Who Gets the Goodies? Overlapping Interests and the Geography of Aid for Trade Allocation in Bangladesh

Abstract The Sustainable Development Goal principle of “leaving no one behind” has led to increased attention being paid to patterns of intra-country allocation of foreign aid. We contribute to these efforts by considering a particular type of foreign aid, Aid for Trade (AfT), to discern allocation objectives of aid. We match a novel, geo-coded, dataset on over 11,000 Bangladeshi exporting firms to over one thousand AfT project locations in Bangladesh similarly geo-coded by AidData and expanded by ourselves. We use this matched data to employ spatial techniques that evaluate political economy logics of aid allocation, wherein AfT is functionally targeted towards exporting firms, is allocated based on prebendalism, and/or is directed to high poverty areas. Our analysis finds support that AfT is allocated based on functional or prebendalist logics. The results for poverty are more nuanced. When considered in a stand-alone fashion, poverty is associated with a smaller likelihood of allocation. However, some evidence suggests that when the other logics are present, the impact of poverty on allocation becomes positive. These findings suggest that the politics of aid allocation is a nuanced and intricate dance with multiple overlapping or competing logics.


Introduction
Understanding patterns of subnational aid allocation is important in ensuring that no one is "left behind" as the world strives to achieve the Sustainable Development Goals (SDGs) by 2030. While comparative measures of performance and (in)equality in aid allocation have historically been at the inter-state level (Ceriani & Verme, 2014), understanding intra-state (subnational) development performance is equally important to make certain that no pockets of deprivation are glossed over in aggregate country-level efforts and measures. To this end, a burgeoning literature has emerged examining subnational development inputs and outcomes (Carnegie, Howe, Lichtenheld, & Mukhopadhyay, 2019;Civelli, Horowitz, & Teixeira, 2018;Isaksson, 2019;Isaksson & Kotsadam 2018;Kotsadam, Østby, Rustad, Tollefsen, & Urdal, 2018;Mitchell & Soni, 2021;Nunnenkamp, € Ohler, & Sosa Andr es, 2017;Saltnes, Brazys, Lacey, & Pillai, 2020). With specific regards to aid allocation, recent analyses (Briggs, 2017(Briggs, , 2018a(Briggs, , 2018b(Briggs, , 2021Song, Brazys, & Vadlamannati, 2021) have suggested that sub-national distribution patterns may not always be pro-poor. Other work has advanced political-economy rationales for what might instead explain sub-national targeting including preference to political leaders' birth regions (Dreher, Fuchs, Holder, Parks, & Raschky, 2019), prebendalism (Bommer, Dreher, & Perez-Alvarez, 2018), or electoral incentives (Masaki, 2018). This paper adds to these efforts by considering sub-national patterns of aid allocation in Bangladesh. Specifically, the paper considers allocation of the so-called "Aid for Trade" (AfT), an aid initiative that stems from discussions in the World Trade Organization's (WTOs) Doha Development Round of trade talks and which is predicated on the idea that trade can lift individuals from poverty. Broadly, AfT is intended to increase the exporting activity of recipient countries by developing trade-related infrastructure, increasing capacity and compliance with trade-related rules and regulations, and by developing specific export-oriented industries (H€ uhne, Meyer, & Nunnenkamp, 2014). As of 2019, UNIDO currently classifies some 30% of all official development assistance (ODA), valued at over $30 billion annually, as AfT. 1 The focus on AfT allows for an exploration if patterns of aid allocation can follow a functional, or technocratic, logic instead of, or in addition to, other prevailing logics of aid allocation. Most aid allocation literature assumes that allocation decisions follow from donor and/or recipient government interests. Examining the locating of AfT allows us to also consider the role of private (domestic) interests in understanding allocation location decisions. To conduct this analysis, we present a novel, geo-coded, dataset of the population of over 11,000 exporting firms in Bangladesh. This is complemented with an expanded version of AidData's geo-coded aid projects in the country. Bangladesh is an excellent candidate country for this type of study as it has numerous exporting enterprises as well as a reasonable geographic distribution of these firms. While the firm data lacks temporal information, and thus inhibits us from pursuing a causal strategy, the data permit a descriptive evaluation of if firms and projects conform to a functional economic geography explanation of co-location. We are then able to match these observations to other data that could be suggestive of AfT allocation patterns conforming to local need and/or local prebendalism.
