The latest issue of Pediatrics has a study that asks whether autism is associated with digestive problems. The authors compared the medical records of about 100 autistics with about 200 matched controls. The controls came from an area in Minnesota, near the Mayo Clinic, in which almost everyone has a health record on file that the researchers could look at. So the controls are a good sample of the non-autistic population.
The New York Times described the results like this:
The scientists found no differences [should be difference, singular] in the overall frequency of gastrointestinal problems reported by the two groups.
This isn’t quite right. The study found that the proportions of persons in each group to have had at least one digestive problem by age 20 weren’t reliably different. For the autistic kids, the proportion was 77%; for the controls, 72%.
The study design seems fine but the data analysis has a lot of room for improvement. You have an idea you want to test, good; try to test it with one test. The authors boiled down all their data into “at least one problem by age 20″ — that’s just what epidemiologists are told to do — but this was a poor choice. First, there is a ceiling problem. If both groups had percentages in the 90′s, this would be obvious. Better to avoid the ceiling problem. Second. to combine different symptoms with the “at least one” rule is likely to be less sensitive to differences than a combination rule that takes amount into account. The analysis in the article treats someone with 1 problem as equal to someone with 50 problems. No justification is given. Third, it isn’t obvious that it makes sense to combine symptoms this way. What if Symptom 1 and Symptom 2 are uncorrelated? In other words, what if whether you have Symptom 1 doesn’t affect your chances of having Symptom 2? Then to combine them (as the authors do) makes no sense. Factor analysis is how you condense several correlated measures into a few uncorrelated measures.
The study separated digestive problems into five categories (constipation, diarrhea, and three others). In each of the five categories, persons in the autistic group were more likely to report the problem than persons in the control group; in four of the five categories, the difference was significant (with one-tailed p values; the authors misleadingly use two-tailed p values — without making that clear).Â In one of the five categories the difference isn’t anywhere close to significant — which supports the idea that that there are at least two dimensions here: one on which the two groups differ, and one on which they don’t.
In the discussion, the authors, not realizing that four out of five of their problem categories differed significantly in the predicted direction, try to explain away the two differences that were significant with two-tailed p values: in constipation and picky eating. They note that autistic children get more medication that normal children. “Many children with autism are treated with resperidone, and this may result in increased appetite and weight gain,” they write. Why a drug that causes weight gain would cause picky eating isn’t explained and, without explanation, doesn’t make sense. Weight gain — they mean too much weight gain — involves eating too much; picky eating involves eating too little. Nor do the authors explain why their results differed from many previous studies. My take on the paper is that their results confirm previous studies, so that would have been interesting to read.