Spatial Price Adjustment with and without TradeEMMA C. STEPHENS†, EDWARD MABAYA‡, STEPHAN VON CRAMON-TAUBADEL§ and CHRISTOPHER B. BARRETT*
†Pitzer College, Claremont, CA 91711 USA(e-mail: [email protected])‡Emerging Markets Program, Cornell University, Ithaca, NY 14850 USA(e-mail: [email protected])§Department of Agricultural Economics and Rural DevelopmentUniversity of Gottingen, Gottingen, Germany(e-mail: [email protected])*Department of Applied Economics and ManagementCornell University, Ithaca, NY 14850 USA(e-mail: [email protected])
November 2010
Word Count=5824
AbstractIn this paper we investigate the possibility that price transmission between spatiallydistinct markets might vary during periods with and without physical trade flows.We are able to test for differences in price transmission between trade and non-traderegimes by using Generalized Reduced Rank Regression (GRRR) techniques sug-gested by P.R. Hansen (2003). We apply these techniques to semi-weekly price andtrade flow data for tomato markets in Zimbabwe and find that intermarket priceadjustment occurs in both trade and non-trade periods. Indeed, the adjustments aregenerally larger and more rapid in periods without physical trade flows. This findingunderscores the importance of information flow for market performance.
JEL Codes: Q13, R12, C32, P42
Keywords: Spatial price transmission, cointegration, Zimbabwe, reduced rank re-gression, tomatoes.
I. Introduction
A large literature explores the behavior over time of distinct markets that are linked
together in a network. This network can be spatial, as in markets for a commod-
ity within a given region or country, or it may represent other kinds of integration,
perhaps through vertical marketing channels as product is transformed or through
intertemporal arbitrage via storage (Williams & Wright, 1991; Deaton & Laroque,
1996; Brummer, von Cramon-Taubadel, & Zorya, 2009). The primary purpose of
such research is often to determine how quickly markets respond to shocks and how
these shocks transmit through the network via price adjustment. The dominant an-
alytical approach has exploited spatial market equilibrium conditions, described in
detail by Takayama and Judge (1971). Deviations from these equilibrium conditions
yield key information on overall market efficiency. Understanding these dynamics
with respect to food markets may be of particular importance for policy makers in
developing countries (Fackler & Goodwin, 2002), where large subpopulations are em-
ployed in the agricultural sector and where the considerable budget shares devoted to
food expenditures leave many poor households vulnerable to price spikes commonly
associated with market disequilibrium.
The underlying theory of spatial market equilibrium suggests particular patterns
of price behavior under competitive arbitrage, based on the transactions costs as-
sociated with movement of goods between markets and observed trade flows. More
specifically, if physical trade flows occur between markets, these markets are said
to be in competitive spatial equilibrium if and only if the price differential exactly
equals the costs of moving goods between them, such that excess returns to trade
1
are completely exhausted. Further, with constant per unit costs of commerce, any
price change in one market due to a local demand or supply shock should generate
an equal price change in the other market. This strong spatial price transmission is
the familiar Law of One Price.
The absence of trade may also imply that markets are in spatial equilibrium. This
can occur when price differentials exactly equal transactions costs, leaving traders
indifferent between trading and not trading, or when the intermarket price differen-
tials are insufficient to cover the costs of arbitraging between the markets. Where
competitive spatial equilibrium with trade implies strong spatial price transmission,
however, this latter, segmented spatial equilibria is consistent with uncorrelated price
series as well as with the Law of One Price. Thus spatial price transmission dynamics
may be markedly different in periods with and without trade even when markets are
always in competitive spatial equilibrium.
Markets may also be out of equilibrium, with either apparent missed arbitrage
opportunities (i.e., no trade in spite of intermarket price differentials in excess of the
costs of arbitrage) or positive trade in the face of negative returns to arbitrage. The
different possible combinations of trade flows and returns to trade are examined in
detail in the switching regime branch of the spatial price analysis literature, which
finds empirically that markets are frequently out of spatial equilibrium (Baulch, 1997;
Barrett & Li, 2002). However, this literature is limited in its ability to comment on
the actual process of transition between equilibrium and non-equilibrium regimes, as
the approach is inherently static and does not make explicit use of the time series
2
nature of the data at hand.1
There is, however, a large body of work that performs spatial price analysis dy-
namically and includes the possibility of disequilibrium periods, primarily through
the use of threshold autoregressive and vector error correction models to character-
ize price dynamics (e.g. Dercon, 1995; Goodwin & Piggott, 2001; B. E. Hansen &
Seo, 2002). But one of the main, implicit assumptions of these models is the pri-
macy of trade flows in bringing about spatial equilibrium. This assumption has been
largely unexamined thus far in the literature. For the most part, lack of available
complementary price, trade flow and transaction cost data has hampered the ana-
lysts’ ability to test empirically whether or not trade flows are the main mechanisms
behind spatial equilibrium patterns.2 Further, until recently, the appropriate meth-
ods in cointegration analysis necesary to fully compare spatial price dynamics in the
presence of multiple trading regimes did not exist.
Using developments in the literature on structural breaks in cointegration models
(Boswijk & Doornik, 2004; P. R. Hansen, 2003; Johansen & Juselius, 1992; Pesaran
& Shin, 2002), we examine the nature of spatial price adjustment dynamics both with
and without physical trade flows. We apply a flexible structural break model to de-
tailed semi-weekly price, transactions costs and trade flow data from tomato markets
in Zimbabwe and test for the presence of non-linearities in long run equilibrium rela-
tionships and potentially different price transmission patterns under different trading
1Negassa and Myers (2007) offer a dynamic extension to the parity bounds model, but likethe rest of the switching regime literature, their approach relies heavily on strong, atheoreticaldistributional assumptions.
2See Barrett (1996) for a break down of the types of market analysis methods by data classifi-cation.
3
regimes. This method allows for useful characterization of market performance with
respect to price adjustment while relaxing some strong, and typically indefensible,
distributional assumptions on which existing switching regime techniques depend.