In general, we find support that the geography of AfT allocation coincides with both the presence of firms and the presence of the ruling party's local political representation. However, we find that, when considering need at a micro-level, it appears to be the less needy areas that receive AfT, all else equal. This finding coincides with those of Briggs (2017Briggs ( , 2018aBriggs ( , 2018b. That said, we find that poverty may play a conditional role, particularly when considered in conjunction with the presence of exporting firms and members of parliament. Collectively, these correlations suggest that it is important to consider a full combination of logics that may drive the allocation of aid projects. This supports previous arguments that aid allocation decisions are indeed the result of a "development dance" between not only donor actors and their government partners, but also domestic private interests (Swedlund, 2017).

Functional subnational aid allocation
The political economy of aid allocation literature has largely rested on assumptions about the degree of "aid capture" or "donor control" (Milner, Nielson, & Findley, 2016). Donor and recipient countries (and heterogenous interests therein) may often have different preferences over where and how aid is allocated. 2 As Swedlund (2017) argues, reconciling these differences often takes place through an intricate "dance", wherein each party utilizes the resources at their disposal in an attempt to sway allocation decisions in line with their preferences. In practice, this usually involves a donor actor having one (or more) objectives with a particular development project or initiative. 3 Depending on the donor, the particulars of the project, including the Who gets the goodies? 243 locating of project works, are then negotiated with a recipient government, who may have other (at times competing) interests and objectives.
These studies have clustered around two distinct allocation interests. The first are investigations into the salience of recipient government prebendalism in explaining aid allocation patterns. Prebendalism, clientelism and cousinage have long been used as heuristics for understanding the political economy of resource allocation (Dunning & Harrison, 2010;Lewis, 1996;Szeftel, 2000). The development of highly granular, sub-national, data has allowed a fresh look at these issues. These studies have largely found that aid resources are directed to areas that are politically important. Both Briggs (2014) and Jablonski (2014) show that aid in Kenya was directed to areas of partisan and ethnic support. Likewise, Dreher et al. (2019) suggest that Chinese aid tends to favor the birth regions of leaders but show no similar bias for World Bank projects. In contrast, Knutsen and Kotsadam (2020) find that local World Bank aid can increase support for an incumbent but find no such effect for Chinese aid. An interesting corrective, however, is Masaki (2018) who finds that instead of targeting areas of political support, aid targets areas with a high proportion of opposition in Tanzania. Masaki (2018) argues that this is due to the incumbents' limited information on the geographic distribution of swing voters, and thus, is an attempt to sway the preferences of weak opposers.
A second strand of this literature explores the importance of poverty reduction or need orientation as a donor interest in determining aid allocations. As € Ohler, Negre, Smets, Massari, and Bogetic (2019) explain, especially since the advent of SDG goal 10 (inequality), donors have been increasingly interested in targeting their aid to ensure that it reaches the bottom 40 per cent of the population within a country. This targeting is based on the realization that in many cases the inequalities within a country are as stark as those across countries. Accordingly, in order to reduce poverty and ensure "no one is left behind", donors may seek to target particularly needy regions within a country. However, the success of these efforts remains unclear. Indeed, € Ohler et al. (2019) find mixed evidence of subnational poverty targeting and also suggest that seemingly "pro-poor" targeting may simply be more about population density rather than need. Likewise, in a recent study of disaster assistance to Nepal following the 2015 earthquake, Eichenauer, Fuchs, Kunze, and Strobl (2020) find little evidence that need determined aid allocation once accounting for earthquake damage. Briggs (2017Briggs ( , 2018aBriggs ( , 2018b series of investigations, across several countries, find that foreign aid flows to richer rather than poorer regions, in contrast to the stated aims of the donor actors. In investigating what might explain these patterns, Briggs (2021) finds that donors believe that recipient governments are interested in the political targeting of aid, complementing the findings above, but also that donors find it harder to implement aid in remote areas which are also more likely to be poor.
