Disentangling short- and long-term neighbour effects, using both removal and observational methods within a single experiment, has strongly improved our understanding of the driving mechanisms of plant–plant interactions. However, there has been no attempt to assess two important underlying processes of their changes along gradients, either environmental-severity (changes in target performance without neighbours) or neighbour-traits (changes in performance with neighbours) effects, the former previously shown in alpine communities to be involved in competition and the latter in facilitation. We addressed this goal in an experiment conducted in continental saline depressions (sebkhas) from the Mediterranean arid climate of central Tunisia. We quantified short- and long-term effects of dominant shrubs, transplanting three target grass species in open, nurse and removed-nurse microhabitats of two habitats of different salinity levels in height sebkhas. The design extended greographically from central Tunisia to the Libyan border, 500 km southeastward. We used the relative interaction index to calculate short- and long-term effects before and after the dry summer seasons and environmental-severity and neighbour-trait effects. Short-term effects were slightly negative and long-term effects strongly positive before the dry summer season in the two habitats. Short-term effects switched to positive with increasing drought stress, due to an environmental-severity effect, whereas long-term effects decreased due to a neighbour-trait effect. Salinity did not affect neither short- nor long-term shrub effects. Soil moisture measurements showed that both changes were due to vanishing shrub soil engineering-effects during the summer drought. We conclude that an increase in short-term facilitation with increasing drought stress through time, apparently supporting the stress gradient hypothesis, might be due to a decrease in long-term facilitation. Thus, we recommend using, as much as possible, both the removal and observational methods in experiments assessing changes in plant–plant interactions along stress gradients to avoid wrong conclusions.< Réduire