Bioequivalence and Bioavailability Forum

Hi Pankaj,

❝ […] the data shows only one outlier subject by Lund's procedure in 2X2 crossover design study. After removing this outlier subject, when the data is again run for Lund’s test, another outlier subject is observed.

❝

❝ Can we perform Lund’s test in such repetative manner to remove outliers from the data?

Please read the original publication.* IMHO, repetitive testing is statistically doubtful (politely speaking). We once faced a similar case 15+ years back when two outliers were detected and the French agency ‘discovered’ yet another one in the reduced data set. Agencies can do almost everything. I would not do it voluntarily.

Some remarks:

• If you are BE with the complete data set, why bother? Nowadays most agencies are allergic against outlier testing and would even accept a study which shows BE with the outlier and fails without. Bizarre, but current practice.
  
• Did you try to find a cause for the outlier (i.e., non-compliance, subject vomited, bioanalytical problems)?
  
• Outliers inflate the variability and lead to failed studies more often than not. If you fail to show BE with the complete dataset and pass after removal of suspected outlier(s) you are in a gray-zone. It would help if you stated the planned procedure in the protocol. Otherwise it smells of cherry-picking.
  
• Since removal of outlier(s) decreases the CV, why do you want to decrease it even further by performing another ootlier test? Still not BE? Smells fishy to me. What does your protocol say?

❝ Is it acceptable by usfda regulatory subject to redosing study with multiple outlier subjects with control subjects)?

In principle, yes. However, the FDA has a strict policy against removal of outliers in non-replicated studies and accepts re-dosing only if the procedure was stated in the protocol. I would not recommend performing a re-dosing study without controled correspondence with CDER/OGD’s review staff (see this recent draft guidance; lines 644–647).

---

• RE Lund
  Tables for An Approximate Test for Outliers in Linear Models
  Technometrics 17(4), 473–6 (1975)