We help our customers, with the statistical analysis of their data and perform:

• Sample size calculation & randomization cards
• Discrete data analysis (e.g. logistic regression analysis)
• Continuous data analysis (e.g. linear regression)
• Survival analysis
• Mixed models (for continuous and discrete outcomes)
• Dose-finding studies for Phase I Trials (Continual Reassessment Method, Up and Down design)
• Adaptive designs for Phase II and III Trials (Simon’s two stage design, Group sequential design)
• Meta-Analysis
• Propensity scores (genetic matching)
• Preclinical (PK/PD, non-linear mixed models)
• Latent Class Analysis
• …

Continual Reassessment Method
The Continual Reassessment Method (CRM), first developed in the cancer setting by O’Quigley et al. (1990), is a statistical adaptive design aiming to estimate the Maximum Tolerated Dose (MTD) in Phase I Clinical Trials. It has been shown superior to other dose finding designs (e.g.: up and down, 3+3) because it “learns” from information gained at earlier points in the study, is less likely to treat patients at toxic doses, and more likely to treat patients at efficacious doses (Garrett-Mayer, 2006). Its inference was originally Bayesian, though maximum likelihood estimates have been also provided (O’Quigley, 1996). Many proposals of modifications have been thereafter proposed. Notably, it has been then easily extended for the setting of dose finding in phase 2 trials, with the aim of estimating the effective dose (e.g.: ED95) of a new therapy, defined as the 100(1-p)th percentile of the dose-failure relationship (Zohar et al., 2013).

You can find here a very good introductory article: http://arxiv.org/pdf/1011.6251.pdf

Group Sequential Method
An adaptive design is a clinical study design that uses accumulating data to decide how to modify aspects of the study as it continues, without undermining the validity and integrity of the trial. The goal of adaptive designs is to learn from the accumulating data and to apply what is learned as quickly as possible. In such trials, changes are made “by design,” and not on an ad hoc basis; therefore, adaptation is a design feature aimed to enhance the trial, not a remedy for inadequate planning (Galo et al., 2006).

Adaptive designs allow also to change the study hypotheses, such as (1) switching for superiority to non-inferiority; (2) switch from a single hypothesis to a multiple hypothesis or a combined outcome; (3) changing hypotheses due to the switch in study endpoints; (4) dropping ineffective treatment arms; (5) interchange between the null and the alternative hypothesis (Chow & Chang, 2011). Indeed, at the beginning of the clinical trial, the investigator may not have adequate information about the effect size of the treatment, and rather than to continue to conduct an inappropriate powered trial, the sponsor may wish to modify the sample size or stop the clinical trial when there is enough convincing evidence of benefit (=efficacy) or harm (=futility) (Friedman, Fureberg & DeMets, 2010). Moreover, adaptive designs are well suited for economical purposes: they allow an earlier termination of the trial. If the results are positive, the compound may be exploited sooner and if the results are negative, the resources are not wasted (Chow & Chang, 2011).

You can find here a very good review article: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2422839/pdf/1750-1172-3-11.pdf

We wrote an introductory text on adaptive designs for phase III studies. Please have a look here: AdaptiveDesignsPhaseIII

Genetic Matching (propensity score)
Genetic matching is a generalization of propensity score and Mahalanobis distance that maximizes the balance of observed covariates between treated and control groups. The algorithm uses a genetic algorithm to optimize balance as much as possible given the data. The method is nonparametric and does not depend on knowing or estimating the propensity score. The Genetic matching attempts to minimize a measure of the maximum observed discrepancy between the matched treated and control covariates at every iteration of optimization. The algorithm attempts to minimize the largest observed covariate discrepancy at every step and this is accomplished by maximizing the smallest p-value at each step. The algorithm stopped when the difference between the last four solutions was small. We performed a one to one genetic matching with replacement. Last, an absolute standardized difference less than 10% was considered to support the assumption of balance between the groups because it is not affected by the sample size, unlike p-values, and it may be used to compare the relative balance of variables measured in different units.

You will fin here two papers of Pr Sekhon:
http://sekhon.berkeley.edu/papers/GenMatch.pdf
http://www.jstatsoft.org/v42/i07/paper

Handle Missing Data
Incomplete data occur when not all planned measurement are observed. The term dropout refers to the case where all observations on a patient are obtain until a certain point in time, but for whom the remaining measurement are missing. Dealing with missing values is common in clinical trials including several measurement times.

When missing values are present in the data, and one excludes any patient having any missing value on any variable measured in the clinical trial, it has an influence on:
a) Power: the power increases as sample size increases, and, therefore, the power is reduced in case of missing data;
b) Variability: the variability is reduced when the sample size increases, and, therefore, the variability increases in case of missing data;
The greater the percentage of missing data, the stronger the influence on power and variability of the clinical trial, compromising the conclusions drawn from statistical tests. Moreover, patients having missing data may be more likely to have extreme values on the primary and/or secondary outcomes, and neglecting the missing data pattern may result in biased conclusions. It is therefore crucial to make all the possible efforts to minimize the amount of missing data. Nevertheless, despite all the good will of investigators/sponsors, it is likely that some missing data are present in the collected data.

Latent Class Analysis
Latent Class Analysis classifies respondents into mutually exclusive groups with respect to a not directly observed (latent) trait. The Latent Class Analysis tries to determine how many classes, based on a statistical model, can be found in the data. Therefore, it is less arbitrary than traditional clustering techniques (Magidson & Vermunt, 2004). Fit indicators used by researchers are the BIC and the AIC: the lower the AIC/BIC, the better the model.