Having outlined the data which are to be used as the basis of the analysis, this section details some of the pre-estimation data diagnostics the estimation techniques which are used to test the hypotheses presented in Chapter 3.0 .
The first diagnostic indicated that there was at least first-degree autocorrelation within the data, although it rejected the hypothesis that either Life Expectancy or Infant Mortality exhibited a unit root. However, there is evidence that these variables may be subject to a positive time- trend, so each of the fully specified models below include a count variable accounting for elapsed time to control for any trending that may be present.
Given the continuous dependent variable, pooled Ordinary Least Squares (OLS) regression would be a suitable estimator. However, the presence of autocorrelation can present difficulties in the estimation of the standard errors. Two popular alternatives for dealing with serial correlation include the inclusion of a lagged dependent variable as a regressor and a technical correction via an AR(1) process. One issue with lagged dependent variables is the possibility that they may artificially depress the coefficients of the other regressors (Achen 2000), which is particularly true given the high year-over-year persistence of both life expectancy and infant mortality. Therefore, the use of an AR(1) correction is preferable in this case.
In addition to serial-correlation, spatial correlation can also make the estimation of accurate standard errors problematic. It is theoretically likely within this data, given the tendency for natural disasters to strike multiple countries at once or for civil conflicts in one country to affect a neighbor, that there would be some degree of spatial correlation present. One way of correcting for this spatial correlation is the use of panel-corrected standard errors (Beck
and Katz 1995), and this is the option chosen for this analysis. Therefore, the models are estimated via pooled OLS utilizing panel-corrected standard errors with a panel-specific AR(1) correction unless otherwise specified.
Furthermore, it is important to determine whether pooling of the observations is appropriate, and additionally whether there are any factors which may cause responses by any of the aid organizations to deviate from the others. For instance, if one organization were to systematically choose countries that are in greater danger than others, the structural differences between the organizations’ method of case selection could result in biased estimates unless these factors were properly controlled for.
It is worth briefly discussing at this point the question of selection and how to properly control for it in the statistical models. Traditionally discussion of selection in the statistical sense has focused on sample selection, or the methods by which observations are included or excluded from the statistical sample. If observations are omitted from the sample based on a non-random criterion, then it is possible that results obtained from analysis of that sample could be biased due to failure to correct for the exclusionary factor. For example, if a study is run only on high- achieving students, then analysis of which factors influence educational achievement may be biased since students who perform poorly are not part of the sample. This is the type of problem which the traditional Heckman selection model (1979) seeks to solve. By estimating two stages, the first of which calculates the likelihood of progressing into the sample evaluated in the second stage and including that estimated probability as an additional regressor, the Heckman model seeks to incorporate a measure of the factors which differentiate the censored and uncensored observations and reduce bias brought about by splitting the sample in a deliberate, non-random manner.
The selection issue present in the current analysis is different. Selection in this instance does not result in the censoring of data, but instead is realized in the non-random allocation of aid to particular countries based on the decision making of humanitarian organizations. It can be argued that this is a selection issue to the extent that organizations “select” to whom to give aid, but a different one than sample censoring. This type of selection doesn’t affect the probability of being included in the sample, but rather the correlation between other factors which influence both aid allocation and the outcome under analysis. It is, in other words, a case of omitted variable bias, which in this analysis would mean that not accounting for factors which make aid more or less likely (or impact aid’s effectiveness) and health outcomes could lead to biased estimates of aid’s efficacy.
The reason this is the case is because simple non-random distributions of data on the right hand side of an econometric estimation is not selection; it is, in many, the very phenomenon an analyst is trying to test. If one were engaging in cancer research, for example, there would likely be many health issues which would be correlated with both age and cancer. Data on these factors would be non-randomly distributed to the extent that they would increase with age, but this is not a case of sample selection so long as the analyst does not limit the sample to include only elderly patients. Assuming the sample is appropriately distributed with respect to age and controls for age as a factor, however, then the researcher can reduce or even negate any bias which would come about from the positive correlation between a health condition, age, and the onset of cancer. To use an example closer to the current study, Fortna (2004) encounters a similar issue in her assessment of the effectiveness of peacekeeping operations. Without an understanding of the common elements between cases where peacekeepers were sent and how that affects outcomes, her results would likely be biased. However, by controlling for the factors which make
peacekeeping more or less difficult, she can account for the correlation between particular measures of peacekeeping, difficulty, and eventual outcome.
