The Accuracy Of Proxy Means Tests For Immigrant Populations: A Case Study In Colombia

Friday, May 1, 2020
by Will Sims

By Will Sims

Abstract

This paper examines the accuracy of proxy means tests (PMTs) for identifying low-income households among migrant and refugee populations. Specifically, it develops a PMT model based on Colombia’s SISBEN system, and evaluates its ability to identify poverty among recent and established Venezuelan migrants and refugees. It finds that these groups have significantly higher rates of exclusion errors relative to native Colombians, which could prevent them from accessing valuable social services. These findings are robust to a number of specifications, and the issue is not resolved by simply including immigration status within the model. Additionally, occupational downgrading is identified as the most likely mechanism for this effect, as Venezuelan migrants and refugees in Colombia generally have lower returns to education when compared with native Colombians. These results should inspire caution when choosing to use PMTs for targeting, and it is recommended that all policymakers evaluate the accuracy of their PMTs for vulnerable subpopulations prior to implementation.


Introduction

A large share of workers in the Global South are employed in the informal sector, and as a result accurate and reliable data on income, and thus poverty, is frequently unobtainable. Economic theory suggests that there is a correlation between income and asset ownership, meaning linear regression models can be used to predict the income of informal workers based on assets such as vehicle or livestock ownership, as well as more readily observable characteristics such as education level (Grosh 1994). These models, known as Proxy Means Tests (PMTs), have become a common tool over the past few decades for targeting anti-poverty interventions on both a local and national scale. Despite their widespread use, the efficacy of PMTs is still under debate, and much research remains to be done to understand both the theoretical and empirical limits of PMT methodologies.

This paper will make three main contributions to the literature on PMTs. First, it will provide the first academic assessment of PMTs in Colombia since Castañeda’s study in 2005. Second, this paper will assess the accuracy of PMT methods for measuring income in specific sub-populations — in particular, large groups of newly-arrived migrants and refugees. Prior to this work, analysis of sub-populations — including immigrant groups — has been largely absent from the literature on PMTs. Given the increasing size and vulnerability of immigrant populations across the globe, this is a particularly glaring omission. Finally, this paper seeks to contribute to broader debates about the efficacy of PMTs as a tool for targeting anti-poverty interventions. PMTs can be a useful tool in specific circumstances, but given their widespread use in the development and humanitarian communities, a comprehensive understanding of their limitations and drawbacks is essential for effective policy design.

PMTs can be a useful tool in specific circumstances, but given their widespread use in the development and humanitarian communities, a comprehensive understanding of their limitations and drawbacks is essential for effective policy design.

Given the natural variation among individuals, erroneous predictions are an unavoidable side-effect of PMT methods. When discussing poverty targeting, the errors generated by this unexplained variation are typically classified into two types: inclusion errors refer to individuals who are classified as poor or receive a benefit even though their actual income exceeds the model’s income threshold, while exclusion errors represent individuals who are below the threshold but for whom the PMT model predicts income above the threshold. Both types of errors can pose serious challenges for policymakers. While exclusion errors can prevent vulnerable households from receiving needed services, high rates of inclusion errors can result in higher costs for social welfare programs.

If we focus on immigrant populations in particular, two potential mechanisms could influence the accuracy of PMTs. First, immigrants typically engage in what economics literature has termed occupational downgrading — taking jobs for which they are overqualified or overeducated due to barriers to transferring qualifications or lack of formal employment status (Chiswick 1978). This could result in higher rates of exclusion errors, as well-educated immigrants are predicted to have higher incomes than they are able to earn in their new country.

On the other hand, it may take new immigrants time to build up stocks of assets, so predicting income based on assets could underestimate their income, leading to higher inclusion errors. If this is the case, this effect would diminish over time, as new immigrants acquire assets that align with corresponding native-born income levels. Of course, it is also possible that these effects cancel each other out, or that neither of the effects is significant enough to impact the overall accuracy of the PMT model. This paper will primarily be concerned with the occupational downgrading mechanism, as exclusion errors can have a significant negative impact for already vulnerable immigrant populations, and the relatively small size of the immigrant population will likely limit the impact of inclusion errors on the overall solvency of social programs.

