A central goal in monitoring and assessment programs is to detect change early before costly or irreversible damage occurs. To design robust early-warning monitoring programs requires knowledge of indicator response to stress as well as the uncertainty associated with the indicator(s) selected. Using a dataset consisting of four organism groups (fish, macrophytes, benthic diatoms and macroinvertebrates) and catchment, riparian and in-stream physico-chemical variables from 77 mountain and 85 lowland streams we determined the relationships between indicator response and complex environmental gradients. The upper (> 75th percentile) and lower (<25th percentiles) tails of principal component (PC) gradients were used to study the early response of the four organism groups to stress. An organism/metric was considered as an early warning indicator if the response to the short gradients was more robust (higher R-2 values, steeper slope and lower error) than the null model (organism response to the full PC gradient). For mountain streams, both fish and macrophyte CA scores were shown to exhibit an early warning response to the upper tail of the 1st PC gradient when compared to the null model. Five of the eight metrics showed better response to the upper tail of the 2nd PC gradient compared to the null model, while only one metric (macrophyte CA scores) showed improvement when compared to the lower tail of the 2nd PC gradient. For lowland streams all four organism-groups showed better response (CA scores) to the upper tail of the PC gradient when compared to the null model. Only one metric (fish CA scores) regressed against the lower tail of the 2nd PC gradient was found to be more robust than the PC2 null model. These findings indicate that the nonlinear relationships of organism/metric response to stress can be used to select potentially robust early warning indicators for monitoring and assessment.