It also provides different, more complete information about spatial price adjustment
dynamics. In particular, we are able to test empirically for differences in these dy-
namics in periods with and without observed trade flows. The results implicitly
demonstrate the importance of mechanisms other than physical trade flows – such
as information flows – in linking spatially distinct markets and thereby influencing
spatial price adjustment. This reinforces recent findings, such as Jensen (2007), that
more directly demonstrate the impact of information flows on spatial price and trade
patterns. Our application to Zimbabwean tomato markets demonstrates that failure
to allow for such structural differences in intermarket price transmission can lead to
serious errors of inference with respect to market efficiency.
II. Models of Cointegration and Regime-Specific
Error Correction
Cointegration models have been used to great effect in many spatial analysis stud-
ies.3 These studies make use of the fact that it is possible to use competitive spatial
equilibrium conditions to find an error correction representation of the relationships
between prices and transactions costs in two different markets, as we explain below.
Based on this model, one can then estimate parameters that characterize how mar-
3See Fackler and Goodwin (2002) for a detailed summary.
4
ket prices adjust to random shocks that move connected markets out of long run
equilibrium.
Error correction models have the attractive property that they allow for analysis
of both long run and short run dynamics in the presence of a cointegrating rela-
tionship. Short run and long run price adjustment dynamics may be quite different,
especially for markets with significant transactions costs, like those for perishable
commodities in developing countries. Underlying error correction models of spatial
market networks is the idea that price time series may display long run equilib-
rium relationships, and short run deviations from equilibrium should be ‘corrected’
in subsequent time periods through the particular mechanisms that control price
adjustment.
We use a generalization of the error correction model to examine the nature of this
adjustment in periods with and without trade flows. We are particularly interested
in isolating price adjustment in periods without trade flows, in order to establish
whether price adjustment is attributable primarily to physical arbitrage associated
with trade, or whether it is perhaps equally attributable to non-material flows, pre-
sumably of information.4 This possibility of a distinction between the mechanism
behind price adjustment and trade flows is the general reason for the importance
of separating spatial market equilibrium from market integration concepts (Barrett,
2001). Thus, if dynamic spatial price analysis is to be used to identify and under-
stand the consequences of poorly functioning market systems, parsing out relative
4Information flows may, for example, be due to the effect of elaborate networks of tradersoperating in several markets simultaneously or the overall complexity of spatial market networks(Fackler & Tastan, 2008).
5
differences in adjustment due to direct linkages via physical trade flows versus more
indirect connections that exist without them seems a useful and novel exercise.
We study spatial equilibrium and integration by making use of data on prices,
transactions costs and trade flows by treating disruptions and resumptions of trade
as structural breaks in the cointegration relationship and then using the General-
ized Reduced Rank Regression techiques (GRRR) described in P. R. Hansen (2003).
GRRR provides estimates of error correction model parameters for structurally dis-
tinct periods, in our case, with and without observed trade. This method generates
estimates of both the long run equilibrium and speed of price adjustment to tempo-
rary disequilibria when linkages between markets are in part due to the behavior of
traders who operate in both markets and take advantage of arbitrage opportunities,
as well as estimates of these relationships when commodities do not move between
markets, which may be attributable to other mechanisms more indirect than physical
trade. The main innovation of GRRR is that it provides a straightforward method for
comparing all of the model parameters in these two regimes using nested hypothesis
tests, which was not previously possible using more conventional tests for structural
breaks due to the fact the parameters in the long run cointegration relationship have a
non-standard distribution (P. R. Hansen, 2003; Johansen, 1988). It also differs from
threshold cointegration models models (Balke & Fomby, 1997; B. E. Hansen & Seo,
2002) in that it takes structural breaks as known and exogenously determined and
incorporates this extra information directly, rather than indirectly using the magni-
tude of the estimated long run error term to identify distinct regimes. While many
other studies have examined different aspects of non-linear cointegration, GRRR is
6
a comprehensive framework for testing multiple hypotheses about both the long run
and short run parameters that was not previously available to researchers.
The model’s theoretical foundation is the standard competitive equilibrium rela-
tionship between prices in spatially distinct markets. The equilibrium conditions are
a function of both the returns to trade and the observed trade flows. With positive
trade flows, two markets (source market i and destination market j) are said to be
in competitive equilibrium if the price margin (mijt = Pjt − Pit) is exactly equal to
the cost to transfer goods between the markets, τijt.5 At this point, traders are indif-
ferent between moving and not moving goods between the two markets. Note that
these costs need not be symmetric. Including non-trading periods into the definition,
the generalized conditions for efficient spatial arbitrage (Barrett, 2001; Takayama &
Judge, 1971) are:
Pjt
≤ Pit + τijt if qij = 0 (a)
= Pit + τijt if q ∈ (0, qij) (b)
(1)
where qij represents trade flows from market i to market j and qij represents a
potential trade flow ceiling, due to either a quota or some other imposed restriction.
Note that in equation (1), the relationship that prevails between prices in different
markets depends upon the observed trade flows, qij. However, as we show below, it
is possible for other mechanisms to affect this relationship.
When the equilibrium relationship between Pjt and Pit binds with equality (as in
1(b)), this equality can be used to develop an error correction representation (Engle
5Subscripts indicate the direction of trade flows. For example, mijt = Pjt − Pit, indicates theintermarket price differential for goods flowing from market i to market j in period t. Similarly,transactions costs to move commodities from market i to market j are listed as τijt.
7
& Granger, 1987) of the dynamic relationship between market prices and transactions
costs over time, with 1(b) characterizing the long run equilibrium (i.e., the cointe-
grating relationship) between markets. However, in the short run, shocks to prices
and/or transactions costs may cause temporary deviations from this equilibrium. In
a dynamic context, the long run difference between destination and source market
prices, et ≡ Pjt−Pit, should be a stationary process, in that any temporary deviation
from long run spatial market equilibrium is expected to disappear over time.6 For
market prices, this process can be expressed quite generally as the residual from the
linear estimated relationship in (1b) (Engle & Granger, 1987):
Pjt − B1Pit ≡ et(B) (2)
with et thus representing an I(0), stationary random variable.7 Analysis of this resid-
ual allows one to establish exactly how long it takes for intermarket price adjustment
to return markets to their long run equilibrium state.
