29231: L Beaulieu (comment) Errors in the Lancet study (fwd)

[Bob Corbett <corbetre@webster.edu>]

From: Laurent Beaulieu <laubeaulieu@wanadoo.fr>

The study published by The Lancet shows that there have been many human rights 
violations, especially sexual assaults. And so it is useful. But it contains 
some blatant errors.
 Imagine that you make a survey about the activities of the Haitians.  1000 
people have been randomly selected among 8 millions, and Rene Preval is one of 
them. How will you extraploate to the entire population ? Will you calculate 
that there are 1 x 8000 = 8000 presidents in Haiti ? No, because you have got 
another information : there is only one president in Haiti.
The Lancet study contains such kind of gross errors.
Look at the table 3  "Proportion of individuals in zone that were victimized 
during the period assessed extrapolated to entire population for the zone". 
According to this table, 0 % of the entire population has been murdered in 
Martissant/Grand Ravine and in Lavil/Belair! Having found no murder in their 
sample in these zones, the authors of the study have extrapolated "0%" for the 
entire population of the zones. It is stupid, because everybody knows that many 
murders have been documented in these areas. Instead of "0%", they should have 
written : "We do not know".
Look at the table 2 : "Percentages of specified violations committed by 
specified group, extrapolated to the greater Port-au-Prince area". According to 
this table, in the greater Port-au-Princie, during 22 months, nobody has been 
murdered by "partisans of anti Lavalas groups", "Lavalas members or partisans" 
and "foreign soldiers". Nobody has been murdered by theses groups in the sample 
....so Kolbe and Hutson conclude : "none" in Port-au-Prince. They should have 
written : "We do not know".

The first error of Kolbe and Hutson has been the size of their sample (only 
1260 households). It might be the good size to study the age of the population, 
but  not to study human rights violations, especially murders or kidnappings, 
i.e. "rare" events. The second error has been extrapolating (making 
projections) when it is impossible. The third error is writing dubious 
comments.
They  extrapolate  the data of the survey, even when it is a nonsense. In their 
sample, they have got 23 murders, 5 (or 4 ?) of them being commited by police. 
What can you extrapolate from 5 ? Nothing, of course. They have found 5 in 
their sample ; with another sample, they would have found zero, or 10 or 
anything else. But they absolutely want to make calculations. In every 
scientifical book about surveys, you can read that you cannot make classical 
extrapolations when the figures are too small. Meanwihile Kolbe and Hutson 
calculate very seriously (table 2) that  between 2,7% and 40,7% of the murders 
have been committed by "HNP and other government security forces" (more exactly 
: the probability for the number of murders committed by this group to belong 
to the interval [2,7%, 40,7%] is 95% ; and they assume a Poisson distribution 
with the highest probability around 21,7%). Even if we accept this hazardous 
extrapolation, it is clear that the "confidence interval" is so wide (from 2,7% 
to 40,7% !) that the information provided is very small. But  Kolbe and Hutson 
write bluntly that "officers in HNP and members of other government security 
forces were identified by respondents as committing a SUBSTANCIAL proportion of 
sexual assaults and murders". For the murders, this assertion is only based on 
5 or 4 murders documented in the survey ! Several reports have shown that 
police officers have killed a lot of people, but not this survey.

There are other gaps between the tables and the comments. They write : "our 
estimates suggest that about 8000 individuals were murdered, with almost half 
of the perpetrators identified as political actors". "About 8000 murders" ? 
According to table 1, the probability for the number of murders to belong to 
the interval [5000, 12000] is 95%. And, for various reasons, this result 
suggests that the probability for the number of murders to belong to the 
interval [6500, 9500] is 67%. Only 67%. "Half of the perpetrators identified as 
political actors" ? Read table 1 and you will see that this is a very hazardous 
assertion.

Many people have been surprised that no murder, no sexual assault and no 
"kidnapping or extra-judicial detention" have been attributed by respondents to 
"Lavalas members or partisans". I do not think that the data were invented. 
There are probably several explanations. One of them is that during "Bagdad 
Operation", media called the armed bands "bandits" . Another explanatition is 
that for the victims, murderers, rapers and kidnappers are "criminals", not 
political activists. It is an answer to gang leader William Baptiste, known as 
Ti Blan, who recently said : "We are not criminals, we are political militants 
".