Poverty Dynamics and Poverty Traps among Refugee and Host Communities in Uganda

Abstract This paper analyses poverty dynamics and checks for the existence of poverty traps among refugee and host communities living close to each other in Uganda. Although some non-linearities emerge in asset dynamics, there is convergence towards one stable equilibrium for the whole sample that suggests the existence of a structural poverty trap. However, households are quite heterogeneous: when analysing refugees and hosts separately, refugees converge to a lower own-group equilibrium than hosts. The household size and education are asset growth enablers for both communities. Noticeably, access to land, past history and social cohesion are also significant correlates of refugees’ asset dynamics. From a policy perspective, structural poverty traps are bad news, because standard anti-poverty interventions would not unlock the trap. Our results stress the need of more structural approaches aimed at promoting economic growth in the whole area where refugee and host communities live, targeting both communities.


Introduction
The majority of refugees worldwide are hosted in low-and middle-income countries (World Bank, 2023).Uganda, with 1.5 million refugees, is the largest refugee-hosting country in Africa (Atamanov, Beltramo, Waita, & Yoshida, 2021;UNHCR., 2022).Refugees in Uganda can aspire to livelihoods beyond the humanitarian assistance thanks to the country's advanced refugee policy that aims at promoting refugees' self-reliance. 1 However, environmental, economic or health shocks can worsen the refugees' (and hosts') already precarious situation, increasing the risk of being trapped into poverty (World Bank, 2017).
While a recent literature has analysed the interaction between refugee and host communities (Alix-Garcia, Walker, Bartlett, Onder, & Sanghi, 2018;Alix-Garcia & Saah, 2010;Ayenew, 2021;d'Errico, Mariani, Pietrelli, & Rosati, 2022;Kadigo & Maystadt, 2023;Kreibaum, 2016;Zhu et al., 2023), there is a lack of studies exploring how poverty persistence and dynamics develop among refugees and hosts (Jacobsen, 2012).The existing literature suggests there are heterogeneous effects and possibly trade-offs among social groups as well as over time (Verme & Schuettler, 2021).Specifically, theoretical considerations and empirical evidence concur to two main predictions: (i) refugees, being very poor, may be all structurally in poverty; (ii) hosts can either benefit or be penalized by the arrival of refugees, though generally they can aspire to higher steady states than refugees'.The idea around which this paper is built is that refugees can be trapped into poverty because they have specific vulnerabilities that curtail their ability to exploit economic opportunities (Jacobsen, 2012;Stojetz & Br€ uck, 2021;World Bank, 2017).The paper addresses two research questions: Given the proximity and the interaction between refugees and hosts, how do the wealth dynamics of these two groups differ?Does a poverty trap exist and, if so, for whom?To address these questions, we adopt the poverty trap framework originally proposed by Carter and Barrett (2006).
The contribution of this work to the existing literature is twofold.First, we provide empirical evidence on poverty traps in a novel context 2 thanks to a panel dataset that surveyed both refugees and hosts between 2017 and 2021.Second, we disentangle the wealth dynamics of refugees and hosts by focusing on group-specific vulnerabilities that are key for dynamic equilibria and by accounting for the factors that may affect wealth accumulation such as assistance and shocks.Understanding the dynamics of assets can shed light on refugees and hosts' prospects and help designing effective policies to alleviate poverty.
We find evidence of a single low-level asset equilibrium when considering the host and refugee populations as a whole, indicating a structural poverty trap.When analysing refugees and hosts separately, the two groups tend to different single low-level equilibria with hosts achieving a higher own-group equilibrium than refugees.Disaggregating the population across various dimensions highlights the importance of geography and selected household characteristics that drive the dynamics.Asset growth-enabling factors for both refugees and hosts are household size and education.In addition, for refugees other factors associated with asset growth are managing a larger land size, having an enterprise income and not getting transfers, while assetreducing factors are being displaced because of famine or natural hazards, having experienced violence, the time spent in settlements, and weak social cohesion (i.e.low trust, rare social or business interactions).
The paper is organized as follows.Section 2 reviews the key literature on poverty traps highlighting the possible mechanisms at work in the case of refugees and hosts.Section 3 introduces the estimation methods.Section 4 describes the data.Section 5 presents the results and deals with attrition.Section 6 provides additional robustness checks (i.e. a different dataset length, asset index specifications and estimators).Section 7 discusses the main findings.Section 8 concludes.

