Abstract We use spatial maize prices in Mozambique to measure transport cost reductions, attribute these reductions to road distance and road quality and assess to what extent producers, traders and consumers benefit. For identification we exploit a unique natural experiment, the construction of a new road bridge over the Zambezi River, which connects the north and south of Mozambique. The applied methodology allows for potentially oligopolistic traders with spatially varying mark-ups. Estimations are based on monthly maize prices, in 22 markets, for 5 years before and after the introduction of the bridge. Estimates of transport cost reductions, averaged over routes, vary from 3% to 7%, with large heterogeneity between routes, and roughly for two-third due to road distance and for one-third due to road quality. On average benefits of trade cost reductions are equally shared between traders and consumers, but for larger distances, a larger part accrues to traders. The evidence also indicates a reduction in prices in destination markets due to the bridge. Results are supported by observed transport cost data, robust for non-random bridge placement and strict source-destination rules. 1. Introduction Trade costs are important for developing countries, both for rural development and economic growth. Most sub-Saharan African countries are plagued with high to extremely high trade costs, which has major implications for the operation of these economies. Since the seminal theoretical work of Key et al. (2000) and de Janvry and Sadoulet (2006) there is a theoretical basis for explaining the rationality of subsistence farming and the key role of transport costs in this outcome, where transport costs affect both input and output prices. Their framework goes a long way in explaining low input levels, low productivity and low technological progress in agriculture. Various contributions have supplied empirical evidence supporting the idea that transport costs constitute a major cause for subsistence farming (see, for example, Omamo, 1998). Conversely, low trade costs or major reductions in trade costs lead to improvements in the operation of markets, moving agricultural produce more easily from low-price rural surplus areas to high-price urban deficit areas and increasing the welfare of consumers. A wide range of studies have empirically investigated to what extent infrastructure—be it road infrastructure (roads, bridges, etc.), rail infrastructure (railroads) or communication infrastructure (ICT, mobile phones)—have improved the efficiency of markets, raised the welfare of households and alleviated poverty. Results so far are mixed, suggesting in some cases that both producers and consumers realise welfare gains (see, for example, Jensen, 2007), while in others welfare gains for farmers are not evident and traders are likely to benefit most (see, for example, Fafchamps and Minten, 2012; Zant, 2019). Improved operation of markets, notably of food markets, will, nevertheless, always be helpful in increasing food security. The topic of the current study is related to several lines of research in the empirical literature on trade costs, on the impact of transport infrastructure (Donaldson, 2018; Banerjee et al., 2012; Casaburi et al. 2013; Brooks and Donovan, 2020; Volpe Martincus et al., 2017; Zant 2018), on the operation of markets (Tostão and Brorsen, 2005; Cirera and Arndt, 2008; Zant 2013, 2019), on prices and transport cost (Minten and Kyle, 1999), on transport costs and behaviour of households (Jacoby, 2000; Renkow et al., 2004; Jacoby and Minten, 2009) and on the impact of (rail) infrastructure on crop prices (Zant, 2018). In the current work, we use the theoretical framework on trade costs with potentially oligopolistic traders, developed by Atkin and Donaldson (2015) and investigate to what extent a new bridge leads to lower trade costs, exploiting the spatial variation of maize prices between major maize markets in Mozambique. Maize is the most widely traded staple food in Mozambique, thereby allowing meaningful inferences. We use monthly maize price data of 22 major maize markets. Estimations take account of spatially varying trader mark-ups and the empirical strategy addresses identification of source markets and destination markets and homogeneity of the traded product. For identification we exploit the introduction of a road bridge over the Zambezi River, in August 2009 between Caia and Chimuara, which removed a major obstruction to north–south trade and created the required variation in shortest trading routes between markets. The price data extend 5 years before and after the introduction of the new bridge. We find a reduction of transport costs, averaged over routes, varying from 3% to 7%, and roughly for two-third due to road distance and for one-third due to road quality. Benefits are mainly captured by traders and consumers, on average in more or less equal proportions, but with longer trading distances increasingly by traders. In the remainder of this paper we discuss, in Section 2, the maize market in Mozambique, the Mozambique road network and transport services and the role of the bridges crossing the Zambezi River in the Mozambique economy. In Section 3 we explain the theory underlying the empirical estimation, present details on data and data sources, set out the empirical strategy and discuss identification. In Section 4 we present the impact estimations, and the parallel trend test, we verify estimated impacts with trade cost data and present robustness checks. In Section 5 we attribute cost reductions to road distance and road quality and quantify the distribution of benefits of transport cost reductions. We conclude with a summary of findings in Section 6. 2. Maize markets, maize prices, road network and the Zambezi River1 2.1 Maize markets We use maize spatial prices to estimate transport costs. Maize is the most important staple food of Mozambique: it is widely produced, marketed and consumed. In all provinces two thirds of rural households produce maize. Despite widespread subsistence farming—only around 30% of maize production is traded on the market (Tschirley et al., 2006)—maize is three times more marketed than cassava. Also, maize has a budget share similar to all other staple foods2 jointly (Tschirley et al., 2006). The calorie share of maize in the average Mozambique diet ranges from 25% to 39%, corresponding with a per capita (annual) consumption of 60–85 kg, although, particularly in the south and in the Maputo region, the maize share is lower due to substitution with rice (Tschirley et al., 2006). Agriculture in Mozambique is rain-fed. The northern and central provinces have better rainfall distribution and better soil fertility, while the Southern provinces have less favourable weather conditions (see Figure A2). As a result, production of maize is concentrated in the central and northern part (specifically Manica, Tete, Zambezia and Niassa). Occasional drought and flooding cause sizable fluctuations in production between years and corresponding hikes in prices. Combining production and population by geographical area3, with per capita dietary consumption is a simple and straightforward way to find surplus and deficit regions and an indication of potential maize trade flows (see Figure A2). The provinces Niassa in the north and Manica and Tete in the center are typical maize surplus areas and thereby key source areas, while most coastal areas and particularly Maputo province in the south are major deficit areas and thereby key destinations of maize trade. Long-distance maize trade in Mozambique is dominated by informal itinerant traders and—to a smaller extent—by large-scale assemblers (Zovala, 2014; De Vletter and Polana, 2001). Informal itinerant traders purchase and collect maize from rural markets or directly from large farmers, right after harvest, with their own working capital, and arrange truck transport themselves once a sufficient number of bags with maize is collected. The bulk of the maize traded in assembly and retail markets in Mozambique is supplied by these informal traders. Maize available for sale in markets in Maputo (Xiquelene and others) is primarily sourced from the central region, Nhamatanda, Chimoio or Manica, around 1100 km by road (Abdula, 2005; SIMA data from 1999–2001), or even as far as Tete, around 1500 km by road from Maputo (Tostão and Brorsen, 2005; SIMA trade flow data from 1998–2001). Transport cost data from SIMA pertain to routes between markets as far as 2300 km apart. Southern Mozambique, and the Maputo-Matola area in particular, also rely on South Africa as a supplier of maize (see Haggblade et al., 2008), where prices often are lower (see Figure A7). 2.2 Maize prices Maize market prices by region move in a parallel fashion indicating that regions are connected by trade (see Figure A3). Overall, prices in the south are higher than in the centre and north, reflecting the difference between surplus and deficit areas. Compared to rural areas, urban markets, with higher levels of demand, typically have higher average price levels, but lower price variability, which is related to more storage and more alternatives for substitution. There is a distinct systematic pattern in maize prices over the season (see Figure A4), with high peaks between January and March, a short but sharp drop in April–May when the new harvest enters the market, followed by a long gradual increase towards the end of the calendar year. During the lean season—characterised by low or even no supply—prices rise to levels more than twice as high as prices directly after harvesting. These movements in prices over the season are common for staple food prices in sub-Saharan countries (see Kaminski et al., 2016; Zant, 2019). Because seasonality between source and destination markets is not fully aligned, seasonality also shows up in spatial price differences (see Figure A5): spatial price differences peak from April to June, directly after harvest when trading activity is high. Jointly with variation in supply, these fluctuations in prices and spatial price differences point to periods with varying potential, possibly not fully exploited, benefits from trade between regions. 2.3 Road network, transport services and bridges across the Zambezi River Maize is mainly transported by trucks. Overall, the road infrastructure in Mozambique is not well developed. However, the trunk-road network, connecting main cities and towns (including all maize markets identified in the current study) is in a reasonably good shape during the period of study, after major improvement over the preceding decades (see Dominguez-Torres and Briceño-Garmendia, 2011). Other roads, notably secondary roads and feeder roads, are in a poor condition and, especially during the rainy season, many of these roads cannot always be used. Also road density is extremely low, even compared to other sub-Saharan countries (see Dominguez-Torres and Briceño-Garmendia, 2011). In summary, the main trunk-road network appears to be a relatively stable vehicle for long-distance maize trade during the period of study. The Zambezi is a large river, with a ‘within Mozambique’ length of around 800 km, running from west to east and splitting the country into a northern and southern part4. At the time of writing (2017) there are four bridges across the Zambezi River. A bridge at Tete—a suspension bridge in operation since 1973—is integrated in the highway network and forms a major gateway—in the northern direction—to Zambia, Malawi and the northern part of Mozambique, and—in the southern direction—to Zimbabwe, South-Africa and the southern part of Mozambique. For a long time this bridge has been the only way to cross the Zambezi, and, as a result, the bridge has become a bottleneck in road transport plagued by severe congestion. In the course of the past decade an additional bridge in Tete has been constructed, which opened for traffic in 2014. Around 250 km downstream from Tete to the southeast, there is another bridge, a railway bridge, spanning the lower Zambezi River between the towns of Vila de Sena and Mutarara. This bridge was originally built by the Portuguese in 1934 with the purpose to connect Malawi and the Moatize coal fields near Tete to the port of Beira. Although not located on a primary highway, it provided an alternative route to cross the Zambezi River, next to the bridge at Tete and the former ferry between Caia and Chimuara. After the ending of the civil war (1992), the bridge was converted to a single-lane bridge for vehicle traffic with assistance of USAID. More recently, in October 