Lancet data on Iraq mortality

Tim Lambert has made available raw data from the Lancet Iraq study.

(My contribution)

I have run quick fixed-effects and random-effects poisson models on the data, which suggest a post-war mortality rate of between 80% and 250% above the pre-war level (i.e. a multiplier of between approx 1.8 and 3.5).

This approach allows for regional differences in mortality rates that are unobserved but time-constant.

Overdispersion

The poisson model assumes no "over-dispersion", i.e. differences above those controlled for by the explanatory variables, such as differences by region in effect of the war and aftermath. The negative binomial model allows for overdispersion, and with Falluja in the data set, has a CI that includes no increased mortality. Excluding Falluja reduces the variance, and gives a CI from 8% to 120% increase (random effects).

• Data and Stata code
• Full results

Summary of random-effects negative binomial model

Random-effects negative binomial regression     Number of obs      =        64
Group variable (i): place                       Number of groups   =        32

Random effects u_i ~ Beta                       Obs per group: min =         2
                                                               avg =       2.0
                                                               max =         2

                                                Wald chi2(1)       =      5.71
Log likelihood  = -115.06857                    Prob > chi2        =    0.0168

------------------------------------------------------------------------------
     ndeaths |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
     prepost |   .4340295   .1816011     2.39   0.017     .0780979     .789961
       _cons |   4.928958   1416.548     0.00   0.997    -2771.455    2781.313
personmonths | (exposure)
-------------+----------------------------------------------------------------
       /ln_r |   15.34534   1416.541                     -2761.024    2791.715
       /ln_s |   2.240766   .7886273                       .695085    3.786447
-------------+----------------------------------------------------------------
           r |    4617379   6.54e+09                             0           .
           s |    9.40053   7.413514                      2.003879    44.09944
------------------------------------------------------------------------------
Likelihood-ratio test vs. pooled: chibar2(01) =     1.47 Prob>=chibar2 = 0.112

. lincom prepost,irr

 ( 1)  [ndeaths]prepost = 0

------------------------------------------------------------------------------
     ndeaths |        IRR   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         (1) |   1.543464   .2802947     2.39   0.017     1.081229     2.20331
------------------------------------------------------------------------------

. 
. hausman fixed_no_falluja .

                 ---- Coefficients ----
             |      (b)          (B)            (b-B)     sqrt(diag(V_b-V_B))
             |  fixed_no_f~a      .          Difference          S.E.
-------------+----------------------------------------------------------------
     prepost |    .4178304     .4340295       -.0161991        .0820209
------------------------------------------------------------------------------
                          b = consistent under Ho and Ha; obtained from xtn_fe
           B = inconsistent under Ha, efficient under Ho; obtained from xtn_re

    Test:  Ho:  difference in coefficients not systematic

                  chi2(1) = (b-B)'[(V_b-V_B)^(-1)](b-B)
                          =        0.04
                Prob>chi2 =      0.8434