The relative amount of variability in memory strength for targets vs. lures in recognition memory is commonly measured using the receiver operating characteristic (ROC) procedure, in which participants are given either a bias manipulation or are instructed to give confidence ratings to probe items. A near universal finding is that targets have higher variability than lures. Ratcliff and Starns (2009) questioned the conclusions of the ROC procedure by demonstrating that accounting for decision noise within a response time model yields different conclusions about relative memory evidence than the ROC procedure yields. In an attempt to better understand the source of the discrepancy, we applied models that include different sources of decision noise, including both the diffusion decision model (DDM) and the linear ballistic accumulator (LBA) model, which either include or lack within-trial noise in evidence accumulation, and compared their estimates of the ratio of standard deviations to those from ROC analysis. Each method produced dramatically different estimates of the relative variability of target items, with the LBA even indicating equal variance in some cases. This stands in contrast to prior work suggesting that the DDM and LBA produce largely similar estimates of relevant model parameters, such as drift rate, boundary separation, and nondecision time. Parameter validation using data from Starns’s (2014) numerosity discrimination data demonstrated that only the DDM was able to correctly reproduce the evidence ratios in the data. These results suggest that the DDM may be providing a more accurate account of lure-to-target variability evidence ratios in recognition memory.