In sub-Saharan Africa and in Madagascar, upland rice is mostly grown for subsistence by resource-poor farmers in low-input production systems with low yields. In this context, soil nitrogen availability is a major limiting factor. To determine the appropriate breeding strategy to improve nitrogen use efficiency (NUE), the genetic variability and the level of G×N interaction of NUE need to be evaluated. NUE and its components are complex quantitative traits and the study of their relationship with other simpler traits should help understand the mechanisms involved in NUE and identify ways to improve it. A study was conducted at mid-altitude in Madagascar using 13 adapted tropical japonica rice varieties over three cropping seasons with two contrasted levels of applied N (high N with 90–120kgNha−1 as urea and low N without mineral fertilizer) in rainfed upland conditions. Year×N and Year×G interactions were significant due to contrasted rainfall distributions across the three cropping seasons. Agronomic efficiencies of N fertilization were low (11.5kg grain kg−1N in the best year) because of N loss through leaching or volatilization. Our results also suggest gaseous loss of N by plants between flowering and harvest, particularly under the high N treatment. The experiment revealed significant genetic variability for NUE, nitrogen uptake efficiency (NUPE) and nitrogen utilization efficiency (NUTE) in both high N and low N treatments but a low level of G×N interaction. The variation in NUPE accounted for more of the variation in NUE than the variation in NUTE. There was no correlation between NUPE and NUTE, either under high N or under low N. We found no negative relationship between grain yield and grain N concentration under low N. The relationship between NUE and agronomic and N-related traits differed from one year to the next, illustrating the plasticity of the contribution of the different agronomic traits to NUE as a function of the contrasted climatic conditions (particularly the rainfall distribution pattern). However, two traits, the number of panicles per m2 and the harvest index were consistently positively correlated with NUE. The difference between total N uptake at flowering and at harvest was positively correlated with NUE, particularly under the low N treatment.