If et truly represents a long run, stationary process, then market prices are said
to be cointegrated time series. In the spatial market analysis literature, cointegrated
prices are often (if controversially) taken as indicators of the degree of market effi-
ciency in a given network.8
6Typically, the long run relationship is thought to exist between market prices, therefore wethus ignore any possible endogenous relationship between prices and transactions costs in thisstudy. Coleman (2009) provides a model of spatial arbitrage that incorporates the effects of tradeflows on the cost of transport, however this model is not appropriate to the data we use.
7As has been pointed out by Dercon (1995) and others (Goodwin & Piggott, 2001), the long-runequilibrium relationship between prices may involve price movements that are proportional to oneanother, rather than one-for-one. Hence the inclusion of the B parameter in equation (2).
8Barrett (1996), Fackler and Goodwin (2002) and others have noted that cointegration is neithernecessary nor sufficient for market efficiency, and that violations of some common assumptions
8
Regime-specific cointegration
For linear cointegration, an error correction representation of the relationship be-
tween prices in a source and destination market can be written as:
(∆Pjt
∆Pit
)= AB′Xt−1 +
∑d
Γd
(∆Pjt−d
∆Pit−d
)+ CZt + ut, ut ∼ N(0,
∑2x2
) (3)
In equation (3), Xt−1 = (Pj,t−1, Pi,t−1)′ are the (lagged) variables in the long run
equilibrium relationship included in the error correction model, and the B (a 2x1
matrix) parameters form the cointegration vector:9
Pjt −B1Pit ≡ et(B) ≡ B′Xt (4)
The parameters in A (a 2x1 matrix) govern the speed of adjustment of each market’s
price from short-term shocks back to the long run equilibrium. The Γd matrices (2x2)
capture the autocorrelation in the system between the price series, with d indicating
the lag length(s) included in the estimation. Zt includes any additional, exogenous
explanatory variables, like seasonal dummies or other factors.10
The model in equation (3) is appropriate only under the assumption that discon-
tinuities in trade flows do not matter to price adjustment dynamics, as it specifies
underlying cointegration models, such as stationary transactions costs and continuous trade flows,may be more to blame for the frequent rejection of efficiency found in the literature than an absenceof efficiency itself. We attempt to address some of these issues in this work.
9The cointegration vector in (4) is typically normalized on Pj in order to identify B.10Due to the particular way that bus fares and fuel prices are determined in Zimbabwe, we
combine them to account for intermarket transfer costs and can include it as an exogenous variablein Zt in our estimation. We also include an unrestricted constant in Zt to preserve degrees offreedom in our small sample.
9
that only one set of parameters govern behavior for the entire length of the price
series, instead of distinguishing between periods with and without trade flows.
Recognizing that the model in equation (3) may be too restrictive to represent
market relationships with multiple regimes, threshold cointegration models, initially
proposed by Balke and Fomby (1997), have been used to test spatial market equi-
librium in the presence of unobserved constant transactions costs (Dercon & van
Campenhout, 1999; Goodwin & Piggott, 2001; B. E. Hansen & Seo, 2002). These
transactions costs divide spatial market data into distinct regimes, within which the
spatial price adjustment dynamics may vary. Further, it is often assumed that no
error correction occurs for prices within a price band defined by the transactions
costs(described by equation (5) below):
|Pjt − Pit| ≤ τijt (5)
since arbitrage is not profitable and thus trade flows should not occur to bring prices
into long run equilibrium.11
Markets linked by trade flows are potentially different from those not linked by
trade flows, as the conditions in relation (1) based on spatial equilibrium theory
suggest. Not only might the speed of adjustment and short run parameters differ
between trade and no trade periods, but there may also be different mechanisms at
play effectively linking market prices in periods with and without trade in long run
equilibrium relationships.
11Equation (5) is just a reformulation of the efficient spatial arbitrage condition shown in (1a)that prevails when trade flows do not occur due to excessively high transactions costs.
10
This is precisely the effect we wish to explore in order to better understand the
role physical trade flows play in bringing about spatial competitive equilibrium via
price adjustment. Evidence of cointegration between markets in non-trade periods
suggests that the role of physical trade may be overstated in the literature and that
information flows that may impact market function and efficiency independent of
trading activity may be underappreciated. Allowing for complete variation in param-
eters across trading regimes is therefore critical to our analysis, however asymptotic
theory and tools for tesing for structural breaks in the cointegration vector itself have
not been available until relatively recently (Gonzalo & Pitarakis, 2006; P. R. Hansen,
2003).
P. R. Hansen (2003) presents a method to test for these prospective non-linearities
that enables us to specify and estimate the following model:
(∆Pjt
∆Pit
)=
(Atrade ∗B′tradeXt−1 +
∑d
Γtraded
(∆Pjt−d
∆Pit−d
)+ CtradeZt
)Itradet +(
Ano trade ∗B′no tradeXt−1 +∑
d
Γno traded
(∆Pjt−d
∆Pit−d
)+ Cno tradeZt
)Ino tradet + ut
(6)
where the Itradet and Ino trade
t are indicator functions for the specific trade regime into
which each time period, t, falls.12
The specification in (6) allows for potential differences in both long run price re-
lationships between markets under different trading regimes (the B parameters), as
12Siklos and Granger (1997) initially proposed the concept of regime-sensitive cointegration inthe context of interest-rate parity and structural breaks in monetary policy rules.
11
well as in how quickly the markets respond to shocks in each regime (the A parame-
ters) and the effects of autocorrelation and other factors (the Γ and C parameters).
With this model and necessary data on trade flows, as well as standard price and
transaction cost time series, we are able to examine empirically a key assumption in
the literature: that the Ano trade parameters equal zero when there are no trade flows
because physical arbitrage is commonly assumed to be the mechanism that returns
markets to equilibrium (Balke & Fomby, 1997, pg. 629).13
Therefore, our main hypothesis is that Ano trade = 0 and all the action in spatial
price adjustment occurs only during trading periods. Also we would like to test
whether the cointegrating vector is the same in both regimes (i.e., Btrade = Bno trade).