2006, the bridge was completely closed to vehicle traffic for rehabilitation and (re-)conversion into a rail bridge again: it was re-opened as a rail bridge in August 2009, the same month the Caia—Chimuara bridge was opened (see below). It is not clear to what extent the Vila de Sena—Mutarara bridge was effectively used as a road bridge, before it was closed for traffic to be rehabilitated as a railway bridge in October 2006. The Vila de Sena—Mutarara bridge is, however, not integrated in the major trunk-road network of Mozambique (see Figure A1), which makes it not useful for transport with trucks5. On 1 August 2009 a new road bridge over the Zambezi River opened, between Caia and Chimuara, linking Sofala and Zambezia provinces in the centre of the country. Tete, where the other road bridges over the Zambezi are located, is around 300 km kilometres to the northwest. The bridge is part of the main north–south highway and connects major commercial centers in the north (e.g. Nacala, Nampula) and south (e.g. Beira, Maputo). The construction of the bridge began in March 2006. The new bridge at Caia has replaced a ferry service between Caia and Chimuara that operated until the completion of the bridge6. In the period before the introduction of the bridge, the ferry was widely perceived as inefficient due to long waiting times and extensive queues of trucks, causing high risks of spoiling perishable crops in case of food transport. Truck transport often encountered delays for days or weeks for the ferry trip between Caia and Chimuara. Engine breakdowns and sensitivity to tides made the ferry connection unreliable. Tostão and Brorsen (2005) report: ‘…in early 2001 the ferry was shut down for nearly two months because the Zambezi River was flooding, and in July 2001 the ferry service was interrupted again because there was not enough water in the river’. We conclude that uncertainties in the availability of transport by ferry across the Zambezi have obstructed systematic trade. Transport by ferry therefore is unlikely to have had a major impact on local commodity markets. Despite the bridges, crossing the Zambezi River has been a barrier to trade and transport by road: before the opening of the Caia–Chimuara bridge major domestic trade flows of maize ran from the central area to the outer south, south of the Zambezi River, while north of the Zambezi River the coastal cities were supplied by the inland production centers in the north. 3. Theory, data, empirical strategy and identification 3.1 Underlying theory We propose the following set-up for spatial prices, taken over, with minor changes, from Atkin and Donaldson (2015). Let pd and po denote price at destination and origin locations, then $$ \begin{equation} {p}_d={p}_o+\tau \left({X}_{od}\right)+{\mu}_{od},\kern0.75em \end{equation}$$(1) where the price in a destination location (pd) is the sum of the price at the origin location (po), transport costs τ(Xod) and a mark-up (μod). In terms of spatial price differences and following Atkin and Donaldson (2015), we re-write this equation as follows: $$ \begin{equation} {p}_d-{p}_o=\tau \left({X}_{od}\right)+{\mu}_{od}\left({c}_{od},{\phi}_{od},{D}_{od}\right), \end{equation}$$(2) indicating that the markup is a function of the traders’ marginal cost |${c}_{od}$|, the competitive environment faced by traders, summarised by the competitiveness index |${\phi}_{od}$|, and demand conditions |${D}_{od}$|. Equation (2) indicates, and this is its key message, that mark-ups |$\mu (..)$| and transport costs |$\tau ({X}_{od})$| are correlated: traders’ marginal cost enters both the transport costs and the mark-up. Using this expression, the effect of a small change in a transport cost shifter xd on the spatial price differences follows: $$ \begin{align} \frac{d\left({p}_d-{p}_o\right)}{d\left({x}_{od}\right)}&=\left(1+\frac{\partial \mu }{\partial {c}_{od}}\right).\frac{\partial \tau \left({X}_{od}\right)}{\partial {x}_{od}}+\frac{\partial \mu }{\partial {\phi}_{od}}.\frac{\partial {\phi}_{od}}{\partial {x}_{od}}+\frac{\partial \mu }{\partial {D}_{od}}.\frac{\partial {D}_{od}}{\partial {x}_{od}} \nonumber\\ &={\rho}_{od}\frac{\partial \tau \left({X}_{od}\right)}{\partial {x}_{od}}+\frac{\partial \mu }{\partial {\phi}_{od}}.\frac{\partial {\phi}_{od}}{\partial {x}_{od}}+\frac{\partial \mu }{\partial {D}_{od}}.\frac{\partial {D}_{od}}{\partial {x}_{od}}, \end{align}$$(3) where |${\rho}_{od}$| is known as the pass-through rate. The pass-through rate is defined as the effect of traders’ marginal cost on prices while holding competitiveness (|${\phi}_{od}$|) fixed. Atkin and Donaldson (2015) show that, in general, the pass-through rate is determined by competitiveness and the curvature of the demand curve. Most of the empirical literature refers to a set-up where mark-ups are independent of costs (|$\partial \mu /\partial {c}_{od}=0$|) resulting in |${\rho}_{od}=1$|. In this special case spatial price differences reflect transport costs and can be used directly in estimations. Alternatively, in all imperfectly competitive settings, mark-ups will be correlated with costs (|$\partial \mu /\partial {c}_{od}>0$|, and |${\rho}_{od}\ne 1$|), some of the marginal cost will be passed through to prices (hence |${\rho}_{od}>0$|), but whether a lower (|${\rho}_{od}<1$|) or higher (|${\rho}_{od}>1$|) than complete pass-through is optimal is not determined. The second and third term in the derivative expression (equation ((3): |$\partial \mu /\partial {\phi}_{od}.\partial {\phi}_{od}/\partial {x}_{od}$| and |$\partial \mu /\partial {D}_{od}.\partial {D}_{od}/\partial {x}_{od}$|) capture that mark-ups vary across locations because of differences in competitive conditions and because of differences in preferences (respectively |${\phi}_{od}$|, |${D}_{od}$|). Atkin and Donaldson (2015) argue that fairly reasonable assumptions—competitiveness |${\phi}_{od}$| may vary across locations but is fixed within a location, and consumer preferences (|${D}_{od}$|) are such that the curvature of the slope of the inverse demand curve is constant (Bulow-Pfleiderer demand)—are sufficient for constant pass-through rates |${\rho}_{od}$|. The previous price spatial difference equation, jointly with these assumptions can be re-written as follows: $$ \begin{equation} {p}_d-{p}_o={\rho}_{od}\tau \left({X}_{od}\right)+\left(1-{\rho}_{od}\right)\left({a}_d-{p}_o\right), \end{equation}$$(4) where |${a}_d$| is a demand shifter. A constant pass-through turns out to be particularly useful to estimate trade costs in the presence of varying mark-ups: this expression shows that the pass-through rate |${\rho}_{od}$| and the demand shifter |${a}_d$| are sufficient to control for the bias arising because of unobserved preferences (|${D}_{od}$|) and market structure (|${\phi}_{od}$|). Introducing time, and assuming that the pass-through rate |${\rho}_{od}$| varies across source-destination combinations, but is fixed over time, and taking account of variable sources of products, we may write, after some reshuffling: $$ \begin{equation} {\mathrm{p}}_{dt}={\rho}_{od}{p}_{ot}+{\rho}_{od}\tau \left({X}_{od t}\right)+\left(1-{\rho}_{od}\right){a}_{dt}. \end{equation}$$(5) In order to extract the pass-through rate from this equation we need to have data on transport costs (|$\tau ({X}_{odt})$|) and local demand shifters (|${a}_{dt}$|). As these variables are not known to the researcher, we assume that both can be approximated with a local time invariant factor, a local trend factor and a residual factor, or formally: $$ \begin{equation} \tau \left({X}_{odt}\right)={\beta}_{1 od}+{\beta}_{2 od} trend+{\zeta}_{odt} \end{equation}$$(6) $$ \begin{equation} {a}_{dt}={\alpha}_{1d}+{\alpha}_{2d} trend+{\nu}_{dt}. \end{equation}$$(7) Combining the last three equations yield: $$ \begin{equation} {p}_{dt}={\rho}_{od}{p}_{ot}+{\gamma}_{1 od}+{\gamma}_{2 od} trend+{\varepsilon}_{dt}. \end{equation}$$(8) Now we can estimate the pass-through rate |${\rho}_{od}$| for each source-destination combination. This is the first step in adjusting spatial price differences to extract transport costs from these price data. The second step is to use estimated values for pass-through rates by source-destination combination, to construct adjusted variables and use these adjusted variables to estimate determinants of transport costs. Hence, we estimate $$ \begin{equation} \left({p}_{dt}-\hat{\rho_{od}}{p}_{ot}\right)/\hat{\rho_{od}}= \tau \left({X}_{od t}\right)+{\alpha}_{1d}\left(\left(1-\hat{\rho_{od}}\right)/\hat{\rho_{od}}\right)+{\alpha}_{2d} trend\left(\left(1-\hat{\rho_{od}}\right)/\hat{\rho_{od}}\right)+{\varepsilon}_{dt,} \end{equation}$$(9) where a hat on the pass-through rates |${\rho}_{od}$| denotes its estimated value. The presented framework, the resulting two-step procedure and notably equations (8) and (9), are the backbone of the empirical estimations. The conceptual framework raises various empirical issues: in the first place we need to ascertain that the ‘source market of the price differential’ is the market where the product is actually sourced. We have identified source and destination markets in the first place by exploiting information on source and destination from the direction of trade in trade cost data: further details on this procedure are in the Appendix. Next, we need to find a convincing set of variables that reflects transport costs (|$\tau ({X}_{odt})$|). We use road distance and road quality as major determinants of real transport costs: details on construction and source are in the data section. Finally, prices should pertain to a homogeneous good with limited quality differences. From the background section we know that white maize grain is produced, consumed and traded throughout Mozambique. Maize prices are recorded without specifications for quality: apparently one kg of white maize grain, in Maputo in the south or Nampula in the north, or in 2005 or 2014, is not different in terms of quality. The trade cost data also only have a marginal number of observations of two-way trade7, a type of trade that is associated with differences in quality. Therefore, and without denying possible quality differences, we assume that the requirement of a homogenous product is satisfied and potential bias due to unobserved quality differences in the estimations is negligible. 3.2 Data and data sources The empirical work is almost completely based on prices: there is virtually no systematic and reliable information on trade flows. Market prices for maize are sourced from Sistema de Informação de Mercados Agrícolas de Moçambique (SIMA; www.masa.gov.mz/sima), from the weekly publication Quente-Quente8. We use in particular weekly retail prices of white maize grain9, averaged to monthly data, and originally recorded for 27 markets10 for the period from January 2005 to December 2014. This period covers a timespan of around 5 years with and without bridge. Price data are collected by interviewing randomly selected traders in each market. There are missing observations in the price data, even in the monthly averages but, fortunately, these are small in number and also not unexpected11. The construction of spatial price differences—the price dispersion measure that is used in the estimations—blows up the number of available observations to large numbers12. We further exploit a small, but nevertheless unique sample of transport cost data, also available from SIMA. Collection of these data is organised similarly to the collection of price data, by asking quotations from randomly selected traders in major source and destination markets. Transport costs are specified by date, route (i.e. source market and destination market), product and weight of the bags transported. We only use transport cost data for maize grain. Unfortunately, transport costs are recorded for the period from August 2001 to December 2010, with nearly three quarter of the observations before 2005. After 2010—and for undisclosed reasons—the publication of tables with transport cost data becomes less frequent and also focuses less on long-distance trade. Moreover, available transport costs series are also only for a limited number of routes (see Table A2). In all we have 832 observations of transport costs that cover the period of study (January 2005–December 2014) to a limited extent (see Figure A4). Also, there are no observations of transport costs after July 2009 of routes that cross the new bridge. Road distance and travel time are both taken from Google Maps, at the time of writing (the first version of) this study (2017). Since our study period is from January 2005 to December 2014, this may entail measurement error: we do not incorporate construction of new roads, and road maintenance and rehabilitation of existing roads (other than changes due to the introduction and rehabilitation of the bridges). To a certain extent common developments in road infrastructure are controlled for through the differences-in-differences (DiD) estimation strategy: common time varying shocks and trends are digested by the time fixed effects. Fortunately, the trunk road network—the road network that connects all major the markets identified in this study—is well maintained and has not changed much during the period of study (see Dominguez-Torres and Briceño-Garmendia, 2011). A few other variables are used for descriptive purposes, as covariates in the estimations and to estimate propensity scores: data of population by city or town are from three censuses (1997, 2007, 2016) from the Instituto Nacional de Estatistica Moçambique. Population data for intermediate months and years are constructed by interpolation. Fuel prices are from the International Financial Statistics from the IMF. Maize production by province is from Trabalho de Inquérito Agrícola/Anuario de Estatistica Agararia, Ministry of Agriculture. Maize production data are incomplete: missing years are constructed (see Table A1). We have used quotations of the nearest SAFEX white maize futures contract as representative for maize prices in South Africa. These series are taken from the SAFEX web site and converted to Mozambique meticais with the help of IMF/IFS monthly exchange rates (period average, domestic currency per US dollars). Finally, we use market prices of maize grain from Lilongwe, Malawi, taken from FAO, also converted to Mozambique meticais with the help of IMF/IFS monthly exchange rates. Lilongwe maize market prices are assumed to represent price (demand) developments in the Malawi maize market. 