Dealing with missing data

For various reasons there are often participants in a trial for whom a complete set of data is not collected. This is the province of missing data. When conducting efficacy analyses we need to address this issue, and the way(s) in which it is addressed can influence the regulatory reviewers interpretation of the analyses presented. The issue of missing data is problematic in clinical research because humans have complex lives. Human participants may choose to leave a study early or be unable to attend a specific visit, both situations leading to missing data. Nonclinical research involves tighter experimental control in which the subjects (animals) do not have the ability voluntarily to leave the study early. 

Piantadosi (2005) observed that there are only three generic analytic approaches to addressing the issue of missing data  

Disregard the observations that contain a missing value. 

Disregard a particular variable if it has a high frequency of missing values. 

Replace the missing values by some appropriate value. 

---

There are a number of alternative approaches to dealing with missing data in common practice. Amongst these are the following ... 

In all cases and particularly where the extent of missing data is substantial, several analyses will usually be undertaken to assess the sensitivity of the conclusions to the method used to handle missing data. If the conclusions are fairly consistent across these different analyses then we are in a good position. If, however, our conclusions are seen to change, or to depend heavily on the method used for dealing with missing data, then the validity of those conclusions will be drawn into question. 

A number of approaches to dealing with missing data are described by Molenberghs and Kenward (2007). 

While there are widely accepted methodologies for dealing with missing data, it is certainly... 

Accurate results may be obtained by maximum likelihood (ML) estimation or Bayesian estimation if one is using a formal probability model (e.g., a normal model) and the missing values are MAR when dealing with missing data. Since both ML and Bayesian approaches rely on the complete data likelihood, the function linking the observed and missing data to the model parameters, the probability model is key. 

Data Stating which data are to be included, the data input format required by NONMEM, how to deal with missing data, and how to handle outhers... 

Walczak, B. Massart, D.L. (2001). Dealing with missing data Part I and Part II. Chemometrics and Intelligent Laboratory System. Vol. 58, pp. 15-27 and pp. 29-42. ISSN 0169-7439... 

B. Walczak, D. Massart, Dealing with missing data Part i, Chemometrics and Intelligent Laboratoy Systems 58 (2001) 15-27. 

F. Arteaga, A. Ferrer, Dealing with missing data in mspc several methods, different interpretations, some examples. Journal of Chemometrics 16 (2002) 408-418. 

An interesting feature of the SOM is its ability to deal with missing data. Some of the components of the data vectors may not be available for all data items, or may not be applicable or defined. The simplest solution when dealing with such incomplete components would be to discard the incomplete variables or incomplete data items completely, but in this way we will lose useful information. In the case of the SOM, the problem of missing data may be treated as follows When choosing the winning unit, the input vector can be compared with the reference vectors nti using only those components that are... 

---