FAQ

Q. What is the correct citation to use for SOLAS for Missing Data Analysis?

A. The correct citation for SOLAS for Missing Data Analysis is as follows:
Statistical Solutions Ltd. (2001) SOLAS for Missing Data Analysis. Statistical Solutions, Cork, Ireland

Q. In the SOLAS™ brochure, the description of the Predictive Model-based multiple imputation says that “the predictive information contained in a user-specified set of covariates is used to predict the missing values “¦”. What about the situation where there are also missing values in these covariates?

A. SOLAS™ provides three options to deal with this situation:

Option 1:

Use hot-deck imputation to impute the covariate(s). Any missing values occurring in the covariate(s) will be imputed by hot-decking on some other variable in the dataset. This method will work well if the variable chosen to hot-deck with is highly correlated with the covariate.

Option 2:

2. Include a missingness indicator variable for this covariate in the regression pool. The indicator method is based on the following statistical model for y:

y=b0 + b01(1-R1+ K + b0p(1-Rp) + K + b1R1×1 + K + bpRpxp + e; with e~N(0,s2)

In this model, the term:

bjR jx j is zero when x j is missing and is equal to bjx j when x j is observed.

When x j is missing, the intercept term is adjusted by the term:

b0 j (1-Rj )

If a covariate x j is completely observed, then the corresponding term b0 j (1-Rj )

disappears.

Option 3:

Exclude the cases that have missing values in the covariate(s) from the analysis. If you choose this option, then any case that has missing values in the covariate(s) is just excluded from the imputation.

Q. I have just installed the new version of SOLAS™. I ran a couple of multiple imputations, and I noticed that the RollUp Editor, that was a feature in version 1.1, is no longer available. Why is this?

A.  One of the new enhancements in the latest version of SOLAS™ is that any statistical analysis that you run on a set of multiply imputed datasets is automatically combined for you and presented in a separate output report. For example, if your complete-data analysis was going to be a two-group t-test, then you would run the t-test on the set of 5 multiply imputed datasets, and the output would consist of 5 separate output pages (one for each of the 5 imputed datasets) and one overall set of combined results.

In version 1.1, the only statistic that was automatically combined by the program was the mean. To combine any other parameter estimate, it was necessary to manually enter the 5 parameter estimates and their standard errors into the RollUp Editor in order to calculate the combined repeated imputation inference.

Q. I am trying to run a multiple imputation in SOLAS™. I have 5 variables in my dataset that I want to impute and 2 other variables that I want to use as covariates. These two variables contain a few missing values so when I try to use them as covariates SOLAS™ warns me that the cases that contain missing values in the covariates are going to be excluded form the imputation. This means that some of the missing values in the 5 variables that I want to impute do not get filled in. Is there anything that I can do to get around this?

A.  SOLAS™ requires that the variables that are to be used as covariates be fully observed, otherwise the cases that contain missing values will be excluded form the analysis. If your covariates only contain a small amount of missing data, then it would be reasonable to use one of the other imputation techniques available (such as Group Mean Imputation or Hotdeck Imputation) to fill in these missing values before proceeding to multiple imputation on the other variables.

Q.  I am using SOLAS™ for Missing Data Analysis to impute for missing values in my clinical trial dataset. I have several longitudinal variables, each made up of a baseline period and four post-baseline periods. I want to use the Last Value Carried Forward technique to impute any missing values, but I want to make sure that baseline measurements are never carried forward. Is there any way to tell the program not to impute if a patient is missing for all periods except the baseline?

A.  There is a very simple solution to this problem. When you define your longitudinal variable in SOLAS™, simply exclude the baseline period from the set. This will mean that any case that is missing in all periods except the baseline will not have any baseline value to be carried forward, and so, will not be imputed.

Q. I am using the Hot-Deck Imputation technique in SOLAS™ to impute the missing values in my dataset. The problem that I have is that all of my data are continuous with up to four decimal places. This makes it almost impossible to find “matching respondents”.

A.  When your variables have many decimal places, it can make finding matching respondents more difficult, especially of you have specified several sort variables. To avoid this problem, you could use the INT(var) transformation to create truncated copies of your variables and use these as you sort variables instead.