Contemporary Colombia is an ideal context to examine these questions because PMTs have become a significant element of the country’s approach to both social welfare and poverty alleviation. The Sistema de Selección de Beneficiarios para Programas Sociales (SISBEN), a system of proxy means testing, has been used in Colombia since 1994. It is used to allocate services such as subsidized health insurance and conditional cash transfers (Vélez, Castaño, and Deutsch 1998; Attanasio and Mesnard 2006). Furthermore, in the past few years, Colombia has experienced an influx of more than a million migrants and refugees from Venezuela (UNHCR 2019). The Colombian government is currently using SISBEN in order to provide these migrants and refugees with subsidized healthcare and other social services (El Tiempo 2018; CONPES 2018). While this paper will focus on a theoretical PMT model (the current SISBEN formula is confidential), the concerns that it raises are applicable to any PMT model currently in use by policymakers. 

Colombia’s recent influx of Venezuelan migrants and refugees also presents a particularly interesting case for statistical reasons. First, Venezuelan migrants and refugees are a compelling comparison class because they have arrived recently and in large numbers, lending them legibility within the data. Second, Colombia and Venezuela are relatively similar in terms of shared history and culture, so it is less likely that divergent patterns in asset ownership are driven by cultural preferences. Third, Venezuelan immigration to Colombia is catalyzed by a political and economic crisis, so the migrants and refugees likely represent a broader cross-section of education and pre-migration earnings than typical economic migrants (UNHCR 2019). In terms of broader validity, these unique characteristics of the immigrant population may limit the cross-group and cross-country applicability of our conclusions. That said, we can also speculate that this analysis will likely underestimate the inaccuracy of PMTs relative to other contexts.

 

Previous Analysis of PMT Accuracy

The theoretical basis for PMT methods is relatively straightforward: higher-income households tend to own more assets, so rough surveys of observable assets can be used to predict income in situations where reliable income data is unavailable. Some research has found that PMT targeting programs can effectively identify poor households, and PMTs have outperformed other targeting methods such as community-based targeting in some studies (Grosh et al. 2008). Furthermore, research suggests that PMTs do not distort consumption, as asset accumulation habits do not change significantly in response to the implementation of PMTs (Banerjee et al. 2018). Based on findings like these, PMTs have become a recommended approach to poverty targeting within international development organizations like the World Bank (World Bank 2017).

PMTs have become a recommended approach to poverty targeting within international development organizations like the World Bank. 

Other research has questioned the overall accuracy of PMT methods, suggesting that high error rates in the model can undermine both their effectiveness and political viability (Kidd and Athias  2019). Even the most accurate PMT models report R2 values between 0.3 and 0.6, leaving 40 to 70 percent of total variation in income unexplained (Kidd and Wylde 2011).  They also point to the inherent trade-off between the size of the targeted population and PMT accuracy — narrower bands of income or consumption mathematically increase the probability of exclusion errors. Finally, these statistical issues with the model are compounded during the implementation of PMTs, including inaccuracy over time, problems with falsification of responses, and recipient dissatisfaction. Despite these criticisms, lack of consensus around alternative methods has allowed PMTs to remain a common approach to poverty targeting.

Latin America in particular has a long history of PMT projects and research. Much of the original development of PMT methods by Margaret Grosh and The World Bank took place in Chile in the 1990s (Grosh 1994; Grosh and Baker 1995). Since then, there have been numerous evaluations of PMT models across Latin America, including studies in Peru (Johannsen 2005) and Bolivia (Klasen and Lange 2015) that found exclusion error rates ranging from 8 to 26 percent. Going beyond models to evaluate implementation, Castañeda and Lindhert compared actual PMT performance in Mexico, Brazil, Costa Rica, Chile, and the United States (Castañeda and Lindhert 2005). They found rates of exclusion as high as 84 percent for Chile’s PASIS old-age pension. Specific to Colombia, a related study by the same authors estimated a best-case exclusion error rate of 19 percent for Colombia’s SISBEN PMT program (Castañeda 2005). The Castañeda and Lindhert research provides essential context for this paper, but the exclusion rates from these evaluations should not be compared directly. Their study also includes the exclusions that result from the uneven implementation of the PMT targeting schemes, while this piece solely addresses exclusions based on the statistical attributes of the model.

