Skip to main content

Advertisement

SpringerLink
Log in

Does it pay to be poor? Testing for systematically underreported GNI estimates

  • Published:
The Review of International Organizations Aims and scope Submit manuscript

Abstract

Coordinating aid distribution to the poorest countries requires identifying which countries are poor. In practice this has meant sorting countries into developmental cohorts on the basis of macroeconomic data, with countries in poorer cohorts gaining access to more and more concessional aid programs. To the extent that governments can influence their macroeconomic data, some, especially those in aid-dependent countries, may prefer to report data that sorts them into lower development cohorts. We term such behavior “aid-seeking data management.” The possibility of data management has substantial implications for aid distribution, and for the use of macroeconomic data in social scientific settings. We look for evidence of aid-seeking data management in the distribution of GNI per capita data around the eligibility threshold for the World Bank’s International Development Association (IDA). Because macroeconomic data are subject to frequent ex post revisions, we separately analyze the heavily revised data available for download from the World Bank’s World Development Indicators and the substantially less revised data that we gleaned from back issues of print edition of World Bank Atlas. We find that the less revised GNI per capita data display patterns that are consistent with aid-seeking data management among aid-dependent countries, and only among aid-dependent countries. This finding is robust to a variety of model specifications, but somewhat sensitive to the exclusion of individual countries from the sample. We find no such evidence in the currently downloadable data, suggesting that whatever biases aid-seeking data management may have generated in early data releases are largely and perhaps entirely wiped away in ex post revisions.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Log in via an institution

Subscribe and save

Springer+ Basic
$34.99 /Month
  • Get 10 units per month
  • Download Article/Chapter or eBook
  • 1 Unit = 1 Article or 1 Chapter
  • Cancel anytime
Subscribe now

Buy Now

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Institutional subscriptions

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6
Fig. 7

Similar content being viewed by others

Notes

  1. While our theoretical frame centers on governments’ willingness to manage development statistics, one could theorize this relationship from the point of view of international institutions’ incentives to reward certain countries with greater access to aid. Our empirical evidence does not obviously rule out that interpretation.

  2. Other relatively dramatic examples of consequential rebasing in Africa include Ghana, which rebased in 2010 and increased GDP by 62.8 % and Kenya, which rebased in 2014 and increased GDP estimates by 25.3 %. Botswana rebased in 2011 and lost 10 % of its GDP estimate in the process.

  3. Being incentivized to deliver economic growth does not necessarily imply a reluctance to manage data. Argentina, for example, was recently censured by the IMF for providing inaccurately rosy inflation and GDP growth data (Economist 2013). Chinese GDP figures are routinely assumed to be overly optimistic (ex. Magnier 2015).

  4. Returning to the Nigerian example, the surprising part of the episode was perhaps not the scale of the revision (brought about by updating the benchmark year from 1990 to 2010), but that the Nigerian political leadership allowed its macroeconomic data to remain so understated for decades (see Jerven 2013a, b). Jerven (2014b) notes speculation that the delay was partly motivated by a desire to qualify for the Heavily Indebted Poor Countries Initiative.

  5. Even without the necessarily intertwined relationship between GNI and GDP, the possibility that governments may influence GNI data raises the possibility that governments may be able to influence other macroeconomic data as well.

  6. As Jerven (2013b) notes, the process by which the data reported by national statistical offices get translated into the data reported by the World Bank is proprietary to the World Bank.

  7. As noted in more detail below, IDA eligibility is not solely determined by a GNI per capita that is below a pre-determined threshold, but placement relative to that threshold is a tremendously important component.

  8. The data used in this paper were downloaded in June 2015.

  9. The IBRD does not assign a risk premium to its loans, so access to these credits is limited to countries that undergo a fairly rigorous qualitative determination of creditworthiness.

