I think it is worthwhile to update James Wimberly’s comparison of surveys of deaths in Iraq. In the table below death tolls have been extrapolated to give a number of deaths due to the war so far.
|Survey||Violent deaths||Excess deaths|
It is interesting to see that the IFHS ends up right in the middle, between the two Lancet studies. If you think that the IFHS study is reasonable then you must conclude that Lancet 1 has been confirmed and the critics of Lancet 1 were wrong.
What about the comparison between Lancet 2 and the IFHS? Some folks are arguing that the larger sample size of the IFHS meaning that its estimate trumps Lancet 2, but the larger sample size just reduces the sampling error. The top of the IFHS confidence interval for violent deaths (220,000) is much smaller than the bottom end of the Lancet 2 interval (420,000) so the difference can only explained by non-sampling error, and here the IFHS isn’t necessarily better. It was too dangerous to visit 11% of their clusters and in the ones they could visit they say that as many of 50% of the violent deaths may have gone unreported. They have attempted to correct for these problems, but the corrections may be not be enough. For example, they use the IBC numbers to estimate the violence in the unvisited areas relative to those they visited. But the most violent areas are going to be too dangerous for reporters as well, so the IBC will tend to undercount deaths in those areas relative to less violent areas. This doesn’t mean that the Lancet 2 estimate doesn’t have non-sampling errors as well — Iraq at present is just a really difficult place to survey. I think that the best we can do is guess that the real number of violent deaths in Iraq to date is somewhere in the range 300,000 to 1,0000,000. Even the lowest number in the range is a horrendous death toll.
The numbers for excess deaths are for Lancet 2 and the IFHS are closer to each other, which suggests that some of the difference in violent death rates could be deaths being misclassified (either as violent by Lancet 2, or non-violent by IFHS). Since the Lancet usually checked death certificates, it seems that they would be less likely to misclassify deaths.
Notes: I did the calculations a little differently from Wimberly. He used the published totals on a given date from the IBC to scale the figures — I used the actual total number of deaths in the current IBC data base for a given time period. For example, the ILCS covered the first 13 months, so it is scaled by (IBC current total)/(IBC total for first 13 months). In addition, I only used IBC scaling for violent deaths. For excess non-violent deaths I assumed that the death rate was constant.
The IFHS did not publish an estimate for excess deaths, so I computed it from the rates in the paper. The IFHS FAQ states:
Further analysis would be needed to calculate an estimate of the number of such deaths and to assess how large the mortality increase due to non-violent causes is, after taking into account that reporting of deaths longer ago is less complete.
Taking this factor into account would reduce the IFHS excess deaths estimate in the table above.