what has been written in the death certificate In the first case the accuracy of report will depend on how clear her memory of event is (this is an issue in any retrospective survey). The second and third cases are of particular importance in the context of CEE/CIS countries where medical staff sometimes misreport pregnancy outcomes. If respondents were influenced by what they were told by medical staff, then the survey results will also be influenced by mis-reporting in the health system (Wuhib, 1998). The omission by respondents of dead children from the birth history can also bias results. One indicator of omission may be the ratio of male to female infant mortality. The expected ratio of male to female deaths is between 1.16 and 1.32 (Curtis 1995). A higher ratio might indicate under-reporting of female deaths. Table 8 shows that both Georgia and Kazakhstan (1995) have a ratio of 1.8 for neonatal mortality, while Georgia, Turkmenistan and Uzbekistan have high ratios for total infant mortality. This may mean that female infant deaths are under-reported in these surveys, and that survey based estimates of infant mortality rates in these countries that understate the true total.14 The very low estimate of infant mortality from the Ukraine RHS underlines the difficulty in collecting information of this kind in surveys, and the large impact that non-reporting error can have on results. As noted in Table 1, the Ukraine RHS produces an infant mortality estimate that is very similar to the official rate. Yet the problems with the Ukraine official data discussed in Sections 2 and 3 of this paper clearly imply that the real infant mortality rate should be higher than the official rate. There seems little doubt 14 There are several other types of non-sampling error (not necessarily specific to CEE/CIS countries) that can affect infant mortality estimates. These are discussed in greater detail in Sullivan, et al, (1994). 22 therefore, that the Ukraine RHS underestimates the true level of infant mortality in that country.15 Table 8: Sex differentials in infant mortality Infant mortality rate Neonatal mortality rate Post-neonatal mortality rate Mortality ratio Male/Female Male Female Male Source: Kazakhstan DHS 1995 and 1999, Uzbekistan DHS 1996, Kyrgyzstan DHS 1997, Georgia and Romania RHS 1999, Armenia and Turkmenistan DHS 2000, and Azerbaijan RHS 2001. See Appendix I. Account also needs to be taken of sampling error. As infant death is typically a rare event, sampling error for survey based mortality estimates can be substantial, implying great deal of uncertainty regarding the true population rate. Some uncertainty in estimates, exemplified by the width of the confidence intervals, might be caused by small sample sizes. Table 9 presents standard errors and sample sizes for countries in the region and for some developing countries outside the region with similar levels of infant mortality. Because fertility levels in the countries of the Caucasus and Central Asia are generally lower than those in many other developing countries, the number of births from which infant mortality rates are estimated also tends to be lower. The infant mortality rate estimated from the 1989 Kenya Demographic and Health Survey, for example, is 61 per 1,000 live births, similar to that in Kazakhstan. However, the sample of infants from which the Kenya infant mortality rate is calculated is 7,319, compared with 1,501 in the case of Kazakhstan. The 95 per cent confidence interval surrounding the Kenyan estimate is narrower, ranging from 52.2 to 69.3 per 1,000 live births, compared with a range of 45.3 to 78.6 for Kazakhstan. 16 Nonetheless in the case of Kazakhstan and most other countries in the region, the lower bound of 15 These worries are echoed by the US Centres for Disease Control, which sponsored the survey. The Child Mortality module on the Ukraine 1999 RHS produced results that were inconsistent with expectations, for example a very low stillbirths rate and an extremely high ratio of post-neonatal deaths. While overall, the survey was judged to be satisfactory, these problems may have been the result of inadequate training of survey interviewers for the reproductive history module (personal communication, Howard Goldberg, CDC). 16 Sample size is not the only factor in determining confidence intervals. Re-weighting of the sample may also increase sampling variability. All DHSs and RHSs require re-weighting. 23 the confidence interval is still notably above the official infant mortality rate for the time period corresponding to the survey (Table 1 shoes the official rates). Exceptions are Ukraine (for reasons discussed above) and Romania, where the official rate (21 deaths per 1,000 live births) falls just within the lower bound of the 95 per cent confidence interval for the survey estimate (20 per 1,000 live births).17 Table 9: Survey sizes and sample errors Infant mortality rate Standard error 95 per cent confidence interval upper bound 95 per cent confidence interval lower bound Number of women interviewed Total fertility rate a Number of births b Turkmenistan, 1996-2000 73.9 5.4 84.7 63.1 7919 2.9 a Total fertility rate 15-49 for the three years preceding survey, expressed per woman (Ukraine - 2 years preceding survey) b Number of births in last five years preceding survey. In most survey reports, only the weighted