3.3 Empirical strategy We postulate that real transport costs are empirically determined by road distance and road quality: these two variables jointly represent transport costs (τ(Xodt)). Following the conceptual framework set out in the Underlying Theory section and formalised in equation (9), we further take account of variations in mark-ups across space. In summary, to measure the impact of the Caia–Chimuara bridge on the determinants of transport costs, we start with the following DiD specification: $$ \left({p}_{jt}/\hat{\rho_{jk}}\right)-{p}_{kt}={\eta}_0+{\eta}_1{roaddistance}_{jk t}+{\eta}_2{roadquality}_{jk t} $$ $$ +{\theta}_{jk}\left(\left(1-\hat{\rho_{jk}}\right)/\hat{\rho_{jk}}\right)+{\omega}_j trend\left(\left(1-\hat{\rho_{jk}}\right)/\hat{\rho_{jk}}\right) $$ $$ \begin{equation} +{\boldsymbol{X}}_{jkt}\gamma +{\psi}_t+{\varphi}_{jm}+{\chi}_{km}+{\varepsilon}_{jkt}\ \end{equation}$$(10) In this equation |${roaddistance}_{jkt}$| is the shortest road distance between markets j and k at time t. We expect that an increase in road distance increases transport costs, and hence |${\eta}_1>0$|. The variable |${roadquality}_{jkt}$| is specified as the number of kilometres realised per hour of travel time or average speed: a high value indicates a good road quality, a low value a poor road quality. We expect that increased road quality decreases transport costs, and hence |${\eta}_2<0$|. The variables road distance and road quality are the real determinants of transport costs. The identification of market pairs that realise a change in shortest road distance due to the bridge (the intervention pairs) is documented below. The RHS variables in the second line of equation (10), jointly with the dependent variable, contain the transformations of spatial price difference, trade pair fixed effect and trend that account for variations in mark-ups across space. In the third line, the vector Xjt represents variables that possibly influence the (adjusted) spatial price difference, such as rainfall in source areas and foreign maize prices. Parameters |${\psi}_t$| are time fixed effects (months), |${\varphi}_{jm}$| and |${\chi}_{km}$| represent within-year seasonality in source and destination market, and take the value 1 for each month (Jan., Feb., etc.) and zero elsewhere and |${\varepsilon}_{jkt}$| is an error term. The parameters of interest are |${\eta}_1$| and |${\eta}_2$|, which measure the contribution of road distance and road quality to trade costs. The first stage regression required to find trade pair-specific estimates of the pass-through rate (|${\rho}_{jk}$|) comes straight from the conceptual framework and is estimated for each combination of source and destination separately: $$ \begin{equation} {p}_{jt}={\rho}_{jk}{p}_{kt}+{\gamma}_{1 jk}+{\gamma}_{2 jk} trend+{\varphi}_{jm}+{season}_{jt}+{\varepsilon}_{jt}, \end{equation}$$(11) where |${season}_{jt}$| represents between-year seasonality. Agricultural crop prices are notorious for seasonality, and maize prices in Mozambique are no exception. Seasonality in prices also translates into seasonality in spatial price differences (see section on Maize prices). Moreover, despite the large common component in seasonality in maize prices and price differences, there is considerable variation in timing and amplitude, both within market(pair)s, between years and between market(pair)s, for the same years. To accommodate for this seasonality, we have included within and between year seasonality, in both estimation steps, and—for estimations of spatial price differences—for both source and destination markets. 3.4 Identifying intervention pairs The key variable that governs impact is the minimum road distance between markets. The pre-intervention period is the period prior to August 2009, the month when the bridge became operational. Prior to August 2009 any road transport crossing the Zambezi River is assumed to be directed through Tete and crosses the Zambezi via the bridge in Tete. Intervention market pairs are those market pairs that realise a reduction in minimum road distance, from August 2009 onwards, as a result of the introduction of the bridge. By way of example: maize transport from Alto Molocue, north of the Zambezi, to Beira, south of the Zambezi (see Figure A1) involved a 1268-km (18 hours, 39 minutes) journey before and a 749-km (11 hours, 5 minutes) journey after the introduction of the bridge, a decrease of more than 40% in both road distance and travel time. Likewise, maize transport from Chimoio, south of the Zambezi, to Quelimane, north of the Zambezi, involved a 1005-km journey before and a 564-km journey after the introduction of the bridge, again, a decrease of more than 40% in road distance13. The scheme below summarizes the number of observations of spatial price differences for the different groups14. . Before August 2009 . August 2009 and later . Intervention pair 1947 2640 Non-intervention pair 3548 4684 . Before August 2009 . August 2009 and later . Intervention pair 1947 2640 Non-intervention pair 3548 4684 Open in new tab . Before August 2009 . August 2009 and later . Intervention pair 1947 2640 Non-intervention pair 3548 4684 . Before August 2009 . August 2009 and later . Intervention pair 1947 2640 Non-intervention pair 3548 4684 Open in new tab The identification strategy assumes implicitly that freight across the Zambezi River that is transported either through the Vila de Sena–Mutarara bridge (before March 2006, when it operated as a road bridge), or through the ferry that operated between Caia and Chimuara before August 2009, or through any other crossing of the Zambezi River, concerns relatively small quantities and has a negligible impact on maize markets, maize market prices and geographical price dispersion. Several sources confirm the poor operation of the ferry (see Section 1), while the Vila de Sena–Mutarara bridge is not integrated in the road network. Note also that, if this assumption is violated, the estimated reduction in transport costs is a lower bound. Finally, we are aware that local shocks that affect both sides of the Zambezi River in a different way may confound impact estimates or, more general, we acknowledge that effects from time varying unobservables cannot be ruled out. For several reasons (parallel trend test, propensity score matching (PSM)) it is useful to re-estimate the DiD specification with a binary impact variable, rather than with the variables road distance and road quality. We estimate the following: $$ {p}_{jt}/\hat{\rho_{jk}}-{p}_{kt}={\eta}_0+{\eta}_1{bridge}_{jk t}$$ $$ +{\theta}_{jk}\left(\left(1-\hat{\rho_{jk}}\right)/\hat{\rho_{jk}}\right)+{\omega}_j trend\left(\left(1-\hat{\rho_{jk}}\right)/\hat{\rho_{jk}}\right) $$ $$ \begin{equation} +{\boldsymbol{X}}_{jkt}\gamma +{\psi}_t+{\varphi}_{jm}+{\chi}_{km}+{\varepsilon}_{jkt},\kern0.75em \end{equation}$$(12) where |${bridge}_{jkt}$| is equal to 1 in period t if the shortest route from j to k runs via the Caia-Chimuara bridge when this bridge was operational, and zero otherwise. All other variables are the same as in equation (10). Impact is now expected to have a negative coefficient reflecting the reduction in transport costs, averaged over routes. 4. Pass-through rates, parallel trend, impact, verification and robustness 4.1 Estimating pass-through rates The estimation results of the first stage regression—the estimations needed to construct the pass-through rate by market pair—are not reported but available from the author15. We do show the estimates of the pass-through rate plotted against road distance, for all combinations of source and destination (see Figure A9). In line with the estimates of pass-through rates in Atkin and Donaldson (2015), our estimates of pass-through rates are lower for further away source markets. The restriction that the pass-through rate needs to be positive is never violated. Also, only with one exception, pass-through rates are below one. The average pass-through rate is 0.544, and, hence, given the incomplete pass-through there is clear evidence of imperfect competition. Atkin and Donaldson (2015) indicate that, if anything, estimates of the pass-through rate are likely to be biassed upward, when shocks at source are correlated with shocks at destination. Especially in the case of agricultural commodities, and more so than in the case of imported goods (the Atkin and Donaldson data), common shocks are likely (e.g. timing of harvest, weather shocks). Bias arising from common shocks is mitigated by including within and between year seasonality (|${\varphi}_{jm},{season}_t$|). Remaining bias supports the claim that estimated pass-through rate represent a lower limit of imperfect competition. Also, and despite attempts to mitigate bias, we cannot rule out that bias, introduced by using estimated coefficients at both side of the impact equation, may affect results. 4.2 The parallel trend assumption Now we can transform variables with estimated pass-through rates (|$\hat{\rho_{jk}}$|) and estimate equation (10). However, prior to presenting these estimation results we investigate the parallel trend assumption. For causal inference the DiD approach requires that pre-intervention outcomes of intervention and control groups develop along a parallel trend. Following standard practise (see Autor, 2003) we test this parallel trend assumption by including a complete set of interactions of market pair dummies of those market pairs that benefit from the bridge, both before and after the opening of the bridge, with time-period dummies. Formally we estimate the following: $$ {p}_{jt}/\hat{\rho_{jk}}-{p}_{kt}={\eta}_0+\varSigma\ {\eta}_{1 time}\big({intervention\ pairs}_{jk t}\ x\ time\big) $$ $$ +{\theta}_{jk}\left(\left(1-\hat{\rho_{jk}}\right)/\hat{\rho_{jk}}\right)+{\omega}_j trend\left(\left(1-\hat{\rho_{jk}}\right)/\hat{\rho_{jk}}\right) $$ $$ \begin{equation} +{\psi}_t+{\varphi}_{jm}+{\chi}_{km}+{\varepsilon}_{jkt},\kern3.5em \end{equation}$$(13) where |${intervention\ pairs}_{jkt}$| are the market pairs, in all time periods, that benefit from the introduction of the bridge, and time is a time indicator variable, either month, quarter or year. If coefficients of the interaction terms before the introduction of the bridge are statistically insignificant, all time trends and shocks are absorbed by the time fixed effects, for both intervention and non-intervention pairs, and the estimation result supports the parallel trend assumption. Jointly with the parallel trend, the graphical evidence also shows if coefficients are statistically significant (and negative) after the introduction of the bridge, and offers information on the dynamic path of impact, notably if it is stable, or how it varies over time. The outcome of this exercise, shown in Figure 1, confirms negative and statistically significant impacts after, and insignificant impacts before the introduction of the new bridge. Consequently, on the basis of the figure we cannot reject the hypothesis of a parallel trend in the pre-treatment period for intervention and non-intervention observations. The figure further confirms the consistency of impacts of a reduction of around 10% over the years after 2009. Figure 1 Open in new tabDownload slide Testing for a Common Trend in the Pre-Treatment Period. Note: The dotted lines indicate 95% confidence intervals. Figure 1 Open in new tabDownload slide Testing for a Common Trend in the Pre-Treatment Period. Note: The dotted lines indicate 95% confidence intervals. 