Trade and transport costs: evidence from hurricane Sandy · 2016. 8. 3. · hurricane Sandy, a...
50
Trade and transport costs: evidence from hurricane Sandy Pierre Cotterlaz * and Emanuele Mazzini †‡ July 29, 2016 Preliminary version. Please do not cite. This paper may change in the near future. Abstract We use the massive disruptions on the Northeastern US road network provoked by hurricane Sandy at the end of October 2012 as a natural experiment to identify the effect of transport costs on trade flows. This allows us to isolate the effect of transport costs from the effect of other potential trade costs. In this way we are able to provide interesting insights towards the comprehension of the mechanisms underlying the strong negative effect of distance on trade flows. Indeed, existing evidence suggests that the distance elasticity of trade flows cannot be entirely explained by transport costs. Other elements, which broadly go under the name of “dark trade costs” matter as well. In our baseline specification, we find that the part of the distance elasticity of trade flows that can be attributed to transport costs is -0.06, lower than the total distance elasticity of trade flows (-0.84). We interpret this as evidence that, even within a country, transport costs are not the only component of trade costs. Moreover, in order to recover the change in transport costs that Sandy caused in the affected areas, we implement a method inspired from the indirect inference estimator, i.e we estimate the value of the change in transport costs for which the change in origin and destination fixed effects estimated from simulated data matches the one estimated from real data. * Department of Economics and LIEPP, Sciences Po, [email protected]. † Department of Economics, Sciences Po, [email protected] ‡ This work is supported by a public grant overseen by the French National Research Agency (ANR) as part of the “Investissements d’Avenir” program LIEPP (reference: ANR-11-LABX-0091, ANR-11-IDEX-0005-02). We are indebted to our supervisor Thierry Mayer for continued advice and support and to Florian Oswald for his precious suggestions. We thank seminar participants at Sciences Po for many useful suggestions, as well as Donna Miller (DelDOT), Ulysses Jacks (MassDOT), Meg Frampton (NJDOT), Mary Shevlin (NYDOT), Sophia Spencer (NCDOT), Lindsey Pitten (PA GOV), Chris Kinsey (WVDOH) and Jim Lambert (WV Division of Highways) for providing useful data about Sandy related disruptions on the road network. All remaining errors are ours. 1
Transcript of Trade and transport costs: evidence from hurricane Sandy · 2016. 8. 3. · hurricane Sandy, a...
Trade and transport costs:evidence from hurricane Sandy
Pierre Cotterlaz∗ and Emanuele Mazzini† ‡
July 29, 2016
Preliminary version. Please do not cite.This paper may change in the near future.
Abstract
We use the massive disruptions on the Northeastern US road network provokedby hurricane Sandy at the end of October 2012 as a natural experiment to identifythe effect of transport costs on trade flows. This allows us to isolate the effect oftransport costs from the effect of other potential trade costs. In this way we are able toprovide interesting insights towards the comprehension of the mechanisms underlyingthe strong negative effect of distance on trade flows. Indeed, existing evidence suggeststhat the distance elasticity of trade flows cannot be entirely explained by transportcosts. Other elements, which broadly go under the name of “dark trade costs” matteras well. In our baseline specification, we find that the part of the distance elasticityof trade flows that can be attributed to transport costs is -0.06, lower than the totaldistance elasticity of trade flows (-0.84). We interpret this as evidence that, even withina country, transport costs are not the only component of trade costs. Moreover, in orderto recover the change in transport costs that Sandy caused in the affected areas, weimplement a method inspired from the indirect inference estimator, i.e we estimate thevalue of the change in transport costs for which the change in origin and destinationfixed effects estimated from simulated data matches the one estimated from real data.
∗Department of Economics and LIEPP, Sciences Po, [email protected].†Department of Economics, Sciences Po, [email protected]‡This work is supported by a public grant overseen by the French National Research Agency (ANR) as part
of the “Investissements d’Avenir” program LIEPP (reference: ANR-11-LABX-0091, ANR-11-IDEX-0005-02).We are indebted to our supervisor Thierry Mayer for continued advice and support and to Florian Oswaldfor his precious suggestions. We thank seminar participants at Sciences Po for many useful suggestions, aswell as Donna Miller (DelDOT), Ulysses Jacks (MassDOT), Meg Frampton (NJDOT), Mary Shevlin (NYDOT),Sophia Spencer (NCDOT), Lindsey Pitten (PA GOV), Chris Kinsey (WVDOH) and Jim Lambert (WV Divisionof Highways) for providing useful data about Sandy related disruptions on the road network. All remainingerrors are ours.
1
1 Introduction
Among the numerous challenges faced by the study of the consequences of transport costson trade, the main one is that transport prices result from the interaction between de-mand and supply for transport. On the demand side, the spatial distribution of economicactivity responds to transport prices. On the supply side, transport infrastructure shouldadjust to demand for transport in the in the long-run. This endogeneity of transport pricesimplies that simply regressing trade flows on transport costs might not give an accurateidea of the role of transport costs. In this paper, we overcome this issue by consideringhurricane Sandy, a storm that hit the Northeastern US at the end of October 2012, as anexogenous shock on transport costs. The storm represents an ideal natural experiment be-cause it caused heavy damages on the transport infrastructure of the affected areas, andthis constitutes a shifter on transport supply that we use for our analysis.
The exogeneity of the shock additionally turns out to be useful to isolate the effect oftransport costs from that of other potential trade costs, because in this setting the varia-tions of the former are uncorrelated with those of the latter. In this way we are able toprovide interesting insights towards the comprehension of the mechanisms underlying thestrong negative effect of distance on trade flows. Indeed, while this effect of distance ontrade flows is by now a well-established empirical fact (see, for instance, the meta-analysisperformed by Disdier and Head (2008) and more recently by Head and Mayer (2014)), itremains much harder to explain its magnitude.
One of the first determinants that comes to mind is the role of transport costs. However,papers that have estimated the elasticity of transport costs with respect to distance (seeHummels (2007) and Limao and Venables (2001)) find values that are not large enoughto explain the whole effect of distance on trade flows. Since transport costs alone can onlypartially explain the effect of distance on trade flows, there must be channels other thantransport costs through which distance increases trade costs. This points to the existenceof a set of spatial frictions, for which Head and Mayer (2013) coined the term “dark tradecosts”. Suggested sources for these frictions include for example differences in cultureand tastes, legacy of past trade costs and the spatial decay of information. Indeed, manycontributions have highlighted the importance of other factors to explain trade flows, ontop of transport costs. For instance Felbermayr and Toubal (2010) use scores at the Euro-vision Song Contest to show that cultural proximity affects trade flows and is negativelycorrelated with distance.
However, the existence of dark trade costs within countries has not received a large
2
share of attention yet. This is nevertheless not a trivial question because we are temptedto think that the mechanisms proposed to explain the dark trade costs should have lessrelevance in an intra-national framework compared to an international one. For instance,culture and tastes are more similar within a country than across countries, and the spa-tial decay of information should also be less intense. Additionally, tariffs and the so-called“grey trade costs” of crossing borders are absent. Despite the relevance of the issue, weare aware of few papers that have attempted to investigate whether transport costs can beconsidered as responsible of the whole distance effect for intra-national trade flows. Hor-tacsu, Martínez-Jerez and Douglas (2009) study eBay transaction within the 48 continentalstates of the US and find that distance has a negative effect on trade, even after controllingfor shipping costs. This suggests that information explains at least part of the negativedistance elasticity. Lameli et al. (2015) find a significant effect of local dialects on Germantrade. Given that institutions and communication costs are similar, they suggest that thiseffect reflects cultural ties. Finally, Wrona (2015) identifies a border effect between East-and West-Japan that he attributes to the structure of social and business networks.
We use the increase in transport costs induced by Sandy to determine a lower bound forthe equivalent change in distance. We find that transport costs do not account for the wholedistance elasticity of trade flows within the US, and therefore confirm the importance ofdark trade costs. In our preferred specification, the part of the distance elasticity of tradeflows that can be attributed to transport costs is -0.06, while the total distance elasticity oftrade flows is -0.84.
This study makes use of a natural experiment to isolate the transport costs componentsin the distance elasticity of trade, an approach that, to the best of our knowledge, was firstlyimplemented by Feyrer (2011), who used the closing of the Suez Canal between 1967 and1975. His work has been subsequently deepened by Hugot and Umana Dajud (2015), whostudied the effects of the opening of the Suez and Panama canals on trade and performeda welfare analysis by simulating counterfactual trade flows, had these canals not opened.The choice of a natural disaster as exogenous shock on transport costs was initiated byVolpe Martincus and Blyde (2013), who show that the 2010 earthquake in Chile had anegative effect on the exports of firms for which the optimal route was in the affectedareas. In the same vein, Volpe Martincus et al. (2014) use the closure of the San MartínInternational Bridge1 from 2006 to 2010 because of a protest action (not related to tradeissues). However, these two papers do not attempt to relate their estimates to the distance
1The San Martín International Bridge is a bridge on the Uruguay river, linking Argentina and Uruguay.
3
elasticity of trade.Our final contribution relates to the necessity to overcome the lack of data about
changes in transport costs. This required us to compute them by using a GIS software,spatial data on the location of the damages, and, most notably, a parameter giving themagnitude of the increase in transport costs in the affected areas. We develop a method torecover this parameter based on a structural gravity model for trade flows. We make use ofa so-called "indirect inference" estimator, i.e. we find the value of the increase in transportcosts for which the distance between some parameters we measure in the observed dataand the same parameters measured in simulated data is minimized. To our knowledge, weare the first to implement such an estimation method.
Our paper is organized as follows: we begin by describing hurricane Sandy and itseffects on transport infrastructure and trade flows (section 2). Section 3 presents the theo-retical framework we use and our empirical specifications. We then present our data, andthe way we compute the bilateral changes in transport costs induced by Sandy (section 4).Section 5 is dedicated to the presentation of our indirect inference estimator. Finally, wepresent our main results along with some robustness checks (section 6).
2 Hurricane Sandy
2.1 Characteristics of the hurricane
Sandy is a hurricane2 that hit the Northeastern US at the end of October 2012. It can beconsidered as an extraordinary storm for the US, both because of its impact and charac-teristics. Sandy made landfall on October 29th, 2012 near Brigantine in New Jersey. Atthis time, it was not at the peak of its intensity (this peak was reached over Cuba), but itwas still very intense and wide, with tropical storm-force winds extending 805 km from thecenter of circulation just prior to landfall.3 The storm’s angle of approach (a perpendicularone) and the fact that the landfall coincided with high tide and full moon also explain the
2Note that when it reached the US territory, the storm was technically speaking no longer a hurricane,but rather a “post-tropical” cyclone, because it lacked the typical strong thunderstorm activity near its centerand had lost its eye. Nevertheless, throughout this paper we will use the term hurricane to refer to it sincethis is the common practice.