Poverty traps mechanisms for refugees and hosts
Poverty traps are self-reinforcing mechanisms that reproduce poverty and make it persistent (Azariadis & Stachurski, 2005).They can be conceptualized either as single equilibrium or as multiple equilibria poverty traps (Barrett, Garg, & McBride, 2016).In the former case, an economic unitindividual, household, country, etc.converges on a unique dynamic equilibrium beneath the poverty line.Differences in structural characteristics, such as differential access to high/low productivity technologies, may give rise to club convergence where specific units converge over time to group-specific single levels of well-being (Durlauf, 1996;Galor, 1996).Figure 1, panel a, shows this case in a graph that maps current wellbeing (i.e.assets) to expected future well-being.The dynamic equilibrium for each of the two groups is represented by the intersection of the solid growth curves with the diagonal locus of points reflecting values equal on both axes.For the less (better) endowed group, represented by the lower (higher) growth curve, the dynamic equilibrium is a persistently poor (non-poor) equilibrium P (N) below (above) the static asset poverty line, w , and dynamic asset poverty line, w Ã : More often poverty traps are conceptualized as a model where both poor and non-poor equilibria exist (Galor & Zeira, 1993).Graphically, multiple equilibria poverty traps are represented in the form of an S-shaped curve (Figure 1, panel b) in which starting conditions determine whether households will converge to a higher-level equilibrium, N, or to a low-level equilibrium, P, creating two regimes of accumulation.The existence of multiple stable states implies Poverty traps arise when there are some exclusionary mechanisms at play that limit households' asset accumulation.In the case of refugees, there are basically four main mechanisms, namely: asset loss (physical or social), trauma and psychological stress, geography, and institutional factors.Some of them can also affect hosts, possibly trapping them into poverty.
The destruction, theft and abandonment of physical assets at home or while fleeing home because of conflicts or emergencies is the most evident mechanism leading to a poverty trap for the refugees (World Bank, 2017).Conflicts and humanitarian emergencies can also significantly impair the accumulation of human capital (Grimard & Laszlo, 2014;Islam, Ouch, Smyth, & Wang, 2016;Weldeegzie, 2017) and increase poverty through the disruption of social capital links (Grant, 2010) and the reduction of off-farm opportunities (Mercier, Ngenzebuke, & Verwimp, 2020).
The trauma and psychological stress experienced by refugees and internally displaced people can lead to a decline in aspirations and to an overall sense of hopelessness, which are found to be detrimental to economic activities (Banerjee & Duflo, 2007).Depression and experience of violence among refugees and internally displaced persons is found to fuel pessimistic beliefs, increase the likelihood of being in poverty (Moya & Carter, 2019), and raise the risk of a depression poverty trap (de Quidt & Haushofer, 2018;Haushofer, 2019).
Geography can be another poverty trap mechanism.Refugee settlements' location characteristicsentailing not only agroecological features and infrastructure, but also economic factors such as physical access to services, job opportunities and social relations (Grant, 2010) can give rise to a spatial poverty trap.This mechanism can apply also to hosts.The interaction between the refugee and host communities can determine a change in the access to natural resources, infrastructure, services, job opportunities as well as social relations within the host community and between the two communities.This can strengthen existing poverty dynamics possibly leading to a poverty trap also for hosts.
Finally, institutional and legal barriers can affect the refugee status and hinder their integration prospects (Azariadis & Stachurski, 2005;Barrett & Carter, 2013;Carter & May, 2001;Sartorius et al., 2013;Zhang, 2017).Social institutions such as kinship systems, community organizations, and informal networks, greatly affect poverty outcomes.Discrimination based on gender, ethnicity, race, religion, or social status can lead to social exclusion and lock people, and specifically refugees, in a poverty trap (Sartorius et al., 2013).Refugee inflows may exacerbate existing problems and impact social cohesion.Social cohesion is indeed associated with safety and productivity, and consequently on asset accumulation; conversely, its absence is associated with spatial segregation, social tensions, and stronger competition over resources (World Bank, 2017).This applies to refugees as well as hosts.The resulting degree of social cohesion will depend on prejudice, political discourses, cultural proximity, perceived justice about aid distribution and service delivery and policy inclusivity (World Bank, 2017).
All these mechanisms may apply differently to refugees and hosts, determining different results in terms of poverty persistence and dynamics between the two groups.Specifically, we expect that the multiplicity and self-reinforcing mechanisms at work in the case of refugees determines a more critical condition for this group vis-a-vis hosts, making a poverty trap for the former more likely.At the same time, refugees could negatively influence hosts, at least in the short run, especially those strata of the host population featuring skills and product orientation similar to refugees', increasing the competition between the two groups (Betts, Chaara, Omata, & Sterck, 2019).

