Multiple Imputation: Better than Single Imputation in Pain Studies?

Missing data is one of the genuine challenges we face during analysis of clinical trials. The main impact of missing data is that it can infuse bias in results which reduces the chance of getting the appropriate interpretation. Hence proper knowledge of techniques for handling missing data is crucial. A common method of handling this problem is by imputing missing values. There can be single or multiple imputations. While single imputation methods like LOCF or BOCF have been widely used, multiple imputation is of a comparatively recent origin but is gaining popularity. The main reason behind its popularity is it overcomes some of the drawbacks of single imputation. The current paper presents comparative effectiveness of each with the help of real life cases involving pain studies. INTRODUCTION The most serious concern of missing data is that it can introduce bias into the results of a clinical study, particularly if the data are missing not at random. Missing data causes the loss of information and in turn loss of statistical power. So the major challenge for these pain studies becomes interpreting the results. It is often preferred to put maximum effort at the design stage to achieve complete capture of all data to minimize the bias and maximize the power (e.g. continue collecting pain outcomes even after subject discontinuation). However, even with the best designed study missing data is still a potential issue. As such, various imputation methods are often employed when handling data of this nature. There are a few standard procedures to impute missing data. Single imputation usually identifies a particular record for a subject, e.g. baseline or just the previous non-missing value and repeats it for the missing data points. Multiple imputation uses a predicted value for a given subject and time point using statistical modelling of available data. Osteoarthritis (OA) is a painful and debilitating musculoskeletal disease that is characterized by intra-articular (IA) inflammation. In these studies pain is typically measured on a daily basis following study drug injection for a prespecified length of time. Pain intensity is generally measured by self-assessment with validated methods such as the visual analogue scale (VAS) or numeric al rating scale (NRS, 0-10). Depending on the length of follow-up of these studies and the fact that pain assessments are collected daily, one potential problem is the frequent occurrence of missing data. This paper will illustrate several different imputation methods that are possible for handling missing data in pain studies. It will also explore the challenges faced in both of the imputation techniques, single and multiple, using a comparative study. REGULATORY GUIDANCE There is no universally applicable method of handling missing data recommended by any regulatory body. Existing regulatory guidance does not have specific instruction on how to address the problem of missing data. Realizing this gap the U.S. Food and Drug Administration (FDA) created a panel on Handling of Missing Data in Clinical Trials. This panel prepared a report which summarized how to reduce the amount of missing data and appropriate statistical methods to address missing data for analysis. Table 1 displays some selected recommendations from that report (The Prevention and Treatment of Missing Data in Clinical Trials, Panel on Handling Missing Data in Clinical Trials; National Research Council)