3Tropical storm-force winds correspond to a wind speed above 89 km/h, a speed at which, ac-cording to Wikipedia, “Trees are broken off or uprooted, structural damage likely”. Hurricane-force winds (above 118 km/h) extended 280 kilometers. Source: http://www.livescience.com/24380-hurricane-sandy-status-data.html
4
exceptional effects of Sandy, since these two element led to record storm tide heights (thestorm surge combined with astronomical tide) and the surge’s large waves were drivendirectly into the coastal cities.
After landfall, Sandy slowed down and weakened, but its broad size nevertheless led todisruptions across the Eastern and Midwestern US, as well as Southeastern Canada. Highwinds, heavy rains and accumulating snows were recorded because of Sandy’s remnantsmoving through southern Pennsylvania. In the central Appalachian Mountains, blizzardconditions occurred and strong winds spread into the Ohio Valley and the Great Lakes onOctober 31st. The hurricane completely dissipated in Eastern Canada in the next two days.The National Oceanic and Atmospheric Administration (NOAA) estimated that the entirearea affected by the winds during the track extended over more than 5 million squarekilometres and that more than 60 million people were directly affected across 24 states(Mildenhall et al. (2013), p. 13).
The exceptional intensity and width of this storm explains the high death toll: accordingto Blake et al. (2013), 72 people died in the Northeastern US as a direct consequenceof the hurricane. Economic losses were also large: AON Benfield, a leading insuranceintermediary, evaluates the total gross direct economic cost of hurricane Sandy as highas 68 billions USD (Mildenhall et al. (2013), p. 38), thus making it the second-costliesthurricane ever recorded in the USA (Katrina in 2005 being the first one) (Mildenhall et al.(2013), p.45).
2.2 Disruptions for road transport
Hurricane Sandy caused a sizable temporary increase in transport costs for all the goodsthat had to transit in the Northeastern US because it severely damaged transport infrastruc-tures in this area. Anecdotal evidence suggests that it was very difficult to circulate afterthe passage of the hurricane, as the words of James Hadden (project manager for the 511Traffic and Travel Information Program at the New Jersey Department of Transportation)perfectly summarize: “Roadway damage was beyond what anyone could have imagined”.4
Main roads and highways were closed or re-routed due to flooding, downed trees, downedwires or debris while trains and long-distance bus companies were forced to suspend oper-ations across the Northeast for several days.5
This claim is supported by data we obtained from the New Jersey (NJ) and the New
4WRTM Workshop and Stakeholder Meeting, September 25, 20135See, for instance http://wtop.com/news/2012/10/amtrak-bus-lines-cancel-service-for-sandy/.
5
York (NY) Department of Transportation (henceforth DOT), which contains informationabout all the disruptions on the highway network recorded by the NJ DOT and the NYDOT during hurricane Sandy and its immediate aftermath.6 The magnitude of the dam-ages is confirmed by the figures given in table 1, where we classified the disruptions bycategories and selected the categories that were unequivocally due to hurricane Sandy.Sandy caused overall more than 1000 disruptions on the highway network in the two con-sidered states. Downed trees, signals, wires or poles represented the most important causeof disruptions (607 occurrences), followed by flooding (136 occurrences) and other genericweather related events (96 occurrences). Other major sources of disruptions include de-bris, emergency interventions and overturned vehicles.
Table 1: Hurricane related disruptions in New Jersey and New York
Type # NJ % NJ # NY % NY # NJ & NYDowned pole, wires, signal, tree 345 33.0% 262 19.2% 607Flooding 89 8.5% 47 3.5% 136Weather related 39 3.7% 57 4.2% 96Debris 6 0.6% 46 3.4% 52Overturned vehicles 19 1.8% 0 0.0% 19Emergency interventions 4 0.4% 14 1.0% 18Fuel/diesel shortage 15 1.4% 2 0.2% 17Electrical damages 1 0.1% 0 0.0% 1Totals 518 54.6% 428 33.2% 946Note: Data are sorted according to the total number of accidents in NY and NJ.Percentages are calculated by considering the total number of disruptions asreported by the NY and NJ DoT over the period Oct. 27th - Nov. 2nd, which is1045 for NJ and 1364 for NY.
Figure 1 shows how the number of disruptions recorded by the NJ and the NY DOTevolved over time. We make a distinction between the disruptions related to the hurri-cane and the other types of disruptions. The figure points out that the former reached aremarkable peak on October 30th, when the number of disruptions recorded was about400.7 The figure also shows that there was a large number of newly recorded disruptionsattributable to the hurricane, especially in its immediate aftermath, hence suggesting that
6To be precise, data extend until November 7th, 2012 for NJ but only until November 2nd, 2012 in thecase of NY.
7Note that these dates are the date at which the DOT record the disruptions, which might be a bitdelayed compared to the date at which the disruption actually started because the DOT might not be able toinstantaneously gather the whole information about the hurricane related damages.
6
the storm caused an unusual increase in the total amount of disruptions. Figures 13 and14 in Appendix C show the same information for NJ and NY separately.
Figure 1: Number of incidents recorded by NJ DOT and NY DOT
total disruptions
other disruptions
hurricane related
010
020
030
040
050
0#
disr
uptio
ns
Oct 27 Oct 29 Oct 31 Nov 2
Finally, the map in figure 2 indicates the location of the recorded disruptions. It isvaluable because it points out that the disruptions affected the road network in a widearea, quite far inland, and were not limited to the coast.8
The situation for motorists was also exacerbated by a shortage of gasoline, which led torationing and price gouging in some instances. This shortage resulted from the combina-tion of several elements on the production side: refineries in the affected areas were eithershut down or run at reduced capacity, pipelines had to be closed for safety precautions, im-ports from the New York harbour were limited, and the disruptions on highways hamperedthe ability of trucks to deliver fuel when they were able to obtain it. The US Department ofEnergy additionally reported that about 8.5 million customers were without power duringor after hurricane Sandy. Among them, many had to use electric generators requiring fuel,which led to a demand spike, while some were gas stations, unable to serve customersbecause of power outages.9 The ultimate consequence is that filling stations saw lines that
8For a zoom on New York City and the border with NJ see Appendix B, figure 10.9http://www.eia.gov/todayinenergy/detail.cfm?id=8730
7
Figure 2: Disruptions due to Hurricane Sandy in New Jersey and New York
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!( !(!(
!(!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(
!(
!(!(!(!(
!(!(!(!(!(!(
!(
!(
!(
!(!(
!(!(!(!(
!(
!(
!(!(
!(!(
!(
!(!(!(
!(!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(!(!(!(!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(!( !(
!(!(!(!(!(!(!(
!(
!(!(!(
!(!(
!(!(!(
!(!(!(
!(
!(
!(
!( !(!(!(!( !(
!(
!(!(
!(
!(
!(
!(!(!(
!(!(!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
!(
!(
!(!(!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(!(
!(!(!(
!(
!(
!(
!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(!(!(
!(
!(
!(!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!( !(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!( !(
!(!(
!(
!(
!(
!(!(
!(
!(!(
!(!(
!(!(
!(
!(
!(
!(
!( !(
!(
!(!(!(
!(!(
!(!(!(
!(
!(!(!(
!(
!(
!( !(!(
!(
!(
!(
!(
!( !( !(
!(!(
!(!( !(
!(
!(
!(
!(!(!(
!(!(!(
!(
!(
!( !(!(!(!(
!(!(!(!(!(!(!(!(!(
!(!(!(!(!(
!(
!(
!(!(!(
!(
!(
!(
!(!(
!(!(!(
!(!(
!(!(!(!(!(
!(!(!(
!(
!(!(
!(
!(!(
!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(!(
!(
!(!(
!(!(
!(
!(!(
!(
!(!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(!(!(!(
!(!(
!(!(
!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(!(!(!(
!(
!(
!(
!(!(
!(
!(
!(!(
!(
!(!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!( !(
!(
!(
!(
!(!(
!(
!( !(
!(
!(!(
!(
!(
!(!(!(
!(
!(!(
!(
!(!(!(
!(
!(
!(!(
!(!(!(!(!(!(
!(
!(
!(
!(
!(!(
!(!( !(
!(
!(
!(
!(
!(!(!(
!(
!(!(!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(
!( !(!(
!(
!(
!(!(!(
!(
!(
!(!(
!(
!(
!(!(
!(!(
!(
!(
!(
!(!(!(
!(!(
!(
!(
!(
!(
!(!(
!(!(!(
!(
!(
!( !(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(!(!(!(
!(
!(
!(!(!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(!(
!(
!(!(
!(!(
!(
!(
!(
!( !(
!(
!(!(!(
!(
!(
!(!(!(
!(
!(
!(
!(!( !(!(
!(!(!(
!(!(
!(!(
!(!(
!(!(!(!(
!(
!(
!( !(!(
!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(!(!(
!(
!(
!(!(
!(
!(
!(!(!(
!(
!(!(
!(
!(!(
!(
!(!(
!(
!(!(
!(
!(!(
!( !(!( !(
!(
!(
!(
!(!(
!(!(
!( !(!(!(!( !(
!(
!(
!(!(
!(!(!(!(
!(
!(
!(
!(
!(
!(!(!(
!(!(!(
!( !(!(
!(
!( !(
!(
!(!(
!(
!(
!(
!(!(
!(!(
!(!(!(!(!(
!(
!(!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(
!(!( !(!(
!(!(
!(
!(
!(!(
!(
!(
!(!(!(
!(
!(!(
!(!(!(!(!(!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!( !(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!( !(
!(!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(!(!(!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(!(!(
!(
!(!(
!(
!( !(!(!(
!(
!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
Note: Red lines show the highway network. Blue points represent disruptions recorded by NJ orNY DOT. Yellow surfaces correspond to Core-Based Statistical Areas (CBSA).
were even miles long in some cases, forcing people to wait for several hours.10
Table 2: Availability of gasoline in the New York City metropolitan area
Station Response Nov. 2 Nov. 3 Nov. 4 Nov. 5 Nov. 6 Nov. 7 Nov. 8 Nov. 9No gasoline supply 10% 28% 24% 38% 28% 28% 21% 21%Gasoline availability 33% 45% 59% 62% 66% 62% 72% 72%No power at station 3% 3% 0% 0% 0% 0% 0% 0%No contact w/station 53% 24% 17% 0% 7% 10% 7% 7%Source: EIA emergency survey
Table 2, where we report the results of a survey implemented by the US Energy Infor-mation Administration (EIA) on gas stations in the New York City metropolitan area from
10See, for instance, http://money.cnn.com/2012/11/01/news/economy/gas-stations-supply-sandy/.
8
the 2nd up to the 9th of November, illustrates the exasperated situation. At the beginningof the considered period, gasoline was available in only one third of gas stations. The othergas stations either had no supply, no power, or were unable to reply. Most notably, even ifthe share of station with no gasoline supply declined over the considered week, on Novem-ber 9th the full supply of fuel was still not restored: 21% of gas stations had no gasolinesupply. Finally, the EIA is careful to mention that the reported figures are “[. . . ] not de-signed to reflect the specific experience of more severely affected areas”, which suggeststhat the shortage could have been locally more severe. 11
2.3 Effects on trade flows
We will detail the source of our data on trade flows in section 4.1. For now, let us give abrief view of the behavior of trade flows during the 4th quarter of 2012, the one in whichSandy took place.