However, recent literature has also begun to consider interests beyond donor institutions and (central) recipient governments. Song et al. (2021) consider how the allocation of World Bank education projects in India was determined, amongst other things, by areas heavily populated by politically empowered Scheduled Caste and Scheduled Tribe members. Likewise, as Dionne (2017) demonstrates in discussing the allocation and implementation of AIDS interventions, donor and central government aims may ultimately be frustrated at the local level by authority figures with different preferences or concerns. The importance of "third wheel" actors in the development dance complicates the logics of aid allocation but may also provide an important missing piece. As the "third wheels" are, in most instances, the ultimate beneficiaries of development interventions it only stands to reason that they will try to sway the allocation of those resources, either through directly engaging and petitioning the donors or government actors, or by using informal influence or public pressure.
In this paper, we focus on the possibility that export-oriented firms may try to sway the allocation of Aid for Trade. While AfT is a broad church, many AfT projects may confer specific benefits on firms be it by lowering production or transportation costs via improved infrastructure, increasing productivity by upskilling labor forces, or via industry or sector-specific support (Brazys & Lightfoot, 2016). Focusing on AfT allows us to consider if aid is allocated along functional lines. This notion harkens to an earlier discussion which understood functional aid as that which is based on "non-controversial … technical problem solving" rather than "political" motivations of "security, sovereignty, and prestige" (Riggs, 1980, pp. 330). Aid allocated in a "functional" manner is, ostensibly, tasked with discharging an objective and technical purpose.
While other work such as Marty, Dolan, Leu, and Runfola (2017), Lordemus (2019), and Song et al. (2021) moves in the direction of functional, technical, aid allocation by assessing if health and education aid flow to areas in greatest need of improvements in those areas, respectively, it is difficult to extract health or education need from general poverty as the latter is likely to be highly correlated with the former. However, by restricting the focus to "Aid for Trade" projects, one can partially strip out the element of personal need as the ultimate beneficiaries, the "third wheels", of AfT are firms and not individuals. While firms attempting to influence siting is still inherently political, using aid to increase the productivity of firms is mainly a technical exercise. For AfT to be efficient and effective, it needs to benefit firms. The technical efficiency of AfT is, in most instances, related to the physical proximity of firms and projects as discussed in Brazys and Elkink (2021). Much AfT is infrastructural, and to utilize most infrastructure (roads or utilities) firms need to be sufficiently proximate. Even for infrastructure which may serve a broader catchment, like an air or seaport, firms that are closer to that infrastructure are more likely to benefit from it as they are more likely to have the connective infrastructure necessary to fully utilize the resource. Likewise, even "softer" AfT, like worker training programs, are likely to be of less use to firms farther afield as labor pools tend to be local. If AfT locates near firms we think it is reasonable to assume that the proximate firms would have likely had some influence in securing the allocation, even if those efforts are limited to the firm providing information to donors and/or governments of the particular infrastructure or training needs in their area.
Yet, in line with Swedlund's (2017) work, it seems unreasonable to assume that any single allocation preference will crowd out others. In other words, donors never have full "control", nor can aid ever be completely "captured", nor can "third wheels" override the interests of those other principals. Using that recognition, we posit that aid allocations will be most pronounced in areas where diverse interests overlap. While different interests may be more or less effective in different settings, none are likely to be completely ineffectual in any setting. Thus, with respect to our logics above, we would expect that areas that have high degrees of poverty (reflecting donor control), are in important political constituencies (reflecting capture and prebendalist interests), and are home to exporting firms (representing a functionalist logic) will be most likely to also be home to AfT projects. Of note, both the prebendalist and need orientation logics are likely to be general for all types of aid. In contrast, the functionalist logic for exporting firms is likely to be exclusive to AfT.

