There is an increasing trend in recent OCR research to improve recognition performance by combining several complementary algorithms. Numerous successful applications of classifier combination have been described in the literature. However, most of the combination algorithms neglect the fact that classifier performances are dependent on various pattern and image characteristics. More effective combination can be achieved if that dependency information is used to dynamically combine classifiers. Two types of dynamic selections are distinguished. The postconditioned selection seeks better approximation to unconditioned classifier output distribution. The preconditioned selection captures the variations in the density function of classifier outputs conditioned on the inputs. Although both types of selections have the potential to improve combination performance, we argue that preconditioned selections have lower error bounds than postconditioned selections. The difficulties of applying preconditioned selections are identifying characteristic variables and estimating their effects on classifier outputs. A solution using neural network to achieve preconditioned selection is suggested. Two examples on handprinted digit recognition are used to illustrate the potential gain.