Methodology
We use three complementary methods usually employed in the poverty traps literature, namely parametric, non-parametric and semi-parametric methods.Non-parametric regressions analyse the relationship between assets A at time t and assets at t-1, without imposing any pre-defined functional forms.This approach is very flexible, but it only estimates a bivariate relation (Adato, Carter, & May, 2006;Barrett et al., 2006;Lybbert, Barrett, Desta, & Coppock, 2004) as follows: where the error term e it is assumed to be normally and identically distributed with zero mean and constant variance.This method assumes that the function to be estimated is smooth, covariates are uncorrelated with the error term, and all households are in the same accumulation regime.Some of these assumptions are heroic.Therefore, we rely on this method only for exploratory purposes and in combination with parametric regression.Parametric regressions allow to study non-linearities while controlling for other factors (Giesbert & Schindler, 2012;McKay & Perge, 2013;Naschold, 2013), using OLS, fixed effects (FE) or random effects (RE) panel estimators: where asset change is a function of the fourth polynomial expansion of assets at t − 1 to capture non-linearities 3 (Naschold, 2012(Naschold, , 2013), household's lagged characteristics, X, and the interaction of district and year fixed effects l ðiÂtÞ : A negative b 1 means general convergence towards the equilibrium, i.e. those poorer in assets accumulate assets faster (Carter, Little, Mogues, & Negatu, 2007).The b 2 , b 3 and b 4 coefficients, if significantly different from zero, indicate non-linearities in the asset accumulation process (Walelign, Charlery, & Pouliot, 2021).The household characteristics vector controls for socio-demographics (with some variables squared to account for possible non-linearities), location characteristics and shocks.Surveyrelated controls such as the different survey time of wave 1 (cf.Section 4.1) and the interactions between wave and districts are included as well.The analysis is carried out for the whole population as well as separately for refugee and host subpopulations (Naschold, 2012).The choice of the parametric regression estimator is not straightforward.Having more than two periods, it is possible to look both at the asset change of each subperiod with one-period lagged assets (RE or FE) and at the long difference between the last and the first wave, while controlling for initial assets (OLS).The latter (OLS) does not exploit the panel structure of the data, but is consistent with the idea of poverty traps that depends on initial conditions.A panel Poverty dynamics and poverty traps 383 estimator (FE or RE) instead captures the period-by-period relation of lagged assets with asset changes rather than overall asset convergence.Since the length of the panel is rather short (2017-2021), we decide to exploit both the panel data structure for understanding the adjustments from the previous periods and the longest possible time span as usual in the poverty traps literature.In the OLS case, standard errors are corrected for generic heteroskedasticity.In the panel case, standard errors are clustered at the household level.
Finally, we test the robustness of our results with semi-parametric regressions, which combine the advantages of the previous two approaches, i.e. flexibility and controlling for covariates (Naschold, 2012(Naschold, , 2013)): The relation between current and lagged assets is estimated non-parametrically, while households' characteristics enter the equation parametrically.

Survey description
Most refugees in Uganda come from neighbouring countries, mainly from South Sudan and the DRC, while the remainder come from Burundi, Somalia, Rwanda and Eritrea (World Bank, 2019).They mostly live in settlements concentrated in 12 districts in the Northern (West Nile) and Western regions. 4Settlements are highly heterogeneous in terms of size, country of origin, share in the total district population, and other socio-demographic characteristics (Tables A.1 and A.2 in the Supplementary materials).By and large, the districts in the Northern region -Adjumani, Arua, Koboko, Moyo, Lamwo and Yumbehost almost two thirds of the total refugee population, coming mostly from South Sudan 2-3 years before the survey, with the notable exception of older and larger settlements (e.g., Adjumani).Instead, the districts in the Western region -Isingiro, Kamwenge, Kyegegwa, Kibuube and Kiryandongohost mostly refugee from the DRC (except for Isingiro that hosts also refugees from Burundi), that came to Uganda more than five years before the survey.In some of these districts especially in the North (Adjumani and Moyo), the refugee population accounts for a significant share of total district population.
We use data from the FAO-RIMA's Uganda Refugee and Host Communities Panel Survey (d 'Errico et al., 2022) that spans over 10 districts and 13 settlements.During wave 1, interviews were conducted at three different points in time -2017, 2018 and 2019 (Figure 2).Households were interviewed again in 2019, 2020 and 2021 (Table 1).The final sample consists of 20,079 observations (9,128 considering only the balanced panel).The sample design was based on two clusters: refugee communities and host communities, with the primary sampling unit being the settlement or the village and the second sampling unit being the household.The sample is representative at the settlement level and for the two communities.
Panel attrition is a common issue in panels as households might refuse to answer again or might have moved away from their initial location and be hard to track.Not surprisingly, the attrition rate is quite high in the FAO-RIMA dataset (Table 2) because of the fragility of the context ( € Ozler, C ¸elik, Cunningham, Cuevas, & Parisotto, 2021) and because of the high number of waves.Nonetheless, over the four waves some dropout households re-entered the panel (Table 2).

Data description
The FAO-RIMA survey collects information on a broad range of topics including: socio-demographics such as the refugee status, the age of household head, the average education of household members, the gender and marriage status of household head, the size of the household, income generating activities, formal and informal transfers, loans, food consumption and coping strategies; location characteristics such as distances from the agricultural market, petty trading market and schools; 5 and self-reported shocks such as Covid-19 in 2020 and 2021 and environmental shocks in all waves.Uganda experienced some floods over the period of analysis  Poverty dynamics and poverty traps 385 that only partially involved our sample's locations.Nonetheless we control for local events such as abundant and scarce rainfalls, exploiting the georeferenced coordinates of each household and the availability of this information from third sources.Specifically, we created two variables using the values of the SPEI index during the growing season, 6 namely: scarce rainfalls if the SPEI index was below 1 standard deviation and abundant rainfalls if the SPEI was above 1 standard deviation from the long-term average. 7 To represent household's wealth, we build a tradable asset index (Giesbert & Schindler, 2012) that includes several durables and tools (radio, TV, bicycle, solar panel, cooker, box, table, chair, bed, mattress, animals, hoe, axe, shovel, pickaxe, sickle, slasher) as well as land size (Table A.3 in the Supplementary materials). 8Aggregation is done via principal components analysis (Sahn & Stifel, 2000) and the index is normalized between 0 and 1. Focusing on tradable assets is more suited for studying asset dynamics over short periods of time as it is the case of our panel.However, we compute also other asset indexes using different combinations of assets and alternative aggregation methods.The first includes both productive and non-productive assets (Giesbert & Schindler, 2012;Naschold, 2012Naschold, , 2013;;Walelign et al., 2021) (Table A.4 in the Supplementary materials).The second is a livelihood index a la Adato ( 2006) including all types of assets that predict household consumption.In principle, no approach is better than others (Naschold, 2013), but in order to keep the analysis as clear as possible, we use the indexes other than the tradable index only as robustness checks.

