Let me add a couple of thoughts. I'm posting them here, as they are too long for a comment.
A good SLA study should not only provide empirical results, but also a model that explains them. The results of the study you quoted are uncontroversial, and I'm pretty sure they could be reproduced. But that doesn't solve the question of the usefulness of radicals in Chinese character pedagogy. The problem I see here is that radicals (in the sense of 部首) cannot be considered a psychologically realistic category. Learning radicals does seem to speed up learning, but most likely it does it because, accidentally, they partially overlap with some useful category, such as semantic components or phonetic components.
Different radicals belong to different categories:
- some radicals are arbitrarily chosen semantic components1,
- some radicals are phonetic components2,
- some radicals are graphical components (without any clear phonetic or semantic value),
- some radicals are strokes3.
At least some of these categories are likely to be psychologically realistic and learning them probably speeds up character learning. But radicals make up an arbitrarily chosen set of components and strokes:
- some strokes are radicals (e.g. 丨,丿) and some are not (乀),
- some phonetic components are radicals (臼) and some are not (e.g. 甫),
- some semantic components are radicals (e.g. 言 in 話) and some are not (e.g. 正 in 歪). The list can go on and on.
Radicals are really nothing more than an indexing tool for looking up characters; after all, the name 部首 simply means ‘section headers’ (in Xu Shen's Shuowen Jiezi dictionary).
If we were to believe that radicals are an important learning tool, by the very fact that they are radicals, teaching of 讓 should be focused on the radical 言. But it's quite clear that the part that causes most trouble is the component 襄. Similarly, correct recognition of the non-radical 甫 is probably a greater problem than the recognition of the radical 勹 in 匍. And the fact that 止 is the radical of 歪 ‘crooked’ isn't going to help with learning that character. What is needed is the recognition of its two semantic components: 不 ‘not’ and 正 ‘straight’.
My conclusion: it'd be much more useful to ask whether learning semantic components/phonetic components/graphical components/individual strokes can speed up learning the characters. I don't have answers to all these questions, but I have hypotheses about some of them. Sets of radicals that are used for indexing dictionaries is a mixed bag of arbitrarily chosen members of the above groups, and is therefore far from the optimal set that should be learned in advance. But I must admit that learning radicals is probably still better than learning characters without paying any attention to their structure at all.
1. e.g. the radical of 尖 ‘pointed’ is 小 ‘small’, but there is no reason why it shouldn't be 大 ‘big’; both components clearly have a semantic function here
2. e.g. in Shuowen Jiezi, 几 is said to be the radical of 鳧 and is indicated as a phonetic component; in modern radical systems, 臼 is both the radical and the phonetic component of 舅, etc.
3. e.g. 丿 is the radical of 乖, but it can't be really called a component: the most sensible decomposition of 乖 is 千 + 北, and 乖 without 丿 is not a component