4.3 Estimating impact of road distance and road quality Next, we proceed with estimating the basic specification that identifies the impact of road distance and road quality on trade costs (equation (10)). Equations are estimated with OLS and include market pair and time fixed effects, and source and destination specific seasonality. Following standard practise (see Bertrand et al., 2004) standard errors in the estimation are clustered at the level of market pairs and the intervention, which is especially important for a DiD estimation with a limited number of shocks. Estimations are based on a few specific samples. We drop, in all estimations, observations that are more than 1800 km apart (road distance): this avoids estimation results driven by trade pairs that are extremely far apart, and thereby less relevant16. Next, we take account of seasonality with two additional sample adjustments. During the lean season trade is minimal, markets are thin and prices fluctuate wildly. Under these circumstances spatial price differences tend to be less informative about transport costs. We focus therefore on the marketing season, the months following harvest, since most trade takes place during these months. In particular we estimate with observations from April to November (Table 1, columns 2 and 4) and from April to July (Table 1, column 3). Finally, we restrict the sample period before the introduction of the new and rehabilitated bridges to the period that is fully ‘without bridge’ facilities (see Section 1), aiming at a clean identification strategy (Table 1, column 4). Table 1 Impact of distance and road quality Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . . Sample . (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 1800 km road distancea Excluding Dec to Marb Excluding Aug to Marb Only 34 months beforec ln(road distance) 0.180*** (0.058) 0.222*** (0.062) 0.243** (0.098) 0.323*** (0.070) ln(road quality) −0.401*** (0.122) −0.629*** (0.123) −1.122*** (0.212) −0.706*** (0.153) Covariates No No No No Adj R2 0.831 0.828 0.842 0.824 Observations 11008 7392 3630 6374 ln(road distance) 0.198*** (0.062) 0.227*** (0.067) 0.222** (0.100) 0.273*** (0.085) ln(road quality) −0.430*** (0.124) −0.678*** (0.124) −1.099*** (0.212) −0.707*** (0.167) Covariates Yes Yes Yes Yes Adj R2 0.832 0.828 0.843 0.825 Observations 11008 7392 3630 6374 Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . . Sample . (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 1800 km road distancea Excluding Dec to Marb Excluding Aug to Marb Only 34 months beforec ln(road distance) 0.180*** (0.058) 0.222*** (0.062) 0.243** (0.098) 0.323*** (0.070) ln(road quality) −0.401*** (0.122) −0.629*** (0.123) −1.122*** (0.212) −0.706*** (0.153) Covariates No No No No Adj R2 0.831 0.828 0.842 0.824 Observations 11008 7392 3630 6374 ln(road distance) 0.198*** (0.062) 0.227*** (0.067) 0.222** (0.100) 0.273*** (0.085) ln(road quality) −0.430*** (0.124) −0.678*** (0.124) −1.099*** (0.212) −0.707*** (0.167) Covariates Yes Yes Yes Yes Adj R2 0.832 0.828 0.843 0.825 Observations 11008 7392 3630 6374 All equations are estimated with OLS and include pass-through corrected trends, market pair and time fixed effects and source and destination specific seasonality. Robust standard errors, clustered by group (market pair, before/after and intervention/control), are in brackets next to the coefficient. Covariates are Lilongwe maize market prices, SAFEX spot prices and lagged seasonal rainfall. aObservations with road distances larger than 1800 km are omitted. bRestricted to marketing months April–November (2) and restricted to marketing months April–July (3). cThe period ‘without bridges’ is from October 2006, the closing of the Mutarara bridge for rehabilitation works, to July 2009, 34 months before 1 August 2009, the opening of the Caia bridge (which also coincided with the completion of the rehabilitation of the Mutarara bridge). ∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01. Open in new tab Table 1 Impact of distance and road quality Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . . Sample . (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 1800 km road distancea Excluding Dec to Marb Excluding Aug to Marb Only 34 months beforec ln(road distance) 0.180*** (0.058) 0.222*** (0.062) 0.243** (0.098) 0.323*** (0.070) ln(road quality) −0.401*** (0.122) −0.629*** (0.123) −1.122*** (0.212) −0.706*** (0.153) Covariates No No No No Adj R2 0.831 0.828 0.842 0.824 Observations 11008 7392 3630 6374 ln(road distance) 0.198*** (0.062) 0.227*** (0.067) 0.222** (0.100) 0.273*** (0.085) ln(road quality) −0.430*** (0.124) −0.678*** (0.124) −1.099*** (0.212) −0.707*** (0.167) Covariates Yes Yes Yes Yes Adj R2 0.832 0.828 0.843 0.825 Observations 11008 7392 3630 6374 Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . . Sample . (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 1800 km road distancea Excluding Dec to Marb Excluding Aug to Marb Only 34 months beforec ln(road distance) 0.180*** (0.058) 0.222*** (0.062) 0.243** (0.098) 0.323*** (0.070) ln(road quality) −0.401*** (0.122) −0.629*** (0.123) −1.122*** (0.212) −0.706*** (0.153) Covariates No No No No Adj R2 0.831 0.828 0.842 0.824 Observations 11008 7392 3630 6374 ln(road distance) 0.198*** (0.062) 0.227*** (0.067) 0.222** (0.100) 0.273*** (0.085) ln(road quality) −0.430*** (0.124) −0.678*** (0.124) −1.099*** (0.212) −0.707*** (0.167) Covariates Yes Yes Yes Yes Adj R2 0.832 0.828 0.843 0.825 Observations 11008 7392 3630 6374 All equations are estimated with OLS and include pass-through corrected trends, market pair and time fixed effects and source and destination specific seasonality. Robust standard errors, clustered by group (market pair, before/after and intervention/control), are in brackets next to the coefficient. Covariates are Lilongwe maize market prices, SAFEX spot prices and lagged seasonal rainfall. aObservations with road distances larger than 1800 km are omitted. bRestricted to marketing months April–November (2) and restricted to marketing months April–July (3). cThe period ‘without bridges’ is from October 2006, the closing of the Mutarara bridge for rehabilitation works, to July 2009, 34 months before 1 August 2009, the opening of the Caia bridge (which also coincided with the completion of the rehabilitation of the Mutarara bridge). ∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01. Open in new tab Following the specification formulated in equation (10), we next include X variables in order to investigate if the previous estimation results are robust for the inclusion of covariates. For this purpose we have included, in the first place, foreign maize prices. Prices in neighbouring countries, especially in South Africa and Malawi, are likely to affect supply and demand in Mozambique. Note that South Africa and Malawi play a different role: South Africa is a major source of maize grain imports, while Malawi is a major destination of maize grain exports (see Zavale, 2014). The influence of foreign prices depends on whether maize is imported from or exported to foreign markets. South African maize prices are mostly below Maputo maize prices, even if import tariffs are accounted for (see Figure A7)17. Imports from South Africa into the southern Mozambique terminal maize markets generate a larger domestic supply, exerting a downward pressure on prices and, ceteris paribus, on spatial price differences. Therefore, we expect a positive impact of South-African maize prices on (domestic) spatial price differences. For Malawi maize prices the reverse applies (see Figure A8): Malawi is an outlet for Mozambique surplus maize and generally higher Malawi maize prices make exports to Malawi attractive, decrease supply and increase prices in source markets, and thereby dampen spatial price differences. Hence, we expect a negative effect of Malawi prices on Mozambique spatial price differences. We have used quotations of the nearest SAFEX white maize contract as indicator of South-African maize prices and Lilongwe market maize prices as indicator of Malawi maize prices, both converted to Mozambique meticais18. We have further included lagged seasonal rainfall. As good rains during the growing season improve harvest outcomes and supply in source areas, this is likely to decrease market prices during the subsequent marketing season and thereby, ceteris paribus, increase spatial price differences. We expect a positive impact of lagged seasonal rainfall on spatial price differences. Several alternative covariates are considered but most of these have serious drawbacks. Selected estimations of the basic specification (equation 10) without covariates—reported in the upper panel of Table 1—show several interesting outcomes. In the first place the estimations confirm the importance of the coefficients of road distance and road quality: these coefficients are both statistically significant and have the expected signs in all samples. The estimated coefficients are relatively stable with substantially different samples. Trends interacted with pass-through rates are also significant in all estimations (not shown). Estimations including covariates, reported in the lower panel of Table 1, confirm previous estimates of impact variables: coefficients of road distance and road quality are both statistically significant and have a similar size. The performance of covariates is mixed: we find lagged seasonal rainfall to consistently have a negative impact on spatial price differences, while coefficients of Lilongwe maize market prices and SAFEX spot prices are only occasionally statistically significant with the expected sign. We attribute these mixed results to the fact that effects will be limited in geographical scope, for a limited set of transport routes and may have lagged and indirect responses. We have nevertheless maintained these covariates. 4.4 Verification of impact with observed transport costs data The ultimate test of the estimations reported in the previous tables and of the proposed methodological framework is to estimate a similar relationship with observed transport costs, rather than (adjusted) spatial price differences. If the applied framework is an adequate technique to control for spatially varying mark-ups, and thereby a justification to use (adjusted) spatial price differences for the estimation of transport costs, estimation with observed transport costs should generate similar results. Although observed transport costs data are generally not or only fragmentary available, we do have a small number of observations on transport costs for Mozambique (see Data and data sources). Apart from conceptual reservations against trade costs data obtained from surveys (these data allegedly do not reflect marginal costs, and do not represent all costs incurred by traders: see Atkin and Donaldson, 2015; Dillon and Dambro, 2017), our trade cost data do not contain observations after July 2009 of routes that cross the new bridge. In order to verify the estimations based on spatial price differences, we estimate the following DiD equation, a simplified version of equation (10): $$ \begin{equation} {tc}_{jkt}={\eta}_0+{\eta}_1{roaddistance}_{jkt}+{\eta}_2 roadquali\mathrm{t}{y}_{jkt}+{\psi}_t+{\varepsilon}_{jkt.