Case selection, data and methods
Our study focuses on AfT allocation in Bangladesh. Bangladesh is a prime candidate to consider the relationships discussed above as it has a large and geographically dispersed population of exporting firms, has a recent history of meaningful political opposition, but also has a large degree of poverty accentuated by intra-country inequality. Recent findings show the existence of 'poverty pockets' throughout the country despite the general reduction in poverty rates over the last decade, which is often attributed to policy bias as well as geographical exclusion (Alam & Iqbal, 2016;GED, 2013;Sen & Ali, 2015). 4 Other analysis also shows that in recent years economic growth was not associated with substantial poverty reduction (Hill & Cevallos, 2019; Hill & Genoni, 2019). Recent years have also seen the country slip from 143rd to 149th out of Who gets the goodies? 245 180 countries on Transparency International's Corruption Perceptions Index, suggesting a weak state of governance. 5 A 2014 study focusing on the state of governance in Bangladesh discusses the confrontational politics and political influence on other institutions affecting democratic governance and service delivery in the country (see SOG, 2014). Additionally, although recently graduated, until 2018 Bangladesh had been the most populous of the least developed countries in the world. 6 Yet despite this size and importance, Bangladesh remains relatively understudied in terms of the political economy of aid allocation and effectiveness. 7 The mechanisms for aid allocation in Bangladesh are inherently political. In general, external resources and assistance are negotiated and channeled through the Economic Relations Division (ERD) under the Ministry of Finance. The ERD is responsible for the formulation of national policies regarding foreign assistance and organizes the development forum meeting with donors, whereby the financing gap is consulted and negotiated with the former. When projects are proposed, executing agencies work jointly with the ERD to assess if the project fits with the government's development vision and priorities. The project preparation is guided by government planning documents including the five-year plans as well as the long-term perspective plan. The project ideas and plans are discussed with development partners, taking into account the priorities set out in these plans. It is then required that foreign aid funded projects or programs have a development project proforma (DPP) or a technical assistance project proforma (TAPP). These proforma describe the detailed work plan and time frame, outlining the inputs as well as outputs of the project. The preparation of these documents is done by the local implementing agency in consultation with the relevant donors/partners. However, these proforma are marked for approval at the highest political level by the Executive Committee of the National Economic Council (ECNEC) which is comprised of the Prime Minister and all other ministers whose portfolios include economic issues. This ensures ultimate domestic political control over the objectives and siting of foreign-funded projects (Government of Bangladesh, 2020).
AfT projects fall most directly under the purview of the Ministry of Commerce (MoC) via a Technical Assistance Working Group headed by the Director General, WTO Cell of the MoC. This group includes representatives from the government (related ministries and departments) as well as from the private sector (including business chambers and trade bodies). The MoC organizes consultations/dialogues with national stakeholders on AfT involving representatives from the government ministries, donors, chambers of commerce and trade bodies as well as local think tanks. In this platform, stakeholders put forward suggestions on trade facilitation issues, removing obstacles to trade growth potential, etc. The institutionalized consultative process gives a plausible mechanism by which functionalist interests may influence the siting of AfT projects (Government of Bangladesh, 2020; OECD, 2012).
To look for spatial relationships between AfT and our measures of functional, prebendalist and poverty allocation logics in Bangladesh, we draw on data from a diverse range of sources. First, as "outcome" data, we utilize AidData's (2016) "Bangladesh Selected Donors" database. This database contains 288 projects at 3,641 unique project locations. The "Aid for Trade" designation has been critiqued as overly broad, with many projects officially classified as such having little, if any, discernible relationship with export activity (Brazys & Lightfoot, 2016). Accordingly, we pared the projects by using a textual algorithm that searches project sector names and descriptions for terms indicating the projects are either trade-related infrastructure, trade-related industry development (including technical training), or related to customs or trade procedures. Descriptive details, including project names, and the full search algorithm are available in the Supplementary Materials. These efforts resulted in the identification of a total of 1,111 project locations at precision code "3" or better spread across 62 projects. In the robustness checks below we further restrict AfT projects to those at precision code "2" or better, equivalent to 25 km precision. The distribution of these projects is represented graphically in Map 1.