However, in order to examine these hypotheses, we need a method that accounts
for the possibility of regime-specific cointegrating relationships. This is because, if
adjustment is taking place in non-trade periods, then a cointegrating relationship
must be present. However, as it is clearly not due to the effect of trade flows,
it could be distinct from the relationship that prevails during trading periods and
should therefore be estimated separately.14
Note that we are assuming that a cointegration vector is defined in non-trade
periods. Although many of the studies in the spatial price adjustment literature
specifically model prices within the price band as a random walk (and often assume
that no trade takes place, although actual trade flow data are typically not avail-
13Our null hypothesis is an extension of the null found in other studies that spatial price adjust-ment with linear cointegration models, where inability to reject the null of a zero-valued adjustmentparameter is taken to mean that prices in a market are weakly exogenous to other prices in thespatial network (Fackler & Goodwin, 2002).
14Also, assuming linearity in the case of nonlinear long-run equilibrium relationships results ininconsistent estimation of the (misspecified) cointegration vector (Gonzalo & Pitarakis, 2006).
12
able), we are unable to include this possibility in our analysis, given the estimation
technique we plan to use. However, it has been observed that periods without trade
may still be consistent with a long run equilibrium relationship, for example, if all ar-
bitrage opportunities are exhausted such that traders are indifferent between trading
and not (Barrett & Li, 2002; Dercon & van Campenhout, 1999). Other recent work
on spatial price transmission in Ghana also finds error correction in non-trade periods
(Ihle, Amikuzuno, & von Cramon-Taubadel, 2010). The Generalized Reduced Rank
Regression method we use begins with the assumption of cointegration and then
allows for relatively straight-forward tests of whether the equilibrium relationship in
non-trade periods is distinct from that in the periods with trade.
III. Data
We estimate the regime-specific cointegration model specified in (6) with semi-weekly
data between January and December 2001 on prices, transactions costs and trade
flows for tomatoes in three important spot markets in Zimbabwe (T = 104). The
geographical distribution of these markets is shown in Figure 1. The data collected
include unit prices for tomatoes (in Zimbabwean dollars ($Z) per crate), intermarket
transfer costs as a proxy for transactions costs ($Z/crate), estimated from fuel prices
and bus fares, and a measure of trade flows between markets. The data were collected
through direct semi-weekly interviews with tomato traders and transporters over 52
weeks in 2001, as well as through monthly surveys of bus operators. The tomato
prices used represent the average price for the best grade of tomato currently available
13
in the market. The trade flow variable used to identify the structural breaks is a
binary indicator of the presence of tomatoes in a destination market from the set
of main source markets as observed by survey enumerators. During this period
in Zimbabwe, tomatoes were primarily transported as excess cargo on busses, the
pricing of which was administratively imposed by government, resolving what might
otherwise be a potential problem of endogenous intermarket transfer costs.15 These
data also have the attractive feature that they are available even when there are no
tomato trade flows, allowing estimation of the error correction term in the non-trade
periods in our model. More details on data collection can be found in Mabaya (2003)
and Mabaya (2004).
The primary players in these spot markets are traders who have purchased pro-
duce from local smallholder farmers and then sell the tomatoes in the market mostly
to low- or middle-income urban consumers. This constitutes a relatively informal
marketing structure, with more formal contract farming transactions occurring di-
rectly between private wholesalers or other large buyers (such as supermarkets or
public institutions like schools) and larger commercial farmers in other venues.
Tomato prices are more volatile in these spot markets than in the formal sector,
due to the absence of formal contracts and produce perishability.
We study price transmission between the following three directional market
pairs:16 Harare to Gweru, Harare to Bulawayo and Gweru to Bulawayo.17 Fig-
15Thus we treat these costs as an additional, exogenous control variable as explained above.16Only market dyads are considered in our analysis and we do not take into explicit account the
possible higher order effects of trade in a network with multiple linkages. See Fackler and Tastan(2008) on this issue.
17Harare is the capital of Zimbabwe and is close to major tomato production areas, which mayexplain the fact that there are no trade flow reversals observed during the sample year.
14
ures 2-4 plot out the price and trade flow data for the three market pairs. These
market pairs display the key characteristics necessary to test our hypotheses, in that
the relevant prices are non-stationary and integrated of order one, and that a suffi-
cient balance of observations of trade and non-trade periods were observed for each
pair.18 The market dyads in this study also display regime-specific cointegration
and seem most suitable for the analysis we wish to conduct. Table 1 presents the
Johansen trace tests for cointegration between market prices for each dyad as well as
the results of cointegration tests on the trade and non-trade period subsamples. As
can be seen, prices in both market pairs appear to be cointegrated overall as well as
within each trade regime, for the most part (tests for the Harare-Bulawayo pair offer
weaker support). However, in order to test our hypotheses of interest, GRRR must
be employed to be able to statistically compare the two regimes, as the distribution
of the B parameters is non-standard in reduced rank regressions.
Although data is available on tomato prices in several other major urban spot
markets in Zimbabwe in the same sample that we have used, there is no inter-market
trade flow of tomatoes between them. These spot markets are instead typically sup-
plied by nearby agricultural production regions. Thus we are not able to incorporate
these markets explicitly into our analysis that compares trade and non-trade peri-
ods. However our results potentially have more general implications for spatial price
adjustment dynamics for these markets in Zimbabwe, as do we find evidence for
cointegration in four out of the eight market pairs that had no trade flows between
them at any point during the year of the study. Moreover, these pairs all contain
18Augmented Dickey-Fuller test statistics were used to determine the integration order of thedifferent time series. Details are available from the authors by request.
15
the Chitungwiza market. Chitungwiza prices are strongly cointegrated with Harare
prices, which are also cointegrated with these other spot markets. While evidence
of a cointegration relationship without trade flows (usually in the absence of data
on transactions costs and trade flows) has previously been taken as evidence of the
limitations of cointegration analysis to understand the efficiency of a spatial mar-
ket network (Barrett, 1996), our current findings (in combination with our specific
knowledge of trade regimes and transaction costs) suggest that some of these re-
lationships might in fact represent real long run adjustment dynamics even in the
absence of trade.
