Anyone else find older architectures (vgg16) perform better than newer models?

I recently tried training some data using a range of different architectures. Was surprised that vgg16 seemed to perform consistently better than resnet.

Is it just my data is quite simple so does not benefit from a complex model? Or should vgg16 be considered as a candidate in other circumstances too. It is often described as if it were “something we used to use in the old days before better architectures came along”.