Repeated rounds of nationally representative surveys are an important source of information on changes in the welfare of the population. In particular, policymakers and donors in many developing countries rely heavily on the Demographic and Health Surveys (DHS) to provide information on levels and trends in indicators of the health status of the population, including child survival. The reliability of observed trends, however, depends strongly on the comparability across survey rounds of the sampling strategy and of the format of questions and how interviews ask them. In Madagascar, the most recent (2003/4) DHS indicated very sharp declines in rates of infant and under-five mortality compared with the previous survey from 1997. However, retrospective under-one and under-five mortality data in 1997 and 2003/4 for the same calendar years also show large differences, suggesting that this trend may be spurious. We employ a range of descriptive and multivariate approaches to investigate the issue. Despite evidence of significant interviewer recording errors (with respect to date of birth and age at death) in 2003/4, the most likely source of problems is that the two samples differ: comparisons of time-invariant characteristics of households and of women suggests that the later DHS sampled a somewhat wealthier (hence lower mortality) population. Corrections to the data using hazard survival model estimates are discussed. These suggest a much more modest reduction in infant and under-five mortality than indicated by the raw data for the two surveys