Despite this breadth of research, analysis of the accuracy of PMTs for immigrant subpopulations has been largely absent from this discourse. 

Despite this breadth of research, analysis of the accuracy of PMTs for immigrant subpopulations has been largely absent from this discourse.  For migrants in particular, the Progresa/Oportunidades/Prospera conditional cash transfer program in Mexico utilizes a PMT model that includes questions on the immigration status of household members, but the focus is on the role of immigration as a determinant of remittances for non-migrant household members (Skoufias, Davis, and Behrman 1999; Orozco and Hubert 2005). Similarly, the PMT survey for the Bolsa Familia in Brazil includes questions about internal migration, but does not identify foreign immigrants (Castañeda and Lindhert 2005). We have found no evaluations that focus on the accuracy of PMTs applied to immigrant populations, either in Latin America or in other regions. This is a particularly glaring omission as the growing ubiquity of PMTs as a development tool has coincided with increased flows of immigrants and refugees around the globe.

In spite of its focus on the Colombian context, this paper will seek to contribute to both regional debates on PMT effectiveness, as well as the discourse on poverty targeting more generally. Globally, refugee and migrant populations are growing in size and increasingly vulnerable. Excluding these populations from social protection programs due to PMT selection methods risks undermining the effectiveness of these programs.

 

Methodology and Data

The data for this study is drawn from Colombia’s Gran Encuesta Integrada de Hogares (GEIH) — a nation-wide household survey performed by the Departamento Administrativo Nacional de Estadística (DANE), the Colombian national statistics office. This data was compiled from three distinct datasets:

  • GEIH Viviendas y Hogares: individual data on household assets
  • Medición de Pobreza Monetaria y Desigualdad: household data on income and individual data on education
  • GEIH Immigration Data: individual data on individual birthplace and place of residence 5 years and 1 year ago

The unique identifier variables Directorio and Secuencia P were used to link households across the three datasets. In addition, the Factor de Expansión variable was used to assign survey weights in the model. The datasets were compiled for 2018 — the most recent year for which data was available.

Income in the GEIH is self-reported, and thus can run into the same issues with misrepresentation that PMTs are designed to avoid. This issue is unavoidable in any analysis of income in the informal sector, but is mitigated in this case by a number of factors. First, the GEIH data is collected by DANE explicitly for research purposes and not for use in any targeting program. This reduces the incentive for participants to lie about their income in order to receive benefits. Second, the GEIH survey is highly detailed, and explicitly asks for incomes from work, rent, social programs, and other sources, reducing the risk that respondents misremember or misreport their incomes. Missing values are imputed by DANE based on similar households only where strictly necessary.

Two additional key variables, immigration status and education level, were only available at the individual level. For these variables, households were classified based on the status of their household head. Household-head status was self-reported by respondents. The creation of a more detailed household-level index for immigration status or education was considered, but decided against due to risks of researcher bias in the creation of the index.

Migration and citizenship status were also not reported directly and had to be defined by the researcher. The GEIH survey includes questions on the respondents’ place of birth, as well as where they lived 5 years and 1 year ago. These questions were used to classify the immigration statuses in Table 1:

Table 1 shows how the four groups are defined.

Unlike Venezuela, Colombia does not grant birthright citizenship, so there is a chance that individuals born in the country do not have citizenship status. Nevertheless, we believe that the “Colombian” status is representative of individuals who have assets, income, and education characteristics that are indistinguishable in aggregate from Colombian citizens. “Venezuelan Residents” were included in order to characterize the impact of any characteristics specific to Venezuelan citizenship, but distinct from immigration status. Venezuelan “Recent” and “Established” migrants and refugees were divided in order to evaluate the potential impact of asset accumulation over time.