First, let us have a look at the share of trade flows for which the origin or the destinationis affected by the hurricane. Again, we will detail later on (section 4.2.2) what we define asan "affected" region. This gives us an idea of the share of trade flows that could potentiallybe affected. These shares are plotted in figure 3. We find that this share is far fromnegligible and that all industries could see part of their flow affected.
Secondly, we compare the evolution over time of trade flows in affected and unaffectedareas. We aggregate all trade flows originating from or arriving to affected areas anddetrend this aggregate by regressing its values for the three first quarters on the timevariable (quarter) and taking the difference between the predicted value and the observedvalue. We use a similar approach for all flows originating from or arriving to the unaffectedareas. These detrended aggregate flows are plotted in figure 4.
They fall in both groups during the 4th quarter, but the fall is more dramatic for flowsfor which the origin or the destination corresponds to an affected area. The fact thattrade flows between non affected CFS areas also decreases could be explained by a globaldownward trend for the 4th quarter, and also maybe by the fact that the transport costsmight have increased for them too, if the optimal path between them and their tradepartners goes through the IA/PA counties.
We insist on the fact that the effects highlighted by figure 4 embed not only the effectsof transport costs, but also all the direct economic damages caused by the hurricane. Our
11The “New York City Metropolitan Area Retail Motor Gasoline Supply Report” is available at http://www.eia.gov/special/disruptions/hurricane/sandy/gasoline_updates.cfm.
9
Figure 3: Share of trade flows whose origin or destination is an affected area by NAICS.
0 20 40 60 80 100Share of trade
314336
42474245
313333321316
4238312311315327
4249332212324322
551114326
42374231
335334
42364248Mean4233
33149314234
339337325
424142355111323
42464239424445414232
4543142424243
Note: The red part corresponds to the share of trade for which the origin or the destination is anaffected area. The blue part corresponds to the share of trade for which neither the origin nor thedestination is an affected area. Data is aggregated over the 4th quarter. Note that this does notcorrespond to the share of trade flows for which the origin or the destination is an affected areaduring Sandy. The signification of the NAICS codes is detailed in table 13.
identification strategy, that we will present in the next section, allows us to isolate the effectof transport costs.
10
Figure 4: Detrended trade flows in affected and unaffected areas.
IA/PA
Not IA/PA
-.08
-.06
-.04
-.02
0
Qtr. 1 Qtr. 2 Qtr. 3 Qtr. 4
11
3 Identification strategy
3.1 Model
To estimate the impact of transport costs on trade flows, we use a structural gravity frame-work. Under structural gravity, the trade flow from location i to location n at time t (de-noted Xnit) takes the following form:
Xnit =YitΩit
Xnt
Φnt
φnit (1)
where Yit =∑
nXnit is the value of production in CFS area i, while Xnt =∑
iXnit is thevalue of expenditure in CFS area n. Our empirical counterparts are the total value of goodsleaving from i (including those dispatched in the same CFS area) and the total value ofgoods arriving in n (including those from the same CFS area) respectively. Ωit and Φnt aremultilateral resistance terms defined as:
Ωit =∑l
φlitXlt
Φlt
(2)
Φnt =∑l
φnltYltΩlt
(3)
φnit is a bilateral resistance term, related to the trade costs τnit in the following way:
φnit = (τnit)εηnit (4)
where ε is the elasticity of trade flows with respect to trade costs and ηnit is an error termuncorrelated with τnit.
Our interest lies in the effect of distance on trade costs, and its decomposition betweena part related to transport costs and a part related to dark trade costs. For this purpose, werewrite the total trade cost τnit as the product of two multiplicative components: a time-varying component capturing the transport costs (Tnit), and a time-unvarying componentcapturing the dark trade costs (Cni).
τnit = TnitCni (5)
Combes and Lafourcade (2005) argue that transport costs depend on two additive com-ponents, each of them increasing with distance. The first one is proportional to distance,
12
as it reflects elements such as fuel consumption, vehicle maintenance operating costs andtolls. The second one represents the cost of time and is proportional to the duration of thetrip, since it reflects costs such as the driver’s wage and accommodation. We could evenconsider a broader range of time-related trade costs since, as emphasized in Hummels andSchaur (2013), time might also be costly because it creates a delay between the momentwhen the production decision is made and the one when the product is sold, and marketconditions might change during this interval. In the case of intermediary goods, such delaymight even jeopardize the whole supply chain. Because the duration of the trip increaseswith distance, the time component of trade costs is an increasing function of distance.
For the sake of tractability, we depart from the above-mentioned additive structure andmodel the effect of distance on transport costs in a simpler way, assuming the followingfunctional form:
Tnit = (dnit)αν1nit (6)
where dnit is the road distance between origin and destination and ν1nit is an error termuncorrelated with dnit. In our specific setting the distance is time varying because hurricaneSandy creates some disruptions that are equivalent to an increase in distance. α is theelasticity of transport costs with respect to distance, capturing both the direct and thetime-related effects of distance.
The dark trade costs are also an increasing function of distance. They do not have a tsubscript because we consider them as time invariant, at least within the period taken intoconsideration for our estimation (year 2012). Note that this assumption of time invarianceis not necessary for our estimation, because given the exogenous nature of the shock weconsider, the potential time variations in dark trade costs would be uncorrelated with thetime variations in transport costs. More formally, dark trade costs take the following form:
Cni = (dni)γν2ni (7)
where dni is the average road distance between i and n, and ν2ni is an error term uncorre-lated with dni.. Dark trade costs depend on distance because of the mechanisms underlyingthem, as evoked in the introduction. For instance, if the spatial decay of information is oneof the factors behind dark trade costs, then we expect them to increase with distance. Theerror term ν2ni captures the unobserved determinants of dark trade costs.
Using the functional forms of transport costs (equation (6)) and dark trade costs (equa-
13
tion (7)), we can rewrite the trade costs (equation (5)) as:
τnit = dnitα(dni)
γνnit (8)
We plug this into φnit (equation (4)) and take logs:
ln(φnit) = εα ln(dnit) + εγ ln(dni) + εnit (9)
where εnit is an error term including both νni and ηnit. It is uncorrelated with dnit and dni.From this equation, we observe that the distance elasticity of trade flows can be decom-posed in two additive components:
• εα : the transport cost effect
• εγ : the dark trade cost effect
Our objective is to estimate εα. This can be done easily if the assumption that the darktrade costs are time invariant holds, because we can use the time variation in transportcosts created by Sandy to identify εα. The effects of hurricane Sandy are equivalent to atemporary increase in distance for goods transiting through the devastated region, and thisallows us to estimate the effect of a change in distance on commodity flows. Given that thedark trade costs remained unaffected by the hurricane, we know that only the transportcosts played a role in the distance effect we estimate.
Going back to equation (8), notice that if we consider the flow during each of the threefirst quarters of the year, then dnit = dni because there is no time variation (no disruptionscaused by the hurricane). As a consequence, equation (8) rewrites:
τni = (dni)α+γ
= (dni)ρ (10)
where ρ ≡ α + γ. This implies:
ln(φni) = ερ ln(dni) (11)
where ερ is the total distance elasticity of trade flows.Finally, we we make our theoretical framework more general by allowing each indus-
try to face a different bilateral resistance term, consistently with the idea that different
14
industries might face different error terms. The structural gravity equation rewrites:
Xnist =YistΩist
Xnst
Φnst
φnist (12)
where the subscript s refers to industry. Keeping the same functional forms for trade costs,transport costs and dark trade costs, bud adding an industry specific error term ηnist to thebilateral resistance term, we can write: φnist =
(dnit
α(dni)γνni
)εηnist. Note that an alterna-
tive formulation would be to allow the functional form of dark trade costs to vary acrosssectors. In terms of estimation, both formulations would result in the same specification.
3.2 Specifications
Our baseline specification derives from equation (12). By taking logs and plugging (9), weobtain:
ln(Xnist) = ln(Yist)− ln(Ωist) + ln(Xnst)− ln(Φnst)
+ εα ln(dnit) + εγ ln(dni) + εnist
This can be estimated with the standard fixed effect estimation method for gravity equa-tions, which in this setting is specified as:
ln(Xnist) = FEnis + FEnst + FEist + β1 ln(dnit) + εnist (13)
where β1 corresponds to εα in equation (9). FEnst is a time varying destination fixed effect,which captures Xnst and Φnst. Similarly, FEist is a time varying origin fixed effect capturingXist and Φist. FEnis is a dyad-industry fixed effect capturing the dark trade costs.
Note that with such a specification, εα is correctly estimated whatever the functionalform of Cni. Indeed, the dyad-industry fixed effects capture all the time-unvarying com-ponents of trade costs so that, as long as the "dark trade costs" are time-unvarying, as weassumed in our model, they will be absorbed by FEnis. But this assumption is not even re-quired because even if the dark trade costs are not time unvarying, their variations shouldbe uncorrelated with the variation in transport costs, thanks to the random nature of theshock on transport costs that we use.
Coming back to the origin and destination time varying fixed effects, it is important tonotice that they entirely capture the local economic effects of the hurricane through theabsorption of Xnst and Xist. As a consequence, the inclusion of such origin and destination
15
time varying fixed effects allows us to rule out the possibility that the estimated coefficientwould not exclusively reflect the transport costs effect, but also include local economiceffects from the hurricane, like for instance damages to the production facilities or poweroutages. In order to better understand why such localized effects of the hurricane willnot affect the coefficient we estimate, let us consider the case of the firms located in NewJersey. Production decreased in the aftermath of the hurricane, and some plants had tostop production. But given the random nature of the destruction, we expect the relativechange in export value to be the same whatever the destination. As a consequence, this"direct effect" of Sandy on trade (less local economic activity) will be entirely captured bythe 4th quarter origin fixed effect of New Jersey. The same reasoning can be made for firmsor customers located in New Jersey and importing shipments from other parts of the US(the local economic damages would be captured by the New-York 4th quarter destinationfixed effect).
As a consequence, once the fixed effects detailed in equation (13) are included, weare confident that the error term is indeed uncorrelated with our explanatory variable andthus that the estimates we provide are unbiased. We can think of no variable that would becorrelated with dnit and Xnist and omitted in our specification. Reverse causality is not anissue and the fact that Sandy is a natural disaster rules out any possibility of simultaneousdetermination.
However, since we computed the explanatory variable ourselves, we cannot rule out thepossibility of some imprecision in its variations. Even though we have no reason to believethat such imprecision would result in a bias, as a robustness check, in some specificationswe exclude certain observations from our samples, for which the change in distance mightbe less accurately computed. More precisely, in the second column of each table of resultswe exclude the flows that occur within the same CFS area because, as explained in section4.2, for these flows we could not apply our usual least cost path algorithm (instead wesimply considered that the transport costs were multiplied by κ during the Sandy state ofthe world). In the third column of the tables of results, we include only the flows occurringbetween urban CFS areas. Indeed, the remainders correspond to larger CFS areas andtherefore their weighted centroid might be a poor proxy for the origin or destination pointof the shipment. Finally, in the fourth column, we apply both restrictions simultaneously.
