Introduction to Clinical Trials, Clinical Trial Designs, and Statistical Terminology Used for Predictive Biomarker Research and Validation

Abstract

This chapter provides an introduction to clinical trial designs and analysis techniques used in evaluating new experimental drugs for cancer patients. It also provides an overview of two major types of biomarkers, prognostic and predictive, that are commonly used in oncology. The chapter closes with descriptions of different clinical trial designs that incorporate, or discover, potential biomarkers.

Glossary

Adjusted (or multivariable hazards ratio (HR)

A multivariable Cox model allows the evaluation of the association of multiple variables on the outcome (e.g., survival). This allows a more accurate assessment of the relationship of a variable of interest to overall survival by accounting for other variables that may be associated with survival. For example, when evaluating the association of a biomarker with survival, a treatment variable may be added to the model. This would allow the evaluation of the association of the biomarker with survival after accounting for the association of treatment with survival. The hazard ratio for a variable from a multivariable Cox model is referred to as a multivariable HR or an adjusted HR.

Continuous (bio)marker

A continuous biomarker is one that has an infinite number of possibilities; in other words, it can take on any value between its minimum and maximum value if it could be measured to any desired degree of precision. An example of a continuous biomarker is PSA level for prostate cancer. The minimum value is 0 and there is no absolute maximum. If PSA could be measured to any desired degree of precision, all nonnegative values are possible.

Cox proportional hazards model

A Cox proportional hazards model is a regression technique for time-to-event data (e.g., survival) where there is censoring (when some patients are alive at the time of analysis). It is a way to evaluate the association of a variable with the time-to-event outcome such as survival. The method is semi-parametric; that is, it does not assume a model for t survival but does assume that the effect of a variable on survival is constant over time. The association is measured by a hazard ratio (HR) where HR = 1 means no association, a HR <1 means increasing values of the variable reduces the chance of death, and HR >1 means that increasing value of the variable increases the chance of death.

Dichotomous (bio)marker

A dichotomous biomarker is one that takes one of two possible values. It is used to split patient cohorts into two categories or groups. An example of a dichotomous biomarker is estrogen receptor (ER) status for women with breast cancer: ER positive versus ER negative.

Log-rank test

A log-rank test is used to compare the survival distributions of two or more groups. The null hypothesis is that there is no difference among the groups. If the p-value is significant (e.g., less than 0.05), this is evidence that the groups have different survival experiences. Note this is only a test for a difference among the survival experiences and does not provide an estimate regarding the size of the differences between any two groups.

Meta-analysis

A meta-analysis encompasses techniques for combining data from multiple studies. An underlying assumption is that the treatment effect is consistent across studies and combining results across studies yields increased power. Most meta-analysis approaches essentially compute a weighted average from the results of the individual studies, and larger studies tend to be given more weight.

Randomization or random assignment

In randomized trials, the participants are assigned by chance to the treatment groups (arms) rather than by choice. Randomization serves to make the groups similar with respect to variables (e.g., patient characteristics, tumor traits) other than the treatment. This means if differences are observed for the outcome variable (primary endpoint), it can be attributable to the treatment since the groups balanced for the other variables. Randomization is accomplished with a chance procedure (e.g., flipping a coin) or a random number generator.

Stratification variable

A stratification variable in a clinical trial is a variable that is used to group patients into strata corresponding to the values of the variable. Randomization is performed separately within each stratum. An example of a stratification variable is whether a patient has disease in his/her lymph nodes or not (e.g., lymph node status with values of lymph node positive and lymph node negative). Variables selected for stratification are those where it is important there is no imbalance between the treatment arms because they are highly prognostic of outcome.

Type I error

Type I error is the error that occurs when the null hypothesis is rejected although it is true. It is a false-positive result. For example, suppose in reality there is no difference between the experimental treatment and standard of care with respect to overall survival. However, a clinical trial is performed, and it is found that the treatment arm had superior survival compared to the standard of care arm with a p-value of 0.03. The investigators conclude that the experimental treatment is better than the standard of care. In reality, this is an incorrect conclusion and an example of a type I error. (Note that the investigators would not know that their conclusion is incorrect.)

Univariable hazards ratio (HR)

A univariable hazard ratio is the ratio of hazard rates for an event (e.g., death) corresponding to the different values of one variable of interest. For example, in a Cox model that contains only a treatment variable (experimental versus control), a HR = 0.50 for survival indicates that patients in the treatment group die at half the rate per unit of time as patients in the control group.