Descriptive statistics
Table 3 shows some descriptive statistics for refugee and host households by wave.On average, host households are larger in size and their heads are older and slightly more educated than refugee households'.Refugees' average land size is significantly smaller than hosts'.Per capita expenditure and income are very low for both groups, though on average hosts report higher values and decreasing over time. 9Formal transfers represent the largest source of income for refugees, while the main income sources for hosts are enterprise, wage and crop income (Figure 3).Refugees' average income was greater than hosts' in 2019, due to transfers.In 2021 there was a general worsening in both groups' conditions (less land, less livestock, less assets, less enterprise activities, less income per capita, less dietary diversity, higher coping strategy index), likely due to the protracted Covid-19 crisis (Squarcina & Romano, 2023).
Kernel density functions of the tradable index show how refugees and hosts' wealth is distributed (Figure 4): refugees are more concentrated at the lower end of the distribution as they own fewer assets.Over time, there is a slight improvement for both communities 10 though between wave 3 and 4 we observe a worsening especially for refugees.

Non-parametric regression
Local polynomial smooth recursion functions (Equation (1)) show that there is only one stable equilibrium for the whole population at about 0.2 asset scores (Figure 5).This is a stable equilibrium because the curve crosses the 45-degree line from above.Scatter points above (below) the 45-degree line represent households whose position improve (worsen) over time.Refugee scatter points (blue dots) are more concentrated around the lower left corner, as they have lower assets than hosts.
Relaxing the assumption that all households are in the same accumulation regime, we run the local polynomial regression separately for refugees and hosts (Figure 6).Refugees converge to a lower equilibrium at around 0.11 asset scores while hosts converge to a higher equilibrium (0.21 asset scores). 11As these are own-group equilibria, a transition from one to another is Poverty dynamics and poverty traps 387 unfeasible.We therefore exclude the existence of multiple equilibria in the whole population.Using different asset indexes confirms this result (cf.Section 6.2).
The different location of the equilibrium for refugees and hosts can be explained by refugees' lower physical asset endowments.This means that refugees, owning fewer durables, agricultural tools, animals and smaller plots have a lower production capacity, less buffer resources to cope with shocks, less collateral, hence less capacity to make investments, not only in assets but also  in human capital. 12This may affect the future prospects for the youngest household members, paving the way for an intergenerational poverty trap.
To further explore the heterogeneity in the sample, we report the different equilibria for refugees and hosts according to various households' initial characteristics (Table 4) (Giesbert & Schindler, 2012;Walelign et al., 2021).Refugees that tend towards higher-than-average equilibria have higher educated heads, larger households, larger plots of land, receive no transfers, own an enterprise, came to Uganda because of persecution/human rights violation and originate primarily from South Sudan.Refugees that converge to lower equilibria have smaller households, live in urban areas, were already displaced for more than 48 months at wave 1, moved  Poverty dynamics and poverty traps 389 Ã The equilibria for the social cohesion variables refer to the period wave 2 to wave 4 as these variables were not available at wave 1. Social cohesion variables are: improvement over time of relationships with other groups in their community, level of trust with the other community, status of relationships within their community, frequency of the interaction, feeling comfortable in interacting with thier group, frequency of interaction with the other community's vendors and businesses (Ugandan nationals for refugees and refugees for Ugandan nationals), and sense of belonging to the community.
because of famine and natural hazards, and originate from Burundi and DRC.Hosts show less heterogeneity and seem to converge to similar equilibria, except those with a small (large) household and female (male) heads who tend to a lower (higher) equilibrium.Social cohesion plays an important role for refugees. 13Refugees' lower-than-average equilibria are associated with low trust, rare interactions, and rare business interactions with other groups in their community, as well not feeling comfortable in these interactions.For hosts, social cohesion does not seem to make a difference, being the equilibria not significantly different from the average.
The equilibria are slightly different across districts, although the overall dynamics look similar (Figure 7).For refugees, Isingiro and Kikuube districts (West) show below-average equilibria.This result is associated primarily to lower-than-average land and livestock (Table A