Our firm data is a novel, geo-coded, dataset of the population of Bangladesh exporting firms. 8 Full details of the sourcing and geo-coding of this data are available in the Supplementary Materials. In total, we were able to geo-code 11,115 firms, comprising 99.9% of the directory. The exporter directory also contained useful information on the firm sector. Bangladesh's export sector is dominated by apparel-related firms, particularly those in the readymade garment (RMG) industry with 8,297 (75%) of firms in these sectors. Other major industries include software (667 firms), handicrafts (441 firms), and food-related products (523 firms). The geographic distribution of these firms, by sector, is provided in Map 2. Circle size reflects the natural log of the number of firms within a given geographic level (neighbourhood, sub-district or district). While the map indicates clear sectoral and geographic clustering, especially around the two major metropolitan areas of Dhaka and Chittagong, it also shows a reasonable amount of dispersion of firms around the country.
To observe patterns of allocation, we spatially joined our variables. Our primary analyses utilize 5,160 base geographic areas at the administrative four level (ADM4) (known in Bangladesh as unions). Using polygon shapefiles we determined how many AfT projects or firms are located within each polygon. 9 We find between 1 and 9 AfT projects in 766 of the 5,160 ADM4 units. For our main analysis, we use this information to create a binary indicator that equals "1" if the ADM4 unit was home to any AfT project as our outcome variable. In the extensions below we also utilize the full count of the AfT projects.
Similarly, our main correlate of interest is a binary indicator, Exporter, coded as "1" if an ADM4 unit has any exporting firm within its borders. We find between 1 and 415 exporting firms located in 451 of the 5,160 ADM4 units. Of these 451 units with exporting firms, we find 112 (24.8%) also have an AfT project location. In contrast, 655 (13.9%) of the 4,710 ADM4 units without an exporting firm had an AfT project location. In the extensions in the Appendix, we also use Count indicators, utilizing the count of AfT projects and the natural log of the count of firms (lnExporterCount) as several unions have an outsized number of firms.
To evaluate the prebendalism logic we turn to constituency-level data on elections to the Bangladesh parliament, the Jatiya Sangsad. We code each ADM4 unit with a binary indicator that equals "1" if it was in a constituency that was represented by a government MP (Gov MP) from the 2008 election (Kollman, Hicken, Caramani, Backer, & Lublin, 2016). 10 We find that, after the 2008 election, 4,453 (87%) of the ADM4 units were represented by a government MP. We use the results from the 2008 election as the parliamentary elections since that year have not had any meaningful opposition. 11 As such, the 2008 results give us an indication of those areas that did and did not support the ruling party when the opposition contested the election under a caretaker government.
Our primary poverty data is the "Relative Wealth Index" from Facebook's "Data for Good" compiled by Chi, Fang, Chatterjee, and Blumenstock (2021). 12 A detailed description of the data is available in that manuscript but, briefly, the data are compiled from satellite images of nighttime lights, mobile phone usage, topographic data and Facebook connectivity data, and ground-truthed using information from household-level demographic surveys. This data is compiled at the 2.4 km grid level. We identify these grid points within ADM4 units and then collapse the mean value for all grid cells which fall within an ADM4 unit. 13 This variable ranges from À0.946 to 2.003 in our data, with a larger value representing a higher level of relative wealth within Bangladesh and it is presented in Map 3 below. 14 In our primary models, we create a binary indicator variable which equals "1" for ADM4 units below the median wealth level and "0" for those above it. In the appendices, we also consider models using the continuous measure and with the measure split into deciles.