Other immigration statuses, including individuals who were born in neither Colombia nor Venezuela, were included when calculating the PMT models, but excluded from the overall analysis of model accuracy. These other immigrant categories are heterogeneous, including everything from temporary workers from neighboring South-American countries, to European and American retirees with high wealth and no income. In order to focus the analysis on the question of interest, this paper will focus only on comparisons of Colombians, long-term Venezuelan residents of Colombia, and the two types of Venezuelan migrants and refugees.

 

Descriptive Analysis

There are significant differences between the different categories of Venezuelan migrants and refugees and their Colombian counterparts in terms of age, education, and income. Table 2 breaks out the number of surveyed individuals in both the training and testing partitions, as well as population statistics by subgroup:

Table 2 shows summary statistics of the four defined groups.

Age is the mean age reported by respondents in years. Income is the median income, reported in thousands of Colombian pesos (1 million Colombian pesos is equivalent to approximately $300 USD at current exchange rates). Education is the mean highest grade level completed. For age and education, the p-values are generated by a Tukey’s Test for pairwise differences in ANOVA, compared against Colombians as a baseline. For income, the p-values are generated by a pairwise Mood’s median test, again compared against Colombians.

The results of these tests suggest that as a cohort both recent and established Venezuelan migrants and refugees are generally younger, poorer, and better-educated than either Venezuelan-born or native-born Colombian residents. 

The results of these tests suggest that as a cohort both recent and established Venezuelan migrants and refugees are generally younger, poorer, and better-educated than either Venezuelan-born or native-born Colombian residents. This lends credence to the possibility that Venezuelan migrants and refugees are experiencing occupational downgrading, as they have lower earnings than native Colombians despite higher levels of education. Furthermore, the relative similarity in incomes between Colombians and Venezuelan Residents suggests that there is nothing inherent to being born in Venezuela that is causing this effect. Finally, the difference between incomes for Venezuelan Established Migrants and Venezuelan Recent Migrants suggests that there may be a process of adaptation occurring, as Venezuelans are gradually able to establish themselves in the Colombian labor market. The effect of these differences on the PMT model is the focus of the next section.

 

Model Fit

The analytic approach undertaken for this paper is fairly straightforward. First, I estimate an ordinary least squares PMT model for all Colombian residents and migrants in the training partition of the dataset (a randomly-selected 25 percent of the 2018 data). The dependent variable for the model is the log of household income, and the independent variables are drawn from both previous research on PMTs (Hanna and Olken) and Colombia’s own SISBEN III PMT model for household targeting:

ln(Ydhi) =  α + βdθd + βhγh + βiδi + εdhi

In this regression equation, θd is a categorical variable for Departamento (the equivalent of states in Colombia). γh is a vector of household-level asset and housing quality variables, including household size and access to utilities like electricity and running water. Finally, δi is a set of variables representing characteristics of the self-identified household-head, including years of education and health insurance status. An abridged regression output is included in Table 3, and the full list of variables and regression results is available in Appendix 1.

Table 3 shows partial results for the main regression specification.

The model overall has an adjusted R2 value of .422. While this is not particularly high, it is well within the range of PMT models that are currently in use in development contexts. The model found statistically significant coefficients on a majority of the variables, including nearly every level of the Department factor. None of the variables included in the model had an unexpected sign or magnitude. Some of the largest and most significant effects came from flooring material in the home, car and motorcycle ownership, and household head education level. Overall, the model was considered to have sufficient predictive power to justify its use in the evaluation of PMT accuracy in this case study.

Overall, the model was considered to have sufficient predictive power to justify its use in the evaluation of PMT accuracy in this case study.

The model was then used to generate a predicted income for each household in the testing partition. The testing households were then classified as exclusion error, inclusion error, or correctly categorized, based on the relationship between the model-predicted income and the actual status of the household’s income relative to the DANE-estimated poverty line. The error rates for Venezuelan groups were compared with Colombians using a simple z-test for difference-in-proportions. Several robustness checks were performed to establish the reliability of the results across different subsets of the data.

