Magnetic resonance imaging (MRI) is a revolutionary tool in medical imaging, which plays an important role in clinical diagnosis. Compressive sensing (CS) has shown great potential in significantly reducing the acquisition time of MRI scanning. However, how to improve the reconstruction quality with limited k-space data is still a challenge. MRI images are featured with large area of smooth regions, sharp edges and rich textures. Motivated by these facts, we propose a nonlocal autoregressive model (NAM) for CS MRI reconstruction. Nonlocal similarity between image patches is exploited as a regularization term to constrain the nonlocal feature in MRI images, which is very helpful in preserving edge sharpness. While an autoregressive regularization term is employed to describe the linear correlation between neighboring pixels, which preserves more spatial details. Different from previous work, we reconstruct an MRI image patch utilizing correlations both among patches and among neighboring pixels. Extensive experimental results demonstrate that our method outperforms mainstream methods in MRI reconstruction in terms of both subjective quality and objective quality.