16
4 Data
4.1 Trade flows
Data on trade flows comes from the Public Use Microdata (PUM) file of the 2012 Com-modity Flow Survey (hereafter CFS). This survey is realized every five years by the Bureauof Transportation Statistics (hereafter BTS) and the Census Bureau. The 2012 CFS cov-ers approximately 60 000 establishments in mining, manufacturing, wholesale, auxiliaries,and selected retail and services trade industries (a list of all the included industries can befound in the appendix, table 13). Once an establishment is selected, it receives a question-naire every quarter. Reply is mandatory and the answers have to be exact, which ensuresgood quality of the data.
In total, the CFS records 4,547,661 shipments in 2012. The term “shipment” is de-fined by the Census Bureau and the BTS as “a single movement of goods, commodities, orproducts from an establishment to a single customer or to another establishment owned oroperated by the same company as the originating establishment (e.g., a warehouse, distri-bution center, or retail or wholesale outlet)’12. The shipments included in the CFS thereforedo not necessarily correspond to trade transactions. However, it is common practice to usethe CFS to proxy trade flows, as for instance in Duranton, Morrow and Turner (2014).For each shipment, we have information about twenty variables such as the origin and thedestination areas, the transport mode, the NAICS code of the product as well as the valueand the weight of the shipment.
The CFS distinguishes five single modes and five multiple modes of transportation.Table ?? provides an exhaustive list of the modes of transport included in the CFS andsummarizes their characteristics in terms of trade flows. It shows that shipments carried bytruck represent by far the large majority of commodity flows. This remark drives the choiceof our sample: we select only shipments carried by truck because road transport is the onlymode of transport for which we have enough observations to complete our analysis.13
For each shipment, we know the “CFS area” where it originated and arrived. Datainclude 129 different CFS areas. Out of these, 83 correspond either to a MetropolitanStatistical Area (MSA) or a Combined Statistical Area (CSA), while the remaining 46 are“remainders”, namely portions of states that are not within a MSA/CSA. We do not include
12Source: 2012 CFS Summary Report, p. ix13To be precise, we include in “road transport” all the goods that are shipped by truck, be this “for-hire
truck”, “private truck” or simply “truck”.
17
Table 3: Transportation modes
Modes ValuesPercents
from totalValue Weight Obs. Value Weight Obs.
All modes 13,837,786 11,291,584 4,546,970 100.0 100.0 100.0Single mode 11,869,117 10,864,393 3,354,106 85.8 96.2 73.8
Truck 10,123,625 8,048,389 3,231,969 73.2 71.3 71.1For-hire truck 6,497,910 4,291,261 1,613,317 47.0 38.0 35.5Private truck 3,624,027 3,754,398 1,618,282 26.2 33.2 35.6
Rail 430,203 1,536,294 38,458 3.1 13.6 0.8Water 198,192 417,915 3,691 1.4 3.7 0.1Air (incl truck & air) 438,009 4,542 68,809 3.2 0.0 1.5Pipeline 433,481 507,032 3,673 3.1 4.5 0.1
Multiple mode 1,968,669 427,191 1,192,864 14.2 3.8 26.2Parcel, USPS, or courier 1,687,586 28,514 1,165,297 12.2 0.3 25.6Truck and rail 189,271 166,900 19,070 1.4 1.5 0.4Truck and water 14,257 30,069 2,498 0.1 0.3 0.1Rail and water 2,341 34,268 200 0.0 0.3 0.0Other modes 21,352 78,357 963 0.2 0.7 0.0
Note: totals may not sum due to the censoring of some observations for confidentiality reasons.Shipment values are expressed in M$, weight in thousand US tons (kt). NB: 1 US ton correspondsto 907.185 kg. Source: own calculations based on 2012 CFS data.
in our sample the two CFS areas that are not within the continental USA, Hawaï andAlaska, because road distances cannot be computed for these areas. After this selection, oursample gathers 3,143,535 observations in 45 NAICS codes. Figure 5 shows the CFS areasused in our analysis. Red areas correspond to CSA/MSA, while green areas correspondto “remainders”. A list of all the CFS areas can be found in the appendix, in table 11 for“urban” CFS areas and table 12 for the “remainders”.
Figures 6 and 7 represent respectively the total export value and the total import valueby CFS area, normalized by the surface of each CFS area. Unsurprisingly, we see that urbanCFS area have a higher "trade intensity", which is consistent with economic activity beingspatially concentrated in these areas. Table 4 presents additionally some summary statisticsfor the most relevant variables. The aggregation is done alternatively by origin CFS area,by destination CFS area and by dyad-NAICS, which is the dimension at which our modeland our regressions will be specified, as will be seen in the next sections.
18
Table 4: Descriptive statistics - 129 origin and destination CFS areas; 218,133 dyad-industries
Variable Mean Std. Dev. Median Min MaxOrigin CFS Area
Value 74,305.0 78,071.8 51,891.3 2,386.0 542,063.8Weight 60,607.9 54,325.0 43,634.9 3,046.6 297,872.3# shipments 34,201.7 40,002.4 23,329.5 1,521.3 316,378.7Value per shipment 2.48 1.10 2.38 0.42 7.13# obs. 24,107.4 14,047.5 20,993 2,072 88,195Value per obs. 2.72 1.36 2.49 0.57 8.24Share of zeros 12.26 2.89 12.20 5.71 20.89shipment distance CFS 13.12 3.93 12.20 2.14 26.95
Destination CFS AreaValue 74,305.0 70,187.4 52,581.6 6,822.9 442,231.5Weight 60,607.9 54,012.4 42,287.9 4,054.5 295,574.5# shipments 34,201.7 32,613.9 25,419.1 2,918.9 207,611.3Value per shipment 2.34 0.88 2.28 0.46 5.15# obs. 24,107.4 15,497.3 19,517 4,285 93,380Value per obs. 2.82 1.05 2.66 1.18 9.09Share of zeros 13.49 4.00 14.17 4.30 24.80shipment distance CFS 12.89 4.89 11.01 6.26 29.45
Dyad - industryValue 43.9 543.9 3.1 0 131,371.1Weight 35.8 686.2 0.5 0 136,257.5# shipments 20.2 394.7 0.7 0 122,862.8Value per shipment 16.47 226.44 4.16 0 51,160# obs. 14.3 73.4 2.0 1 5,028Value per obs. 4.16 24.21 0.95 0 4,821.9Share of zeros 41.46 29.84 50.00 0 75shipment distance CFS 15.18 11.70 12.34 0 56.55
Note: shipment values and value per obs. are expressed in M$; weight is inthousand US tons (kt); # shipments and value per shipment are in thousands;distance is in hundreds of km.
19
Figure 5: CFS areas and their weighted centroids
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
!( !(!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(!(!(
!(
Note: Red surfaces correspond to urban CFS areas, while green surfaces represent rural CFS areas.Dark blue points are the weighted centroids of the CFS areas.
4.2 Determination of transport costs
The computation of distances between CFS areas requires the conversion of each of theseareas to a single point. We therefore computed weighted centroids, where the weights arebased on population. More precisely, we first determine the centroid of all the countiesconstituting a given CFS area and assign a weight to these centroids, equal to the popu-lation of their county. The coordinates of the weighted centroid of the CFS area are thengiven by the weighted average of the coordinates of the centroids of each county withinthe considered CFS area. We use population weights because they can be considered as aproxy for economic activity, and it is more likely that a shipment leaves from (or arrivesto) a place where economic activity is more intense. As a final step, since we are interestedin the computation of road transport costs, we determine the point on a road which is theclosest to the calculated weighted centroid and use this point in our distance computation.This final adjustment is necessary because the method we use for the computation of trans-port costs, which we will describe more precisely in sections 4.2.1 and 4.2.2, can only work
20
Figure 6: Total export value by CFS area
Note: Export value is divided by the surface of the CFS area.
if both the origin and the destination points are on a road.Throughout this paper, we will distinguish two "states of the world". The first state of the
world corresponds to the situation during hurricane Sandy and its immediate aftermath,where the road network is heavily affected. We will often refer to it as the Sandy stateof the world and use the superscript S to denote it. The second state of the world, whichwe will refer to as the "normal" state of the world, corresponds to the rest of the year. Nounusual disruption affects the road network. We will use the superscript N to denote thisstate of the world.
We need to compute two distinct set of bilateral transport costs: the transport costsin the normal state of the world (denoted TNni ) and the transport costs during the Sandystate of the world (denoted T Sni). What we are interested in is the change in transport costsbetween these two states of the world, not their absolute value.
21
Figure 7: Total import value by CFS area
Note: Import value is divided by the surface of the CFS area.
4.2.1 Transport costs during the normal state of the world
TNni is computed via a GIS software which has a built-in least cost path algorithm allowingto find the shortest path between all pairs of CFS areas and to return the associated cost(which is our measure of distance). Running the algorithm requires the creation of araster (a grid) representing the road network. Geographical data on the US road networkcome from Natural Earth14. The map includes only major roads, which is an advantagefor our analysis because it corresponds approximately to the US National Truck Network, anetwork of approved state and interstate highways for commercial truck drivers15.
We turn this map of the road network into a raster where each cell can take one of twovalues: either infinite value, if there is no road in the cell, or value 1 if there is a road inthe cell. This value corresponds to the transport cost to go through the cell. The fact that
14Natural Earth is a website supported by the North American Cartographic Information Society (NACIS)proposing public domain maps at different scales on various themes: http://www.naturalearthdata.com
15The National Truck Network includes almost all of the Interstate Highway System and other specifiednon-Interstate highways.
22
we choose a value of 1 is a normalization, yet this does not involve any arbitrary choicebecause in our analysis we will be interested in the relative variation in transport costs(over time and/or across cities), not in their absolute value.
4.2.2 Transport costs during Sandy
The computation of transport costs during Sandy requires the creation of a new road raster,which differs from the one evoked above by the fact that road cells in the areas affectedby the hurricane now have a value of κ, and not 1. We come back to the determination ofκ in section 5. κ corresponds to the increase in transport costs occurring in affected areasbecause of the hurricane. For instance, if κ = 5 , it means that it is five times more costlyto go through cells in the affected areas during the hurricane.
To determine the areas that were affected by the Hurricane, we used geographic datafrom the Federal Emergency Management Agency (hereafter FEMA). We selected all thecounties that benefited from Public Assistance (PA), or Individual Assistance (IA). PublicAssistance is a program through which the FEMA provides a grant to “fund the repair,restoration, reconstruction or replacement of a public facility or infrastructure damaged ordestroyed by a disaster”, while IA provides a federal funding “to individuals and familieswho have sustained losses due to disasters”16. The counties that benefited from IA or PAare represented in blue in figure 8. Note that sixteen CFS areas are at least partly in thedevastated zone, and the zone is large enough for increases in transport costs to occur evenfor pairs of CFS areas that were not directly hit by the Hurricane, but for which the optimalitinerary goes through the affected area.
