Stones were used to sample macroinvertebrates and characterise microhabitats at monthly or bimonthly intervals in six Ecuadorian streams covering a gradient in four different stability measures and other stream characteristics. The physical variables current velocity, water depth, horizontal position, embeddedness and size were measured to characterise stone microhabitats and presumed to be affected by or related to physical impact during hydrological disturbances. My first objective was to analyse how density, the number of families and a richness measure (residuals from a power regression of families vs. individuals) were related to the physical characteristics of individual stone habitats. My second objective was to quantify temporal variability in fauna-stone relationships and to analyse if such variability was related to overall stability of stream reaches. Partial Least Squares (PLS) multiple regression analyses showed high temporal variability between sampling dates in factor loadings of specific stone micro habitat variables. In spite of this, there was a clear negative effect of depth and a positive effect of current on density and number of families. Stone size was consistently negatively related to density and positively related to number of families. Patterns were less clear for richness residuals. Simple linear regressions of fauna vs. stone parameters generally confirmed the results reached by the PLS analysis, although few of the regressions were significant. For all fauna-stone regressions the variability in slopes was much higher among sampling dates within streams (temporal variability) than among streams (spatial variability), and significant slopes were even inverted on different sampling dates. Although the coefficients of variation (CV) of slopes of a given combination of fauna parameter and stone variable from different sampling dates (n=9-11) were rarely correlated to any of the measures of stream stability, this study has demonstrated high temporal variability in fauna-stone relationships (CV's of regression slopes). Consequently, temporally un-replicated studies of such relationships do not necessarily reveal general patterns.