A prominent finding in recognition memory is that studied items are associated with more variability in memory strength than new items. Here, we test 3 competing theories for why this occurs--the "encoding variability," "attention failure", and "recollection" accounts. Distinguishing among these theories is critical because each provides a fundamentally different account of the processes underlying recognition memory. The encoding variability and attention failure accounts propose that old item variance will be unaffected by retrieval manipulations because the processes producing this effect are ascribed to encoding. The recollection account predicts that both encoding and retrieval manipulations that preferentially affect recollection will affect memory variability. These contrasting predictions were tested by examining the effect of response speeding (Experiment 1), dividing attention at retrieval (Experiment 2), context reinstatement (Experiment 3), and increased test delay (Experiment 4) on recognition performance. The results of all 4 experiments confirm the predictions of the recollection account and are inconsistent with the encoding variability account. The evidence supporting the attention failure account is mixed, with 2 of the 4 experiments confirming the account and 2 disconfirming the account. These results indicate that encoding variability and attention failure are insufficient accounts of memory variance and provide support for the recollection account. Several alternative theoretical accounts of the results are also considered.