Visit Window for Longitudinal Studies

In the previous post, the time window was discussed. The time window is an issue in studies with relatively shorter durations (such as the clinical pharmacology studies and clinical trials in analgesic drugs). The measures outside the time window can still be used in analysis by using the actual time in calculations. The impact is relatively neglectable.   

In clinical trials with a longitudinal design where the study subjects are followed up at pre-specified schedules for a long period of time, the visit window is an issue to be handled. The visit schedules may not be followed due to the reasons of 1) patient side (for example, travel arrangement); 2) investigator side (for example, investigator may have other responsibilities and not able to see the patient on a specific day); 3) the study procedure cannot be performed on the specified visit day (for example, the CT, MRI, … may not be arranged at the exact visit date).

It is common that the visit window is allowed and specified in the protocol, in this way, not every out of window visit will be recorded as the protocol deviation. The visit window needs to be protocol specific and needs to specified according to the length of the study and interval of the study visits. A study with every four-week visit schedule may have a visit window +/- 1 week; a study with every six-month visit schedule may have a visit window +/- a month.

It is inevitable to have always some subjects with visit outside the visit window. The study will also allow having unscheduled visits and early termination visit (the early termination visit can occur at any time during the study). These visits and measures outside the visit window will need to be used in the analysis so that the missing data can be minimized, and the more representative measures are selected to be used in the analysis. This needs to be handled through programming (usually in analysis dataset programming or CDISC ADaM dataset programming). Generally, two steps are needed:

Specify the Visit Slotting Algorithm

Based on the actual visit date, allocate the actual measure to the specific visit. An example below is for a six-month (24 weeks exactly) study. The visit slotting algorithm can be specified as the following:

Nominal Visit (Scheduled Visit)	Scheduled Visit Day (Study Day)	Slotting Intervals
Baseline	1	-3 to 1 day
Week 1	8	2 to 11 days
Week 2	15	12 to 22 days
Week 4	29	23 to 43 days
Week 8	57	44 to 71 days
Week 12	85	72 to 99 days
Week 16	113	100 to 127 days
Week 20	141	128 to 155 days
Week 24	169	156 to 14 days after the last treatment dose

Selecting the Value for Analysis 

After all the observations have been slotted based on the algorithm above, it is very possible that there are multiple valid observations for an assessment within an assigned analysis visit. Only one of these observations will be used for summary statistics and analyses. In the ADaM program, this is to determine the ANL01FL (‘Y’ if the observation is selected for analysis)

The observation to be used is determined using the following hierarchy (in decreasing order):

The observation closest to the target study day

The latter observation, if 2 observations are equally close to the target study day

Impact on the Missing Data Imputation

If the missing value needs to be imputed, the imputation should be implemented after the above two steps. For missing values where the last observation carried forward (LOCF) algorithm is applied, it is always the last valid observation on treatment carried forward, even though this might not be the observation obtained by the above hierarchy and used in the summaries by visit window.

Impact on analyses using a mixed model such as MMRM (mixed model repeat measure)

After the slotting and determining the analysis flag (ANL01FL), the mixed model will be based on the data with ANL01FL =’Y’ (i.e., some observations within a visit window may be excluded from the analysis).  Can we use all observations and use the actual visit day in the mixed model analysis? If you try, you may run into the issue that the model does not converge.