are unlikely to exhibit Giffen properties. Our approach uses individuallevel
data and focuses on parsing the data by an appropriate measure of wealth. These techniques
have never been applied to other likely candidates for Giffen behavior. We believe there
is nothing unusual or unique about the Chinese case; anecdotal evidence suggests that similar
dietary patterns to those observed here are found among much of the world’s poor (with staples
including maize, millet, sorghum, and cassava, in addition to rice and wheat).43
Thus, it may be
better to interpret our study as the first evidence of Giffen behavior rather than the only possible
evidence. In fact, it is even possible that if the right data were available, the claims by Marshall
and Samuelson that bread and potatoes were Giffen would be verified.
These results have important policy implications. The most obvious is that the patterns of
substitution away from calorie-dense foods such as rice or wheat implied by Giffen behavior and
observed in the data suggest that programs designed to improve nutrition may have little impact
for subsistence-zone consumers; this is particularly important given that nearly two-thirds of our
sample of urban poor households fall in this zone. While the subsidy would still have a positive
impact on utility for these households, to the extent that the aim of the policy is to improve nutrition,
it may have a much weaker effect than intended (theoretically, it is possible they will have
no effect at all, or even reduce nutrition). A full, detailed theoretical and empirical analysis of the
nutritional implications of price subsidies is beyond the scope of the present paper, and is instead
addressed in a companion paper, Jensen and Miller (2008).
In identifying the relevance of the three consumption zones, the results also point to an important
nonlinear effect of programs designed to improve nutrition. Such programs are unlikely to
significantly improve nutrition among subsistence households unless they effectively move them
out of the subsistence zone, since within this zone households will spend additional income
on improving the taste of their meals, rather than their nutritional content. Only if the size of
the benefit is large enough to move a subsistence-zone household into the standard zone will
it substantially improve nutrition. Thus, programs that appear not to improve nutrition might
have a significant effect on these households if their benefits were increased. However, the same
nonlinearities have the opposite implications for households in the calorie-deprived zone. For
these households, increasing the value of the benefits will improve nutrition only up to a point,
after which some households will be in the subsistence zone and the nutrition improvement will
level off, and possibly even decline. This creates a challenge for program targeting, as these
nonlinearities suggest programs will be effective only if the size of the subsidy is tailored to
the household’s consumption zone. From a program evaluation perspective, this leveling off or
decline in calories could make the average effect of a program appear small, masking larger
changes among specific groups unless the heterogeneous response is taken into consideration.44
Thus, recognizing the heterogeneity of poor households is important for the design, evaluation,
and targeting of such programs. It is also important to note that the difficulties facing subsidies
43
There is ample quantitative evidence of these patterns as well. For example, Shankar Subramanian and Deaton
(1996) find that the poorest households in rural Maharashtra, India, on average receive 53 percent of their calories from
sorghum (and spend on average 20 percent of their total household budgets on this food).
44
In fact, the two effects could offset each other from an evaluation perspective; relatively small benefits improve
nutrition among the calorie-deprived but decrease it among subsistence consumers, while larger benefits begin to
reduce calories among the formerly calorie-deprived but increase them among the formerly subsistence consumers.
Thus, the heterogeneous response could mask significant gains among subsets of consumers.
September 20081576 THE AMERICAN ECONOMIC REVIEW
or price controls would apply to almost any other program whose goal is to improve nutrition. In
particular, any program that results in an increase in real wealth, whether cash transfers, employment
schemes, in-kind transfers, or food stamps, will generate similar substitution effects.
The results are also important in light of the large literature concerned with the income elasticity
of demand for calories (see John Strauss and Duncan Thomas 1995; Deaton 1997). Those
policymakers who wish to increase nutrition but believe the income elasticity is small might
be inclined to suggest, as alternatives to cash transfers, subsidies or price controls that directly
encourage increased consumption of nutritious foods. Similar policy prescriptions may arise if
there is a concern, often stated, that simply giving cash is not desirable because households may
spend it on other luxuries (food or nonfood). However, our results suggest even these price-based
policies will face similar difficulties by virtue of the large wealth effects they create.
It is also worth noting that the results present somewhat more positive news about household
vulnerability. Policymakers are often concerned that households highly dependent on staple
goods may experience nutrition declines when the prices of those goods increase, either transitorily
during a crisis or permanently with the liberalization of markets, or as a result of trade agreements,
and a number of programs are designed to protect against this very possibility. Our results
suggest that in fact households in the subsistence zone are able to buffer their staple consumption
against such changes quite well (unless perhaps the price changes are very large). However, there
may still be a value in offering such assistance to households in the calorie-deprived zone, or
perhaps even households in the subsistence zone (though the justification would have to be on
general welfare, rather than nutritional, grounds). But here, again, the heterogeneity of consumers
is important to take into consideration.
Finally, although historically the primary importance of the Giffen phenomenon has been
pedagogical, we believe it also serves a more fundamental role in economic theory. The neoclassical
model of the consumer is one in which the consumer maximizes stable preferences subject
to a budget constraint, and in recent years this model has come increasingly under attack. These
objections run from the simplistic, “people don’t maximize” arguments to the sophisticated criticisms
found in psychology and behavioral economics. The possibility of Giffen behavior is a
clear, complex, nuanced prediction of the neoclassical model. Where naïve intuition suggests
that consumers should respond to a price increase by consuming less of the good in question,
consumer theory suggests that a sophisticated consumer, mindful of the interplay between nutritional
and budgetary concerns, might increase consumption. Further, the theory tells us exactly
when and where we should expect to find Giffen behavior, and where we should not. To our
knowledge, no other theory predicts the pattern of behavior implicit in the Giffen phenomenon,
i.e., the sign of the price elasticity changing from negative to positive and back to negative as
very poor consumers’ wealth increases. Thus, while economists’ failure to document Giffen
behavior in the past has been interpreted as a criticism of the approach, our finding of Giffen
behavior provides a type of vindication. Giffen behavior is predicted by the neoclassical model
and no other. While psychological and behavioral theories help to account for some areas of economic
behavior, in the case of the Giffen phenomenon, and of the consumption behavior of the
extremely poor more generally, the standard model appears to be the right one.
