The objective of the current study was to develop and evaluate the internal predictability for level C and A in vitro -in vivo correlation (IVIVC) models for prototype modified-release (MR) dosage forms of metformin. In vitro dissolution data for metformin were collected for 22 h using a USP II (paddle) method. In vivo plasma concentration data were obtained from 8 healthy volunteers after administration of immediate-release (IR) and MR dosage forms of metformin. Linear level C IVIVC models were developed using dissolution data at 2.0 and 4.0 h and in vitro mean dissolution time (MDT). A deconvolution-based level A model was attempted through a correlation of percent in vivo input obtained through deconvolution and percent in vitro dissolution obtained experimentally. Further, basic and extended convolution level A IVIVC models were attempted for metformin. Internal predictability for the IVIVC models was assessed by comparing observed and predicted values for C-max and AUC(INF). The results suggest that highly predictive level C models with prediction errors (%PE) of <5% could be developed. Mean percent in vivo input for metformin was incomplete from all formulations and did not exceed 35% of dose. The deconvolution-based level A models for all MR formulations were curvilinear. However, a unique IVIVC model applicable to all MR formulations could not be developed using the deconvolution approach. The basic convolution level A model, which used in vitro dissolution as the in vivo input, had %PE values as high as 103%. Using an extended convolution approach, which modeled the absorption of metformin using a Hill function, a level A IVIVC model with %PE as low as 11% was developed. In conclusion, the current work indicates that level C and A IVIVC models with good internal predictability may be developed for a permeability- and absorption window-limited drug such as metformin. (C) 2001 Wiley-Liss, Inc. and the American Pharmaceutical Association.