Some of the factors which should impact aid’s effectiveness are included in the above discussion as included regressors. For example, if aid was only sent to poor countries, and being poor also results in lower life expectancy, then failure to control for income would downwardly bias the estimates of aid effectiveness. To assess other factors which could engender a similar result, Figure 4-3 presents diagrams detailing NGO response by health indicator and region. Each point shows an instance where a single aid organization was present in order to see how the three may differ in their response.59 If the organizations’ response profiles were similar, one would expect to see fairly close clustering of aid missions. However, from these diagrams it is clear that this is not the case. Rather, Oxfam appears to choose particularly challenging environments, at least in terms of the current level of health. For both life expectancy and infant mortality they tend to cluster towards the countries which exhibit poor health performance. MSF demonstrates similar tendencies in regards to life expectancy, but tend to operate in countries with average or better than average infant mortality. Finally, ICRC appears to end up in countries which have higher levels of life expectancy but lower levels of infant mortality. With regards to spatial dispersal, MSF and Oxfam both show tendencies to operate in Asia and Africa. The ICRC is less prevalent in these areas, but has a greater presence in the Middle East than either of the other two organizations. These regions also tend to show some clustering of health performance as well, with Africa clearly having the worst life expectancies but performing better in the area of infant mortality. Asia is slightly better or approximately even in both categories,
while the Middle East is the top performer of the three in life expectancy but lags behind in infant mortality.
4-3 NGO Response by Health and Region 40 50 60 70 80 L if e Exp e ct a n cy (Ye a rs)
Europe Middle East Africa Asia Americas
MSF ICRC Oxfam 20 40 60 80 1 00 1 20 Inf an t M or tal it y ( D ea th s p er ' 00 0)
Europe Middle East Africa Asia Americas
From this information two issues become apparent, which is that there are clear trends in specific NGO responses with regards to both health performance and regional presence, which suggests that two further estimation adjustments must be undertaken. The first is that some accommodation of the differences in starting life expectancy must be made, or else the estimates may be biased by the tendency of Oxfam (and to a lesser degree MSF) to operate in countries whose starting life expectancy is on average lower than the countries in which the ICRC is present. One way this might express itself would be by observing an upward bias on the ICRC coefficient and a downward bias for Oxfam and MSF, since estimation in levels would mean that ICRC was on average associated with higher life expectancies. To remedy this situation, the dependent variables are differenced so that the measure reflects the future two-year change in life expectancy or infant mortality.60 This specification has a number of advantages. The first is that it removes any differences in initial life expectancy, leaving the organizations on a level playing field. Additionally, it accounts for the fact that the aid organizations may not have an immediate impact, but may it may take some time for the effects of their efforts to be realized.
Given the general stability in life expectancy, one question about differencing the variable may be whether there is sufficient variation to provide useful information. Fortunately there does appear to be a surprising amount of difference in estimated life expectancy values over a two-year period. The mean change is close to zero at 0.02, but the standard deviation is 0.05. The minimum value is -0.29, and the maximum is 0.39. These aren’t tremendously large figures, but it does indicate that life expectancy is more variable than may otherwise be expected. This dispersal of values can also be seen in Figure 4-4, which plots the two year forward change in life expectancy relative to its initial value in levels.
60 Therefore, for aid observed in year t
4-4 Two Year Change in Life Expectancy
The regional clustering of both NGO responses as well as health performance also suggests that it is important to account for inter-regional differences. Therefore, regional dummies for Africa, Asia and the Middle East are included as regressors.
Finally, it is possible that some factors not included in the model estimates may influence public health. In order to account for the unobserved factors, the models are estimated with fixed effects using a Least Squares Dummy Variables (LSDV) approach, wherein a dummy variable for each country in the analysis is included to account for country-specific time-invariant
-.4 -.2 0 .2 .4 T w o Y ea r C ha ng e in L if e E x pe c tan c y 40 50 60 70 80 Life Expectancy
characteristics.61 The coefficients on both the regional dummies and country-dummies are omitted from the tables due to space considerations.
To review, the models are therefore estimated via OLS using panel-corrected standard errors to account for spatial correlation, a panel-specific AR(1) correction to deal with temporal correlation, a time-trend, and regional and country dummies to measure regional- and country- fixed effects.