While the AidData contains information on the timing of AfT projects, unfortunately, we currently have no such data on the timing of either the establishment or start/resumption of exporting activity by the firms. As such, we are unable to use spatial-temporal approaches which might enable us to look for a causal relationship between AfT and exporting firm presence. In particular, we cannot say if the location of exporting firms preceded or succeeded the presence of the AfT project(s), although related data, discussed in the Supplementary Materials, is suggestive that most firms would have been operating prior to the time AfT projects were allocated. Likewise, there is no temporal variation in the poverty data, nor is there meaningful temporal variation in electoral data since 2008. The poverty data, in particular, is mainly compiled from data gathered after the bulk of the AfT allocations. While poverty (wealth) is likely to be geographically persistent, or at least slow to change, this timing also means that we cannot discount that AfT may have served to reduce poverty. 15 Accordingly, this analysis should not be understood as a causal analysis. However, we believe that documenting co-location between AfT projects and exporting firms, government constituencies, and levels of poverty can provide useful insights that are indicative of different types of aid allocation decision-making processes.
Our main estimation technique is a linear probability model. We use an OLS estimator with robust standard errors clustered at the unit of analysis (ADM4 in the main models) for ease of interpretation, although we check our results with non-linear estimators in the robustness section. 16 We first present a baseline model with each indicator of the allocation logics. We note that including all indicators in the same model means that the interpretation of each is conditional on the presence of the other covariates. In the robustness checks, we examine each logic in isolation. We then introduce a model with a measure of distance to the nearest major metropolitan area (either Dhaka or Chittagong) as firms and AfT projects cluster in these areas, and a model that also includes population (at the ADM3 level) to test the robustness of the baseline results to these indicators of (population) density, as more dense areas are simply more likely to receive both aid projects but also have exporting firms. Our results are qualitatively unchanged by the inclusion/exclusion of these measures, but we retain the distance measure in the remaining models as a control as it is precise at the ADM4 level. We then introduce separate two-way interaction effects before modelling the full three-way interaction between exporting firm presence, political representation and poverty. Our reduced form baseline model is: where Y is our binary indicator of area (ADM4 unit) i having an AfT project. This outcome indicator is regressed on our Exporter, Poverty, and MP indicators and a measure of Distance of the area from the nearest metropolitan area of 500,000 or more people (effectively Dhaka or Chittagong) and e i are standard errors clustered at the ADM4 level.

Results
Our main results are presented in Table 1. The baseline model (1) shows that the prebendalist (MP) and funcationalist (Exporter) logics are positively associated with AfT projects and statistically significant at (at least) the 1% level when conditioned on the other covariates. Substantively, the larger relationship is with exporting firms. The probability of an AfT project's presence in a given ADM4 unit is 8.3 per cent points higher when that ADM4 unit also has an exporting firm (p-value 0.000). That compares to a 4.5 per cent point increase being in an electoral district with a ruling party MP (p-value 0.001). However, in contrast to our need orientation logic expectation, we see that poverty is negatively associated with AfT allocation, where areas below the median poverty level see a 4.0 per cent point decrease in the likelihood of a co-located AfT project. While this result is consistent with Briggs (2017Briggs ( , 2018aBriggs ( , 2018b and others' findings, we again stress that the timing of our data is such that the poverty measure is mostly drawn on data gathered after the AfT allocation and, as such, another explanation for the result is that the allocation of AfT led to an increase in wealth. These baseline results are substantively unchanged when including distance to cities (model 2) or distance to cities and population (model 3) as controls.
