The AE Time to Event report screens all adverse events by performing log-rank and Wilcoxon tests between treatment groups. The time to first occurrence of the adverse event is used as the response.
Note: If you run this analysis using data from a blinded study or from a study that has only one ARM, you generate a different set of output. See
AE Time to Event (One ARM) for a detailed description
The AE Time to Event report screens all adverse events by performing log-rank and Wilcoxon tests between treatment groups. The time to first occurrence of the adverse event is used as the response.
Running this report with the Nicardipine sample setting generates the tabbed
Results shown below.
The Volcano Plots section contains the following elements:
The Y axis represents the
-log10(raw
p-value) from either the
Log-Rank test (
left) or the
Wilcoxon test (
right). In most instances, these two tests coincide. To interpret this axis, consider the following.
1. a p-value of 1 equals 0 on the -log
10 scale,
2. a p-value of 0.1 equals 1 on the -log
10 scale,
3. a p-value of 0.01 equals 2 on the -log
10 scale,
4. a p-value of 0.001 equals 3 on the -log
10 scale,
5. a p-value of 0.0001 equals 4 on the -log
10 scale.
The X axis represents the maximum computed distance between the two curves, and
color indicates which treatment was likely to have the event sooner. For example,
Phlebitis occurs at approximately 0.27 on the
X axis and is
green. This means that at least one time point in the
Kaplan-Meier and Hazard Plots, the risk of having this particular event was 27% more likely for Nicardipine versus Placebo.
Events above the dotted red line are considered statistically significant after adjusting for multiple comparisons. Adjustment is applied using the linear step-up method of Benjamini and Hochberg (1995) to control the false discovery rate.
•
|
Fit Incidence Density Model: Click to fit selected events to a generalized linear model. This button is surfaced when either Incidence density based on first event only or Incidence density based on all events is specified as the Response Type.
|
•
|
Click the Options arrow to reopen the completed report dialog used to generate this output.
|
The Term Level is determined by the coding dictionary for the Event or Intervention domain of interest, typically these levels follow the MedDRA dictionary. You must indicate how each adverse event is named and the level at which the event is considered. For example, selecting
Reported Term reports the event specified by the actual event term as reported in the AE domain.
Available variables include Planned, which is selected when the treatments patients received exactly match what was planned and
Actual, which is selected when treatment deviates from what was planned.
You can also specify a variable other than the ARM or
TRTxxP (planned treatment) or
ACTARM or
TRTxxA (actual treatment) from the CDISC models as a surrogate variable to serve as a comparator. Finally, you can select
None to plot the data without segregating it by a treatment variable.
Analysis can consider all events or only those that emerge at specific times before, during, or after the trial period. For example, selecting On treatment events as the
Event Type includes only those events that occur on or after the first dose of study drug and at or before the last dose of drug (+ the offset for end of dosing).
If you choose to Ignore available treatment emergent flags, the analysis includes all adverse events that occur on or after day 1 of the study.
Check the Treatment end date is equivalent to the start date if the treatment end date (
EXTENDTC) is missing from the data. In this case, it is assumed that all treatments were given on the same day and that the treatment start date can be used instead.
By default this report examines the time to the first event as the Response Type, In this analysis, a test of differences of time to first event across treatment groups is conducted using both
Log-Rank and Wilcoxon statistics. Alternatively, you can opt to use incidence densities. Incidence densities compare the proportion of occurrences of events across treatment groups. Data for incidence density, based on the first event only, are aggregated to the treatment level whereas data for incidence density based on all events are aggregated to the subject level.
Use the Include events with overall percent occurrence greater than option to specify a threshold for reporting adverse events. Only events that occur above the entered threshold (in terms of overall percent of occurrence) are displayed in the reports. This value is set to 0 by default, so that all adverse event occurrences are included.
If there is a supplemental domain (SUPPAE) associated with your study, you can opt to merge the non-standard data contained therein into your data.
The Treatment Control Level is specified as either
“Placebo” or
“Pbo”, depending on the value found in your data, by default. However, if your control is defined differently you can use the text box to specify the control level is identified in your study.
You can opt to assess interventions across the entire study (specified by default). Alternatively, you can use the Trial Time Windows option to limit it to selected time points or intervals. By default, time is measured in days. However, you can change the
Time Scale to measure time in weeks. This option is useful for assessing report graphics for exceptionally long studies.
Choose the Distribution the event counts are assumed to follow. Negative binomial fits an extra parameter that can account for over-dispersion.
Specify Fixed Effects by which to model the mean of the response variable. The effects can include any mixture of class variables (as specified above) or continuous covariates. Do not specify the treatment variable as it is automatically included.
Random Effects are typically comprised of class variables and their interactions that are used to model the covariance structure of the response variable. Commonly used random effects are SITEID (for multi-center trials) and STUDYID (for data assembled from multiple studies).
The Gradient Convergence Criterion is a number that is used for determining the convergence for iterations on gradient to numerically solve the GIMMIX model. The specified number here are multiplied by 10
-8 and set to NLOPTIONS statement in SAS PROC GLIMMIX.
The Color Theme option enables you to specify the color scheme of the report graphics.