As I once again read in Struck By Lightning by Jeffery S. Rosenthal, even seemingly solid statistical data can be and often is marred by bias. Rosenthal gave an example of a disease that causes death in 50% of those infected however the other 50% recover on their own. He presents a scenario in which a new drug company wants statistical data with a p-value of at most 5% that their product improves the condition of those infected. Even with a ineffective drug, the company can promote it by using what is known as sampling bias. What the company can do is choose test subjects that already show signs of recovery or could consider them cured when they are still somewhat ill. To avoid sampling bias, test subjects must be chosen randomly so that every subject has the same probability of recovering or not recovering from the disease. The drug company could also use reporting bias in which they focus on reports that support their claims and ignore any negative side affects or flaws. This way the product would seem to be highly successful even if only a small portion of test subjects show improvement (they would 'forget' to mention that the p-value is over the statistically significant 5%). Lastly the questionable drug company can resort to publication bias. They can hire 100 independent experts to conduct 100 independent studies on the effects of their drug. Due to the large number of tests, one study is bound to support the company's claim. This one study will be paraded throughout advertisements and the other 99 studies--that showed little to no improvement as a result of the drug-- will be disregarded. Although statistics portray the idea of objectivity and accuracy, biases are always present whether through who the test subjects are or through which data the company chooses to present to the consumer.
The Paradoxicality
No comments:
Post a Comment