CBER Selecting the Appropriate Statistical Distribution for a Primary Analysis P. Lachenbruch
CBER A Study of Xeroderma Pigmentosa (XP) A characteristic of XP is the formation of Actinic Keratoses (AK s ) Multiple lesions appear haphazardly on a patient’s back The rate of appearance may not be the same for different patients
CBER Background Analysis: Rank Sum test. Late in study the Statistical Analysis Plan (SAP) was amended to use Poisson regression Unclear if stepwise selection of covariates was planned a priori
CBER Study Results Poisson regression analysis showed highly significant treatment difference (p=0.009) adjusting for baseline AK, age, and age x treatment interaction (stepwise selection) All these effects were highly significant. Substantial outlier problem
CBER Assumptions Each patient has the same incidence rate, per area unit. Chance of more than one AK in small area unit is negligible. Non-overlapping lesions are independent, that is, lesions occurring in one area of the body are not affected by those occurring in another area.
CBER Outliers Outliers are observations that are jarringly different from the remainder of the data May be multiple outliers If frequency is large, this may be evidence that we have a mixture distribution. Can substantially affect analysis
CBER Analyses Two-Sample Wilcoxon rank-sum (Mann-Whitney) test trt | obs rank sum expected | | Combined| unadjusted variance adjustment for ties adjusted variance Ho: ak12tot(trt==0) = ak12tot(trt==1) z = Prob > |z| =
CBER Distribution of AK Data at Baseline (Stem and Leaf) (Yarosh et al, Lancet) Lead | Trailing digits 0* | // 4* | 27 // 10* | 0 oops! Lead | Trailing digits 0* | // 4* | 27 // 10* | 0 oops!
CBER Distribution of 12 Month AK Total Data (Stem and Leaf). stem ak12tot,w(10) Lead| Trailing digits 0* | * | * | 3* | 7 // 7* | 1 8* | 9 // 19*| 3 same patient - in placebo group. stem ak12tot,w(10) Lead| Trailing digits 0* | * | * | 3* | 7 // 7* | 1 8* | 9 // 19*| 3 same patient - in placebo group
CBER Results of Poisson Analyses Poisson regression Number of obs = 29 LR chi2(3) = Prob > chi2 = Log likelihood = Pseudo R2 = ak12tot | Coef. Std. Err. z P>|z| [95% Conf. Interval] age | trt | akb | _cons | G-O-F in control group, 2 = with 8 d.f. G-O-F in treatment group, 2 =682.5 with 19 d.f.
CBER Permutation Test Procedure: Scramble treatment codes and redo analysis. Repeat many (5,000?) times. Count number of times the coefficient for treatment exceeds the observed value.
CBER Command and Output. permute trt "permpois trt ak12tot age akb" rtrt=rtrt rage=rage rakb=rakb,reps(5000) d command: permpois trt ak12tot age akb statistics: rtrt = rtrt rage = rage rakb = rakb permute var: trt Monte Carlo permutation statistics Number of obs = 30 Replications = T | T(obs) c n p=c/n SE(p) rtrt | rage | rakb | Note: c = #{|T| >= |T(obs)|} I deleted the confidence intervals for the proportions
CBER Permutation Tests (2) Poisson with 5000 Replications Treatment: p = 0.57 Age: p = 0.62 AK Baseline: p = 0.28 All significant results disappear
CBER Results of Poisson Analysis Sponsor found that all terms were highly significant (including the treatment x age interaction). We reproduced this analysis. We also did a Poisson goodness-of-fit test that strongly rejected the assumption of a Poisson distribution. What does a highly significant result mean when the model is wrong?
CBER Conclusions The data are poorly fit by both Poisson and Negative Binomial distributions Permutation tests suggest no treatment effect unless treatment by age interaction is included Justification of interaction term by stepwise procedure is exploratory Outliers are a problem and can affect the conclusions.
CBER Conclusions (2) The results of the study are based on exploratory data analysis. The analysis is based on wrong assumptions of the data. Our analyses based on distribution free tests do not agree with the sponsor’s results. The results based on appropriate assumptions do not support approval of the product.
CBER Suggestions Conduct a phase II study to determine appropriate covariates. Need to use appropriate inclusion / exclusion criteria. Stratification. a priori specification of full analysis
CBER Reference Yarosh D. et al., "Effect of topically applied T4 endonuclease V in liposomes on skin cancer in xeroderma pigmentosum: a randomised study" Lancet 357: , 2001.
CBER The End
CBER Grid on “Back”
CBER The Data | sex trt akb ak12tot| | | | F | | M | | F | | F | | F | | | | M | | F | | M | | M | | M | | | | F | | F | | F | | F | | F | | | | sex trt akb ak12tot| | F | | F | | M | | F | | M | | | | F | | F | | F | | F | | F | | | | M | | F | | F | | F | | M... |
CBER Descriptive Statistics (1) Baseline AK N Mean SD Control Treatment Months Total AK Control Treatment
CBER Descriptive Statistics (2) Baseline AK Median Min Max Control Treatment Months Total AK Control Treatment
CBER Negative Binomial Model Need a model that allows for individual variability. Negative binomial distribution assumes that each patient has Poisson, but incidence rate varies according to a gamma distribution. Treatment: p = 0.64 Age: p = 0.45 AK Baseline: p = Age x Treat: p <0.001 Main effect of treatment is not interpretable. Need to look at effects separately by age.
CBER Negative Binomial Results This model shows only that the baseline AK and age x treatment effects are significant factors. It also gives a test for whether the data are Poisson; the test rejects the Poisson Distribution: p< A test based on chisquare test (obs - exp) suggests that these data are not negative binomial.