A system is considered where single analog data are quantized to be transmitted over a digital communication system. Channel coding schemes are investigated that are designed to provide a low mean square error (MSE) in the analog data instead of providing large minimum distances.In case of transmission errors, the influence of a bit error in the quantized word on the analog value is much larger when it corrupts the most sigificant bit compared to the least sigificant bit. Accordingly, a channel code should provide more protection to the foremost bits than to the backmost bits of a quantized word, more precisely, it should offer a continuous increase of protection from the last to the first bit.In this thesis it is shown that a channel code with such properties can be obtained by truncation of a convolutional code. Usually, the convolutional encoding procedure starts in a predefined state, therefore, the decoder has a-priori information about the first states which leads to an increased protection of the foremost bits. Furthermore, the truncation of convolutional codes leads to a low protection of the backmost bits. It is shown that when truncated convolutional codes are applied on short information sequences they show the required bitwise increase of protection from the last to the first information bit. Such codes may not provide high minimum distances, instead they show a very good performance with respect to the MSE. Simulation results are presented that show that truncated convolutional codes usually have better performance than standard block codes, other linear unequal error protection block codes, and constructions of redundant index assignments.In order to gain more control of the bitwise increase of protection, a new class of timevarying convolutional codes is developed. The unequal error protection capabilities of such codes are examined and it is shown that a huge performance gain is possible when this method is used to adapt the code to channel conditions.Furthermore, a new approximation formula is developed which determines the MSE based on the generator matrix of a code. This approximation is used to compare the performance of codes and to determine MSE-optimal codes for several different dimensions.