IV. Estimation Strategy
Estimating and testing for trade regime-specific cointegration presents an econo-
metric challenge. Under the null hypothesis, the speed of adjustment parameters
in the non-trading periods (Ano trade) should equal zero and prices in non-trade pe-
riods should not show any tendency towards long-run adjustment. During trade
periods, there is a unique long-run relationship under the null, et, which is a sta-
tionary process. The alternative hypothesis is that the trade and no trade regimes
have statistically significant speeds of adjustment in both trading regimes and/or
different long-run relationships. Given our desire to compare the relative strength of
the impact on price adjustment of physical flow of goods between markets with the
impact of other forces that operate in non-trading periods, we need to test the linear
cointegration model (shown in (3)) against one in which the alternative is non-linear
16
cointegration in the sense that the long run relationship may differ across regimes
(6).
Our estimation strategy is as follows. First, we follow P. R. Hansen (2003) and
estimate equation (6) using GRRR and analyze the Atrade, no trade and Btrade, no trade
parameters for each regime.19 We then estimate a restricted model, which imposes
that A and B are identical across trade regimes. As GRRR is used to examine non-
linear adjustment processes with structural breaks, conventional maximum likelihood
techniques are not appropriate for parameter estimation. To overcome this difficulty,
GRRR makes use of the fact that the parameters can be estimated iteratively, and
with specific sets of parameters held constant, estimation reduces to a set of sim-
pler GLS problems for which there are analytical solutions. Oberhofer and Kmenta
(1974) first outlined an iterative algorithim for these kinds of estimation problems.
P. R. Hansen (2003) follows Oberhofer and Kmenta and then demonstrates that
the likelihood ratio statistic obtained after estimating the system with parameter
restrictions can determine whether the restricted model is appropriate for the data
and describes the distribution of this test statistic for these reduced rank regressions.
V. Results
The parameter estimates from equation (6) are presented in Tables 2-4 for the three
market pairs. We used the Akaike Information Criterion and the Hannan-Quinn’s
Criterion on the linear model to both determine that one lag of the dependent variable
19We also allow the C parameters on transaction costs to vary across trade regimes, although weconstrain the Γ parameters and the constant, due to our small sample size.
17
was appropriate for each market pair in the associated vector autoregression and also
provide additional model selection information for jointly determining the lag length
and the cointegration rank of both the linear and the non-linear cointegration models,
following Baltagi and Wang (2007).
We present the unrestricted and restricted parameters and the likelihood ratio
test results for equality between the trade and no trade parameters for each market
pair. P. R. Hansen (2003) demonstrates that the speed of adjustment A, along
with the Γ and C parameters have a Gaussian distribution, so standard errors can
be easily calculated. The B parameters have a mixed Gaussian distribution, so
standard errors are not presented, however the likelihood ratio test does not depend
on the standard errors and can still be performed. The likelihood ratio test has a chi-
squared distribution, with the degrees of freedom equal to the number of parameter
restrictions (which for each market pair equals 6). Table 5 presents the predicted
half-lives for each market pair and trade regime.20
For the Harare to Gweru market dyad, in trading periods, the speed of adjustment
is significantly different from zero in both the source and destination markets, which
is what is typically assumed in most spatial price transmission models. However,
in non-trade periods, statistically significant price adjustments also occur in both
markets. This signals qualitatively different spatial price adjustment mechanisms
20For ease of interpretation, the speed of adjustment parameters can be expressed as a half-life,Thalf , which indicates how long it takes for half of the deviation from long run equilibrium to becorrected. Thalf = ln(0.5)/ln(1 − |ak|), k = {Pj , Pi},where ak is the parameter estimate from thematrix A for the destination price (k = Pj) and the source price (k = Pi) equation, respectively(Dercon and Van Campenhout (1999)). In all of our estimates, since our time observations aresemi-weekly, Thalf has been multiplied by 3.5 (days/semi-week) to convert the half-life into unitsof days, rather than ‘semi-weeks.’
18
under trade and no trade regimes, an impression confirmed by the likelihood ratio
test statistic, which strongly rejects the null of similar mechanisms across trade
regimes (p=0.011). Thus the data do not support linear cointegration; the data
favor modeling price transmission for this market pair as two separate processes
during trade and non-trade periods. The estimated half-life for price adjustment
in the source market is likewise shorter during the non-trade regime, underscoring
that price adjustment does not seem to be driven exclusively by physical arbitrage
activity. The long run relationship between the prices also changes between trade
regimes, with prices more closely proportional to one another (0.977) during non-
trade periods.
The results are broadly similar for the Harare to Bulawayo and Gweru to Bu-
lawayo market pairs (Tables 3 and 4, respectively). The likelihood ratio test also
strongly rejects the restricted model or equivalent adjustment processes across trade
regimes for these two market pairs (p=0.017 and p=0.013, respectively). The error
correction process appear distinct for trade and non-trade periods in all three market
pairs. Between Harare and Bulawayo, the Bulawayo prices do most of the adjusting,
primarily during non-trade periods (the adjustment parameter for the trade periods
is significant at the 20% level). Intermarket transfer costs only seem to influence
destination market prices (i.e. Bulawayo) and in a manner we would expect by
raising destination market prices (CPj> 0). Intermarket transfer costs also affect
destination market prices during periods both with and without trade.
Between Gweru and Bulawayo, Bulawayo prices adjust in all periods, while Gweru
prices only seem to adjust to shocks when there are physical trade flows. In this
19
market dyad, intermarket transfer costs only affect the source market (Gweru), again
in an expected manner by lowering source market prices (CPi< 0).
If one were to assume linear cointegration between these markets (as reflected in
the rightmost columns of Table 5), the estimated adjustment speed would generally
appear slower than that estimated for the trade regimes separately. The implication
is that failure to allow for potential differences in the spatial price adjustment mech-
anism between trade and no trade periods can understate the speed at which price
convergence occurs.
