Amazon PayPhrase wasn't a good system for them to study.
By default, Amazon PayPhrase recommends a random pairing of two words. I bet that most users didn't bother changing their recommended passphrase. It also affected user behavior: users are more inclined to pick two-word pairings or other super simple passphrases if that's what's presented to them initially. Amazon PayPhrase also discourages users from making traditional non-dictionary passwords, which is very different from most other password systems. This, along with the fact that no two passphrases are allowed to be the same, makes their passphrases highly predictable.
I think this study says more about user behavior in regards to using the Amazon PayPhrase system than it does about multi-word password security in general.