In this paper, we investigate absorbing sets, responsible of error floors in Low Density Parity Check codes. We look for a concise, quantitative way to rate the absorbing sets' dangerousness. Based on a simplified model for iterative decoding evolution, we show that absorbing sets exhibit a threshold behavior. An absorbing set with at least one channel log-likelihood-ratio below the threshold can stop the convergence towards the right codeword. Otherwise convergence is guaranteed. We show that absorbing sets with negative thresholds can be deactivated simply using proper saturation levels. We propose an efficient algorithm to compute thresholds.