BACKGROUND: Monitoring the time trends in socioeconomic inequalities in mortality by cause is a key public health issue. The aim of this study was to compare methods to measure social inequalities in cause-specific mortality in the French population aged 25-55 years. More specifically, it compares bias and precision related to the use of occupational class declared at the last census (linked data) to the one declared at the time of death on the death certificate (unlinked data). METHODS: We used a representative sample of 1% of the French population. Causes of death were obtained by direct linkage with the French national death registry. Occupational class was classified into eight categories. Taking professionals and managers as the reference, relative risks of mortality by cause and their 95% confidence intervals were estimated using Poisson models for the 1983-1989, 1991-1997, and 2000-2006 periods. The relative risks were calculated with both linked data and exhaustive unlinked data. RESULTS: Over the 2000-2006 period, occupational classes declared at census and on the death certificate were consistent for half of the deaths. Relative risks for manual workers were found to be similar between the two approaches over the 1983-1989 and 1991-1997 periods, and higher for the unlinked approach over the 2000-2006 period. Over the latter period, the order and magnitude of relative risks varied similarly by occupational class and cause of death for both approaches. Confidence intervals obtained from linked data were wide. CONCLUSION: Occupational class derived from the death certificate must be used with caution as a measure for epidemiological purposes and the available linked data do not allow accurate estimates of social inequalities in cause-specific mortality. Other solutions should be considered in order to improve the follow-up of social inequalities in mortality. This would require the collection of educational level on the death certificate or the linkage of the cause of death database with other exhaustive and informative databases.