There are several ways to interpret the finding that prices are cointegrated with-
out trade flows. It is possible that tomato prices in the source and destination markets
for all three pairs were influenced by a common, exogenous factor, such as overall
inflation or seasonality. However, Mabaya (2003) still found general evidence for
this ‘segmented integration’ between the different price series even after controlling
for both of these processes. Another possibility is that participants in the different
spot markets engage in formula pricing (Tomek & Robinson, 1990) during non-trade
periods, using information about tomato prices in larger markets as a baseline to
adjust local spot prices. Informal conversations with traders during the data collec-
tion period revealed that information about Harare prices was readily available to
traders in smaller markets through bus drivers and passengers coming from the cap-
ital. As well, in some, but not all, of the spot markets, survey respondents indicated
that they used this information to set prices when there were relatively few market
participants (Mabaya, 2003).
Furthermore, trade flows under domestic market conditions (unlike international
20
trade) are difficult to capture clearly. Local spot markets draw produce from ‘catch-
ment areas ’ (McNew, 1996; Mabaya, 2003) made up of farms within a defined radius
of the urban market. In some instances, farmers may reside in the intersection of two
(or more) catchment areas and may supply to multiple urban spot markets, depend-
ing on potential profits. The activities of such farmers will serve to link prices in all
of the markets in which they participate, regardless of whether a trader subsequently
physically ships produce between the urban spot markets. Finally, prices between
two markets where there are no physical trade flows may also be cointegrated if both
markets trade with a common third market (Fackler & Tastan, 2008). The appar-
ent cointegration between the tomato prices in Chitungwiza and several other spot
markets, despite the fact that no produce was observed to move between them, may
be informal evidence for this last possible mechanism. This indirect evidence seems
to suggest one plausible mechanism by which price transmission may occur in the
absence of trade flows and supports our contention that developing a method for
comparing trade and non-trade periods is important in understanding the behavior
of these markets more completely. However without additional data we are not able
to further explore all of these particular information channels in more detail.
VI. Conclusions
Using a newly developed test for non-linear cointegration analysis, we specify and
estimate a trade regime-sensitive cointegration model and analyze the similarities
and differences between spatial price adjustment dynamics in tomato markets in
21
Zimbabwe in periods with and without physical trade flows. It appears that er-
ror correction between price margins and transactions costs occurs both with and
without trade. This provides clear evidence that the mechanisms governing spatial
price adjustment dynamics extend well beyond the physical arbitrage activities of
traders, and thus that analysis of market integration and efficiency must allow for
the multiple price transmission mechanisms that may be at work in a given network.
In particular, in this instance, by ignoring the possibility of multiple spatial pricing
equilibria, time series analysis based on linear cointegration would have suggested
that tomato markets in Zimbabwe adjust more slowly than they do and would not
reflect the quite distinct processes that appear to govern price transmission during
non-trade periods. Our more general method, by accounting for trade and non-trade
regime dynamics separately, leads to the conclusion that these markets are actually
very efficient in transmitting market signals across space through price adjustment
dynamics. Given the very different policy implications of each of these analyses, it
seems clear that testing and accounting for non-linearities in long run equilibrium
states characterized by trade flows is important. It also opens up questions as to the
precise nature of spatial price adjustment in non-trade periods, which has been given
very little attention in the spatial price analysis literature, despite the acknowledged
importance of trade flows on spatial price adjustment (Barrett & Li, 2002).
For market studies with the full complement of prices, transactions costs and
trade flow time series, our method should provide a useful means to characterize the
time series behavior of these variables and to avoid the possible mistaken inference
likely to result from spatial price analysis with trade and non-trade regime data mixed
22
together. This should serve as an additional tool for researchers and policy makers
working to understand spatial price transmission and the competitive performance
of commodity markets.
23
Table 1Johansen trace tests for cointegration between Harare and Gweru, Harare and
Bulawayo and Gweru to Bulawayo market pairs
Market Pair Rank Trade Non-Trade Whole Sample†
Harare-Gweru 0 20.555 (0.044) 19.595 (0.060) 20.002 (0.053)1 8.838 (0.058) 4.907 (0.304) 8.384 (0.071)
T=64 T=40 T=104Harare-Bulawayo 0 12.493 (0.414) 17.592 (0.113) 17.084 (0.131)
1 3.853 (0.447) 5.362 (0.255) 7.485 (0.108)T=48 T=56 T=104
Gweru-Bulawayo 0 17.610 (0.112) 25.761 (0.007) 24.787 (0.010)1 5.471 (0.244) 5.588 (0.233) 7.294 (0.114)
T=58 T=46 T=104† p-values are reported in parentheses. The underlying model contained one lag
and a restricted constant in the cointegration vector.
28
Table
2P
aram
eter
esti
mat
esof
equ
atio
n(6
)-
Har
are
toG
wer
um
arke
tpa
ir
Tra
de
Regim
eN
oT
rade
Regim
eR
est
rict
ed
(θT
=θN
T)
Eq.
Para
mete
rC
oeffi
cien
t†Std
.C
oeffi
cien
tStd
.C
oeffi
cien
tStd
.(θ
T)
Err
.(θ
NT
)E
rr.
Err
.[e
t−1(B
)]B
1-0
.642
-0.9
77-1
.012
[Pjt
]A
Pj
-0.1
95**
*0.
081
-0.1
94*
0.11
9-0
.026
0.05
1C
Pj−
tran
s.co
sts
-0.2
000.
514
-0.8
96**
0.52
0-0
.374
0.50
5C
Pj−
con
s.93
.793
*62
.987
93.7
93*
62.9
8754
.099
64.4
34Γ
Pj,P
jt−
10.
060
0.10
60.
060
0.10
6-0
.002
0.10
9Γ
Pj,P
i t−
1-0
.058
0.08
6-0
.058
0.08
60.
001
0.08
8
[Pit]
AP
i0.
184
**0.
098
0.39
0**
*0.
144
0.18
8**
**0.
059
CP
i−
tran
s.co
sts
-0.7
120.
621
-0.3
250.
628
-0.8
54*
0.58
3C
Pi−
con
s.47
.924
76.1
3647
.924
76.1
3696
.217
74.3
42Γ
Pi,P
jt−
1-0
.087
0.12
8-0
.087
0.12
8-0
.063
0.12
6Γ
Pi,P
i t−
10.