  10. The following summary describes loan terms and loan eligibility as they were at the time of the paper’s writing. While the general thrust and intention of IDA policies have not changed substantially over time, the specific terms change annually, and over time these changes can amount to substantial differences. Historical data on loan terms and rates are available from the IDA website < < http://www.worldbank.org/ida/index.html>>

  11. An exception is made for small island nations that lack creditworthiness but have GNIs per capita higher than the IDA cutoff. Countries with such an exception currently include: Cape Verde, Sao Tome and Principe, Marshall Islands, Micronesia, Tuvalu, Vanuatu, Samoa, Tonga, Grenada, Dominica, Saint Vincent, Saint Lucia, and the Maldives.

  12. Ravallion (2013) recounts that the use of GNI-based thresholds at the World Bank dates back to 1971, when it was decided that countries with less than $200 GNI per capita would be allowed to exercise the so-called civil works preference, meaning that they are allowed to offer preference to national suppliers when procuring certain goods and services.

  13. Countries borrowing on blend terms include: Bolivia, Moldova, St. Lucia, Cape Verde, Mongolia, St. Vincent and the Grenadines, Cameroon, Nigeria, Timor-Leste, Congo Rep., Pakistan, Uzbekistan, Dominica, Papua New Guinea, Vietnam, Grenada, Sri Lanka, Zimbabwe.

  14. Flexible spread loans and loans denominated in other currencies can have different interest rates.

  15. Eleven countries have reverse graduated (IDA 2012c, 9).

  16. Nigeria’s exclusion from IDA-only status came about from its ostensible creditworthiness at the IBRD – its GNI per capita was at the time estimated to be substantially under the eligibility threshold.

  17. The standards of national accounts have since been revised three times; thus, there are four versions. In addition to SNA 1952, there are also SNA 1968, SNA 1993, and SNA 2008. However, Ward (2004) argues that “although they pay lip service to the subsequent revisions . . . many countries still adhere to the basic system and its corresponding accounting foundations as first set out” (45).

  18. This was also when the World Bank started converting to US dollars using the Atlas method in order to deal with volatile nominal exchange rates. See Ward (2004, 97–99) for fuller account.

  19. If there is no agreement or if there are no regular close relations between IMF or World Bank missions and the national government, the country and the World Bank may publish different numbers (see Jerven 2013b, Table 1.2, 24–25). These data are often revised and this revision process is documented in the World Bank Group Archives. (Folder 1353586, WB IBRD/IDA DEC-03, Records of the Office of the Vice President, Development Economics, and Chief Economist and later Senior Vice President, Development Economics and Chief Economist (DECVP), World Bank Group Archives, Washington, DC.) The revised figures are not published until the country economist and the World Bank’s International Economics Department have approved them.

  20. Threshold data are available at siteresources.worldbank.org/DATASTATISTICS/Resources/OGHIST.xls.

  21. For the years 2012 and 2013 we used historical data releases available from the World Bank’s website. We use the December release to remain consistent with our Atlas-derived data.

  22. Designations relative to the IDA eligibility threshold are made on the basis of the data as they exist in the July following the year to which they refer. The World Bank Atlas reflects estimates in December that year. So, for example, the initial official GNI per capita estimates for 2000 are set in July 2001, but the Atlas records those data as they exist in December 2001. Recent releases of historical estimates suggest that there is often no difference between the July and December estimates, but for a substantial number of countries the estimates do change between July and August, typically by small amounts. We see no reason based on the data we have access to suspect that there are systematic patterns in these July-December revisions, but this assertion is better tested with more comprehensive data than what is available to the authors at the time of writing.

  23. The World Bank’s measure of aid per capita is stated in current dollars, which we converted into constant 2005 dollars.

  24. Among country-years in our sample within $300 of the IDA threshold, the correlation between aid per capita and aid as a percentage of GNI is .89. Among country-years in our sample within $100 of the IDA threshold, the correlation climbs to .96.