 

Results

PMT Error Rates

The first assessment of this model is a simple comparison of exclusion error rates between different immigrant groups, relative to the poverty line estimated by DANE. The poverty line varies for households based on their family size and estimated costs of basic nutrition, ranging from 167,000 to 304,000 Colombian Pesos per month (approximately $50-$90 USD at current exchange rates), with an average of 271,000 Colombian Pesos  ($80 USD). Households were considered ’Correct’ if their PMT-predicted income correctly classified them as above or below their DANE-estimated poverty line. Exclusion errors represent individuals who have actual incomes below the DANE-estimated poverty line, but predicted incomes above the line. Inclusion errors are the converse. 

Table 4 compares exclusion and inclusion rates for each of the four groups.

The rates of exclusion errors are nearly identical for both Established and Recent Venezuelan Migrants, and both groups have a rate of exclusion error more than twice as high as the error rate for Colombians. Venezuelan residents also have slightly higher rates of exclusion error. This would suggest that the PMT is overestimating the wealth of both Recent and Established Venezuelan Migrants, but after more than 5 years of residence in Colombia the effect begins to dissipate. The rates of inclusion error were similar across immigration-status groups, justifying our focus on exclusion errors.

To verify these observed differences in exclusion error rates, the exclusion error rates for immigrant groups were compared to Colombians in a pairwise two-sample difference of proportions z-test. The results of these tests are detailed in Table 5, including a 95 percent confidence interval for the difference in proportions:

Table 5 focuses on the exclusion errors.

Since these results are statistically significant at the .01 significance level, we can conclude that Recent and Established Venezuelan Migrants below the poverty line have higher rates of exclusion from the PMT model relative to Colombians. Venezuelan Residents who have been in Colombia for at least 5 years also have slightly higher rates of errors, but this difference is only significant at the p<0.1 level. In terms of practical implications, while a 16 percent exclusion error rate may be acceptable for the PMT model in general, this model excludes approximately one in three Venezuelan migrants and refugees below the poverty line threshold.

...this model excludes approximately one in three Venezuelan migrants and refugees below the poverty line threshold.

Robustness Check 1: Near-Poverty Error Rates

We might be concerned that these error rates are simply driven by the fact that there are a higher share of migrants and refugees clustered around the poverty line, which would mechanically increase the probability of exclusion errors. This is further reinforced by our previous finding that Colombians have higher median incomes than Venezuelan migrants. In order to address this issue, the analysis was rerun using a subset of the data restricted to only cases that have actual income less than 200 percent above the poverty line (Attempts were made to take an even smaller subset of incomes as a robustness check, but the resulting sample sizes were signicantly smaller, resulting in concerns about the power of the resulting tests). This resulted in the subset in Table 6, which captures approximately the bottom quartile to third of the income distribution — in line with other PMT targeting thresholds.

Table 6 shows details of the bottom third of the income distribution.

These results also indicate that concerns about the distribution of incomes across immigrant groups was well-founded. Colombians within the specified income range are a much smaller share of the overall tes t data compared to the immigrant populations. Once this subset is taken, however, the mean income across groups is comparable, suggesting that income distribution effects have been mitigated. The subset was then assessed using the same methods as the test-data PMT:

Table 7 show the error rate for the bottom third of the income distribution.

As expected, the error rates in Table 7 are higher across the board for the restricted dataset than for the whole income distribution. Nevertheless, the rates are still significantly higher for the three Venezuelan groups. This suggests that even within this restricted cohort, there are systemic factors that are reducing the accuracy of the PMT model for Venezuelan migrants and refugees. Furthermore, these results imply that for migrants and refugees with incomes near the poverty line, the PMT model is actually more inaccurate than accurate. With exclusion error rates higher than 50 percent, we would expect that a series of fair random coin-flips could categorize migrants and refugees at a roughly similar rate to the PMT model. In practical terms, this inaccuracy would result in high levels of exclusion of migrants and refugees from social programs that rely on PMT targeting.