With the new road raster, the procedure to determine the transport costs during Sandyis identical to the one we described in subsection 4.2.1. For trade flows occurring withinCFS areas, and for which at least part of the CFS area is made of an IA/PA county, weconsider that transport costs have been multiplied by κ.
4.2.3 From transport costs to distance
Now that we computed TNni and T Sni, we can compute the average quarterly trade costs.During the three first quarters, there is no hurricane Sandy, so the state of the world isalways normal. Hence, for t = 1, 2, 3, Tnit = TNni . In the 4th quarter, the state of the
16Both definitions are taken from the official FEMA website: http://www.fema.gov/news-release/2015/07/20/understanding-individual-assistance-and-public-assistance, consulted on January 20th,2016
23
Figure 8: Counties benefiting from IA or PA after Sandy
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(!(!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!( !(
!(
!(
!(
!(
!(
!(
!(!(!(
Note: Blue surfaces correspond to the areas that benefited from IA or PA after hurricane Sandy.
world can be Sandy or normal, so the trade costs will be a mixture of TNni and T Sni. Moreprecisely, we assume that the probability for a firm to be willing to send a shipment in agiven day is always the same. Hence, considering the 4th quarter of 2012, the probabilitythat the firm is willing to send the shipment during the Sandy state of the world is givenby the ratio of the disaster duration over the quarter duration (92 days). We consider thatthe Sandy state of the world lasts for ten days. The choice of this duration is justified byanecdotal evidence and by the fact that a few days after Hurricane Sandy, a snow stormaffected approximately the same area (the so-called “November 2012 nor’easter”). Wedenote this ratio by χ (hence χ = 10/92). The transport costs between any pair of CFSareas for the 4th semester are thus given by:
Tni,t=4 = χ T Sni + (1− χ) TNni
Knowing the change in Tni, we want to find its equivalent in terms of dni. For that, weneed to set the value of one parameter : ε. We set ε = −5, which corresponds to the meanvalue found in the literature when structural gravity is used (cf. Head and Mayer (2014)).Once ε is set, we can find the value of ρ because we can estimate ερ from a cross-sectional
24
gravity regression, as explained in section 3.1. The regression table is given in section 6.2(table 6): we find that ερ = −0.84. With ε = −5, this implies ρ = 0, 17
In section 3.1, we showed that ρ = α + γ. Moreover, γ ≥ 0. As a consequence α ≤ 0.17
so that 1α≥ 5.88. From Tnit = dnit
α (equation (6)), we get:
dnit = Tnit1α
with 1α≥ 5.88. Taking 1
α= 5.88 gives us a lower bound for the distance equivalent of the
change in transport costs, which we will use as regressor.
5 Estimation of the overcost parameter (κ)
In section 4.2.2, we described how we used an "overcost parameter", denoted κ, to com-pute the change in transport costs during Sandy. This parameter is crucial for our finalresults, since increasing the value of κ leads to a decrease in our estimate of εα. Therefore,we determine a lower bound for its value, using an estimation method inspired from indi-rect inference. The intuition behind this estimation method is the following: we simulatetrade flows with different magnitudes of transport costs increase in the affected areas anddetermine the changes in multilateral resistance terms (MRT) they imply, relying on thestructural gravity equations presented in section 3.1. We can then compare the changesin MRT obtained from the simulated data to the ones estimated with real data, and findthe magnitude of transport costs increase that minimizes the distance between these twovectors. Our results suggest that a lower bound for κ is 6.
A bit more formally, our estimation proceeds as follows. Firstly, we estimate FEit andFEnt in the real data and we take the difference between fixed effects in the 4th quarterand fixed effects in other quarters, which gives us a vector of parameters, which we denoteθ0. Then, we can simulate trade flows for any value of κ and estimate FEit and FEnt in thesimulated data. Taking again the difference between fixed effects in the 4th quarter andfixed effects in other quarters, we get a new vector of parameters whose values depend onκ and which we denote θ(κ). Our estimate is the value of κ such that θ0 and θ(κ) are "asclose as possible".
25
5.1 A simplified model of the effect of Sandy on trade costs
The first step is to express the bilateral trade costs during Sandy (T Sni) as a function of thebilateral trade costs in normal times (TNni ). Note that throughout this section, we will keepusing the distinction between the Sandy state of the world and the normal state of theworld introduced in section 4.2, with the same superscript convention (N for normal, S forSandy). Note also that, in order for our estimation method to be computationally feasible,we have to consider a less general version of equation (1) by modeling the flows at thedyad level (instead of the dyad-industry level).
Let’s begin by rewriting the bilateral transport cost (Tni) as the product of the averagetransport cost per km (denoted E(τni)) and the total road distance between i and n (gni):
Tni = E(Tni)gni (14)
During Sandy, the average transport cost per km changes. Consistently with what weexplained in section ??, we define κ as the ratio between the transport cost per km duringSandy in affected areas and the normal transport cost per km:
κ =E(Tni|ID = 1)
E(Tni|ID = 0)
where ID is a dummy variable equal to one when the concerned road segment is within anaffected county (IA/PA county) and the state of the world is Sandy. The average transportcost per km during Sandy is given by:
E(T Sni) = sniE(Tni|ID = 1) + (1− sni)E(Tni|ID = 0)
where sni is the proportion of road segments on the itinerary between i and n that arewithin the affected areas (i.e. road distance within IA/PA counties divided by total roaddistance of the itinerary). This can be rewritten as:
E(T Sni) = sniκE(Tni|ID = 0) + (1− sni)E(Tni|ID = 0)
= (sni(κ− 1) + 1)E(Tni|ID = 0)(15)
In the normal state of the world, ID = 0 for all road segments, so the average transport
26
cost is equal to the average transport cost in non affected areas:
E(TNni ) = E(Tni|ID = 0) (16)
Moreover, given that the dark trade costs are time unvarying, the relative variation intrade costs (τni) between the two states of the world is equal to the relative variation intransport costs (Tni):
τSniτNni
=T SniTNni
(17)
Combining equations (17), (14) and (15), the relationship between trade costs duringSandy and trade costs in normal times can be expressed as:
τSni = (sni(κ− 1) + 1)τNni (18)
Ideally, we should redetermine the optimal path for each value of κ, because both gni
and sni are affected by κ. However, for technical feasibility reasons, we have to disregardthis path adjustment, so that both sni and gni are constant. In other words, for the estima-tion of κ (and only for this part of our work), we assume that agents do not change theirpath, whatever the value of κ. While not realistic, this hypothesis does not compromise ourmain results because we deliberately choose the path in such a way that κ will be underes-timated. This downward bias on κ is obtained by using an upper bound for sni instead ofthe real value of sni. Indeed, overestimating the share of itinerary affected by the hurricaneleads to overestimate the effect of κ on trade flows and as a consequence to underestimateκ. To find an upper bound for sni, we modify our road raster so that the cost of passingthrough a road cell in an IA/PA area is 10−6 whereas this cost for a road cell outside IA/PAareas is left unaltered at 1. As a consequence, the least cost path algorithm will choose thepath that includes the largest possible share within IA/PA areas. We plot the distributionof sni in figure 9, in the appendix.
5.2 Simulate trade flows
5.2.1 In the normal state of the world
In order to simulate trade flows, we need to set the value of one parameter : ε. We setε = −5, for reasons that we explained in section 4.2.3. The choice of ε = −5 impliesρ = 0, 17, again for reasons detailed in section 4.2.3. Knowing ρ, we obtain τNni from
27
equation (10) :τNni = (dni)
ρ (19)
From the bilateral trade costs, we obtain the bilateral resistance term :
φNni = (τNni )ε
Given these bilateral resistance terms, we compute the multilateral resistance terms(ΩN
i and ΦNn ) that solve equations (2) and (3). For this, we use the contraction mapping
algorithm proposed by Head and Mayer17. Then, we obtain the simulated trade flows innormal times (XN
ni) using the structural gravity equation (1):
XNni =
Xi
ΩNi
Xn
ΦNn
φNni (20)
where for simplicity we omitted the subscript t, but we obtain a distinct simulated flow foreach considered trimester. Indeed, Xi and Xn are time varying.
5.2.2 During Sandy
Again the starting point is to compute the bilateral trade costs during Sandy. From equation(18), we know that for any value of κ, the bilateral trade costs during Sandy are a simplefunction of the bilateral trade costs in the normal state of the world. As a consequence, wetake the τNni simulated from equation (19) and multiply them by (sni(κ−1)+1) to obtain thebilateral trade costs during Sandy (τSni). Once we have τSni, we apply the same procedureas above to obtain first φSni, then ΩS
i and ΦSn, and finally the simulated trade flows during
Sandy (XSni).
Note that we slightly depart from structural gravity in the sense that we do not imposeXi =
∑nXni and Xn =
∑iXni. We take Xi and Xn as exogenous and set their values at the
level observed in the data. Our simulated trade flows therefore correspond to the modulartrade impact (MTI) of Sandy, and not to its general equilibrium trade impact (GETI) (cf.Head and Mayer (2014)).
17https://sites.google.com/site/hiegravity/stata-programs
28
5.2.3 Simulated trade flows
In the first three quarters there are no disruptions, hence the simulated trade flows aresimply equal to the simulated normal time flows which we obtain from equation (20). Fort = 1, 2, 3: Xnit = XN
nit.In the 4th quarter, the simulated trade flow is a weighted average of the simulated
Sandy trade flow and the simulated normal trade flow. For t = 4:
Xni,t=4 = χXSni,t=4 + (1− χ)XN
ni,t=4
where χ is the ratio of the duration of Sandy’s related disruptions over the duration of the4th quarter. For reasons explained in section XX, we set χ = 10/92
Given that we observe 11,73% of zero trade flows in our data, we select for each quarterthe 11,73% of flows that have the lowest values and set them to zero.
5.3 Indirect inference estimator
The idea behind indirect inference is to estimate parameters from an "auxiliary model" andfind the value of the parameter(s) of interest for which the auxiliary model parametersfrom the simulated data match the auxiliary model parameters from the real data. Here,our auxiliary model is the traditional fixed effect estimation of gravity equations, which weuse to estimate FEit and FEnt.
ln(Xnit) = FEni + FEnt + FEit + εnit (21)
For each origin and each destination, we compute the difference between the fixedeffect in the 4th quarter and the fixed effect in each other quarter, i.e. FEi,t=1 − FEi,t=4,FEi,t=2 − FEi,t=4, FEi,t=3 − FEi,t=4, FEn,t=1 − FEn,t=4, FEn,t=2 − FEn,t=4 and FEn,t=3 − FEn,t=4.