We present the results from the interaction models (4-7) graphically in Figure 1. As we are most interested in the impact of "overlapping logics" of allocation, our primary interest in the allocations is the contrast between the predictive margins of the "on" state when both (or all three) logics are present (that is, 1,1 or 1,1,1) compared to the "off" state where none of the logics are present (that is, 0,0 or 0,0,0). As such, we take the pairwise comparisons of these predictive margins and plot them in Figure 1. As seen there, the contrasts are positive and, in all cases expect the MP Ã Poverty interaction, the contrast in the predictive margin is larger than the marginal effect of any individual interaction component as shown in Table 1, Model 3. However, the contrast is only statistically significant for the Exporter Ã MP interaction. In this latter instance, the contrast of the predictive margins is 12.1 per cent points, suggesting that the likelihood of an ADM4 unit with both an exporting firm and MP representation is 12.1 per cent points more likely to receive any AfT project, compared to an ADM4 unit with neither. While this is larger than the regression coefficient of either stand alone component from Table 1, Model 3 (7.5 (exporter) or 4.2 (MP)), the impact is only incrementally larger than the sum of those parts. Figure 1. Contrasts of two and three way interactions ("on" state compared to "off" state). Notes: Where the red dot indicates the contrast of the pairwise comparison of the predictive margins of the "on" state (1,1 or 1,1,1 respectively) compared to the "off" state (0,0 or 0,0,0 respectively). 95% (90%) confidence intervals are given by the thin (thick) bars.
Accordingly, while this is indicative of additive allocation logics, at least between MP and exporter, there does not appear to be any great catalyst of multiple logics leading to a "sum greater than the parts" type of outcome. For the interactions with poverty (both two-way and the three-way interactions) the contrast is not statistically significant, which is perhaps unsurprising given the finding above that poverty, as a stand alone logic, is associated with a reduced likelihood of AfT receipt. Accordingly, while we find that the "third-wheel" agency of the exporting firms is likely influencing allocation decisions, there is only qualified evidence that this impact is magnified when "overlapping" interests are present.

Extensions and robustness checks
We subject our findings to several extensions and robustness checks. While we describe those approaches and results in full in the Supplementary Materials, in brief, we include models that utilize counts of AfT projects and firms, rather than binary indicators; we disaggregate the firm data by apparel and non-apparel sectors; we exclude observations from the major metropolitan areas; we use a non-linear (logit) estimator; we consider each logic in a "stand-alone" model without including the other logics or controls; we use the natural log of our distance control; we account for spatial auto-correlation with Conley (1999) Standard Errors; we estimate models with ADM1 and ADM2 dummies (fixed-effects) and with multi-level mixed effects; we consider models alternative definitions of AfT and using all aid projects; we use only the nonimputed RWI data; and, finally, we examine models using only AfT projects coded at AidData precision code "2" or better. These results are presented in Table A1-A5 and A7, and Figure  A5, in the Supplementary Materials. While we discuss the results fully there, our primary results are substantively unchanged with any of these extensions or robustness checks. However, given the surprising finding on the poverty logic, we investigate that finding further in two ways. First, we collapse our data and conduct the analysis at the ADM3 level. We do this, in part, to also utilize an alternative measure of poverty; the World Bank's subnational estimates of poverty 17 which utilizes data from the 2010 Bangladesh Poverty Maps 18 , the 2011 Census of Population and Housing 19 and the World Food Programme's 2012 undernutrition maps 20 . We use the "extreme poverty headcount ratio" measure in this data which shows the percentage of the local population that lives below the official national lower poverty line. We collapse the RWI data to the ADM3 level and conduct the analysis at that level with that data as well. Given the loss of information in collapsing the data, we use the natural log of the count of AfT projects and firms in the ADM3 units, as well as the continuous measures of poverty. For the member of parliament measure, we again use a binary indicator that captures if the ADM3 unit is represented by a MP in government.
The results in the baseline model using the RWI data (Table A5, Model 1) show positive relationships between both exporting firms and members of parliament and AfT projects, although only the exporting firm relationship is statistically distinct from zero, and then only at the 10% level. However, the starker change is on the wealth (poverty) measure, which now shows that a one-unit change in that measure (an increase in poverty) increases the likelihood of an AfT project in the ADM3 unit by 3.8 per cent points. We discuss this flipped sign further in the conclusions. The three-way model, however, is more in line with our existing results, with an insignificant interaction. We then also consider results at the ADM3 level using the World Bank Poverty Map measure of poverty. The results from the baseline model (Table A5, model 3) are consistent with those using the RWI. The poverty measure is again positively associated with AfT projects, although now both the exporting firm and MP relationships are statistically indistinct from zero.