} \end{equation}$$(14) In this equation all pass-through rate adjusted variables are omitted. Additionally, since we do not have post-intervention transport cost observations, there is no variation of road distance and road quality over the years and by implication impacts are only identified by cross-sectional variation. Therefore, we need to drop market pair fixed effects. Moreover, we cannot cluster standard errors at the level of the intervention, because we lack observations after the opening of the bridge: hence, we cluster errors by market pair. Finally, in order to preserve statistical power, market specific seasonality is omitted (and, hence, we drop |${\varphi}_{jm}$| and |${\chi}_{km}$|): this seasonality, to a certain degree, is absorbed by time fixed effects. Selected estimation results, reported in Table 2, show impacts of road distance and road quality that are statistically significant and have the expected sign. Estimation results are also robust to large variations in samples. On the whole, the size of the coefficients is slightly larger (in absolute terms) relative to the adjusted price differences estimations in Table 1, but not far off-the-mark: at standard levels of confidence coefficients of Table 2 and those of Table 1 are statistically in the same range. Difference may be due to differences in the samples (see Figure A10). Restricting road distance further or adding confounders like seasonality brings the coefficients still closer to the coefficients in Table 1. We conclude that the available transport cost data offer reasonable support for the applied framework to use (adjusted) spatial price differences for the estimation of transport costs. Table 2 Verifying Impact of Distance and Road Quality with Transport Cost Data Dependent variable: ln(tcjk) . Sample (see also note to Table) (1) (2), as (1) (3), as (2) (4), as (2) Variables All observations Within 1800 km road distance Excluding Dec to Mar Excluding Aug to Mar ln(road distance) 0.474*** (0.062) 0.444*** (0.077) 0.456*** (0.072) 0.448*** (0.096) ln(road quality) −0.693** (0.311) −0.663** (0.309) −0.726** (0.306) −0.540* (0.315) Adj R2 0.571 0.536 0.561 0.531 Observations 634 596 419 224 Dependent variable: ln(tcjk) . Sample (see also note to Table) (1) (2), as (1) (3), as (2) (4), as (2) Variables All observations Within 1800 km road distance Excluding Dec to Mar Excluding Aug to Mar ln(road distance) 0.474*** (0.062) 0.444*** (0.077) 0.456*** (0.072) 0.448*** (0.096) ln(road quality) −0.693** (0.311) −0.663** (0.309) −0.726** (0.306) −0.540* (0.315) Adj R2 0.571 0.536 0.561 0.531 Observations 634 596 419 224 Note: The selection of source and destination markets in the transport cost data correspond to the selection explained in detail inSection 2. All equations are estimated with OLS and include time fixed effects. Robust standard errors are clustered by marketpair. ∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01. Open in new tab Table 2 Verifying Impact of Distance and Road Quality with Transport Cost Data Dependent variable: ln(tcjk) . Sample (see also note to Table) (1) (2), as (1) (3), as (2) (4), as (2) Variables All observations Within 1800 km road distance Excluding Dec to Mar Excluding Aug to Mar ln(road distance) 0.474*** (0.062) 0.444*** (0.077) 0.456*** (0.072) 0.448*** (0.096) ln(road quality) −0.693** (0.311) −0.663** (0.309) −0.726** (0.306) −0.540* (0.315) Adj R2 0.571 0.536 0.561 0.531 Observations 634 596 419 224 Dependent variable: ln(tcjk) . Sample (see also note to Table) (1) (2), as (1) (3), as (2) (4), as (2) Variables All observations Within 1800 km road distance Excluding Dec to Mar Excluding Aug to Mar ln(road distance) 0.474*** (0.062) 0.444*** (0.077) 0.456*** (0.072) 0.448*** (0.096) ln(road quality) −0.693** (0.311) −0.663** (0.309) −0.726** (0.306) −0.540* (0.315) Adj R2 0.571 0.536 0.561 0.531 Observations 634 596 419 224 Note: The selection of source and destination markets in the transport cost data correspond to the selection explained in detail inSection 2. All equations are estimated with OLS and include time fixed effects. Robust standard errors are clustered by marketpair. ∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01. Open in new tab 4.5 PSM estimation Since the placement of the bridge is non-random, and, hence, since the selection of market pairs that realise a reduction in road distance and an improvement in road quality is also non-random, a DiD estimation comparing this selection with all other market pairs is potentially suffering from selection bias in time-varying observables. As a result estimated causal impacts of the bridge on transport costs may be biased. In the spatial economics literature the major strategy to address this is to develop plausible instruments that meet the exclusion restriction19. As the proposed techniques cannot be used for the current work, we have implemented Propensity Score Matching (PSM). To allow full comparison of PSM with OLS results we have first estimated the basic specification with a binary impact. Estimations—reported in Table 3—repeat the estimations of Table 1, with the only difference that a binary impact variable is substituted for the road distance and road quality variables. Coefficients of the impact variable, the bridge dummy, are in most cases statistically significant and indicate a reduction in transport costs of 3%–7%. Table 3 Impact of distance and road quality with a binary impact variable Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . Sample (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 1800 km road distancea Excluding Dec to Marb Excluding Aug to Marb only 34 months beforec Bridge (binary) −0.029* (0.017) −0.049** (0.020) −0.058** (0.029) −0.072*** (0.024) Covariates no no no no Adj R2 0.831 0.827 0.841 0.824 Observations 11008 7392 3630 6374 Bridge (binary) −0.039** (0.018) −0.054** (0.021) −0.050* (0.029) −0.060** (0.024) Covariates Yes yes yes yes Adj R2 0.816 0.817 0.834 0.816 Observations 11008 7392 3630 6374 Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . Sample (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 1800 km road distancea Excluding Dec to Marb Excluding Aug to Marb only 34 months beforec Bridge (binary) −0.029* (0.017) −0.049** (0.020) −0.058** (0.029) −0.072*** (0.024) Covariates no no no no Adj R2 0.831 0.827 0.841 0.824 Observations 11008 7392 3630 6374 Bridge (binary) −0.039** (0.018) −0.054** (0.021) −0.050* (0.029) −0.060** (0.024) Covariates Yes yes yes yes Adj R2 0.816 0.817 0.834 0.816 Observations 11008 7392 3630 6374 All equations are estimated with OLS and include pass-through corrected trends, market pair and time fixed effects and source anddestination specific seasonality. Robust standard errors, clustered by group (market pair, before/after and intervention/control), arein brackets next to the coefficient. aObservations with road distances larger than 1800 km are omitted. bRestricted to marketing months April–November (2) and restricted to marketing months April–July (3). cThe period ‘without bridges’ is from October 2006, the closing of the Mutarara bridge for rehabilitation works, to July 2009, 34 months before 1 August 2009, the opening of the Caia bridge (which also coincided with the completion of the rehabilitation of the Mutarara bridge). ∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01. Open in new tab Table 3 Impact of distance and road quality with a binary impact variable Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . Sample (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 1800 km road distancea Excluding Dec to Marb Excluding Aug to Marb only 34 months beforec Bridge (binary) −0.029* (0.017) −0.049** (0.020) −0.058** (0.029) −0.072*** (0.024) Covariates no no no no Adj R2 0.831 0.827 0.841 0.824 Observations 11008 7392 3630 6374 Bridge (binary) −0.039** (0.018) −0.054** (0.021) −0.050* (0.029) −0.060** (0.024) Covariates Yes yes yes yes Adj R2 0.816 0.817 0.834 0.816 Observations 11008 7392 3630 6374 Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . Sample (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 1800 km road distancea Excluding Dec to Marb Excluding Aug to Marb only 34 months beforec Bridge (binary) −0.029* (0.017) −0.049** (0.020) −0.058** (0.029) −0.072*** (0.024) Covariates no no no no Adj R2 0.831 0.827 0.841 0.824 Observations 11008 7392 3630 6374 Bridge (binary) −0.039** (0.018) −0.054** (0.021) −0.050* (0.029) −0.060** (0.024) Covariates Yes yes yes yes Adj R2 0.816 0.817 0.834 0.816 Observations 11008 7392 3630 6374 All equations are estimated with OLS and include pass-through corrected trends, market pair and time fixed effects and source anddestination specific seasonality. Robust standard errors, clustered by group (market pair, before/after and intervention/control), arein brackets next to the coefficient. aObservations with road distances larger than 1800 km are omitted. bRestricted to marketing months April–November (2) and restricted to marketing months April–July (3). cThe period ‘without bridges’ is from October 2006, the closing of the Mutarara bridge for rehabilitation works, to July 2009, 34 months before 1 August 2009, the opening of the Caia bridge (which also coincided with the completion of the rehabilitation of the Mutarara bridge). ∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01. Open in new tab Details on the decisions taken in the PSM estimation (determinants and estimation of propensity score, performance of estimations, matching technique, robustness of matching algorithm, common support) are all reported in the Appendix. Selected estimation results of the PSM (KM) estimations are reported in Table 4. Table 4 Impact of Bridges: PSM, Kernel Matching Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . . Sample . (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 2000 km road distancea Excluding Dec to Marb Excluding Aug to Marb Only 34 months beforec Est. technique PSM/KM PSM/KM PSM/KM PSM/KM Bridge (ATT) −0.095** (0.043) −0.131** (0.051) −0.132** (0.063) −0.122** (0.051) Bridge (ATU) −0.108 −0.127 −0.116 −0.125 Bridge (ATE) −0.101 −0.129 −0.125 −0.123 On support: treated 1216 835 501 832 untreated 932 663 408 669 Off support treated 311 245 166 248 untreated 7053 4702 2953 3586 Observations 9512 6445 3210 5335 Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . . Sample . (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 2000 km road distancea Excluding Dec to Marb Excluding Aug to Marb Only 34 months beforec Est. technique PSM/KM PSM/KM PSM/KM PSM/KM Bridge (ATT) −0.095** (0.043) −0.131** (0.051) −0.132** (0.063) −0.122** (0.051) Bridge (ATU) −0.108 −0.127 −0.116 −0.125 Bridge (ATE) −0.101 −0.129 −0.125 −0.123 On support: treated 1216 835 501 832 untreated 932 663 408 669 Off support treated 311 245 166 248 untreated 7053 4702 2953 3586 Observations 9512 6445 3210 5335 Equations are estimated with propensity score matching (PSM). Estimates of the propensity scores are in the Table A5. Matchingalgorithm: Kernel Matching, Epanechnikov kernel and bandwidth 0.06. Standard errors are in brackets next to the coefficient. aObservations with road distances larger than 1800 km are omitted. bRestricted to marketing months April–November (2) and restricted to marketing months April–July (3). cThe period ‘without bridges’ is from October 2006, the closing of the Mutarara bridge for rehabilitation works, to July 2009, 34 months before 1 August 2009, the opening of the Caia bridge (which also coincided with the completion of the rehabilitation of the Mutarara bridge). ∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01. Open in new tab Table 4 Impact of Bridges: PSM, Kernel Matching Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . . Sample . (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 2000 km road distancea Excluding Dec to Marb Excluding Aug to Marb Only 34 months beforec Est. technique PSM/KM PSM/KM PSM/KM PSM/KM Bridge (ATT) −0.095** (0.043) −0.131** (0.051) −0.132** (0.063) −0.122** (0.051) Bridge (ATU) −0.108 −0.127 −0.116 −0.125 Bridge (ATE) −0.101 −0.129 −0.125 −0.123 On support: treated 1216 835 501 832 untreated 932 663 408 669 Off support treated 311 245 166 248 untreated 7053 4702 2953 3586 Observations 9512 6445 3210 5335 Dependent variable: ln[(pj – |$\hat{\rho_{jk}}$|pk)/|$\hat{\rho_{jk}}$|] . . Sample . (1) (2), as (1) (3), as (1) (4), as (2) Variables Within 2000 km road distancea Excluding Dec to Marb Excluding Aug to Marb Only 34 months beforec Est. technique PSM/KM PSM/KM PSM/KM PSM/KM Bridge (ATT) −0.095** (0.043) −0.131** (0.051) −0.132** (0.063) −0.122** (0.051) Bridge (ATU) −0.108 −0.127 −0.116 −0.125 Bridge (ATE) −0.101 −0.129 −0.125 −0.123 On support: treated 1216 835 501 832 untreated 932 663 408 669 Off support treated 311 245 166 248 untreated 7053 4702 2953 3586 Observations 9512 6445 3210 5335 Equations are estimated with propensity score matching (PSM). Estimates of the propensity scores are in the Table A5. Matchingalgorithm: Kernel Matching, Epanechnikov kernel and bandwidth 0.06. Standard errors are in brackets next to the coefficient. aObservations with road distances larger than 1800 km are omitted. bRestricted to marketing months April–November (2) and restricted to marketing months April–July (3). cThe period ‘without bridges’ is from October 2006, the closing of the Mutarara bridge for rehabilitation works, to July 2009, 34 months before 1 August 2009, the opening of the Caia bridge (which also coincided with the completion of the rehabilitation of the Mutarara bridge). ∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01. Open in new tab The PSM results generate a statistically significant average treatment effect on the treated (ATT) which ranges from −0.09 to −0.13. Estimations with Nearest Neighbour matching generate similar results (see Table A6) as with Kernel Matching and give confidence about the robustness of the matching procedure. PSM results further indicate an average treatment (ATE) that is slightly higher but still reasonably in line with OLS DiD specification with a binary impact variable. In view of the common support analysis (see Appendix) the slightly higher impact is not surprising. We conclude that the PSM results further confirm the impact of bridges on transport costs. 