The initialization of numerical prediction models usually requires the transformation of variables observed in a p-coordinate system into some other coordinate frame of reference (e.g., α-coordinates or Θ-coordinates). Such transformations require the application of interpolation or curve-fitting techniques. The present study demonstrates that the choice of an appropriate interpolation scheme can become a critical issue for the skill of a low-resolution prediction model. First we show that the interpolation scheme, when applied to more than one meteorological variable, should satisfy the balance requirements that exist between these variables. Not all of the currently used schemes meet this condition. Next we provide evidence indicating that interpolation schemes used to convert p-into α-coordinates, and then back into p-coordinates, do not necessarily replicate the original, observed field distributions of these meteorological variables. Such double transformations usually are required, because the numerical output in model coordinates has to be translated back to p-coordinates for verification of model results. Because of the limitations of certain interpolation procedures, even a correct model prediction may exhibit low predictive skill because of errors introduced in this final coordinate transformation process.