The Cytel blog keeps you up to speed with the latest developments in biostatistics and clinical biometrics.
By Ashwini Joshi
For small sample data or rare events data, exact non-parametric tests perform better than asymptotic tests. But they come with the disadvantage of conservativeness. Many corrections have been suggested to reduce this conservativeness but none of them solve the problems entirely. StatXact provides various methods of computing exact p-values. Depending upon the problem at hand, the user can decide which one to use.
Let’s consider a hypothetical example of stratified count data. The example shows two sample data with two strata. Events in Treatment1 are rare as compared to the ones in Treatment0. But the event rates are comparable.
A common challenge of working with small sample sizes is determining proper bias correction methods when evaluating a given set of data. Oftentimes, statisticians depend on large sample sizes to naturally correct for any bias. Small sample sizes, by contrast, require innovations like Firth’s famous bias correction method.
Recently, Cytel statisticians Ashwini Joshi and Sumit Singh gave a talk entitled “How to Reduce Bias in the Estimates of Count Data Regression.” Using a number of case studies and simple to use LogXact PROC software, they demonstrate the ease with which bias correction can be implemented for small sample clinical data.
It is often necessary to pool safety data from late phase studies, in preparation for regulatory submission. Some of our clients have also begun to add Phase 1 safety data to this pool. On some occasions this is required by regulators. In many cases, however, these Phase 1 data simply provide further evidence that a new therapeutic lives up to the promise of safety across patient populations.
When conducting a clinical trial with small or sparse data sets, statistical methods meant for large sample sizes may fail to obtain an accurate interpretation of data. This is where computationally challenging exact methods often come into play.
Exact methods, however, are inferentially conservative in the sense that due to small sample sizes, the actual Type 1 error rate is often smaller than the nominal (intended) rate . There exists an array of strategies to combat this troublesome feature of exact tests, each of which varies along the parameter of computational complexity.