.1).
There are other important characteristics that differentiate the refugee populations in these districts.Refugees in Isingiro, who originate mainly from Burundi, have widely experienced violence in the past, feature one of the longest average permanence in settlements (inflows since 2015) and receive the lowest average annual transfers.Despite the highest share of wage income earners, they are highly vulnerable (low food consumption score and high coping strategy index).Refugees in Kikuube, who originate mainly from DRC, also have small plots of land, low agriculture or enterprise income, receive the second lowest average annual transfers, and have in general low education (Table A.1).
For hosts, Adjumani and Lamwo districts (North) are the ones showing lower-than-average equilibria.On average, host households in Adjumani have younger heads, very low food consumption scores, receive the second highest informal transfers, own fewer assets but have access to relatively larger land size, show low expenditure levels, and are far from agricultural and petty trading markets (Table A.2 in the Supplementary materials).Adjumani is also the district hosting the second largest number of refugees (Government of Uganda OPM, 2019, 2021).

Poverty dynamics and poverty traps 391
Hosts in Lamwo, despite the high frequency of wage and own enterprise incomes and large average land size, have high coping strategy index, low income, low expenditure and low asset levels (Table A.2).

Parametric regression
We estimate the parametric regressions for the whole sample and separately for refugees and hosts (Equation ( 2)), first using OLS for the asset change between wave 1 and wave 4, then using panel estimators (RE and FE) for the one-time lagged asset dynamics.
In the case of the OLS, there is convergence if it is possible to reject the hypothesis that all terms of the polynomial are equal to zero in favour of the alternative that b 1 is between −2 and 0 and b 2-4 coefficients are all equal to zero (Quisumbing & Baulch, 2013).The null is rejected for the whole population and for hosts (Table 5), indicating that non-linearities are relevant, whereas convergence is only found for refugees.Furthermore, b 1 is much larger for refugees.Interestingly, if we sum predicted asset change to lagged assets and plot it against lagged assets (Giesbert & Schindler, 2012;Naschold, 2013), we obtain patterns very similar to the non-parametric ones (Figure 8).There is only one equilibrium for the whole population, but refugees converge to a lower stable equilibrium (0.125 asset scores) than hosts (0.23 asset scores).
Next, we estimate Equation ( 2) using random effects (RE) and fixed effects (FE) panel estimators 14 with all covariates one period-lagged (Table 6).The convergence test signals non-linearities in period-by-period adjustments in all subsamples.The RE variables that are significantly and positively associated with asset growth are age of the household head (at a decreasing rate), squared education, household size, being married, abundant rainfalls.Conversely, being female headed, being a wage earner, being a refugee and distance from school are significantly and negatively associated with asset growth.In the FE regression, asset growth Robust standard errors in parentheses.The dependent variable is the asset difference between the last and the first wave.Controls are three-periods-lagged variables: refugee status, age of the head (and its square), average education level (and its square), female headship, household size (and its square), married head, crop income (dummy), enterprise income (dummy), wage income (dummy), annual formal transfers (US$), annual informal transfers (US$), borrowed money (dummy), distance from agricultural market, petty trading market, primary school, negative and positive SPEI values, agroecological zones, Covid-19-related shock in the household (any), rural, subsamples of wave 1, and the interaction between year and district.ÃÃÃ p < 0.01; ÃÃ p < 0.05.
correlates positively and significantly with head age.Running the FE model separately for refugees and hosts (columns 3 and 4) confirms the presence of group-specific characteristics affecting period-by-period asset growth.For instance, refugees show a positive correlation with squared education and household size, but also with female heads.This result, in contrast with expectations, might indicate an important role of assistance in the settlements. 15Also, average education correlates negatively with asset growth, signalling perhaps a difficulty of adapting skills to the new setting (up to a point).Hosts show a positive correlation with age of the household head and formal transfers, which are quite in line with expectations and a negative correlation with wet weather, indicating the relevance of agriculture for hosts. 16

Dealing with attrition
The attrition between wave 1 and wave 4, i.e. ignoring whether households appear in the intermediate waves or not, amounts to 33% for the whole population (23% for hosts and 41% for refugees). 17Should attrition be correlated with the variable of interest, our results would be biased (Wooldridge, 2010).For instance, we could expect that refugees that dropped out after the first wave are richer or better connected. 18This is not the case for Ugandan refugees.T-tests at wave 1 show that attritor households are more often households with female, younger and less educated heads, generally smaller in size and with higher dependency ratios, less engaged in crop and wage activities, receiving larger amounts of formal assistance, borrowing less, having fewer assets, and spending less.Another test for attrition is to run the regression for the balanced as well as unbalanced samples and compare the coefficients: these are very similar (cf.Table A.5 in the Supplementary materials), signalling that attrition bias might be negligible (Prieto, 2021;Wooldridge, 2010).However, the coefficients of attrition-related auxiliary variables 19 indicate that attrition is somewhat relevant.Attrition probits show that the probability of attrition is not correlated to the asset index (except for hosts from wave 1 to wave 4) and only marginally correlated with some other variables (Table A.6 in the Supplementary materials).
Having ruled out that attrition is fully random, some corrections are needed.A first approach is the Heckman (1976) procedure, which uses a set of instrumental variables that correlate with attrition but not with the error term (selection on unobservables).As for any instrumental variable approach, it is difficult to find appropriate instruments 20  (Baulch & Quisumbing, 2011).Another approach is the inverse probability weights (IPW) correction Poverty dynamics and poverty traps 393 which relies on auxiliary variables that are correlated with both attrition and the outcome variable (selection on observables) 21  (Robins, Rotnitzky, & Zhao, 1995;Wooldridge, 2010).
We implement both the IPW correction (Table 7, columns 2 and 5) and the Heckman model (column 3).Overall, coefficients on the lagged polynomial of assets are quite similar for the weighted and non-weighted sample.This is reassuring: the results we get using the balanced panel apply to the overall sample as well.