We first estimate equation (21) with real data. The estimated differences define avector θ0. This vector is the one we will try to match with our simulated data. Indeed,for simulated trade flows, the values of the change in fixed effects between the 4th quarterand the other quarters depend on κ, so we get a different vector of differences for eachvalue of κ. We call θ(κ) the vector of differences obtained for a given value of κ
We look for κ such that θ0 and θ(κ) are "as close as possible", i.e. the distance betweenθ0 and θ(κ) is minimized. At this point, it is necessary to choose a metric for measuringthe distance between θ0 and θ(κ). We simply rely on the Wald approach with a weighting
29
matrix W equal to the identity matrix.
κ = arg maxκ
(θ0 − θ(κ))′W (θ0 − θ(κ))
With ε = 5 and χ = 10/92, the value of κ that we obtain is 6.49, which we round downto 6. Remember that this is a lower bound for the true value of κ.
6 Results
6.1 Estimates of the transport cost part of distance elasticity (εα)
Estimates of the transport cost part of distance elasticity (εα) obtained using the specifica-tion described in equation (13) (section 3.2) are presented in table 5. With the completesample, we find an elasticity around -0.06. This estimate is robust to the omission of flowsoccurring within the same CFS area, of flows involving a rural CFS area (remainder), or ofboth former types of flows together, although it is estimated somewhat less precisely oncethese restrictions are imposed.
Table 5: Baseline results with κ = 6
(1) (2) (3) (4)VARIABLES Flow Flow Flow Flow
Distance -0.0564** -0.0498* -0.0519 -0.0575(0.0252) (0.0299) (0.0335) (0.0398)
Observations 374,549 358,825 152,254 142,218R-squared 0.773 0.753 0.785 0.756Rural excluded NO NO YES YESWithin CFS excl. NO YES NO YESClustering Dyad-industry Dyad-industry Dyad-industry Dyad-industryN. of clusters 126711 121538 52074 48734R2 0.773 0.753 0.785 0.756
Clustered standard errors in parentheses.Significance levels: *** p<0.01 ** p<0.05 * p<0.1.
30
6.2 Comparison with total distance elasticity (ερ)
We now want to compare this elasticity to the total distance elasticity of trade flows. As canbe seen from equation (11), we can estimate this total distance elasticity (ερ) by estimatingcross-sectional gravity equations for each of the three first quarters (where dni = dni). Wetherefore run three distinct regressions (one per quarter) with the following specification:
ln(Xnis) = FEns + FEis + β2 ln(dni) + εnis
where, according to equation (11), β2 = ερ. Table 6 gives our estimates for each semester.
Table 6: Cross-section results
(1) (2) (3)VARIABLES Flow T1 Flow T2 Flow T3
Distance -0.827*** -0.850*** -0.849***(0.00575) (0.00574) (0.00582)
Observations 122,316 121,065 117,239R-squared 0.469 0.471 0.474
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
These estimates lie in the lower part of the distribution of structural estimates foundin Head and Mayer (2014). This is consistent with the intuition that the spatial frictionsare lower for flows within a country than for international flows. More importantly forour purpose, the average of the estimates is around -0.84, which is much higher thanthe distance elasticity related to transport costs, and confirms the fact that transport costscannot account for the whole distance elasticity of trade flows in the US.
Nevertheless, there remains a last subject of concern that could threaten our claim: ifshipments are postponed, the estimates of εα presented in table 5 exhibit a downward bias,because some of the trade destruction effect of the increase in transport costs during Sandyis offset by an increase in trade after the hurricane. The two remaining subsections aretherefore devoted to presenting evidence going against the hypothesis that shipments havebeen postponed.
31
6.3 Decomposition between intensive and extensive margin
If shipments are postponed, then we expect an increase in the average shipment value forthe affected dyads. Indeed, it should be the case that the goods that could not be shippedbecause of the hurricane are added to shipments taking place after the hurricane, thusincreasing their value. Hence, looking at the decomposition of the trade elasticity betweenan intensive margin (average value per shipment) and an extensive margin (number ofshipments) can inform us about the presence, or the absence, of a postponement effect.More formally, let Nnist denote the number of shipments, then the intensive margin isdefined as Xnist
Nnist, while the extensive margin is Nnist. We therefore estimate the following
equations:
ln
(Xnist
Nnist
)= FEnis + FEnst + FEist + β5 ln(dnit) + εnist (22)
ln (Nnist) = FEnis + FEnst + FEist + β6 ln(dnit) + εnist (23)
Table 7: Intensive margin, with κ = 6
(1) (2) (3) (4)VARIABLES Value per shipmt Value per shipmt Value per shipmt Value per shipmt
Distance 0.0202 0.0299 0.0308 0.0357(0.0224) (0.0268) (0.0297) (0.0356)
Observations 374,549 358,825 152,254 142,218R-squared 0.727 0.720 0.734 0.726Rural excluded NO NO YES YESWithin CFS excl. NO YES NO YESClustering Dyad-industry Dyad-industry Dyad-industry Dyad-industryN. of clusters 126711 121538 52074 48734R2 0.727 0.720 0.734 0.726
Clustered standard errors in parentheses.Significance levels: *** p<0.01 ** p<0.05 * p<0.1.
Results are given in tables 8 and 7. As expected, most of the effect of Sandy wentthrough the extensive margin, a finding consistent with what volpemartincus2013 ob-served after the 2010 Chilean earthquake. The intensive margin was not significantly af-
32
Table 8: Extensive margin, with κ = 6
(1) (2) (3) (4)VARIABLES Nbr of shipmts Nbr of shipmts Nbr of shipmts Nbr of shipmts
Distance -0.0765*** -0.0798*** -0.0827*** -0.0932***(0.0184) (0.0218) (0.0243) (0.0289)
Observations 374,549 358,825 152,254 142,218R-squared 0.857 0.829 0.872 0.836Rural excluded NO NO YES YESWithin CFS excl. NO YES NO YESClustering Dyad-industry Dyad-industry Dyad-industry Dyad-industryN. of clusters 126711 121538 52074 48734R2 0.857 0.829 0.872 0.836
Clustered standard errors in parentheses.Significance levels: *** p<0.01 ** p<0.05 * p<0.1.
fected. We interpret this as evidence against a postponement effect.
6.4 Restriction to most regular shipments
Another way to test for the absence of postponement effect is to focus on shipments thatneed to be sent regularly, and can hardly be delayed. We do this using two differentmethods. The first one consists in selecting the industries that exhibit the most regularshipment patterns (with the implicit assumption that if they send shipments so regularly, itmust mean that it is hard or costly for them to delay shipments). The second method usesthe fact that some shipments are temperature controlled, and therefore cannot be delayed.
6.4.1 Selected industries
We select the industries for which the dyad-industry trade flows have the highest stabilityover the three first quarters. A natural criterion to determine a low volatility is the coef-ficient of variation. However, given the large number of zero trade flows, this coefficientwould be very low for industries that represent a small share of trade flows. Therefore wetake another criterion into account: the share of zero trade flows. We mix these two crite-
33
ria, giving the same weight to each of them and select the ten industries with the highestranking. A list of the selected industries is given in the appendix, table 9, in section D. Thespecification is the same as for the baseline regression (equation (13)). Table 9 displaysour results.
Table 9: Selected industries
(1) (2) (3) (4)VARIABLES Flow Flow Flow Flow
Distance -0.0737** -0.0749* -0.0683 -0.0769(0.0358) (0.0418) (0.0475) (0.0554)
Observations 153,764 149,870 60,696 58,114R-squared 0.757 0.738 0.769 0.739Rural excluded NO NO YES YESWithin CFS excl. NO YES NO YESClustering Dyad-industry Dyad-industry Dyad-industry Dyad-industryN. of clusters 49643 48393 19736 18893R2 0.757 0.738 0.769 0.739
Clustered standard errors in parentheses.Significance levels: *** p<0.01 ** p<0.05 * p<0.1.
We see that the coefficient on distance for the most regular industries is not very dif-ferent from the one estimated with all industries. Moreover, it is still far from the totaldistance elasticity (-0.84).
6.4.2 Temperature controlled shipments
Our CFS data includes a dummy variable indicating whether the shipment is "temperaturecontrolled", i.e. whether it is carried in a vehicle designed to maintain the shipment ata certain temperature18. If a shipment is temperature controlled, it suggests that it tendsto depreciate quickly over time and therefore that it is very costly to postpone it. As a
18A temperature controlled shipment is defined as a shipment that is transported in a vehi-cle or container that regulates the temperature while en route (such as heating and refriger-ation) or maintaining the temperature of the commodity at the time of loading (such as in-sulation). Source: http://www.rita.DOT.gov/bts/sites/rita.DOT.gov.bts/files/publications/commodity_flow_survey/html/def_terms.html, consulted on 27/05/2016
34
consequence, if there is a postponement effect, then dyad-industries in which there aremore temperature controlled shipments should be less affected by this postponement, andthus have a higher distance elasticity, because cancelled shipments cannot be postponed.
We therefore compute the average share of temperature controlled shipments duringthe three first quarters for each dyad-industry and interact this share with distance. Ifthere is a postponement effect, the coefficient on this interaction should be negative andsignificant. We estimate the following specification:
ln(Xnist) = FEnis + FEnst + FEist + β1 ln(dnit) + β2(ln(dnit) ∗ Snis) + εnist (24)
where Snis is the share of temperature controlled shipments during the three first quarters.Table 10 gives the results of this regression. As can be seen, the coefficient on the
interaction term is not significant, which is an additional clue that there is no postponementeffect.
Table 10: Temperature controlled goods
(1) (2) (3) (4)VARIABLES Flow Flow Flow Flow
Distance -0.0544** -0.0471 -0.0468 -0.0552(0.0263) (0.0312) (0.0348) (0.0415)
Distance*Sh. temp. contr. -0.0216 -0.0280 -0.0541 -0.0236(0.0743) (0.0811) (0.105) (0.114)
Observations 374,549 358,825 152,254 142,218R-squared 0.773 0.753 0.785 0.756Rural excluded NO NO YES YESWithin CFS excl. NO YES NO YESClustering Dyad-industry Dyad-industry Dyad-industry Dyad-industryN. of clusters 126711 121538 52074 48734R2 0.773 0.753 0.785 0.756
Clustered standard errors in parentheses.Significance levels: *** p<0.01 ** p<0.05 * p<0.1.
35
7 Conclusion
Using hurricane Sandy as a natural experiment providing exogenous variations in bilateraltransport costs, we are able to estimate the effect of transport costs on trade flows. Ouranalysis requires the estimation of an "overcost parameter" corresponding to the increasein transport costs in the areas affected by Sandy after the hurricane, which we realize usinga method inspired from indirect inference. Finally, we are able to conclude that transportcosts are not sufficient to explain the whole distance elasticity of trade flows within the US,which suggests the existence of dark trade costs even at the intra-national scale.
36
References
Blake, E. S. et al. (2013). Tropical Cyclone Report Hurricane Sandy. National HurricaneCenter.
