Deep learning builds the foundation for many medical image analysis tasks where neuralnetworks are often designed to have a large receptive field to incorporate long spatialdependencies. Recent work has shown that large receptive fields are not always necessaryfor computer vision tasks on natural images. We explore whether this translates to certainmedical imaging tasks such as age and sex prediction from a T1-weighted brain MRI scans.