Robustness checks
To further check the robustness of our results, we test for different specifications, namely excluding the first wave observations, using different asset indexes, and running semi-parametric regressions.

Excluding wave 1 observations
To rule out that the different starting times of the first wave (cf.Section 4.1) are driving our results, we repeat the analysis by excluding the first wave (Table 8).In this case also, we reject the null hypothesis in all subsamples (marginally in the host sample).Predicted equilibria are similar. 22

Different asset indexes
Using a more comprehensive asset index, i.e. adding to items included in the tradable asset index other asset such as the types of toilet and water source, the test cannot reject convergence for the two samples (Table 9, col 1-3).Hosts and refugees again have close but different equilibria (Figure A.2 in the Supplementary materials).Another asset index, built by predicting household expenditure (divided by the poverty line) 23 with asset ownership and socio-demographic characteristics (Adato et al., 2006), indicates general convergence for the host sample but not for refugees (Table 9, col. 4-6  Standard errors clustered at the household level in parentheses.The dependent variable is the one-period asset growth.Controls are one-period lagged variables: refugee status, age of the head (and its square), average education level (and its square), female headship, household size (and its square), married head, crop income (dummy), enterprise income (dummy), wage income (dummy), annual formal transfers (US$), annual informal transfers (US$), borrowed money (dummy), distance from agricultural market, petty trading market, primary school, negative and positive SPEI values, agroecological zones, Covid-19related shock in the household (any), rural, subsamples of wave 1, and the interaction between year and district.ÃÃÃ p < 0.01; ÃÃ p < 0.05; Ã p < 0.10.
Poverty dynamics and poverty traps 395

Semi-parametric regression
Semi-parametric regressionspenalized spline, partially linear regression and lowess smooth (Figure 9)confirm the existence of a single equilibrium for the whole population at 0.18 asset scores (slightly lower than parametric and nonparametric estimates).However, close but different equilibria emerge for refugees (0.11 again) and hosts (0.22).Estimated coefficients are similar to those of the parametric case. 25