Robustness Check 2: Work-Status Error Rates

As a further check, we estimated the amount of variation in the error rates that is attributable to different rates of employment among different immigration subgroups. The GEIH asks respondents how they spent a majority of their time in the previous week, with a majority of cases reporting that they were working or looking for work. As illustrated in Table 7, Colombians are less likely to be working, but also less likely to be looking for work compared with Venezuelans in Colombia. They also have a higher share in the ’Other’ category, which includes students, homemakers, and the disabled.

Table 8 shows rates of working/looking for work, etc. across these groups.

The PMT does a very poor job of prediction for individuals with zero income.

In order to mitigate the effect of these differential work statuses on our evaluation of the PMT’s accuracy, we restricted the exclusion error analysis to only include self-identified workers. Furthermore, we can go beyond work status to restrict the analysis to individuals who have income. In this data “zero income” is different from “not working,” because most non-workers (including students, homemakers, and the disabled) have some source of income, such as rents or transfer programs. The PMT does a very poor job of prediction for individuals with zero income, due to the relatively high intercept term in the model and the fact that every case in the data had at least some housing attributes or asset ownership. Table 8 documents exclusion error rates for various combinations of these data subsets, including the worker-only, positive income, and near-poverty (NP) subsets. As before, p-values are based on pairwise difference in proportions tests against Colombians:

Table 9 shows exclusion error rates for various subsets of the data.

These results suggest that near the poverty line the model is less accurate for workers, and at least some of the low error rates for near-poverty Colombians are attributable to the correct classification of Colombian non-workers. This implies that at least some proportion of the differences in model accuracy is attributable to the higher share of workers in the Venezuelan immigrant populations. This is encouraging, as it seems like rather than being biased against immigrant groups in particular, the model is biased against workers near poverty, a group in which migrants and refugees are over-represented. Unfortunately, while this seems to resolve some of the inaccuracy, it is likely not feasible to include work status in an actual PMT because it would be very easy for respondents to misrepresent their occupational status. Ideally, there would be another observable factor that could be used to decrease exclusion error rates for immigrant populations without relying on easily misrepresented information.

Identifying the Occupational Downgrading Mechanism

In order to isolate the educational downgrading mechanism, we looked at the model’s accuracy across different educational strata. Overall, Venezuelan migrants and refugees have higher levels of education relative to Colombians, with a lower share having a primary school or lower education and a higher share with media (10th and 11th grade) education:

Table 10 shows the accuracy of the model across educational strata.

To roughly assess the returns to education for different immigrant and non-immigrant groups, we can look at a regression of the interaction of immigration and education on income using Colombians as a baseline, controlling for age in order to mitigate the effect of age-differences between the cohorts: 

ln(Yabc) = α +[ βa(EDUCATION LEVEL) * βb(IMMIGRATION STATUS)] + βc(AGE) + ϵabc

We can see in Table 11 that the overall positive coefficient for education level has a weakly negative interaction with Venezuelan Established Migrants and significantly negative interaction with Venezuelan Recent Migrants.

Table 11 shows the results of the regressions interacting education and immigration status.

This suggests that occupational downgrading is indeed occurring within the data: returns to education are lower for Venezuelan migrants and refugees than they are for Colombians. In terms of the PMT model, if we compare rates of exclusion errors across both immigration and education-level subcategories, Table 12 finds lower rates of exclusion error among university graduates across all immigration categories:

Table 12 finds lower rates of exclusion error among university graduates across all immigration categories.

This effect is particularly strong for Colombians. For both Recent and Established Venezuelan Migrants, exclusion errors at the university level are still equal or higher than the exclusion rates for Colombians at any education level. Furthermore, these differences between the rates are higher as we move up education levels. This implies that the observed occupational downgrading is indeed increasing the inaccuracy of the PMT model for Venezuelan migrants and refugees.

Modified PMT Model with Immigration Status Term

A possible solution to this issue is to simply include immigration status in the PMT model itself. To test this approach, the previous regression model was modified to include the βkIMMIGRATION term — a factor variable for immigration status:

ln(Ydhik) = α + βdθd + βhγh + βiδi + βkIMMIGRATION + ϵdhik

The full model result is available in Appendix 2, and relevant coefficients for the immigration-status factor variable are outlined in Table 13, relative to the Colombian reference category:

Table 13 shows relevant coefficients from the regression of the specification including the immigration term.