  25. In unreported tests we re-estimated our models using a measure of aid as a percentage of GNI. The results of these analyses are substantively similar to what is reported in the paper, but their statistical significance is in general more sensitive to arbitrary specification changes. For example, we report the results of McCrary tests using aid per capita at lag lengths of 8, 9, 10, 11 and 12-years. Tests using aid per capita produce a statistically significant break in the density function at any of these lag lengths; tests using aid as a percentage of GNI only result in a statistically significant break in the density function when operationalized at a 8,9, 11 or 12-year lag. Our regression analyses using aid as a percentage of GNI are also more sensitive to specification changes. These distinctions seem (especially in the case of the regression estimates) to be largely driven by the increase in missing data and its distribution, rather than by the values of aid per se.

  26. To illustrate, 92 % of post-1988 observations with a GNI per capita that was within $300 under the IDA eligibility threshold in year t were also under the IDA eligibility threshold in year t-1 while only 20 % of country-year observations with a GNI per capita within $300 over the IDA eligibility threshold in year t were also under the IDA eligibility threshold in year t-1This difference is statistically significant at the .001 level. A Kolmogorov-Smirnov test also suggests statistically significant differences across the two samples.

  27. The correlation between aid per capita and its 10-year lag is .578 in our sample and highly statistically significant.

  28. We also considered using lagged infant mortality, which has been used as an instrument for aid dependence elsewhere in the literature (ex. Burnside and Dollar 2000; Knack 2001) but found it not to be well correlated with current aid per capita in our sample, which is not entirely surprising given the similar development levels across countries within our sample. Regressions demonstrating as much can be found in the appendix.

  29. We experimented with two ways to define this set of countries. One was to exclude countries for which the small island exception applied. The other is to exclude all small island nations with population below 1.5 million, for whom the small island exception could theoretically apply given a severe deterioration of those countries’ macroeconomic circumstances. For example, Barbados and The Bahamas would be included in the former sample but not the latter. In practice these coding rules produce very similar results. We present the results using the former coding as it allows more countries into the sample.

  30. We implement these tests using the DCdensity package in Stata. http://emlab.berkeley.edu/~jmccrary/DCdensity/DCdensity.ado.

  31. Kernel based estimators such as this are highly sensitive to bandwidth selection (see McCrary 2008; Imbens and Lemieux 2008). Our main tests all employ a bandwidth of 300. We note in the text how these estimates compare to estimates derived using alternative bandwidths. Except where otherwise noted, we use the McCrary test’s default binwidth selector, which is equal to 2*sd(runvar)*length(runvar)^(−.5), where the “runvar” refers to the GNI per capita – the IDA threshold. See McCrary (2008).

  32. These non-findings persist in alternative specification using bandwidths of 100, 200, and 400.

  33. We define the median as the median for the whole sample, rather than calculating a separate median for each year. The choice of $500 is atheoretical, but it seems to us a reasonable definition of the relevant comparison countries. In unreported robustness tests we estimated our models using more and less restricted samples from which to take the median, and came to similar results.

  34. These two disadvantages work against each other. Choosing more finely-grained divisions of the data (quartiles, for example) may give a better picture of what is happening in the data, but also creates samples that are even smaller and less reliable.

  35. Note that the slight overlap of the confidence intervals does not preclude there being a statistically significant difference between the two estimates (Schenker and Gentleman 2001).

  36. Threshold_Distance is in principle almost certainly endogenous to the lagged value of aid dependence. However, when we regressed Threshold_Distanceit on aid per capitait-10 for the different samples of data that we employ we failed to find any statistically significant relationship between the two variables. Statistical significance aside, the coefficients on our estimates were quite small, implying that a one-standard-deviation change in aid per capitait-10 corresponds with change in Threshold_Distance of between 4.5 % and .1 % of a standard deviation, depending on the sample. Whatever endogeneity exists in these regressions is thus unlikely to be a substantial problem in our application.

  37. To accommodate negative numbers we add the minimum value within the sample + 1 and take the log of that sum.

  38. This unconditional correlation is substantial (greater than .4 and highly statistically significant) regardless of whether a country was above or below the IDA threshold in the previous year, though it is larger among observations that were above it.

  39. We report models using the nominal rate of growth in Atlas GNI per capita, but the results are not meaningfully changed if we instead use a real growth rate normalized by the SDR inflation rate that dictates changes in the IDA threshold.