Discussion
We find evidence of convergent dynamics with two low-level equilibria, one for refugees and another for hosts.To qualify these equilibria as structural poverty traps, we should show that the 'long-term household welfare equilibrium is below the poverty line' (Naschold, 2013, p. 937).However, it is difficult to document this in the case of refugee and host communities.First, it is not appropriate to use the national or the international poverty lines as we are dealing with a very specific context.In fact, refugees are quite different from the rest of Ugandan population and live in districts that are among the poorest in the country (Development Pathways, 2020).Second, FAO-RIMA survey expenditure and income data are likely underestimated compared to national data because of the long recall period (12 months). 26Nonetheless, using as poverty line the FAO-RIMA median expenditure per capita (0.10 US$/day), we find poverty headcounts consistent with those of Development Pathways (2020).The dependent variable is the asset difference between the last and the first wave (col.1-3) and oneperiod asset growth (col 4-5).Controls are three-periods-lagged variables (col.1-3) or one-period lagged variables (col.[4][5]: refugee status, age of the head (and its square), average education level (and its square), female headship, household size (and its square), married head, crop income (dummy), enterprise income (dummy), wage income (dummy), annual formal transfers (US$), annual informal transfers (US$), borrowed money (dummy), distance from agricultural market, petty trading market, primary school, negative and positive SPEI values, agroecological zones, Covid-19-related shock in the household (any), rural, subsamples of wave 1, and the interaction between year and district.
Finding no evidence of multiple equilibria poverty traps could either mean a true absence of multiple equilibria poverty traps or that we are unable to detect it.The latter may be due to an inaccurate households' assets estimation or a too short time frame or significant attrition.Data for the two groups is collected consistently (i.e.same questionnaire and survey modalities), which is fundamental for comparing them.Nonetheless, one could expect that refugees and hosts' asset bundles differ.For instance, refugees could accumulate relatively more tradable assets, while hosts could accumulate other types of capital (e.g., investments on dwelling or land) which are not well reflected in a tradable asset index.We tend to exclude that the choice of the assets included in the wealth index might be a problem because our results are robust to different specifications of the wealth index.However, the short time frame of analysis could be a problem for the identification of a poverty trap.For this reason, we use the maximum stretch of the panel and show that the results hold even if the period of analysis is shortened.We also use an asset index based only on tradable assets that captures faster accumulation/decumulation dynamics.However, we cannot completely rule this issue out.Vice versa, we are quite confident that households' mobility is not related to the main variable of interest: correcting for non-random attrition does not alter our findings.The rich set of information collected by the survey provides insights into the living standards of refugees and hosts below and above the equilibrium.Households with assets below the equilibrium show statistically significant traits associated with destitution and poverty (i.e.lower expenditure and income, smaller land, lower prevalence of improved water source and improved toilet, higher coping strategy index, lower food consumption score and lower transfers) (Tables A. 6 for refugees and A.7 for hosts).This does not prove that the asset equilibria are below the poverty line, but we argue that the equilibria are effective in separating two groups, with the ones below the equilibrium having much worse standards of living as compared to the ones above.The likelihood of the former groups falling below the poverty line is quite high considering that poverty is rampant in the analysed districts. 27Poverty dynamics and poverty traps 397  Robust standard errors in parentheses.The dependent variable is the asset difference between the fourth and the first wave.Controls are three-periods-lagged variables: refugee status, age of the head (and its square), average education level (and its square), female headship, household size (and its square), married head, crop income (dummy), enterprise income (dummy), wage income (dummy), annual formal transfers (US$), annual informal transfers (US$), borrowed money (dummy), distance from agricultural market, petty trading market, primary school, negative and positive SPEI values, agroecological zones, Covid-19-related shock in the household (any), rural, subsamples of wave 1, and the interaction between year and district.ÃÃÃ p < 0.01.
Although our analysis is not intended to draw any causal inference on the impact of refugees, our results largely concur with the literature that shows that the major channels through which refugees might affect hosts' household welfare in Uganda are the labour market and market creation, with the agriculture sector playing an important role (Alix-Garcia & Saah, 2010;d'Errico et al., 2022;Kadigo & Maystadt, 2023;Maystadt & Verwimp, 2014).Descriptive statistics for households above and below the equilibria confirm these channels.On the labour market, households below the dynamic equilibrium are indeed more involved in wage employment, while wages are lower for these households (Tables A.7 and A .8).This resonates with the findings of Maystadt and Verwimp (2014) who show that refugees represent a supply of cheap labour for commercial farmers and foster competition among agricultural wage workers.
Households above the equilibrium are more involved in market creation as shown by the sales of products, purchase of agricultural inputs and purchase of food, with an important role played by exchanges between hosts and refugees (Tables A. 7 and A.8).This is consistent with the findings of d 'Errico et al. (2022) who show that refugee proximity increases the welfare and the level of economic activity of hosting-community households by generating incentives for economic exchanges.However, while refugees sell their own products mostly in the place they live or in a trade centre, hosts have more differentiated channels, including neighbouring villages and itinerant traders, especially for those above the equilibrium (Tables A. 7 and A.8).This generalizes the results by Alix-Garcia and Saah (2010) who showed that refugees can create a market for agricultural goods, thus benefitting those involved in the production and sale of these goods.
Moreover, households above the dynamic equilibria are more involved in enterprise business and commercial farming (crop sale) than the ones below the equilibrium (Tables A. 7 and A.8).We observe also that the share of host households who moved from selling no crops in wave 1 to crop sales in wave 4 is much larger for the ones above the equilibrium.The opposite, i.e. moving from crop sales in wave 1 to no crop sales in wave 4, is larger for the households below the equilibrium.The same applies for enterprise income switches.This evidence is consistent with Kadigo and Maystadt (2023) who argue that the host households who benefit the most are the ones switching from subsistence to commercial farming.
Taken together, these results emphasize that households above the equilibrium are more likely to exploit the opportunities offered by the interaction between refugee and host communities.While being mere correlations, our results largely confirm the findings of the literature on refugee impact on host household welfare, looking at it from the poverty persistence and dynamics viewpoint.