Combes, P.-P. and M. Lafourcade (2005). “Transport costs: measures, determinants, andregional policy implications for France”. Journal of Economic Geography, 5, pp. 319–349. URL: http://EconPapers.repec.org/RePEc:oup:jecgeo:v:5:y:2005:i:3:p:319-349.
Disdier, A.-C. and K. Head (2008). “The Puzzling Persistence of the Distance Effect onBilateral Trade”. The Review of Economics and Statistics, 90, pp. 37–48. URL: http ://EconPapers.repec.org/RePEc:tpr:restat:v:90:y:2008:i:1:p:37-48.
Duranton, G., P. Morrow and M. Turner (2014). “Roads and Trade: Evidence from the US”.Review of Economic Studies, 81, pp. 681–724. URL: http://EconPapers.repec.org/RePEc:oup:restud:v:81:y:2014:i:2:p:681-724.
Felbermayr, G. and F. Toubal (2010). “Cultural proximity and trade”. European EconomicReview, 54, pp. 279–293. URL: http://EconPapers.repec.org/RePEc:eee:eecrev:v:54:y:2010:i:2:p:279-293.
Feyrer, J. (2011). Distance, Trade, and Income – The 1967 to 1975 Closing of the Suez Canalas a Natural Experiment. 2011 Meeting Papers 1438. Society for Economic Dynamics.URL: http://EconPapers.repec.org/RePEc:red:sed011:1438.
Head, K. and T. Mayer (2013). “What separates us? Sources of resistance to globalization”.Canadian Journal of Economics, 46, pp. 1196–1231. URL: http://EconPapers.repec.org/RePEc:cje:issued:v:46:y:2013:i:4:p:1196-1231.
— (2014). “Gravity Equations: Workhorse,Toolkit, and Cookbook”. In: Handbook of In-ternational Economics. Vol. 4. Elsevier. Chap. Chapter 3, pp. 131–195. URL: http://EconPapers.repec.org/RePEc:eee:intchp:4-131.
Hortacsu, A., F. A. Martínez-Jerez and J. Douglas (2009). “The Geography of Trade inOnline Transactions: Evidence from eBay and MercadoLibre”. American Economic Jour-nal: Microeconomics, 1, pp. 53–74. URL: http://EconPapers.repec.org/RePEc:aea:aejmic:v:1:y:2009:i:1:p:53-74.
Hummels, D. (2007). “Transportation Costs and International Trade in the Second Eraof Globalization”. Journal of Economic Perspectives, 21, pp. 131–154. URL: http : / /EconPapers.repec.org/RePEc:aea:jecper:v:21:y:2007:i:3:p:131-154.
37
Hummels, D. and G. Schaur (2013). “Time as a Trade Barrier”. American Economic Review,103, pp. 2935–59. URL: http://EconPapers.repec.org/RePEc:aea:aecrev:v:103:y:2013:i:7:p:2935-59.
Lameli, A. et al. (2015). “Same Same But Different: Dialects and Trade”. German EconomicReview, 16, pp. 290–306. URL: http://EconPapers.repec.org/RePEc:bla:germec:v:16:y:2015:i:3:p:290-306.
Limao, N. and A. J. Venables (2001). “Infrastructure, geographical disadvantage, transportcosts, and trade”. The World Bank Economic Review, 15, pp. 451–479.
Mildenhall, S. et al. (2013). Hurricane Sandy Event Recap Report. AON Benfield.Volpe Martincus, C. and J. Blyde (2013). “Shaky roads and trembling exports: Assessing
the trade effects of domestic infrastructure using a natural experiment”. Journal of In-ternational Economics, 90, pp. 148–161. URL: http://EconPapers.repec.org/RePEc:eee:inecon:v:90:y:2013:i:1:p:148-161.
Volpe Martincus, C. et al. (2014). “How do transport costs affect firms’ exports? Evi-dence from a vanishing bridge”. Economics Letters, 123, pp. 149–153. URL: http://EconPapers.repec.org/RePEc:eee:ecolet:v:123:y:2014:i:2:p:149-153.
Wrona, J. (2015). Border effects without borders: What divides Japan’s internal trade? DICEDiscussion Papers 185. Düsseldorf Institute for Competition Economics (DICE), Univer-sity of Düsseldorf. URL: http://EconPapers.repec.org/RePEc:zbw:dicedp:185.
38
Appendices
A Data description
Table 11: List of urban CFS areas ranked by export value
CFS Area Exports (M$) # destinations Imports (M$) # originsLos Angeles 542,064 128 442,232 128Dallas 410,957 128 375,901 128Chicago (IL part) 383,354 128 316,610 128New York (NJ part) 247,206 128 200,476 127Houston 206,714 126 220,756 128Atlanta 194,754 129 195,717 127New York (NY part) 191,548 126 244,328 126San Jose 176,665 127 170,256 125San Antonio 159,823 114 173,115 123Detroit 151,106 125 172,690 125Boston (MA part) 132,273 127 161,687 125Minneapolis 129,143 127 118,134 124Columbus 116,392 125 116,407 121Hartford 115,541 115 78,581 112Cleveland 101,044 128 108,495 126Seattle 94,600 119 117,294 125Baltimore 91,526 118 72,175 125Miami 90,327 124 132,269 126Indianapolis 89,229 123 95,822 122Philadelphia (PA part) 88,497 127 105,237 126Milwaukee 88,006 127 74,314 121New York (CT part) 87,837 125 77,141 114Greensboro 84,637 124 64,548 118Denver 80,080 119 81,927 125Philadelphia (NJ part) 79,480 120 52,582 119Phoenix 78,989 110 97,223 124Pittsburgh 78,730 125 84,620 122Nashville 69,271 123 65,462 122Tampa 69,051 115 64,279 120Portland (OR part) 68,264 120 65,871 122Charlotte 67,046 126 64,012 122Birmingham 63,276 124 69,594 118
Continued on next page
39
Table 11 – continued from previous pageCFS Area Exports (M$) #destinations Imports (M$) # originsMemphis 62,386 123 55,414 116St-Louis (MO part) 61,215 126 55,693 123Grand Rapids 61,041 124 49,631 116Salt Lake City 59,108 122 74,972 123Tulsa 58,589 118 43,244 116Austin 58,240 109 73,495 120Louisville 56,714 124 71,496 119Richmond 55,188 117 47,863 116Sacramento 51,993 97 46,030 117Raleigh 51,891 123 41,138 118Cincinnati (OH part) 51,204 124 54,477 123New York (PA part) 51,144 120 35,189 115San Diego 49,932 120 64,954 120Greenville 46,990 127 46,454 117Kansas City (KS part) 45,468 122 33,861 119Fort Wayne 45,301 120 31,329 110Kansas City (MO part) 44,629 122 50,654 119Orlando 44,517 120 52,626 123Jacksonville 42,242 113 41,036 116Buffalo 41,294 122 37,281 113Albany 37,861 116 35,996 112Boston (RI part) 36,842 118 30,254 101Dayton 35,604 118 37,359 114Rochester 34,752 121 33,941 110Cincinnati (KY part) 34,130 115 21,156 106Oklahoma City 33,578 116 47,405 120Knoxville 32,800 115 30,593 115Beaumont 31,019 88 22,055 78New Orleans 30,408 104 41,616 116Washington (VA part) 30,179 103 50,152 115Chicago (IN part) 30,167 117 33,467 102Omaha 28,040 117 29,275 111Wichita 27,777 122 27,268 104Fresno 26,625 98 21,152 99Boston (NH part) 26,451 116 38,625 111Philadelphia (DE part) 23,318 110 28,025 104Baton Rouge 22,783 99 23,319 106Virginia Beach 21,708 115 36,504 118
Continued on next page
40
Table 11 – continued from previous pageCFS Area Exports (M$) #destinations Imports (M$) # originsLas Vegas 21,064 94 32,562 116St-Louis (IL part) 20,841 116 28,686 102Washington (MD part) 19,111 90 42,790 114Savannah 18,533 93 12,985 93Tucson 16,832 99 16,620 105El Paso 16,563 101 22,445 116Laredo 15,770 23 27,639 117Charleston 15,246 112 17,085 100Corpus Christi 13,367 46 19,831 81Mobile 10,940 105 19,520 95Portland (WA part) 9,178 100 9,322 76Lake Charles 4,888 64 6,904 62Washington (DC part) 2,386 18 8,812 78Total 6,395,276 9,398 6,339,951 9,492
Table 12: List of remainder CFS areas ranked by export value
CFS Area Exports (M$) # destinations Imports (M$) # originsRem, of Texas 284,705 126 292,913 127Rem, of Pennsylvania 199,180 129 185,466 127Rem, of Illinois 165,996 126 149,311 125Rem, of Wisconsin 153,842 127 143,950 125Rem, of Iowa 144,175 127 143,787 125Rem, of Ohio 130,466 127 112,213 123Rem, of Indiana 126,331 128 128,343 124Rem, of Mississipi 112,931 127 92,872 124Rem, of N, Carolina 112,255 128 87,318 125Rem, of Arkansas 95,135 126 103,843 124Rem, of Kentucky 93,133 126 91,022 123Rem, of Kansas 92,449 123 83,236 115Rem, of Michigan 92,381 126 88,959 123Rem, of New York 90,642 124 71,392 120Rem, of Alabama 90,361 126 86,790 122Rem, of Georgia 89,352 126 87,019 123Rem, of Virginia 84,327 127 70,645 123Rem, of California 80,732 113 96,672 123
Continued on next page
41
Table 12 – continued from previous pageCFS Area Exports (M$) # destinations Imports (M$) # originsRem, of Tennessee 78,249 127 63,660 123Rem, of Florida 70,242 120 103,297 125Rem, of Missouri 69,771 125 82,783 122Rem, of Minnesota 61,816 124 69,345 114Rem, of S, Carolina 55,104 127 57,704 122Rem, of Nebraska 50,515 119 50,077 112Rem, of Louisiana 49,775 115 67,341 119Rem, of S, Dakota 46,216 119 37,273 106Rem, of Oklahoma 42,004 118 59,348 110Rem, of W, Virginia 37,157 124 46,109 118Rem, of New Mexico 34,349 109 44,012 117Rem, of Washington 34,067 111 40,768 113Rem, of Colorado 31,795 117 42,781 120Rem, of N, Dakota 29,991 106 42,592 107Rem, of Maine 29,566 109 33,835 109Rem, of Idaho 27,306 111 35,520 112Rem, of Connecticut 26,169 98 29,575 82Rem, of Nevada 25,986 106 23,849 111Rem, of Maryland 24,497 114 22,859 104Rem, of Oregon 23,644 109 34,439 109Rem, of Massachussets 20,541 117 28,047 105Rem, of Vermont 17,791 116 20,200 96Rem, of Montana 16,781 86 26,097 108Rem, of Wyoming 12,805 85 20,991 102Rem, of Arizona 11,658 84 20,767 101Rem, of Delaware 9,646 70 6,823 62Rem, of Utah 9,336 69 11,480 87Rem, of New Hampshire 4,900 100 8,071 69Total 3,190,068 5,292 3,245,392 5,198
