Does Aid Promote Electoral Integrity?

ABSTRACT

Since the late 1990s, aid spending for elections has witnessed a dramatic increase. Yet, we lack a comprehensive evaluation of aid effectiveness in this particular programme area. Here, we investigate the impact of aid on electoral integrity using panel data on purpose-disaggregated aid disbursements and a multi-dimensional index of electoral quality from the Varieties of Democracy project. Based on 502 elections in 126 aid-receiving countries during 2002–2015, we estimate a statistically significant effect of election-support ODA on the integrity of elections. The estimated effect is, however, economically small and not very persistent. In the long run, a permanent increase in aid spending by one million US$ leads to an improvement in electoral quality of 1.4 per cent of a standard deviation on the integrity index. We also find that different dimensions of electoral integrity are variably responsive to donor interventions. Additionally, aid spending for elections is subject to diminishing marginal returns, and is less effective at higher levels of development. These findings underline the difficulty of promoting democratic change in countries with adverse structural conditions. Still, donors may improve the cost-effectiveness of electoral assistance programmes by targeting specific countries and prioritising certain types of intervention.

Acknowledgements

We would like to thank the participants of a seminar organised by the U4 Anti-Corruption Resource Centre in Bergen (Norway) in March 2018 (and in particular, Svein-Erik Helle and Tina Soreide) for helpful comments on an earlier version of this paper. Data, replication files and all the results not reported in full in these pages, are available upon request and/or on the corresponding author’s website.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. For instance, article 21(3) of the Universal Declaration of Human Rights, for instance, stipulates that ‘the will of the people shall be the basis of the authority of government; this will shall be expressed in periodic and genuine elections which shall be by universal and equal suffrage and shall be held by secret vote or by equivalent free voting procedures’.

2. For instance, the Perceptions of Electoral Integrity, or PEI, dataset.

3. For instance, the NELDA indicators on electoral quality (Hyde & Marinov, 2012), the Free and Fair Elections variable used by Finkel et al. (2007).

4. For instance, the Freedom House Electoral Process index.

5. v2elfrfair asks respondents the following question: ‘taking all aspects of the pre-election period, election day, and the postelection process into account, would you consider this national election to be free and fair?’.

6. See Bishop and Hoeffler (2016, p. 608) for a similar argument.

7. We construct an index that is slightly more expansive than V-Dem’s own ‘clean elections’ index (v2xel_frefair), which aggregates over eight component indicators – including, oddly, v2elfrfair (V-Dem, 2017).

8. In case more than one election is held in the same year, each election is coded separately. The responses are subsequently aggregated to produce an overall rating for the (election) year.

9. The output of the V-Dem item response model is a distribution of possible true scores, and the reported point estimates are the means of this distribution. Thus, in principle, it is possible to incorporate a measure of uncertainty in the regressions (Bizzarro, Coppedge, & Pemstein, 2016). In practice, however, we do not do this as our dependent variable is a score obtained from factor analysis rather than a single V-Dem variable. In any case, classical measurement error in the dependent variable leads to higher standard errors but does not induce bias in the OLS estimator (Stock & Watson, 2015, p. 370).

10. The uniqueness of each of the 13 variables is always much lower than the usually accepted threshold of 0.6.

11. v2elfrfair, in turn, is highly correlated with the PEI index favoured by Norris (2017, p. 104).

12. See, for instance, the USAID project documents quoted by Finkel et al. (2007, fn., p. 29).

13. The DAC is the OECD’s Development Assistance Committee.

14. Authors’ calculations based on data from the OECD, International Development Statistics, 2018.

15. We also consider alternative specifications such as aid per polling station or the fraction of aid in total election costs. To our best knowledge, however, data on polling stations and election costs are not available across countries and over time.

16. Aid donors may also deliberately punish the countries committing the most egregious violations by withholding aid (Alesina & Weder, 2002). If this is the case, omitting the relevant confounders would bias the coefficient on aid upwards. Past studies, however, find that donors give little consideration to recipient merit when allocating aid (Hoeffler & Outram, 2011). Alternatively, donors may allocate aid resources to the countries where future electoral contests are most likely to be free and fair (Knack, 2004). This pathway would also spuriously generate a positive statistical relationship between aid and integrity. We let the data decide whether the omission of confounding factors leads to downward or upward bias.