051
0.10
40.
051
0.10
40.
033
0.10
2
AIC
17.9
0817
.974
HQ
18.0
7418
.088
Log
Lik
elih
ood
-118
6.76
0-1
195.
109
df86
80L
RS
tati
stic‡
χ2(6
)16
.698
(0.0
11)
†N
ote:∗,∗∗,∗∗∗,∗∗∗∗
indic
ate
sign
ifica
nt
at20
%,
10%
,5%
and
1%re
spec
tive
ly.
Subsc
ripti
refe
rsto
the
sourc
em
arke
tan
dsu
bsc
riptj
refe
rsto
the
des
tinat
ion
mar
ket.
‡p-v
alue
rep
orte
din
par
enth
eses
.
29
Table
3P
aram
eter
esti
mat
esof
equ
atio
n(6
)-
Har
are
toB
ula
way
om
arke
tpa
ir
Tra
de
Regim
eN
oT
rade
Regim
eR
est
rict
ed
(θT
=θN
T)
Eq.
Para
mete
rC
oeffi
cien
t†Std
.C
oeffi
cien
tStd
.C
oeffi
cien
tStd
.(θ
T)
Err
.(θ
NT
)E
rr.
Err
.[e
t−1(B
)]B
1-0
.829
-0.7
191.
705
[Pjt
]A
Pj
-0.1
17*
0.07
0-0
.380
****
0.09
5-0
.076
***
0.03
1C
Pj−
tran
s.co
sts
0.91
4**
*0.
408
0.77
1**
*0.
315
0.00
70.
268
CP
j−
con
s.-9
4.47
8*
60.4
88-9
4.47
8*
60.4
8897
.851
*61
.892
ΓP
j,P
jt−
1-0
.002
0.09
6-0
.002
0.09
6-0
.066
0.09
8Γ
Pj,P
i t−
10.
052
0.12
30.
052
0.12
30.
156
0.13
2
[Pit]
AP
i0.
091
*0.
057
0.07
80.
077
-0.0
55**
*0.
023
CP
i−
tran
s.co
sts
-0.4
260.
331
-0.2
820.
256
-0.0
460.
203
CP
i−
con
s.45
.670
49.1
0845
.670
49.1
0877
.970
**46
.800
ΓP
i,P
jt−
1-0
.082
0.07
8-0
.082
0.07
8-0
.030
0.07
4Γ
Pi,P
i t−
10.
003
0.10
00.
003
0.10
00.
035
0.10
0
AIC
19.1
5519
.008
HQ
19.1
2119
.122
Log
Lik
elih
ood
-124
0.16
3-1
247.
860
df86
80L
RS
tati
stic‡
χ2(6
)15
.396
(0.0
17)
†N
otes
asp
erT
able
2.
30
Table
4P
aram
eter
esti
mat
esof
equ
atio
n(6
)-
Gw
eru
toB
ula
way
om
arke
tpa
ir
Tra
de
Regim
eN
oT
rade
Regim
eR
est
rict
ed
(θT
=θN
T)
Eq.
Para
mete
rC
oeffi
cien
t†Std
.C
oeffi
cien
tStd
.C
oeffi
cien
tStd
.(θ
T)
Err
.(θ
NT
)E
rr.
Err
.[e
t−1(B
)]B
1-1
.140
-0.8
31-1
.146
[Pjt
]A
Pj
-0.2
31**
*0.
096
-0.6
08**
**0.
145
-0.1
85**
*0.
076
CP
j−
tran
s.co
sts
0.64
90.
724
0.57
40.
744
0.23
70.
747
CP
j−
con
s.-2
7.76
977
.452
-27.
769
77.4
52-1
6.36
182
.313
ΓP
j,P
jt−
10.
040
0.10
00.
040
0.10
00.
003
0.10
6Γ
Pj,P
i t−
1-0
.114
0.15
3-0
.114
0.15
30.
063
0.15
9
[Pit]
AP
i0.
142
***
0.06
40.
103
0.09
80.
129
****
0.04
7C
Pi−
tran
s.co
sts
-0.8
74**
0.48
8-0
.971
**0.
571
-0.8
13**
0.46
2C
Pi−
con
s.93
.405
**52
.138
93.4
05**
52.1
3890
.582
**50
.837
ΓP
i,P
jt−
1-0
.008
0.06
7–0
.008
0.06
7-0
.006
0.06
5Γ
Pi,P
i t−
10.
013
0.10
40.
013
0.10
40.
024
0.09
8
AIC
18.4
7818
.539
HQ
18.6
4418
.653
Log
Lik
elih
ood
-121
5.82
1-1
223.
943
df86
80L
RS
tati
stic‡
χ2(6
)16
.244
(0.0
13)
†N
otes
asp
erT
able
2.
31
Table
5Im
plie
dH
alf-
Liv
es(i
nda
ys)
for
Tra
devs
.N
on-T
rade
Reg
imes
and
for
the
Ove
rall
sam
ple
Mark
et
Par.
Tra
de
Reg
ime
No
Tra
de
Reg
ime
Overa
llPair
(Eq.)
(Thalf
=λ
T)
(Thalf
=λ
NT)
(Thalf
=λ)
Est
.C
oeff
.s.
e.E
st.
Coe
ff.
s.e.
Est
.C
oeff
.s.
e.H
re-G
wr
Thalf
(AP
j)
11.1
78**
*5.
201
11.2
50*
7.72
693
.091
187.
878
Hre
-Gw
rT
half
(AP
i)
11.9
36**
7.07
24.
907
***
2.34
811
.633
****
4.06
3H
re-B
yoT
half
(AP
j)
19.5
33*
12.5
375.
075
****
1.63
530
.799
***
13.0
52H
re-B
yoT
half
(AP
i)
25.4
60*
16.8
0030
.069
31.3
0842
.967
***
18.7
84G
wr-
Byo
Thalf
(AP
j)
9.25
5**
*4.
386
2.58
9**
*1.
022
11.8
81**
*5.
439
Gw
r-B
yoT
half
(AP
i)
15.8
90**
*7.
