Prediction-error expansion (PEE) develops Tian's difference expansion technique by performing reversible data embedding in the prediction-error histogram. As the prediction accuracy greatly affects the embedding performance, the PEE-based algorithms depend heavily on image's spatial redundancy. However, one drawback of these algorithms is that the modification of pixels may distort the high correlation in their local regions, and thus affect the prediction accuracy. To deal with this problem, we propose an adjustment process named "compensation" to compensate for the already modified pixels, before using them for prediction. Combining PEE and compensation together, a new reversible image watermarking algorithm is then presented. Comparisons with state-of-the-art works indicate that the compensation technique may benefit reversible watermarking in term of capacity-distortion behavior.