42
Table 13: Descriptive statistics by NAICS codes
SectorNAICS
Value Weight # shipmentsValue per
# obs.Value per
% zero flows # dyads # exp. # imp.Shipment
2007 shipment obs. distanceManufacturing
Mining 212 41,053 1,566,560 71.2 576.2 129,227 0.3 2.9 2,233 124 129 8.4Food manufacturing 311 660,284 400,506 122.9 5,371.8 169,855 3.9 7.3 9,616 129 129 14.9Beverage and tobacco 312 132,326 138,662 21.2 6,240.7 44,631 3.0 13.5 3,786 121 129 11.4Textile mills 313 23,906 5,354 2.6 9,346.8 19,424 1.2 27.6 3,605 109 129 14.6Textile product mills 314 17,666 4,319 6.5 2,716.5 20,056 0.9 31.3 4,081 122 129 14.7Apparel 315 8,018 374 1.3 6,157.4 4,441 1.8 41.1 1,308 95 127 15.8Leather and allied product 316 2,632 308 0.5 5,657.4 3,298 0.8 43.6 1,051 87 127 14.8Wood product 321 66,313 167,786 17.2 3,853.3 92,055 0.7 9.1 5,480 126 129 11.8Paper 322 144,812 103,557 18.0 8,027.0 88,850 1.6 11.6 7,176 121 129 12.7Printing and related activities 323 57,292 18,407 51.8 1,105.9 75,005 0.8 14.6 6,875 129 129 13.3Petroleum and coal products 324 178,434 397,882 27.7 6,433.4 50,680 3.5 13.0 3,853 121 129 10.5Chemical 325 408,786 267,124 49.1 8,318.0 154,292 2.6 9.3 11,014 128 129 15.1Plastics and rubber 326 186,738 48,383 30.6 6,111.1 115,772 1.6 11.8 10,292 127 129 14.8Non-metallic mineral product 327 82,284 581,291 47.2 1,744.4 110,911 0.7 8.6 5,936 129 129 11.3Primary metal 331 189,467 114,581 12.3 15,406.4 61,300 3.1 15.3 6,639 123 129 13.3Fabricated metal product 332 266,445 80,452 67.4 3,954.8 139,980 1.9 10.2 10,129 128 129 14.9Machinery 333 270,753 27,359 27.1 9,991.8 64,297 4.2 21.9 9,999 127 129 15.8Computer and electronic product 334 105,257 2,948 8.9 11,826.7 18,428 5.7 37.5 4,831 126 129 17.0Electrical equipment, appliances 335 87,805 13,273 10.7 8,233.2 36,229 2.4 29.1 7,487 120 129 16.2Transportation equipment 336 453,886 55,024 26.8 16,945.2 51,146 8.9 18.2 6,107 123 129 14.2Furniture and related 337 61,917 12,608 23.7 2,616.5 54,007 1.1 21.4 7,801 127 129 14.7Miscellaneous 339 71,732 6,764 26.2 2,738.9 36,088 2.0 28.7 6,866 128 129 16.5
WholesalersMotor vehicle and parts 4,231 415,850 53,151 680.8 610.8 81,055 5.1 9.7 4,452 127 129 10.6Furniture and home furnishing 4,232 57,211 13,926 45.8 1,249.1 44,206 1.3 16.6 4,312 120 129 14.0Lumber and other construction materials 4,233 108,051 286,664 96.1 1,124.3 112,067 1.0 4.8 3,251 128 129 7.6Commercial equip. 4,234 214,771 15,653 160.5 1,338.5 36,554 5.9 19.8 4,206 125 129 14.1Metal and mineral 4,235 184,693 125,692 77.7 2,377.6 97,199 1.9 8.4 5,023 126 129 10.1Electrical and electronic goods 4,236 265,768 21,182 283.0 939.0 73,628 3.6 12.7 4,998 128 129 13.8Hardware and plumbing 4,237 105,996 15,783 153.8 689.2 103,853 1.0 7.4 4,188 129 129 10.3Machinery, equipment and supplies 4,238 271,964 54,421 226.3 1,201.7 96,765 2.8 11.2 6,002 129 129 12.8Miscellaneous durable goods 4,239 110,164 145,324 50.9 2,162.5 56,882 1.9 14.9 4,943 129 129 12.4Paper and paper products 4,241 80,828 33,270 100.5 803.9 64,898 1.2 8.5 3,142 127 129 8.9Drugs and druggists’ sundries 4,242 300,474 9,946 91.8 3,271.7 22,351 13.4 19.4 2,607 120 129 12.3Apparel and piece goods 4,243 78,127 6,728 28.8 2,713.9 16,462 4.7 29.8 3,214 114 129 15.7Grocery and related 4,244 618,210 295,168 388.5 1,591.5 164,889 3.7 4.4 4,921 129 129 11.5Farm product raw material 4,245 115,921 271,231 18.4 6,300.3 35,241 3.3 7.5 1,436 107 129 10.1Chemical and allied products 4,246 136,334 83,500 100.0 1,362.8 68,623 2.0 12.0 4,867 123 129 11.0Petroleum and petroleum products 4,247 1,186,285 1,076,171 197.3 6,011.4 75,499 15.7 3.1 1,473 127 129 5.0Beer, wine, and distilled alcoholic 4,248 118,976 49,759 128.2 927.7 84,463 1.4 1.9 990 128 129 8.6Miscellaneous non-durable goods 4,249 239,482 147,828 113.9 2,103.0 98,095 2.4 8.1 4,594 127 129 11.5Electronic shopping and mail-order houses 4,541 48,548 5,763 206.7 234.9 8,223 5.9 45.9 2,757 117 129 15.5Warehousing and storage 4,931 1,086,774 247,565 128.2 8,479.8 79,230 13.7 10.5 5,576 124 129 11.5Newspaper, periodical and book 5,111 36,929 10,078 352.2 104.8 29,518 1.3 12.0 1,817 121 129 11.3Direct selling establishments 45,431 35,161 32,262 69.5 506.0 100,204 0.4 0.5 529 127 129 0.8Corporate, subsidiary, and regional offices 551,114 251,020 78,162 42.2 5,943.9 19,981 12.6 22.0 2,670 108 129 11.5Note: Shipment values are expressed in M$; weight is in thousand US tons (kt); # shipments is expressed in millions; shipment distance is in hundreds of Km.
43
Figure 9: Distribution of sni
01.
0e+
042.
0e+
043.
0e+
044.
0e+
04F
requ
ency
0 .2 .4 .6 .8 1Share in IA/PA
44
B Maps
Figure 10: Disruptions due to hurricane Sandy in New Jersey and New York
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!( !(
!(
!(
!(
!(
!(
!(
!( !(!(
!(!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(
!(!(
!(!(!(!(!(!(!(!(
!(
!(
!(!(
!(!(!(!(
!(
!(
!(!(
!(
!(
!(
!(!(!(
!(!(
!(
!(!(!(!(
!(
!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(!(!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!( !(
!(!(!(!(!(!(!(
!(
!(!(!(
!(!(
!(!(!(
!(!(!(
!(
!(
!(
!(!(!(!(!( !(
!(
!(!(
!(
!(
!(
!(!(
!(!(
!(!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
!(
!(
!(!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(!(
!(!(!(
!(
!(
!(
!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(!(!(
!(
!(
!(
!(!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!( !(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!( !(
!(!(
!(
!(
!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(!(!(
!(!(
!(
!(!(
!(
!(!(
!(
!(
!(
!( !(!(
!(
!(
!(
!(
!(!( !(
!(
!(
!(
!( !(!(
!(
!(
!( !(!(
!(!(
!(
!(
!(
!(!(
!(!(!(
!(!(!(
!(!(!(!(!(!(
!(!(
!(!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(!(!(
!(!(
!(
!(!(
!(!(
!(!(
!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(!(
!(
!(!(
!(!(
!(
!(!(
!(
!(!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(!(!(
!(!(
!(!(
!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(!(!(!(
!(
!(
!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!( !(
!(!( !(
!(
!(
!(!(!(
!(
!(!(
!(
!(!(!(
!(
!(
!(!(
!(!(!(!(!(!(
!(
!(
!(
!(
!(!(
!(!( !(
!(
!(
!(
!(
!(!(!(
!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!( !(!(
!(
!(
!(!(!(
!(
!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(
!(!(
!(
!(!(
!(
!(
!(
!(
!(!(
!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(!(!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(!(
!(!(!(!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(!(
!(
!(!(
!(!(
!(
!(
!(
!(!(
!(
!(!(!(
!(
!(
!(!(!(
!(
!(
!(
!(!( !(!(
!(!(
!(
!(!(
!(!(
!(
!(
!(!(!(!(
!(
!(
!(!(!(
!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(!(
!(
!(!(
!(!(
!(
!(!(
!(!(!( !(
!(
!(
!(
!(!(
!(!(
!( !(!(!(!( !(
!(
!(
!(!(
!(!( !(!(
!(
!(
!(
!(
!(
!(
!(!(
!(!(
!(
!( !(!(
!(
!( !(
!(
!(!(
!(
!(
!(
!(!(
!(!(
!(!(!(!(!(
!(
!(
!(
!( !(
!(
!(
!(
!(
!(
!(
!(
!(!( !(!(
!(!(
!(
!(!(
!(
!(
!(!(
!(
!(
!(
!(
!(!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!( !(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(
!(!(
!(!(!(
!(
!(
!(
!(
!(
!(
!(
!(
!(!(!(!(
!(
!(
!(
!(
!(
!(!(
!(
!(
!(!(!(!(
!(
!(
!( !(!(!(
!(
!(
!(
!(
!(
!(!(!(
!(
!(
!(
Note: Red lines show the highway network. Blue points represent disruptions recorded by NJ orNY DoT. Yellow surfaces correspond to Core-Based Statistical Areas (CBSA).
45
Figure 11: Road raster in normal times
!(
Note: Blue cells correspond to a cost of 1. Green cells correspond to an infinite value (i.e., no road).
46
Figure 12: Road raster during Sandy
!(
Note: Blue and green cells have the same cost as in figure 11. Red cells have a cost of κ.
47
C Graphs
Figure 13: Number of disruptions recorded by NJ DOT
total disruptions
hurricane relatedother disruptions
010
020
030
0#
disr
uptio
ns
Oct 27 Oct 29 Oct 31 Nov 2
48
Figure 14: Number of disruptions recorded by NY DOT
total disruptions
other disruptions
hurricane related
050
100
150
200
250
# di
srup
tions
Oct 27 Oct 29 Oct 31 Nov 2
49
D List of industries selected in section 6.4.1
Table 14: List of selected industries
Sector NAICS 2007Food manufacturing 311Wood product manufacturing 321Paper manufacturing 322Chemical manufacturing 325Plastics and rubber products manufacturing 326Nonmetallic mineral product manufacturing 327Primary metal manufacturing 331Fabricated metal product manufacturing 332Grocery and related product merchant wholesalers 4244Petroleum and petroleum products merchant wholesalers 4247
50