  40. Recall that the IDA threshold moves over time, so that observations in different years could have similar values of Threshold_Distance it-1 but quite dissimilar values of GNIPCit-1, and vice-versa.

  41. There is limited (but not entirely non-existent) evidence of a growth slow down as countries approach the IDA threshold in samples limited to country-year observations that are modestly far below it (between $200 and $500 below).

  42. One might reasonably suspect that growth rates should slow down for countries that have already crossed the threshold for two consecutive years, and are therefore one year of too-high GNI per capita away from losing their IDA-only borrowing status. The results reported in Table 6 suggest a lack of evidence for this. In unreported regressions we tested for this dynamic more explicitly and don’t find any convincing evidence for it. It is worth noting, however, that it is extremely rare for a country that has been over the threshold for two years to slip back below it in the third year (it occurs exactly once in our sample.) We don’t have a strong intuition for why this is, though it is possible that the act of crossing the threshold implies that the relevant authorities either were not previously interested in data management, or that they lost the will or capacity to do so. To the extent that they exist, “accidental” crossings – i.e., threshold crossings by countries whose authorities were willing and able to manage their data in order to avoid that outcome – may be rare enough that they don’t drive patterns that are discernable in the aggregate data.

  43. Percentile values change slightly across models; for simplicity we use percentiles based on the sample used in model 4 throughout.

References

Download references

Author information

Authors and Affiliations

  1. University of Michigan, Ann Arbor, MI, USA

    Andrew Kerner & Alison Beatty

  2. Norwegian University of Life Sciences, Akershus, Norway

    Morten Jerven

Authors
  1. Andrew Kerner
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Morten Jerven
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Alison Beatty
    View author publications

    You can also search for this author in PubMed Google Scholar

Corresponding author

Correspondence to Andrew Kerner.

Electronic supplementary material

Below is the link to the electronic supplementary material.

ESM 1

(ZIP 220 kb)

Appendix A

Appendix A

1.1 Predictive power of aid dependence proxies

Appendix Table 1 shows the results of regressions relating to the relative predictive power of two different proxies for aid dependence – the 10-year lag of aid per capita and infant mortality in 1980, which has been used elsewhere in the literature (ex. Burnside and Dollar 2000; Knack 2001). These are OLS regressions with standard errors clustered by country. Our dependent variable is aid per capita in the current period. To better approximate our tests we limit the sample to World Bank members within $300 of the IDA threshold, excluding countries subject to the small islands exemption. GNI per capita is implicitly controlled for as the assignment variable in the McCrary tests (and explicitly so in our regression discontinuity models) and so we control for it explicitly here.

Model One regresses aid per capita on the 1-year lag of GNI per capita. The coefficient on GNI per capita is negative and highly statistically significant. The R2 is .024. Model Two adds the measure of infant mortality in 1980. The coefficient is negative, which is contrary to expectations, and statistically insignificant. The R2 increases slightly, to .047. Model Three replaces infant mortality in 1980 with the 10-year lagged measure of aid per capita. The coefficient is positive and statistically significant, and the R2 increases substantially from .024 in Model One to .255, suggesting that the 10-year lag of the aid-to-GNI ratio is a substantively as well as statistically significant predictor of the level of aid-dependence among the countries in our sample. When both proxies are included in the same model (Model Four), the coefficient estimate on the 10-year lag retains its statistical significance and magnitude, while the estimated coefficient on infant mortality remains statistically significant and continues to carry the wrong sign. Models Five through Eight replicate Models One through Four, but extend the sample to country-years within $1000 of the IDA eligibility threshold. The results are substantively similar and indicate the 10-year lag as the more appropriate proxy.

Table 8 Regressions of aid dependence on proxies for aid dependence
Full size table

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Kerner, A., Jerven, M. & Beatty, A. Does it pay to be poor? Testing for systematically underreported GNI estimates. Rev Int Organ 12, 1–38 (2017). https://doi.org/10.1007/s11558-015-9239-3

Download citation

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s11558-015-9239-3

Keywords

Search

Navigation