4.6 Strict identification of source and destination markets: omitting transit markets The empirical methodology—the Atkin Donaldson (2015) framework—rests on the assumption that prices of specific and identical products can be distinguished, both at origin and at destination. Despite our good intentions to rigorously identify source and destination markets of maize, it is obvious that the decisions we have taken do not connect prices at origin and prices at destination in the way this is implemented in the paper by Atkin and Donaldson (2015), who use meticulously specified product characteristics at barcode level. However, improvement is possible by restricting the estimations to those trade pairs that unambiguously connect source markets with destination markets, and by omitting transit markets. Hence, we drop observations of market pairs, where markets are operating both as destination for a set of markets and as source for a different set of markets (notably Chimoio as source, Tete and Nampula as destination). The estimation results with these sample restrictions confirm previous estimations (see Table A8). Accuracy of the road distance and road quality coefficients improves slightly, with coefficient sizes similar to previous estimations. The number of observations in these estimations drops substantially, indicating that these results may also be considered a more general support for the robustness of previous estimations. 5. Trade costs reduction by route and distribution of benefits 5.1 Trade cost reductions by route To what extent does adjusting for spatially varying mark-ups matter? We can calculate the difference in marginal impact of estimating with and without accounting for spatially varying mark-ups. In other words, we compare marginal cost of road distance and road quality under oligopolistic pricing of traders, with estimated pass-through rates less than 1 (|${\rho}_{od}<1$|), with these marginal costs under competitive pricing of traders, with pass-through rates by assumption equal to 1 (|${\rho}_{od}=1$|). The comparison of these estimations indicates that (absolute) marginal impacts, of both road distance and road quality are larger if accounted for spatially varying mark-ups (estimations available from the author). This outcome corresponds with the non-parametric estimations reported in Atkin and Donaldson (2015). Having established the robustness of the impact estimation in the previous section, we now return to the interpretation of the key estimation results in Table 1. The logarithmic transformation allows an interpretation of the estimated coefficients in terms of elasticities: a 10% reduction in road distance leads to a 1.8%–3.2% reduction in transport costs and a 10% improvement in road quality leads to a 4.0%–11.2% reduction in transport costs. It should be noted that changes in road distance are—in our case—large while changes in road quality are usually modest: the median reduction in road distance is close to 20%, with a maximum of close to 60%, while the median improvement of road quality is around 3%. Estimated elasticities offer insight into the impact of bridges on transport costs, averaged over routes, and how the reduction in transport costs can be attributed to the key determinants of transport costs. Although interesting and useful20, these averages are less informative about realised cost reductions of particular routes that benefit from the new bridge. With the estimated elasticities for road distance and road quality, we are now in the position to quantify benefits by route. For a selection of routes the reduction in trade costs is shown in Table 5. Table 5 Reduction in Trade Costs by Route Due to Bridge . Road distance . Transport cost . Road quality . Transport cost . . . . Source – destination %Δ %Δ %Δ %Δ Total Road distance Road quality Alto Molocue Maputo −16.2% −4.9% 1.4% −1.1% −5.9% 82.2% 17.8% Alto Molocue Beira −40.9% −12.3% −0.6% 0.5% −11.8% (≈100%) Mocuba Maputo −17.7% −5.3% 2.4% −1.8% −7.1% 74.7% 25.3% Ribaue Beira −37.4% −11.2% −0.1% 0.1% −11.1% (≈100%) Nampula Maputo −14.8% −4.4% 3.2% −2.4% −6.8% 64.9% 35.1% Nampula Beira −35.2% −10.6% 1.0% −0.8% −11.3% 93.4% 6.6% Chimoio Nampula −18.5% −5.6% 3.4% −2.6% −8.1% 68.5% 31.5% Gorongosa Nacala −26.8% −8.0% 17.5% −13.1% −21.2% 38.0% 62.0% Manica Nampula −14.5% −4.4% 3.2% −2.4% −6.8% 64.4% 35.6% . Road distance . Transport cost . Road quality . Transport cost . . . . Source – destination %Δ %Δ %Δ %Δ Total Road distance Road quality Alto Molocue Maputo −16.2% −4.9% 1.4% −1.1% −5.9% 82.2% 17.8% Alto Molocue Beira −40.9% −12.3% −0.6% 0.5% −11.8% (≈100%) Mocuba Maputo −17.7% −5.3% 2.4% −1.8% −7.1% 74.7% 25.3% Ribaue Beira −37.4% −11.2% −0.1% 0.1% −11.1% (≈100%) Nampula Maputo −14.8% −4.4% 3.2% −2.4% −6.8% 64.9% 35.1% Nampula Beira −35.2% −10.6% 1.0% −0.8% −11.3% 93.4% 6.6% Chimoio Nampula −18.5% −5.6% 3.4% −2.6% −8.1% 68.5% 31.5% Gorongosa Nacala −26.8% −8.0% 17.5% −13.1% −21.2% 38.0% 62.0% Manica Nampula −14.5% −4.4% 3.2% −2.4% −6.8% 64.4% 35.6% Note: elasticity of transport costs for road distance: +0.30; elasticity of transport costs for road quality: −0.75. Open in new tab Table 5 Reduction in Trade Costs by Route Due to Bridge . Road distance . Transport cost . Road quality . Transport cost . . . . Source – destination %Δ %Δ %Δ %Δ Total Road distance Road quality Alto Molocue Maputo −16.2% −4.9% 1.4% −1.1% −5.9% 82.2% 17.8% Alto Molocue Beira −40.9% −12.3% −0.6% 0.5% −11.8% (≈100%) Mocuba Maputo −17.7% −5.3% 2.4% −1.8% −7.1% 74.7% 25.3% Ribaue Beira −37.4% −11.2% −0.1% 0.1% −11.1% (≈100%) Nampula Maputo −14.8% −4.4% 3.2% −2.4% −6.8% 64.9% 35.1% Nampula Beira −35.2% −10.6% 1.0% −0.8% −11.3% 93.4% 6.6% Chimoio Nampula −18.5% −5.6% 3.4% −2.6% −8.1% 68.5% 31.5% Gorongosa Nacala −26.8% −8.0% 17.5% −13.1% −21.2% 38.0% 62.0% Manica Nampula −14.5% −4.4% 3.2% −2.4% −6.8% 64.4% 35.6% . Road distance . Transport cost . Road quality . Transport cost . . . . Source – destination %Δ %Δ %Δ %Δ Total Road distance Road quality Alto Molocue Maputo −16.2% −4.9% 1.4% −1.1% −5.9% 82.2% 17.8% Alto Molocue Beira −40.9% −12.3% −0.6% 0.5% −11.8% (≈100%) Mocuba Maputo −17.7% −5.3% 2.4% −1.8% −7.1% 74.7% 25.3% Ribaue Beira −37.4% −11.2% −0.1% 0.1% −11.1% (≈100%) Nampula Maputo −14.8% −4.4% 3.2% −2.4% −6.8% 64.9% 35.1% Nampula Beira −35.2% −10.6% 1.0% −0.8% −11.3% 93.4% 6.6% Chimoio Nampula −18.5% −5.6% 3.4% −2.6% −8.1% 68.5% 31.5% Gorongosa Nacala −26.8% −8.0% 17.5% −13.1% −21.2% 38.0% 62.0% Manica Nampula −14.5% −4.4% 3.2% −2.4% −6.8% 64.4% 35.6% Note: elasticity of transport costs for road distance: +0.30; elasticity of transport costs for road quality: −0.75. Open in new tab Total reduction in transport costs due to the bridges ranges from 6% to 21%. In most instances the cost reduction is mainly due to the shorter distance: with a few exceptions (in particular routes to Nacala) change in quality mostly contributes only modestly to transport cost reduction. Overall, roughly two-third of the cost reduction is on account of road distance and one-third on account of road quality. To measure the benefit of the bridge for the total Mozambique maize market, we need to weigh reductions in transport cost per route with the size of freight transported through these routes. Trade flow data are, however, not available. 5.2 Who captures the benefits from reduced transport costs? Allowing for and modelling oligopolistic traders and estimating pass-through rates beg for a welfare analysis. However, without data on trade flows a fully fledged welfare analysis is difficult. In order to measure the distribution of surplus between traders and consumers, we follow, in the first instance, the technique proposed by Atkin and Donaldson (2015). On the basis of several assumptions Atkin and Donaldson (2015) derive the following expression for the ratio of trader surplus and consumer surplus: $$ \begin{equation} {TS}_{od}/{CS}_{od}=1/{\rho}_{od}+\left(1-{\phi}_d\right)/{\phi}_d, \end{equation}$$(15) where |${\rho}_{od}$| is the pass-through rate and |${\phi}_d$| the competitiveness index prevailing in the destination market. Hence, the share of surplus of traders relative to consumers is a simple function of the pass-through rate (|${\rho}_{od}$|) and the competitiveness index (|${\phi}_{od}$|). Estimation of the equilibrium pass-through rate (|$\hat{\rho_{od}}$|) is set out in Section 3 (equation (11)), and estimation results are discussed in Section 4. To find an estimate for the competitiveness index (|${\phi}_{od}$|), we use the following expression (considerably simplified, as we have only one product in the current exercise, as opposed to k products in the paper by Atkin and Donaldson (2015)): $$ \begin{equation} {\Xi}_{od}={\gamma}_0+{\gamma}_{od}+{\varepsilon}_{od}, \end{equation}$$(16) where |${\Xi}_{od}\equiv \mathit{\ln}\Big(\frac{1}{\hat{\rho_{od}}}-1\Big)$|, |${\gamma}_{od}$| is a destination fixed effect and |${\varepsilon}_{od}$| is an error term. A consistent estimate of the competitiveness index |$\hat{\phi_{od}}$| in destination d is calculated as |$\hat{\phi_{od}}={e}^{\hat{-{\gamma}_{od}}}$|. The results of these calculations, the share of trader surplus vis-à-vis consumer surplus, correlated with road distance, are shown in Figure 2. The calculated share of trader versus consumer surplus varies from a low of 0.1 to values close to 6. It should be noted, however, that we expect trade-flows to be larger (smaller) for shorter (longer) road distances: although the bias in benefits towards traders becomes extremely large with larger road distances, the underlying trade flows are likely to be thin. Conversely, it is remarkable that below 1 shares of trader versus consumer surplus occur with a road distance of less than 1000 km. The evidence supports a positive correlation between this share and road distance. On average traders’ surplus is 1 to 3 times consumer surplus for common trading distances, with a wide spread. Figure 2 Open in new tabDownload slide Trader Surplus Vis-à-Vis Consumer Surplus, by Road Distance. Figure 2 Open in new tabDownload slide Trader Surplus Vis-à-Vis Consumer Surplus, by Road Distance. How do these numbers compare with the corresponding numbers for Ethiopia and Nigeria21 in the Atkin and Donaldson (2015) study? Trader versus consumer surplus increases with road distance in all these countries. However, the Ethiopia and Nigeria shares start at levels 1.8 and 2.0, at a road distance of 100 km and increase to 2.1 and 2.4, respectively, at a road distance of 500 km. The Mozambique shares start much lower, at an average level of 1.1 at a road distance of 100 km, to increase to around 1.4 at a road distance of 500 km. Apparently, Ethiopian and Nigerian traders have more market power than Mozambique traders. Also, the Mozambique data extend to a road distance of 2500 km, while the Ethiopia and Nigeria data do not extend beyond 500 km. To assess the distribution of benefits, there is a less complicated, but more intuitive alternative, which exploits the identification strategy adopted in the current study. Maize trade is multilateral: each source market supplies different destination markets, and each destination market is served by several source markets. The impact of the bridge on the price level in a specific market is, consequently, a mixture of different markets. As a result it is complicated to estimate for all markets jointly if benefits are biased towards source markets or destination markets. However, if we assume that a few source markets and a few destination markets dominate in maize trade, while prices in other markets follow, we can restrict