This paper examines the association of the sociodemographic characteristics of women and the unobserved hospital factors with maternal mortality in Kenya using multilevel logistic regression. The data analysed comprise hospital records for 58,151 obstetric admissions in sixteen public hospitals, consisting of 182 maternal deaths. The results show that the probability of maternal mortality depends on both observed factors that are associated with a particular woman and unobserved factors peculiar to the admitting hospital. The individual characteristics observed to have a significant association with maternal mortality include maternal age, antenatal clinic attendance and educational attainment. The hospital variation is observed to be stronger for women with least favourable sociodemographic characteristics. For example, the risk of maternal death at high-risk hospitals for women aged 35 years and above, who had low levels of education, and did not attend antenatal care is about 280 deaths per 1000 admissions. The risk for similar women at low-risk hospitals is about 4 deaths per 1000. To complement results from the analysis of individual patient records, the paper includes findings from hospital staff reports regarding the maternal mortality situation at the hospitals.