Statistical Analysis of Radiological Data Suboptimal in Rheumatoid Arthritis Studies

Data published in Seminars in Arthritis and Rheumatism show significant heterogeneity in the strategy used to analyze radiographic progression in recent clinical trials and observational studies of rheumatoid arthritis (RA), with the majority of the analyses applying simple, suboptimal, or inappropriate methods.

The distribution of progression scores in RA is highly skewed and thus requires the use of advanced statistical analysis techniques, with different methods yielding different outcomes. Therefore, the investigators searched 3 databases to identify RA clinical trials and observational studies that described radiographic analysis techniques and compared ≥2 groups of individuals. Of the 5980 papers that were ultimately identified, a total of 225 were considered eligible for data extraction.

Of the 225 reports selected, 71% reported on clinical trials, while the remainder reported on observational studies. Overall, 43% of the single studies involved a comparison of ≥2 groups. In total, 49% of the studies had a sample size of between 100 and 500 patients. Moreover, 12% of the papers had a sample size of <100 patients, whereas 38% of the papers had a sample size of >500 patients. Nearly all of the studies (97%) included foot radiographs.

In 44% of the studies, the radiographs were read in random order. In addition, in 60% of the studies, 2 readers scored the radiographs. In 7 studies, all of the radiographs were interpreted by 1 reader, with a sample taken by the second reader. Most of the remaining studies used a single reader only. In 36% of the reports, information on the number of readers or the reading order was not reported.

The skewness of radiographic data is a well-known issue that remains unresolved. The non-normal distribution renders it difficult to define and identify the most relevant “center” of the distribution, thus hampering between-group comparisons. Although median and mean values are similar in normally distributed data, they differ in skewed data, rendering it challenging to interpret variations between groups. In the field of RA, most patients will not exhibit progression of damage during the typical trial observation period, with the current focus being on the subgroup of patients who do progress despite optimal treatment.

The researchers concluded that based on their investigation, they recommend Tobit analysis in models with a variation in the lower limit of the scale. In cases in which counts are skewed to the right, a Poisson or negative binomial model is suggested. Counts that contain an excess of zeros are best handled with the use of zero-inflated models, which can be used only with cross-sectional data. In longitudinal studies that include an excess of zeros, use of the 2-part model is best.

Reference

Mahmood S, van Tuyl L, Schoonmade LJ, et al.  Systematic review of rheumatoid arthritis clinical studies: Suboptimal statistical analysis of radiological data [published online February 21, 2019]. Semin Arthritis Rheum. doi:10.1016/j.semarthrit.2019.02.009