estimations to these dominant trade pairs. Selecting market pairs on the basis of the size of trade flows is, however, not possible without data on trade flows. Instead we have selected market pairs that benefit from the bridge on the basis of the largest number of available transport cost observations (see Table A2). Underlying this—admittedly crude approximation—is the assumption that trade-flows are correlated with the number of trade cost observations. We propose that price levels can adequately be described by trends and within and between seasonality, both specified by source or destination. We estimate the following equation: $$ {p}_{jt}={\eta}_0+{\eta}_1\ {bridge}_t\ x\ {market}_{ojt}+{\eta}_2\ {bridge}_t\ x\ {market}_{djt} $$ $$ +{season}_{ojt}+{season}_{djt}+{\varphi}_{ojm}+{\chi}_{djm}+{trend}_{ojt}+{trend}_{djt} $$ $$ \begin{equation} +{\psi}_t+{\zeta}_j+{\varepsilon}_{jt}, \end{equation}$$(17) where subscript o and d refer respectively to source markets (origin markets) and destination markets. As we use, in each estimation, observations of source and destination that potentially benefits from the Caia–Chimuara bridge, we cannot include time fixed effects: this would absorb all impact. To control to some extent for common variation over time—other than through the bridge—we have included year fixed effects (|${\psi}_t$|)22. The coefficients of interest are |${\eta}_1$| and |${\eta}_2$|: with prices converted to log(prices), these coefficients reflect elasticities and we expect |${\eta}_1>{\eta}_2$|, and (|${\eta}_2-{\eta}_1$|) preferably somewhere in the range of the impact estimates on spatial price differences. Estimation results—reported in Table 6—support negative impacts, both for source and destination markets. Most coefficients for destination markets are smaller (more negative) than those for source markets. Hence, prices in both source and destination markets decrease, but they decrease more in destination markets. In many cases (although not all) tests confirm that coefficients of source and destination markets are different. The size of the difference is in many cases on the high side, but not completely out of range. These results point at a bias of reduced transport costs towards destination markets: prices in destination markets reduce stronger than those in source markets. Table 6 Impact on Price Levels, Estimated Pairwise Dependent variable: ln(pj) . Source – destination . Bridge × source . Bridge × destination . R2 . N . p . Nampula – Maxixe −0.160*** (0.050) −0.398*** (0.135) 0.997 227 0.097 Alto Molocue – Maputo −0.174*** (0.059) −0.109*** (0.038) 0.998 215 0.353 Nampula – Maputo −0.162*** (0.050) −0.099** (0.040) 0.998 235 0.309 Alto Molocue – Maxixe −0.168*** (0.061) −0.428*** (0.138) 0.997 206 0.085 Ribaue – Beira −0.141*** (0.043) −0.530*** (0.237) 0.995 188 0.098 Alto Molocue – Beira −0.172*** (0.060) −0.509*** (0.187) 0.995 183 0.073 Nampula – Beira −0.150*** (0.053) −0.469** (0.218) 0.995 206 0.142 Mocuba – Maxixe −0.233*** (0.069) −0.396*** (0.138) 0.996 203 0.291 Gorongosa – Nacala −0.117 (0.146) −0.194** (0.074) 0.995 211 0.633 Dependent variable: ln(pj) . Source – destination . Bridge × source . Bridge × destination . R2 . N . p . Nampula – Maxixe −0.160*** (0.050) −0.398*** (0.135) 0.997 227 0.097 Alto Molocue – Maputo −0.174*** (0.059) −0.109*** (0.038) 0.998 215 0.353 Nampula – Maputo −0.162*** (0.050) −0.099** (0.040) 0.998 235 0.309 Alto Molocue – Maxixe −0.168*** (0.061) −0.428*** (0.138) 0.997 206 0.085 Ribaue – Beira −0.141*** (0.043) −0.530*** (0.237) 0.995 188 0.098 Alto Molocue – Beira −0.172*** (0.060) −0.509*** (0.187) 0.995 183 0.073 Nampula – Beira −0.150*** (0.053) −0.469** (0.218) 0.995 206 0.142 Mocuba – Maxixe −0.233*** (0.069) −0.396*** (0.138) 0.996 203 0.291 Gorongosa – Nacala −0.117 (0.146) −0.194** (0.074) 0.995 211 0.633 All equations are estimated with OLS and include trends, and within and between year seasonality, both by market, and market andyear fixed effects. Robust standard errors are in brackets next to the coefficient. p-values for F-test: coef(bridge x source) =coef(bridge x destination). ∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01. Open in new tab Table 6 Impact on Price Levels, Estimated Pairwise Dependent variable: ln(pj) . Source – destination . Bridge × source . Bridge × destination . R2 . N . p . Nampula – Maxixe −0.160*** (0.050) −0.398*** (0.135) 0.997 227 0.097 Alto Molocue – Maputo −0.174*** (0.059) −0.109*** (0.038) 0.998 215 0.353 Nampula – Maputo −0.162*** (0.050) −0.099** (0.040) 0.998 235 0.309 Alto Molocue – Maxixe −0.168*** (0.061) −0.428*** (0.138) 0.997 206 0.085 Ribaue – Beira −0.141*** (0.043) −0.530*** (0.237) 0.995 188 0.098 Alto Molocue – Beira −0.172*** (0.060) −0.509*** (0.187) 0.995 183 0.073 Nampula – Beira −0.150*** (0.053) −0.469** (0.218) 0.995 206 0.142 Mocuba – Maxixe −0.233*** (0.069) −0.396*** (0.138) 0.996 203 0.291 Gorongosa – Nacala −0.117 (0.146) −0.194** (0.074) 0.995 211 0.633 Dependent variable: ln(pj) . Source – destination . Bridge × source . Bridge × destination . R2 . N . p . Nampula – Maxixe −0.160*** (0.050) −0.398*** (0.135) 0.997 227 0.097 Alto Molocue – Maputo −0.174*** (0.059) −0.109*** (0.038) 0.998 215 0.353 Nampula – Maputo −0.162*** (0.050) −0.099** (0.040) 0.998 235 0.309 Alto Molocue – Maxixe −0.168*** (0.061) −0.428*** (0.138) 0.997 206 0.085 Ribaue – Beira −0.141*** (0.043) −0.530*** (0.237) 0.995 188 0.098 Alto Molocue – Beira −0.172*** (0.060) −0.509*** (0.187) 0.995 183 0.073 Nampula – Beira −0.150*** (0.053) −0.469** (0.218) 0.995 206 0.142 Mocuba – Maxixe −0.233*** (0.069) −0.396*** (0.138) 0.996 203 0.291 Gorongosa – Nacala −0.117 (0.146) −0.194** (0.074) 0.995 211 0.633 All equations are estimated with OLS and include trends, and within and between year seasonality, both by market, and market andyear fixed effects. Robust standard errors are in brackets next to the coefficient. p-values for F-test: coef(bridge x source) =coef(bridge x destination). ∗p < 0.10; ∗∗p < 0.05; ∗∗∗p < 0.01. Open in new tab 6. Conclusion and discussion In this study we have investigated transport cost reduction through infrastructure investment. We have used spatial maize prices to estimate to what extent infrastructure investment leads to transport cost reductions and attribute these reductions to road distance and road quality. We also quantify who captures the benefits of these cost reductions. The applied methodology allows for potentially oligopolistic traders with spatially varying mark-ups, borrowed from Atkin and Donaldson (2015). For identification we exploited the introduction of a new bridge over the Zambezi River, between Caia and Chimuara, in the centre of Mozambique. This event generates the required variation in trading distances between markets, needed to attribute impact to road distance and road quality. The empirical estimations are based on monthly maize prices, for 22 major markets, of which 12 are typical source markets and 10 typical destination markets, for a period stretching from 5 years before to 5 years after the opening of the bridge. Additionally, a limited set of transport cost data is helpful to determine the direction of trade, to identify source and destination markets and to verify the spatial prices based estimation results. The key finding is that, averaged over routes, the bridge has caused a 3%–7% reduction in transport costs. For specific routes this reduction in costs is as large as 21%. Roughly two-third of the cost reduction is due to the reduction in road distance. The evidence points to a bias in the benefits of reduced trade costs towards traders and consumers. On average benefits of trade cost reductions are equally shared between traders and consumers for short distances, but for larger distances a larger part accrues to traders. The evidence also indicates a reduction in prices in destination markets due to the bridge. Results are robust for inclusion of covariates, non-random bridge placement and strict origin–destination rules, and are supported by observed transport cost data. Supplementary material Supplementary material is available at Journal of African Economies online. Footnotes 1 A map of Mozambique and figures on geographical characteristics (population, maize production, rainfall), maize prices and seasonality that underscore the information in this background section are in the Appendix (Figures A1–A5). 2 Staples in Mozambique are maize, rice, cassava, wheat, sorghum, millet, sweet potatoes beans and groundnuts. 3 Data on maize production are available by province. Further geographically disaggregated data are unfortunately not available. 4 See Appendix for a map and more details on the Zambezi river. 5 Whatever is the case, the ‘closing for rehabilitation date’ suggests a period of being closed to traffic of 34 months before the re-opening in August 2009. The coincidental correspondence in timing (of the introduction of a road bridge between Caia and Chimuara, and the completion of the rehabilitation of the railway bridge between Vila de Sena and Mutarara) allows to mark a period in which complete absence of these Zambezi bridges overlaps, creating the required variation in shortest trading routes by road between markets. 6 The 2009 toll for the new bridge is the same as motorists had to pay for using the ferry—800 meticais (in 2017 equivalent to around 30 US dollars) for trucks and 80 meticais for light vehicles. 7 Two-way trade observations in the trade cost data used for the current study are an exception: out of the 95 trade pairs for which (independently recorded) trade cost data are available, 77 market pairs exclusively have one-way trade, while only 9 pairs have two-way trade. Only one case of two-way trade—between Tete and Chimoio—is substantial in number of observations, in both directions: all other cases of two-way trade concern either less than 2 observations per market pair (out of a total of 832) or a clear domination in one direction. Even the Tete–Chimoio trade is one-way for a number of years, suggesting differences in production by season (see also Appendix, Figure A6 in the Appendix for a schematic overview of multilateral trade). 8 SIMA, which started in the 1990s as a USAID/Michigan State University funded initiative, is responsible for collection and distribution of price information on agricultural commodities and distributes weekly price bulletins by email (covering among others farmer organisations, traders), by SIMA’s provincial offices (that further reproduce and distribute this information locally), through the Ministry of Commerce that uses the information in their own bulletins and through regular broadcasts on the national radio and television news (to whom SIMA contractually offers weekly input to market programs). Traders interviews support the effectiveness of the SIMA price information (see ‘In Mozambique, Market Information publishes its 500th weekly bulletin, a Cause for Celebration’, February 2006 posted on the internet (www.masa.gov.mz/sima/). The rollout of the mobile phone infrastructure that started in the 1990s has further improved the dissemination of price information (see Zant, 2019). 9 Quadro 3, Preço e Mudança Percentual a Nível de Mercado Retalhista (MT/kg), Grão de Milho Branco (Table 3, prices and percentage price changes in retail markets (meticais per kg), white maize grain) 10 Alto Molocue, Angoche, Angonia, Beira, Chimoio, Chokwe, Cuamba, Gorongosa, Lichinga, Manica, Maputo, Massinga, Maxixe, Milange, Mocuba, Monapo, Montepuez, Mutarara, Nacala, Nampula, Nhamatanda, Pemba, Quelimane, Ribaue, Tete, Vilanculos en Xai-Xai. Figure A1 shows the locations of these markets in Mozambique. 11 Missing observations are common in agricultural prices series and reflect lack of supply and corresponding absence of transactions. Missing observations are therefore not correlated with the presence of the Caia–Chimuara Zambezi bridge. The share of missing observations is also relatively small: after dropping a few markets of the original data set (Angoche, Lichinga, Milange, Monapo and Vilanculos: no or very few observations) we have around 96% of the potential number of monthly price observations (2533/(22 markets × 12 months × 10 years)). 12 Price data for n markets in a specific month, yields n2 market pair data, of which (n2–n)/2 are economically relevant, as we ignore two-way trade; hence, with 22 markets 1 month without missing observations yields observations for 231 market pairs ((222–22)/2). Note that we restrict the sample of data for estimations to price differences that connect source markets with destination markets (see also Appendix). 