Concluding remarks
We analyse households' asset dynamics in Ugandan refugee camps and neighbouring villages.We find no evidence of multiple equilibria poverty traps.Rather, the whole population tends to a single low-level equilibrium, indicating a structural poverty trap.Looking at refugees and hosts separately, we find that the dynamics are more severe for refugees as their own-group equilibrium is lower.Further disaggregating the population across various dimensions highlights the importance of geography and specific household characteristics.The most important factors associated with asset growth for both refugees and hosts are household size and education.In addition, for refugees other asset growth enabling factors are managing a larger land size, getting their own livelihood from running an enterprise and not getting transfers, while asset-reducing factors are being displaced because of famine or natural hazards, having experienced violence, the time spent in settlements, and weak social cohesion (i.e.low trust, rare social or business interactions).There is evidence in the literature that the interaction with refugees may bring positive effects to the native population.Our results are generally consistent with the literature emphasizing that the channels affecting both hosts' and refugees' welfare are the Poverty dynamics and poverty traps 399 labour market and market creation, with the agriculture sector playing a key role.In such a context, providing the refugees with agricultural land is crucial (Betts et al., 2019;World Bank, 2019).
The best explanation of the results is that all households in the study are in a structural poverty trap (Carter & May, 2001;McKay & Perge, 2013;Naschold, 2013), more severe for refugees.In this case, standard transfer-based approaches might not be effective in determining permanent changes.While standard anti-poverty interventions such as cash and in-kind assistance play a crucial role in addressing immediate food security needs, achieving sustained, longterm improvements in living conditions requires comprehensive structural reforms capable of shifting the overall equilibrium upwards. 28Efforts should be directed at untying the knots that trap entire communities in poverty.In the case of refugees, this could involve tackling possible behavioural traps created by the psychological stress, trauma and hopelessness (Dang, Trinh, & Verme, 2022;Moya & Carter, 2019) or, more broadly, by reducing the impact of the mechanisms that decrease people's ability to sustain themselves, allowing these households to effectively accumulate assets.
The Ugandan approach to refugees has been praised as a very progressive model worldwide because it provides refugees with the right to work and freedom to move.It also provides refugees with agricultural land which has proven to be the most effective tool to improve the standards of living of both refugees and hosts.However, the sustainability of such a model has been called into question because land is becoming scarce.This calls for revisiting land allocation policy, possibly giving it only to the groups who engage in agriculture (e.g.Somali refugees in Uganda generally do not engage in agriculture, cf.Betts et al., 2019).Furthermore, land allocation should be coupled with interventions aiming to increase returns to assets.For instance, investing in public infrastructure can be key in fostering higher asset returns via complementarities (Escobal & Torero, 2005) or fostering higher farm yields through access to agricultural inputs and training.Other interventions should be aimed at opening new livelihood opportunities (Naschold, 2012) or improving households' self-reliance through market creation and social cohesion.
Another policy implication of our findings is the need to address the hosts' and refugees' needs together.Both populations are very poor and tend to very low equilibria.In such a context, standard interventions acting on education, skills, and the labour force have low returns because of the limited set of available economic opportunities for both hosts and refugees.For policies to become effective and a viable substitute to transfers, an expansion of the set of economic opportunities available to refugees is required.As emphasized in similar contexts (Verme et al., 2016), the policy focus must shift beyond social protection for refugees to include economic growth in the whole hosting areas, so that refugees and host communities can share in economic progress.This calls for a closer collaboration between humanitarian and development partners to transform a persistent humanitarian emergence into a development opportunity for all.
In conclusion, our work calls for a more nuanced approach to the Uganda model.We acknowledge that this modelbased on the right to work, access to land, and mobilityhas achieved significant positive results for both refugees and hosts, with important spillovers also beyond the local economy (d 'Errico et al., 2022;Kadigo & Maystadt, 2023;Zhu et al., 2023).However, we have also documented the existence of a non-trivial share of refugees and hosts that are not able to achieve higher standards of living.Failing of addressing the serious constraints trapping them at low level of welfarelack of human, natural, social and financial capitalwill not make possible to lift them out of poverty.

Notes
Poverty dynamics and poverty traps 381 the existence of thresholds (the dynamic asset poverty line w Ã in Figure1 panel b) separating each stable dynamic equilibrium's basin of attraction.

Figure 2 .
Figure 2. Households' location by year of first interview.

Figure 3 .
Figure 3. Income decomposition by refugee status, all years.

Figure 4 .
Figure 4. Kernel density of tradable asset index by wave by refugee status.

Figure 7 .
Figure 7. Local polynomial smooth by district.Note: The dashed lines report the refugees' and hosts' average equilibria of 0.11 and 0.21, respectively.

Figure 8 .
Figure 8. Local polynomial smooth of OLS-predicted current assets and actual lagged assets, refugees and hosts.
), as in the case of the tradable asset index.The owngroup equilibria identified(Figure A.3) are found at −0.34 and −0.1 poverty line units.24

Table 1 .
Sample composition across waves and refugees and hosts subpopulations (row percentages)

Table 2 .
Pattern of observed data throughout the panel

Table 3 .
Mean comparisons over time, refugees and hosts

Table 4 .
Non-parametric regression by groups, wave 1wave 4 unless otherwise specified Bold figures are equilibria statistically significantly different from the average whole sample equilibria (in italics).

Table 6 .
Parametric regression, asset growth from t-1 to t, pooled, RE and FE estimators

Table 6 .
(Continued)The metric of the dry season variable is reversed, i.e. higher values mean drier conditions. a

Table 8 .
Parametric regression, wave 2wave 4, OLS estimator Robust standard errors in parentheses.The dependent variable is the asset difference between the fourth and the second wave.Controls are two-periods-lagged variables: refugee status, age of the head (and its square), average education level (and its square), female headship, household size (and its square), married head, crop income (dummy), enterprise income (dummy), wage income (dummy), annual formal transfers (US$), annual informal transfers (US$), borrowed money (dummy), distance from agricultural market, petty trading market, primary school, negative and positive SPEI values, agroecological zones, Covid-19-related shock in the household (any), rural, subsamples of wave 1, and the interaction between year and district.ÃÃ p < 0.05; Ã p < 0.10.

Table 9 .
Parametric regression using different asset indexes, OLS estimator