We also look at the models using the continuous measure of poverty at the ADM4 level in Table A6. We visualize the impact on the marginal effect of poverty from the three-way interaction model (Table A6, Model 5) in Figure A4. There we see that the marginal effect of poverty, when an ADM4 unit also has an exporting firm and a member of parliament, is again positive. That result is also evident from our main three-way interaction using the binary poverty indicator (Table 1, model 7) as shown in Figure 2. This figure illustrates the contrasts of the pairwise comparison between the states where an exporting firm, member of parliament and poverty (wealth) are present (1(MP),1(Exporter),1 (Poverty)(0(Wealth))) and that in which only poverty (wealth) is present (0(MP), 0(Exporter), 1 (Poverty)(0(Wealth))). The contrast of the poverty state (that is, poverty ¼ "1") in this model is 18.9 per cent points, and statistically different from zero, meaning that adding an exporting firm and a MP to a poverty region increases the likelihood of that region also having an AfT project by 18.9 per cent points when compared to a poverty region without either an exporter or MP. In comparison, the contrast when using the wealth state (that is, poverty ¼ "0") is only 11.1 per cent points. While we are hesitant to read too much into this result as there are 28 pairwise comparisons one could consider in an interaction between three binary variables, combined with the ADM3 results, we think this suggestive of some additional nuance with respect to the poverty allocation logic which we opine on further in the conclusions.

Conclusions
We investigated if the geographic patterns of "Aid for Trade" (AfT) allocation in Bangladesh are consistent with functional, prebendalist and/or poverty-based logics. Overall, using novel data on the locations of over 11,000 exporting firms and over 1,000 AfT project locations across over 5,000 administrative units, we find evidence of significant relationships between the presence of exporting firms (functional) and ruling party representation (predendalist) and the location of AfT projects. However, we find the direction and impact of the poverty (need orientation) logic is more susceptible to both conditional effects and the scale of aggregation. While poverty is negatively associated with AfT allocation as a stand-alone logic when considering the more disaggregated ADM4 spatial level, it appears to be positively associated with allocation at the more aggregated ADM3 spatial level. As we do not have sufficient temporality in our data, however, we cannot discern a causal effect. The association we evidence at the ADM4 level may be due to the fact that wealthier areas are more likely to receive aid, supporting findings from Briggs (2017Briggs ( , 2018aBriggs ( , 2018b and others, or it could be due to the fact that the AfT has made those areas wealthier, suggesting to a "positive" effect of the aid. Likewise, at the ADM4 level, there is some evidence that the marginal impact of poverty on AfT allocation is positive when exporting firms and members of parliament are also in the spatial unit. While needing further investigation, these nuanced results provide some indication that poverty as an allocation logic depends on the degree of spatial precision. Donors may be able to direct allocation, in response to poverty and need, at a higher level of spatial aggregation, but may "lose control" to other allocation logics as spatial precision increases. Alongside, or alternatively, poverty may simply be "last amongst equals" in allocation logics in that it only influences allocation decisions after other logics have been taken into consideration. A research design that more specifically investigated these dynamics would be an interesting step forward. Collectively, these findings contribute to the literature on the political economy of subnational aid allocation by further arguing that aid allocation need not be either-or in terms of donor control and elite capture. Patterns of aid allocation may serve multiple political economy functionsthey may (or may not) satisfy donor preferences for need orientation, be useful to recipient prebendalist interests but also be attuned to the functional interests of third parties. Thus, the "development dance" may be an ensemble affair, rather than a two-partner engagement (Swedlund, 2017).