804
22.2
3022
.183
17.5
26**
*6.
845
32
References
Balke, N. S., & Fomby, T. B. (1997, August). ‘Threshold cointegration.’ International
Economic Review , 38 (3), 627-645.
Baltagi, B., & Wang, Z. (2007, July). ‘Testing for cointegrating rank via model
selection: Evidence from 165 data sets.’ Empirical Economics , 33 (1), 41-49.
Barrett, C. B. (1996, August). ‘Market analysis methods: Are our enriched toolkits
well-suited to enlivened markets?’ American Journal of Agricultural Eco-
nomics , 78 (3), 825-829.
Barrett, C. B. (2001, Spring). ‘Measuring integration and efficiency in international
agricultural markets.’ Review of Agricultural Economics , 23 (1), 19-32.
Barrett, C. B., & Li, J. R. (2002, May). ‘Distinguishing between equilibrium and
integration in spatial price analysis.’ American Journal of Agricultural Eco-
nomics , 84 (2), 292-307.
Baulch, B. (1997, May). ‘Transfer costs, spatial arbitrage, and testing for food market
integration.’ American Journal of Agricultural Economics , 79 (2), 477-487.
Boswijk, H. P., & Doornik, J. A. (2004, November). ‘Identifying, estimating and
testing restricted cointegrated systems: An overview.’ Statistica Neerlandica,
58 (4), 440-465.
Brummer, B., von Cramon-Taubadel, S., & Zorya, S. (2009, June). ‘The impact of
market and policy instability on price transmission between wheat and flour in
ukraine.’ European Review of Agricultural Economics , 36 (2), 203-230.
Coleman, A. (2009, February). ‘A model of spatial arbitrage with transport capacity
constraints and endogenous transport prices.’ American Journal of Agricultural
33
Economics , 1 (1), 42-56.
Deaton, A., & Laroque, G. (1996, October). ‘Competitive storage and commodity
price dynamics.’ Journal of Political Economy , 104 (5), 896-923.
Dercon, S. (1995, October). ‘On market integration and liberalization: Method and
application to Ethiopia.’ Journal of Development Studies , 32 (1), 112-143.
Dercon, S., & van Campenhout, B. (1999). Dynamic price adjustment in spatially
separated food markets with transaction costs. (Working paper. Centre for the
Study of African Economies, Oxford University)
Engle, R. F., & Granger, C. W. J. (1987, March). ‘Cointegration and error-correction:
Representation, estimation and testing.’ Econometrica, 55 (2), 251-276.
Fackler, P. L., & Goodwin, B. K. (2002). ‘Spatial price analysis.’ In B. L. Gardner
& G. C. Rausser (Eds.), Handbook of agricultural economics (p. 972-1024).
Elsevier Science.
Fackler, P. L., & Tastan, H. (2008, February). ‘Estimating the degree of market
integration.’ American Journal of Agricultural Economics , 90 (1), 69-85.
Gonzalo, J., & Pitarakis, J.-Y. (2006, December). ‘Threshold effects in cointegrating
relationships.’ Oxford Bulletin of Economics and Statistics , 68 (s1), 813-833.
Goodwin, B. K., & Piggott, N. E. (2001, May). ‘Spatial market integration in the
presence of threshold effects.’ American Journal of Agricultural Economics ,
83 (2), 302-317.
Hansen, B. E., & Seo, B. (2002, October). ‘Testing for two-regime threshold coin-
tegration in vector error-correction models.’ Journal of Econometrics , 110 (2),
293-318.
34
Hansen, P. R. (2003, June). ‘Structural changes in the cointegrated vector autore-
gressive model.’ Journal of Econometrics , 114 (2), 261-295.
Ihle, R., Amikuzuno, J., & von Cramon-Taubadel, S. (2010). ‘L’integration des
marches avec et sans echanges commerciaux directs: le cas de la tomate au
Ghana.’ Revue d’economie du developpement , 24 (1), 21-46.
Jensen, R. (2007, August). ‘The digital provide: Information (technology), market
performance, and welfare in the south indian fisheries sector.’ Quarterly Journal
of Economics , 122 (3), 879-924.
Johansen, S. (1988, June-September). ‘Statistical analysis of cointegrating vectors.’
Journal of Economic Dynamics and Control , 12 (2-3), 231-254.
Johansen, S., & Juselius, K. (1992, July-September). ‘Testing structural hypotheses
in a multivariate cointegration analysis of the ppp and uip for uk.’ Journal of
Econometrics , 53 (1-3), 211-244.
Mabaya, E. (2003). Smallholder agricultural markets in Zimbabwe: Organization,
spatial integration and equilibrium. Doctoral dissertation, Cornell University.
Mabaya, E. (2004, September-December). ‘Spatial market integration under discon-
tinuous trade: an analysis of urban spot markets for tomatoes in Zimbabwe.’
Economie Rurale, 283-284 .
McNew, K. (1996, April). ‘Spatial market integration: definition, theory and evi-
dence.’ Agricultural and Resource Economics Review , 25 (1), 1-11.
Negassa, A., & Myers, R. J. (2007, May). ‘Estimating policy effects on spatial market
efficiency: An extension of the parity bounds model.’ American Journal of
Agricultural Economics , 89 (2), 338-352.
35
Oberhofer, W., & Kmenta, J. (1974, May). ‘A general procedure for obtaining
maximum likelihood estimates in generalized regression models.’ Econometrica,
42 (3), 579-590.
Pesaran, M. H., & Shin, Y. (2002, February). ‘Long-run structural modeling.’
Econometric Reviews , 21 (1), 49-87.
Siklos, P. L., & Granger, C. W. J. (1997, September). ‘Regime-sensitive cointegration
with an application to interest-rate parity.’ Macroeconomic Dynamics , 1 (3),
640-657.
Takayama, T., & Judge, G. G. (1971). Spatial and temporal price and allocation
models. Amsterdam: North Holland.
Tomek, W. G., & Robinson, K. L. (1990). Agricultural product prices. Cornell
University Press: Ithaca.
Williams, J. C., & Wright, B. D. (1991). Storage and commodity markets. NY:
Cambridge University Press.
36