Jim Newbery has a great post on his blog that talks about choosing/buying devices for testing. While Jim is a developer, a similar question is often asked by my prospective clients when they ask what devices they should support.
Jim initially advises, "I think take some time to find out what your target audience are using" but then the remainder of his post goes on to explain how to use research to determine devices. My experience is that the actual devices that prospective users are using is far more important than research.
Research is often biased in some way or other. For example, if it’s from an ad network then it will only be sampling apps that carry ads and geographic regions that place/use ads for that network. Research based on device sales can also be misleading because it considers what was sold recently rather than the actual types of device (still) in use.
My advice to clients is to, where you can (it’s not always possible), take a sample/survey of your prospective users before you even start a project. Factors such as age ranges, country and even region within a country (cities tend to have more expensive devices) can affect which devices you need to support.