Background/Purpose : Linear Extrapolation (LE) is a frequently applied method to impute missing radiographic data in trials. However, there is frequent critique that LE overestimates overall progression. Therefore, ‘Last Observation Carried Forward’ (LOCF) has been suggested by regulatory bodies as a more conservative method. In the OSKIRA-1 trial (NCT01197521), radiographs were taken at the 12 week (wk) time point, where early escape was possible, in all patients thus providing an excellent opportunity to compare extrapolations based upon LOCF and LE to the truly observed radiographic progression.

Methods : The phase 3 OSKIRA-1 trial enrolled rheumatoid arthritis patients (pts) with an inadequate response to methotrexate. Films of hands and feet were obtained at baseline, wk 12 and 24 in those pts still on study, and were assessed by two readers in random time order using the van der Heijde modified total Sharp score (mTSS). Ten datasets with an artificially, randomly selected sample of 20% missing wk 24 data were created, based upon pts with complete sets of films. First, these missing data were imputed using LE as (mTSS at wk 12 + progression wk 0 – 12, corrected for the actual days between films). Second, the missing data were imputed using LOCF. This approach was iterated for 10 random samples with 50% missing data and 10 random samples with 80% missing data. The datasets obtained with LE and LOCF were compared to the dataset with truly observed data at week 24.

Results : Complete sets of films were available for 579 pts. All our analyses were essentially similar in the 3 treatment arms, so here we present the analysis on pooled data. Mean (SD) observed progression from baseline to wk 24 was 0.33 (2.42). Table 1 shows the average (SD) and range of the mean radiographic progression in the 10 random samples. Using LE, the mean progression estimates were close to the observed data, and not affected by the proportion of missing data. The SD however increased by increasing proportions of missing data. Using LOCF, the mean progression estimates were consistently lower than the observed progression. LOCF increasingly underestimated observed progression by increasing proportions of missing data. As expected, the SD of the LOCF estimates remained stable by increasing proportions of missing data.  

Conclusion : In contrast to LOCF, linear extrapolation gives a more accurate impression of true mean radiographic progression at a group level and is less influenced by the proportion of missing data. Since LE inflates the standard deviation of progression scores, the statistical power to detect a significant difference between active treatment and placebo may decrease by increasing proportions of missingness. LE does therefore not overestimate mean treatment effects and is a more robust method than LOCF in this respect.