This preview shows pages 1–2. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.View Full Document
Unformatted text preview: Bias and Error in Clinical Trials The analytic goal in a clinical trial is to make inferences about some population of people who might receive the intervention. Possible forms of inference: Hypothesis tests: is loss of bone density slowed by giving soy isoflavones, compared to placebo? Point estimate with confidence interval: the rate of bone decline in high-dose soy isoflavones is 8% per year slower than in women receiving placebo (95% CI, 2% to 17%) Decision making: We conclude that the response rate from giving bevacizumab plus radiation warrants further investigation in a randomized Phase III trial. Prediction with prediction intervals: We provide a formula for estimating the likely benefit of a new therapy for metastatic prostate cancer in terms of progression-free survival, as a function of the change in PSA in first week on the treatment. Any inference may have errors . We might be wrong! There are two main sources of error in clinical trials: random variation and bias . Random variation as a source of error Random variation can lead to errors, based on purely chance sources. The expected impact is usually neutral, that is, on average it neither increases nor decreases the evidence favoring a treatment. Here are some typical examples to consider and discuss. 1. Random assignment to treatment. Big advantage: we can quantify this, by probability models. Example: We randomly assign 20 patients with asthma to use a standard inhaler, and 20 to use a new one, and we measure FEV before and after use of inhaler. What kinds of inferences are possible for comparison of standard vs. new? How can we use randomization in our inference? 2. Variation within patient from sources other than treatment: between-day variation, within-day variation, within-sample between-aliquot variation. We can address this by adequate design. Example: Suppose we want to compare blood pressure control under two different drugs. We will measure people during a baseline period, and then again over a follow-up period. Blood pressure is known to have substantial variation between days and within day in the same person. How might we model this, and how might it influence our design considerations? 3. Instrument or laboratory variation in measurements. Example: A key outcome measure in AIDS clinical trials is the number of different kinds of immune system cells in the blood. These are usually measured by flow cytometry: you treat the sample, run it through a machine, and the machine is set to steer cells with certain markers to a different gate, then count how many go through. We recently analyzed data from an experiment where the investigator took a large sample from each of three patients. He split each sample into 18 aliquots. The aliquots were analyzed using 3 different instruments, two reagents (thus 3 2 = 6 conditions), with three replicates per condition. What are the potential sources of variation here? How might you write a model for this?per condition....
View Full Document
This note was uploaded on 12/30/2010 for the course BST 252 taught by Professor Tsodikov during the Winter '06 term at UC Davis.
- Winter '06