13 A number of the shortest transport routes run through Malawi. Estimations that allow for an effect of border-crossing did not generate significant results. Hence, we conclude that costs of crossing the border with Malawi are negligible and can be ignored. 14 Without restrictions, hence, all road distances and all months. 15 We distinguish 12 source and 10 destination markets, leading to 120 regressions, assuming that pass-through rates are constant before and after the opening of the bridge. Further work is needed to relax this assumption. Also various checks on the stability (varying specifications of the pass-through equation) awaits future work. 16 This cut-off distance is arbitrary, but alternative cut-off distances around 1800 generate similar results. 17 Maize imports are subject to a 2.5% import tariff and a 17% Value Added Tax, which is not levied on domestic production (see Zavale, 2014). These import duties, however, do not offset the price difference with Maputo. 18 See Appendix, Figures A7 and A8 for the development of domestic maize prices vis-à-vis SAFEX white maize spot and Lilongwe average maize prices. For Mozambique exports the relevant comparison is domestic post-harvest prices (April–July) with foreign prices (Malawi), while for Mozambique imports this comparison is domestic lean season prices (January–March) with foreign prices (South Africa). 19 The exclusion restriction says that the excluded exogenous variables—the instruments—are correlated with the change in infrastructure, but only affect transport costs through this channel. To meet this requirement the spatial economics literature has proposed the so-called planned route IV, the historical route IV and the inconsequential place approach (Redding and Turner, 2014). 20 To measure the total benefit of the bridge for the Mozambique maize market, we ideally need to weigh reductions in transport cost per route with the size of freight on transported through these routes. Unfortunately, trade flow data are not available. 21 We ignore the evidence of the US reported in the study by Atkin and Donaldson (2015). 22 Using year fixed effects (rather than month fixed effects) also preserves statistical power. 23 In this context it is suggestive that transport of coal from the Moatize fields near Tete, and close to the Zambezi, to the coast is implemented by rail rather than through the Zambezi river. 24 Instruments have to satisfy the exclusion restriction, meaning that the excluded exogenous variables—the instrument—is correlated with the change in infrastructure, but only affects transport costs through this channel. 25 Probit or logit are likely to give similar outcomes. However, the logit distribution has more density mass in the bounds and this corresponds with our empirical setting (see also Caliendo et al., 2005) 26 In selecting variables for the propensity score estimation, we aimed at maintaining the maximum number of observations (to improve power) and focusing on spatial and climate variables to guarantee exogeneity. 27 This bandwidth value is the default value in the STATA routine psmatch2 (E. Leuven and B. Sianesi, 2003, ‘PSMATCH2: Stata module to perform full Mahalanobis and propensity score matching, common support graphing, and covariate imbalance testing’.) 28 |$B=\frac{\big({\overline{X}}_1-{\overline{X}}_0\big)}{\sqrt{\big({V}_1(X)+{V}_0(X)\big)/2}}$| where |${\overline{X}}_1$| (|${\overline{X}}_0$|) and |${V}_1(X)$| (|${V}_0(X)$|) are, respectively, the average and variance of covariate X in the treatment (control) group. The standardised bias, B, is calculated before and after matching, for each covariate X. 29 From the trade cost data we extracted that a few markets are both destination market for a certain group of source markets, and source market for a number of terminal markets (see also Appendix, Figure A6 in the Appendix for a schematic overview of multilateral trade). A clear example is Nampula which is a destination for Alto Molocue, Mocuba and Ribaue, and a source for Maputo, Beira, Maxixe and Xai-Xai. Tete and Chimoio are similar in this respect. Such transit markets are neither a genuine origin nor a genuine destination market—a requirement from the theory—we investigated potential bias in impact estimates if the sample is adjusted for these observations (see Section 4). 30 Destination markets are: Beira, Maputo, Massinga, Maxixe, Nacala, Nampula, Pemba, Quelimane, Tete, Xai-Xai; Source markets are: Angonia, Alto Molocue, Cuamba, Chimoio, Chokwe, Gorongosa, Manica, Mocuba, Montepuez, Mutarara, Nhamatanda, Ribaue. 31 At the same time, it is fair to add that identification of source and destination in so-called barcode level data (see Atkin and Donaldson, 2015; Broda and Weinstein, 2008), where the production factory location of specific domestically manufactured goods and the port of entrance of specific imported goods are recorded, is without doubt more accurate than the approach followed in this study. References Abdula D.C. ( 2005 ) ‘Improving maize marketing and trade policies to promote household food security in Southern Mozambique’, Thesis submitted for the degree of Master of Science, Department of Agricultural Economics, Michigan State University . Atkin D ., Donaldson D. ( 2015 ) ‘ Who’s Getting Globalized? The Size and Implications of Intra-national Trade Costs ’, National Bureau of Economic Research, w21439 . Autor D. H. ( 2003 ) ‘ Outsourcing at Will: The Contribution of Unjust Dismissal Doctrine to the Growth of Employment Outsourcing ’, Journal of Labor Economics , 2003 , 21 : 1 . Google Scholar OpenURL Placeholder Text WorldCat Banerjee A. , Duflo E., Qian N. ( 2012 ) ‘ On the Road: Access to Transportation Infrastructure and Economic Growth in China ’, NBER working paper ’, 17897 . Bertrand M. , Duflo E., Mullainathan S. ( 2004 ) ‘ How Much Should We Trust Difference-in-Differences Estimates? ’, Quarterly Journal of Economics , 119 , 1 : 249 – 275 . Google Scholar Crossref Search ADS WorldCat Broda C.M. , Weinstein D.E. ( 2008 ) ‘ Understanding International Price Differences Using Barcode Data ’, NBER Working Paper 14017 . Brooks W. , Donovan K. ( 2020 ) ‘ Eliminating Uncertainty in Market Access: The Impact of New Bridges in Rural Nicaragua ’, Econometrica , 88 , 5 : 1965 – 1997 . Casaburi L. , Glennerster R., T. Suri ( 2013 ) ‘ Rural Roads and Intermediated Trade: Regression Discontinuity Evidence from Sierra Leone’ , mimeo . Cirera X. , Arndt C. ( 2008 ) ‘ Measuring the Impact of Road Rehabilitation on Spatial Market Efficiency in Maize Markets in Mozambique ’, Agricultural Economics , 39 : 17 – 28 . Google Scholar Crossref Search ADS WorldCat De Janvry A. , Sadoulet E. ( 2006 ) ‘Progress in the Modelling of Rural Households’ Behaviour under Market Failures’, Chapter 8 in De Janvry, A. and R. Kanbur , 2006 . Dillon B. , Dambro C. ( 2017 ) ‘ How Competitive are Crop Markets in Sub-Saharan Africa? , American Journal of Agricultural Economics , 99 , 5 : 1344 – 1361 . Google Scholar Crossref Search ADS PubMed WorldCat Dominguez-Torres C. , Briceño-Garmendia C. ( 2011 ) ‘Mozambique’s Infrastructure: A Continental Perspective’, AICD report, World Bank, Washington . Donaldson D. ( 2018 ) ‘ Railroads of the Raj: Estimating the Impact of Transportation Infrastructure ’, American Economic Review, 108 , 4–5 : 899 – 934 . Google Scholar OpenURL Placeholder Text WorldCat Fafchamps M. , Minten B. ( 2012 ) ‘ Impact of SMS-Based Agricultural Information on Indian Farmers ’, World Bank Economic Review, 27 , February : 1 – 32 . Google Scholar OpenURL Placeholder Text WorldCat Jacoby H. G. ( 2000 ) ‘ Access to Markets and the Benefits of Rural Roads ’, The Economic Journal , 110 : 713 – 37 . Google Scholar Crossref Search ADS WorldCat Jacoby H.G. , Minten B. ( 2009 ) ‘ On Measuring the Benefits of Lower Transport Costs ’, Journal of Development Economics , 89 , 1 : 28 – 38 . Google Scholar Crossref Search ADS WorldCat Jensen R. ( 2007 ) ‘ The Digital Provide: Information (Technology), Market Performance, and Welfare in the South Indian Fisheries Sector ’, Quarterly Journal of Economics , 72 , 3 : 879 – 924 . Google Scholar Crossref Search ADS WorldCat Kaminski J. , Christiaensen L., Gilbert C. L. ( 2016 ) ‘ Seasonality in Local Food Prices and Consumption: Evidence From Tanzania ’, Oxford Economic Papers , 68 , 3 : 736 – 757 . Google Scholar Crossref Search ADS WorldCat Key N. , Sadoulet E., de Janvry A. ( 2000 ) ‘ Transactions Costs and Agricultural Household Supply Response ’, American Journal of Agricultural Economics , 82 , May : 245 – 259 . Google Scholar OpenURL Placeholder Text WorldCat Minten B. , Kyle S. ( 1999 ) ‘ The Effect of Distance and Road Quality on Food Collection, Marketing Margins, and Traders’ Wages: Evidence From the Former Zaire ’, Journal of Development Economics , 60 : 467 – 95 . Google Scholar Crossref Search ADS WorldCat Omamo S. W. ( 1998 ) ‘ Transport Costs and Smallholder Cropping Choices: An Application to Siaya District, Kenya ’, American Journal of Agricultural Economics , 80 : 116 – 23 . Google Scholar Crossref Search ADS WorldCat Redding S. J. , Turner M. A ( 2014 ) ‘ Transportation Costs and the Spatial Organization of Economic Activity ’, National Bureau of Economic Research Working Paper , 20235 . Renkow M. , Hallstrom D.G., Karanja D. D. ( 2004 ) Rural Infrastructure, Transactions Costs and Market Participation in Kenya , Journal of Development Economics , 73 , 1 : 2004 , 349 – 367 . Google Scholar Crossref Search ADS WorldCat Rosenbaum P.R. , Rubin D. B. ( 1985 ) ‘ Constructing a Control Group Using Multivariate Matched Sampling Methods That Incorporate the Propensity Score ’, The American Statistician , 39 , 1 : 33 – 38 . Google Scholar OpenURL Placeholder Text WorldCat Tostão E. , Brorsen B. W. ( 2005 ) ‘ Spatial Price Efficiency in Mozambiques’s Post-Reform Maize Markets ’, Agricultural Economics , 33 : 205 – 14 . Google Scholar Crossref Search ADS WorldCat Tschirley D. , Abdula D., Weber , M. T. ( 2006 ) ‘ Toward Improved Maize Marketing and Trade Policies to Promote Household Food Security in Central and Southern Mozambique ’, Research Report 60E, Ministry of Agriculture, Mozambique . De Vletter F , Polana E. ( 2001 ) ‘ Female Itinerant Maize Traders in Southern Mozambique: A Study of a Higher-End Informal Sector Activity and Its Potential for Poverty Reduction ’, ILO-SAMAT Discussion Paper 17 . Volpe Martincus C. , Carballo J., Cusolito A. ( 2017 ) ‘ Roads, Exports and Employment: Evidence From a Developing Country ’, Journal of Development Economics , 125 : 21 – 39 . Google Scholar Crossref Search ADS WorldCat Zant W. ( 2013 ) ‘ How Is the Liberalization of Food Markets Progressing? Market Integration and Transaction Costs in Subsistence Economies ’, World Bank Economic Review , 27, 1 : 28 – 54 . doi: 10.1093/wber/lhs017 . Google Scholar Crossref Search ADS WorldCat Zant W. ( 2018 ) ‘ Trains, Trade and Transaction Costs: How Does Domestic Trade by Rail affect Market Prices of Malawi Agricultural Commodities? ’, World Bank Economic Review , 32 , 2 : 334 – 356 . Google Scholar Crossref Search ADS WorldCat Zant W. ( 2019 ) ‘ Mobile Phones and Mozambique Traders: What is the Size of Reduced Search Costs and Who Benefits ?’ Tinbergen Discussion Paper 19-047/V, Tinbergen Institute, Amsterdam . Zavale H. ( 2014 ) ‘ Analysis of Price Incentives and Disincentives for Maize in the Republic of Mozambique for the Time Period 2005–2013 ’, MAFAP, FAO, Rome . © The Author(s) 2021. Published by Oxford University Press on behalf of the Centre for the Study of African Economies. This is an Open Access article distributed under the terms of the Creative Commons Attribution NonCommercial-NoDerivs licence (http://creativecommons.org/licenses/by-nc-nd/4.0/), which permits non-commercial reproduction and distribution of the work, in any medium, provided the original work is not altered or transformed in any way, and that the work is properly cited. For commercial re-use, please contact journals.permissions@oup.com
Journal of African Economies – Oxford University Press
Published: Aug 18, 2021
Read and print from thousands of top scholarly journals.
Already have an account? Log in
Bookmark this article. You can see your Bookmarks on your DeepDyve Library.
To save an article, log in first, or sign up for a DeepDyve account if you don’t already have one.
Copy and paste the desired citation format or use the link below to download a file formatted for EndNote
All DeepDyve websites use cookies to improve your online experience. They were placed on your computer when you launched this website. You can change your cookie settings through your browser.