REFERENCES
Baruch, Shmuel, and Yakar Kannai. 2001. “Inferior Goods, Giffen Goods, and Shochu.” In Economics
Essays, A Festschrift for Werner Hildenbrand, ed. Gerard Debreu, Wilhelm Neuefeind, and Walter
Trockel, 9–17. Heidelberg: Springer-Verlag.
Barzel, Yoram, and Wing Suen. 1992. “The Demand Curves for Giffen Goods Are Downward Sloping.”
The Economic Journal, 102(413): 896–905.
VOL. 98 NO. 4 1577Jensen and Miller: Giffen Behavior and Subsistence Consumption
Battalio, Raymond C., John H. Kagel, and Carl A. Kogut. 1991. “Experimental Confirmation of the Existence
of a Giffen Good.” American Economic Review, 81(4): 961–70.
Boland, Lawrence A. 1977. “Giffen Goods, Market Prices and Testability.” Australian Economic Papers,
16(28): 72–85.
Bopp, Anthony E. 1983. “The Demand for Kerosene: A Modern Giffen Good.” Applied Economics, 15(4):
459–67.
Chen, John-Ren. 1994. “The Effects of Land Reform on the Rice Sector and Economic Development in
Taiwan.” World Development, 22(11): 1759–70.
Deaton, Angus. 1997. The Analysis of Household Surveys: A Microeconometric Approach to Development
Policy. Baltimore: Johns Hopkins University Press for the World Bank.
DeGrandpre, R. J., Warren K. Bickel, S. Abu Turab Rizvi, and John R. Hughes. 1993. “Effects of Income
on Drug Choice in Humans.” Journal of the Experimental Analysis of Behavior, 59(3): 483–500.
Dorfman, Robert, Paul A. Samuelson, and Robert M. Solow. 1958. Linear Programming and Economic
Analysis. New York: McGraw Hill.
Dougan, William R. 1982. “Giffen Goods and the Law of Demand.” Journal of Political Economy, 90(4):
809–15.
Dwyer, Gerald P., Jr., and Cotton M. Lindsay. 1984. “Robert Giffen and the Irish Potato.” American Economic
Review, 74(1): 188–92.
Friedman, Jed, and James Levinsohn. 2002. “The Distributional Impacts of Indonesia’s Financial Crisis
on Household Welfare: A ‘Rapid Response’ Methodology.” World Bank Economic Review, 16(3):
397–423.
Gilley, Otis W., and Gordon V. Karels. 1991. “In Search of Giffen Behavior.” Economic Inquiry, 29(1):
182–89.
Halbrendt, Catherine, Francis Tuan, Conrado Gempesaw, and Dimphna Dolk-Etz. 1994. “Rural Chinese
Food Consumption: The Case of Guangdong.” American Journal of Agricultural Economics, 76(4):
794–99.
Jensen, Robert T., and Nolan H. Miller. 2002. “Giffen Behavior in China: Evidence from the China Health
and Nutrition Survey.” Harvard University, John F. Kennedy School of Government Faculty Research
Working Paper RWP–02–014.
Jensen, Robert T., and Nolan H. Miller.  2008. “Do Consumer Price Subsidies Really Improve Nutrition?”
Harvard University, John F. Kennedy School of Government Research Working Paper RWP–08–025.
Koenker, Roger. 1977. “Was Bread Giffen? The Demand for Food in England Circa 1790.” Review of Economics
and Statistics, 59(2): 225–29.
Marshall, Alfred. 1895. Principles of Economics. London: Macmillan.
McDonough, Terrence, and Joseph Eisenhauer. 1995. “Sir Robert Giffen and the Great Potato Famine:
A Discussion of the Role of a Legend in Neoclassical Economics.” Journal of Economic Issues, 29(3):
747–59.
McKenzie, David. 2002. “Are Tortillas a Giffen Good in Mexico?” Economics Bulletin, 15(1): 1–7.
Nachbar, John H. 1998. “The Last Word on Giffen Goods?” Economic Theory, 11(2): 403–12.
Ravallion, Martin. 2007. “Geographic Inequity in a Decentralized Anti-Poverty Program: A Case Study
of China.” The World Bank Policy Research Working Paper Series 4303.
Rosen, Sherwin. 1999. “Potato Paradoxes.” Journal of Political Economy, 107(6): S294–313.
Samuelson, Paul. 1964: Economics. 6th ed. New York: McGraw Hill.
Stigler, George J. 1947. “Notes on the History of the Giffen Paradox.” Journal of Political Economy, 55(2):
152–56.
Stigler, George J. 1966. The Theory of Price. 3rd ed. New York: The Macmillan Company.
Strauss, John, and Thomas Duncan. 1995. “Human Resources: Empirical Modeling of Household and
Family Decisions.” In Handbook of Development Economics,Volume 3A, ed. Jere R. Behrman and T.
N. Srinivasan, 1183–2023. Amsterdam: North-Holland.
Subramanian, Shankar, and Angus Deaton. 1996. “The Demand for Food and Calories.” Journal of Political
Economy, 104(1): 133–62.
Thomas, Duncan, and John Strauss. 1997. “Health and Wages: Evidence on Men and Women in Urban
Brazil.” Journal of Econometrics, 77(1): 159–85.