Effect of development assistance on domestic health expenditures

Although we welcome the attempt by Chunling Lu and colleagues1 to test the hypothesis that development assistance for health might induce a reduction of domestic health expenditures, caution is needed when translating results of this single econometric study into policy recommendations. Uncertainty associated with incomplete cases remains an issue of primary concern for time-series cross-sectional analysis since it can intrinsically bias the estimates. Lu and colleagues acknowledge that the Wooldridge test of zero correlation between the error terms within countries points to the presence of autocorrelation. They suggest that it might be explained by their own computation of the government health expenditure as source (GHE-S), by subtracting the development assistance for health disbursed to government from the government health expenditure as agent (GHE-A), but imputation concerning the missing values might also be a source of such autocorrelation. To take into account the autocorrelation within countries, Lu and colleagues use the Arellano-Bover/ Blundel-Bond linear generalised method of moments, and find that there is no significant evidence of serial correlation in the first-differenced errors at order 2. However, nothing is mentioned about the possibility of having a model with moving average errors. Data used by Lu and colleagues de facto introduce an additional source of autocorrelation since aid programmes devoted to health might differ according to region-specific and country-specific prevailing diseases. Given that the study focuses on the effect of development assistance for health on GHE-S, and that recent increases in such assistance have been mainly driven by funding for disease-targeted programmes, a test for spatial autocorrelation in the context of a fixed-effects panel data model3 would have strengthened the robustness of their results.