twosampt -- See if two sets of censored data are from the same population.
The twosampt task computes several nonparametric two-sample tests for comparing two or more censored data sets, giving a variety of ways to test whether two censored samples are drawn from the same parent population. These are mostly generalizations of standard tests for uncensored data, such as the Wilcoxon and logrank nonparametric two-sample tests. They differ, however, in how the censored data are weighted or "scored" in calculating the statistic. Thus each is more sensitive to differences at one end or the other of the distribution. The Gehan and logrank tests are widely used in biometrics, while some of the others are not.
The two-sample tests are somewhat less restrictive than the Kaplan-Meier estimator, since they seek only to compare two samples rather than determine the true underlying distribution. Because of this, the two-sample tests do not require that the censoring patterns of the two samples are random. The two versions of the Gehan test assume that the censoring patterns of the two samples are the same, but the version with hypergeometric variance is more reliable in case of different censoring patterns. The logrank test results appear to be correct as long as the censoring patterns are not very different. Peto-Prentice seems to be the test least affected by differences in the censoring patterns. There is little known about the limitations of the Peto-Peto test.
If the results of the tests differ significantly, then the Peto-Prentice test is probably the most reliable. The two-sample tests all use different, but reasonable, weightings of the data points, so large discrepancies between the results of the tests indicates that caution should be used in drawing conclusions based on this data.
- input [string]
- Input file(s); this can be a list of files. Following each file name is a list of column names in brackets. Thes column names specify which columns in the file contain the information used by this task. The brackets MUST contain three names in the following format: [censor_indicator, variable, group_indicator]. The censor indicator specifies the censorship of the data point. A 0 indicates a detection, 1 indicates that the data point is a lower limit, and -1 indicates that the point is an upper limit. The variable specifies the column containing the data points. The group indicator specifies to which group the data point belongs. Each file must contain at least two different groups. The group indicators may be any integers. The two groups to be tested are specified by the task parameters first and second. If the input file is an STSDAS table, the names in brackets are the table column names. If the input file is a text file, the names in brackets are the column numbers. A title string will be printed if the input file is a table containing the header parameter TITLE.
- (tests = " ") [string, allowed values: gehan-permute | gehan-hyper |
- logrank | peto-peto | peto-prentice ]
Names of two-sample tests to perform. If this task parameter is blank (the default), the task will perform all of the above tests. Otherwise the task will perform those tests specified as the task parameter. Test names are given as a comma or blank separated list of names. Test names may be abbreviated as long as the abbreviation is unambiguous. Ambiguous abbreviations or unrecognized test names will be ignored.
- (first = 1) [integer]
- The group indicator which specifies the first group of data to use in the two sample test.
- (second = 2) [integer]
- The group indicator which specifies the second group of data to use in the two sample test.
- (kaplan = no) [boolean]
- Is Kaplan-Meier estimator also desired for each group?
This is the same as given by the kmestimate task.
- (verbose = no) [boolean]
- Print detailed output?
1. Apply the two-sample tests to the data in the text file "twosampt.dat". There is a copy of this file in the statistics$data directory (i.e., "statistics$data/twosampt.dat"). The notation [2,3,1] means that the second column is the censor indicator, the third column contains the values, and the first column contains the group numbers. The hidden parameters first and second indicate that the task should compare groups 0 and 1.
cl> twosampt twosampt.dat[2,3,1] first=0 second=1
If none of the data are censored (i.e., all values in the censor column are zero), then the task may crash with a divide by zero error. A workaround is to give an additional value that is marked as a lower or upper limit but with such an extreme value that it does not bias the result.
Type "help statistics option=sys" for a higher-level description of the statistics package.