17. Thus, controlling for these factors explicitly might explain away the influence of ODA spending, leading to post-treatment bias. Thus, in our specifications, we are careful to only control for exogenous institutions – that is, institutional characteristics that are neither the outcome nor the ‘inputs’ of electoral assistance interventions.

18. To purge the influence of time-invariant country traits, we choose to use the FE (within-) estimator, rather than the first-difference estimator. Since the average number of elections held in a country during 2002–2015 is four, differencing and lagging the regression equation would leave us with very few observations.

19. We report robust standard errors clustered at the country level as we find evidence of country-level heteroskedasticity (the p-value of a modified Wald test for group-wise heteroskedasticity is 0.000). Since elections do not necessarily take place in the same year across all countries, we assume the residuals to be cross-sectionally independent.

20. We also experimented with a second lag of ODA, but its estimated coefficient is statistically insignificant.

21. We also estimated model 3 (Table 3) using v2elfrfair and the simple arithmetic mean of the 13 component indicators as the dependent variable (integrity). The coefficient on (the first lag of) ODA is 0.030 (0.018) in a regression for integrity and 0.058 (0.017) in a regression for v2elfrfair.

22. That is, a 100 per cent increase in aid.

23. If the aid-integrity relationship has the following functional form: $I_{it}={\alpha }_{0}lnOD{A}_{it}^{elect}$, the effect of the marginal dollar is $\mathrm{\partial }{I}_{it}/\mathrm{\partial }OD{A}_{it}^{elect}=a_0/OD{A}_{it}^{elect}$. See Uberti (2017) for a discussion of the computation procedure in Stata.

24. This is a country receiving 5.6 million US$ in electoral assistance during election years, and 3.7 million US$ in pre-election years. These averages are based on the 460 observations used in the regression.

25. Results available upon request.

26. Wooldridge’s LM test for autocorrelation in panel data rejects the null that there is no first-order autocorrelation when the dependent variable is integ1 (p-value = 0.002) but cannot reject the null at conventional levels when the dependent variable is integ2 (p-value = 0.139).

27. For the dynamic model ($I_{it}={\phi }_0+{\phi }_1{I}_{it-1}+{\alpha }_{0}lnOD{A}_{it}^{elect}+{\alpha }_{1}lnOD{A}_{it-1}^{elect}$), the long-run effect of the marginal dollar is given by $\mathrm{\partial }{I}_{it}/\mathrm{\partial }OD{A}_{it}^{elect}=\left[\left({\alpha }_0+{\alpha }_1\right)/\left(1-{\phi }_1\right)]\cdot (1/OD{A}_{it}^{elect}\right)$. To derive this expression, note that if $t\to \mathrm{\infty }$, then $t=t-1$.

28. Similar results are obtained by proxying the level of development using the UNDP’s Human Development Index.

29. The full results are available upon request.

30. We did not find the effect of ODA spending on integ2 to depend significantly on the level of development. Full results available upon request.

31. To separate the effects of an exogenous change in aid expenditure today and last year, we would need two highly relevant and uncorrelated instruments (Stock & Watson, 2015, p. 485). We do not have two such instruments.

32. In Model 5 (Panel A), the coefficient on the first lag of ODA loses statistical significance at conventional levels (p-value = 0.124).

33. Donors should also consider prioritising election-year interventions, which are estimated to have a marginally larger impact than interventions implemented in pre-election years. In any case, most (51.4%) election-support ODA is already disbursed during or immediately prior to an election, and the remaining 48.6 per cent in years preceding an election (see Section 3).

Additional information

Funding

This work was supported by the Chr. Michelsen Institute (Bergen, Norway), and by the Norges Forskningsråd (Research Council of Norway) [240505]. The opinions expressed in this article are solely those of the authors.