Statistics never lie

It’s good to count stuff. We can count how frequently someone makes a grammatical error. Or how many speech sounds they find difficult. Or how many times they point to the wrong picture in a comprehension test. Once we’ve done our sums we can get out or look-up table and find out if someone is language-typical or language-impaired. We can find out if an intervention programme is working. We can demonstrate that speech and language therapy works. There’s no need for messy subjective judgments. We can let the data do the speaking. We might even run a few statistical tests which will enable us to quantify our degree of certainty. And we can even count that certainty (using p-values).

There’s no escaping that Speech and Language Therapy needs to embrace quantitative methods. Yet at the same time we risk making the wrong conclusions if we cannot think “beyond” the numbers. One particular pitfall is understanding the difference between causation and correlation. People who live in large houses are more likely to vote Conservative, but this doesn’t mean that the size of one’s house actually determines one’s voting habits. Fortunately, SLT students with their advanced statistical training are able to spot such dodgy claims a mile off. And tear them apart with savage gusto.

But let’s not be too smug. There are plenty of research areas where vigorous debates thrive on the causation versus correlation dilemma. One example is the field of Verbal Working Memory (VWM). VWM tasks involve retaining and recalling linguistic items while simultaneously doing some kind of additional processing task. A classic example is “listening span” where you have to say if sentences are true (the processing task), and then recall the final words from the last few sentences (the recall task). Because performance on VWM is strongly associated with language skill (e.g. comprehension of complex sentences), many have argued that VWM determines language skill. But as every SLT student will know, correlation does not equal causation. An alternative, and I think, far more convincing explanation, is that performance on both language and VWM tasks are influenced by the same underlying factor, which is… err… language. Viewed from this perspective, the idea that VWM determines language skill simply vanishes in a puff of smoke.

To be fair, proponents of VWM accounts have countered these criticisms with some rather sophisticated methods (e.g. structural equation models). No one has yet struck a killer blow and the debate rages on. But stepping back one or two paces, there is a larger lesson to be learned. Running stats is just the starting point. The crucial thing is how we interpret the stats. At the end of the day, though we do lots of complex stuff with numbers, it’s all about the fine art of crafting a convincing and persuasive argument. And in that sense, perhaps we have more in common with the humanities than we’d like to think…

Nick Riches

Leave a Reply