The overall predictive strength of the model has barely increased: Adjusted R2 has gone up by .001, and RSE has gone down by just .007 relative to the model without immigration. The specific coefficients for Venezuelan  migrants, however, are statistically significant at the .01 level. In terms of magnitudes, the coefficient on Recent Migrants is comparable to owning a car, and the coefficient on Established Migrants is comparable to having health insurance. It is unclear from this result exactly why the coefficients are positive — the most plausible scenario is that migrants and refugees occupy a uniquely underprivileged position in regards to their combination of housing and asset variables, so their average income conditional on the other PMT variables requires a positive correction. According to Table 14, due to this low increase in predictive power, the revised model does almost nothing to decrease the level of bias in the exclusion error rates.

Table 14 shows the exlusion rates of the new specification including immigration status.

This result implies that the bias in the model cannot be corrected by simply including immigration status in the model, but requires a reevaluation of the model itself. Optimization to remove this bias will likely require reworking the set of variables included in the model, or providing a post-hoc correction to the PMT predictions for immigrant populations. While this likely involves significant data gathering and pre-testing on the part of the PMT designer, it may be the only way to ensure that vulnerable populations are not excluded from programs in an unfair manner due to statistical errors.

This result implies that the bias in the model cannot be corrected by simply including immigration status in the model, but requires a reevaluation of the model itself. 

 

Policy Prescriptions

The assessment in this paper is a case study of a particular choice of PMT model, applied to a specific threshold in Colombia, compared against a unique subpopulation. That said, it is possible to use this case to draw two general conclusions regarding the distributional equity of PMTs:

  1. Policymakers should be cautious in the implementation of PMTs, avoiding them if at all possible. If the results of this paper were to hold in the implementation of a nationwide PMT program in Colombia, not only would the PMT erroneously exclude significant numbers of Colombians, significant numbers of Venezuelan migrants and refugees would also be excluded due to statistical bias. As a rough calculation, UNHCR estimates that by the end of 2018 there were approximately 1.2 million displaced Venezuelans in Colombia; the 34 percent exclusion error rate found in this model could leave more than 400,000 individuals without benefits and services to which they are entitled. There is a robust evidence base supporting both universal programs and other non-statistical targeting methods, which should be considered alongside PMTs wherever possible (Kidd and Athias 2019).

  2. If PMTs must be used, they should be checked for accuracy within sub-populations prior to implementation. Given the demonstrated possibility for unequal coverage of vulnerable sub-populations within a PMT model, it is important that the evaluation of PMTs include estimates of coverage rates and exclusion errors prior to implementation. This paper proposes a methodology for the evaluation of PMTs for sub-populations that can be drawn upon in other country contexts and for other gender, disability, or ethnic sub-populations. The comparison of exclusion error rates against a baseline category is only a single dimension of PMT bias, but it could be easily identified using this type of pre-analysis. Once identified, it is incumbent upon policymakers to design targeting programs that mitigate or avoid these types of biases.

 

Conclusion

In addition to these recommendations, this case study also identifies several areas for future research. There is ample room to test for this type of PMT exclusion bias across different models, country contexts, and sub-populations to develop a sense of its extent. In terms of immigrant sub-populations, it will be important to see how PMT models hold up over time. PMT models are typically recalculated once every several years, so it is an open question as to whether a model that may be unbiased based on a single year’s results is able to maintain accuracy in the face of a large wave of immigration. Providing an alternative approach to PMT modeling was outside the scope of this paper, so it is incumbent on future research to determine how to best excise exclusion bias from targeting approaches. Finally, as a general project it is important to continue to probe the limitations and biases of PMT models in order to advance a more just and equitable paradigm for economic development and social protection.


About the Author

Will Sims is an MPP student at the University of Michigan's Gerald R. Ford School of Public Policy. He would like to thank Professors John Hanson and Alton Worthington at the University of Michigan, and Andrea Pellandra at UNHCR for their invaluable support in the execution of this project. He can be reached at wlsims@gmail.com.


References

Attanasio, O, & Mesnard, A. (2006). The Impact of a Conditional Cash Transfer Programme on Consumption in Colombia. Journal of Applied Public Economics, Institute for Fiscal Studies.

Banerjee, A. Hanna, R. Olken, B. & Sumarto, S. (2018). The (lack of) Distortionary Effects of Proxy-Means Tests: Results from a Nationwide Experiment in Indonesia. NBER Working Paper No. w25362.

Camacho, A. & Conover, E. (2011). Manipulation of Social Program Eligibility American Economic Journal: Economic Policy.

Castañeda, T. (2005). Targeting Social Spending To The Poor With Proxy–Means Testing: Colombia’s SISBEN System. Social Protection Discussion Paper Series, The World Bank.

Castañeda, T. & Lindert, K. (2005). Designing and Implementing Household Targeting Systems: Lessons from Latin American and The United States. Social Protection Unit, Human Development Network, The World Bank.

Chiswick, Barry R. (1978). The Effect of Americanization on the Earnings of Foreign-born Men. Journal of Political Economy.

CONPES (Consejo Nacional De Política Económica Y Social). (2018) Estrategia Para La Atención De La Migración Desde Venezuela. Departamento Nacional De Planeación.

El Tiempo, (2018). ¿Un venezolano que vive en Colombia puede acceder a salud o al Sisbén?  Journal of Political Economy.

Grosh, M. (1994) Administering Targeted Social Programs in Latin America: From Platitudes to Practice. The International Bank for Reconstruction and Development, The World Bank.

Grosh, M. & Baker, J. (1995) Proxy Means Tests for Targeting Social Programs: Simulations and Speculation. Living Standards Measurement Survey Working Paper, The World Bank.

Grosh, M., Del Ninno, Carlo., Tesliuc, E. & Ouerghi, A. (2008). For Protection and Promotion: The Design and Implementation of Effective Safety Nets. The International Bank for Reconstruction and Development, The World Bank.

Hanna, R. & Olken, B. (2019). Universal Basic Incomes vs. Targeted Transfers: Anti-Poverty Programs in Developing Countries. NBER Working Paper Series.

Johannsen, J. (2006). Operational Poverty Targeting in Peru — Proxy Means Testing with Non-Income Indicators. United Nations Development Programme, International Poverty Centre Working Paper.

Kidd, S. & Athias, D. (2019). Hit and Miss: An assessment of targeting effectiveness in social protection Working Paper, Development Pathways.

Kidd, S. & Wylde, E. (2011). Targeting the Poorest: An assessment of the proxy means test methodology AusAID.

Klasen, S. & Lange, S. (2015). Accuracy and Poverty Impacts of Proxy Means-Tested Transfers: An Empirical Assessment for Bolivia. Courant Research Centre, Georg-August-Universitat Gottingen.

Orozco, M. & Hubert, C. (2005). La Focalizacion En El Programa De Desarrollo Humano Oportunidades De Mexico Social Protection Unit, The World Bank.

Skoufias, E., Davis, B. & Behrman, J. (1999). An Evaluation of the Selection of Beneficiary Households in the Education, Health, and Nutrition Program (PROGRESA) of Mexico.

International Food Policy Research Institute UNHCR. (2019). The 2018 Global Trends Report. United Nations High Commissioner for Refugees.

Vélez, C., Castaño, E. & Deutsch, R. (1998). An Economic Interpretation of Colombia’s SISBEN: A Composite Welfare Index Derived from the Optimal Scaling Algorithm. Poverty and Inequality Advisory Unit, The Inter-American Development Bank.

World Bank. (2017). Closing the Gap: The State of Social Safety Nets, 2017. World Bank Group.
 


Appendix

Appendix table 1 show all the coefficients from the first regression.

 

Appendix table 2 shows the full results for the expanded coefficient.