Thursday, March 15, 2007


Let’s assume that FDA researchers are reviewing a drug for possible approval. They are reviewing results and data from clinical trials etc. Of course there are some side effects , but only from a small percentage of the samples. Let’s look at a model of their decision making process in the context of hypothesis testing. Let’s assume:

Ho: X = Xo ‘drug is harmful’
Ha: X = Xa ‘drug is safe’

If the decision makers make the mistake of releasing a harmful drug, the consequences would be easily identifiable, and could be visibly traced to their decision. They would want to avoid this at all cost. In essence, they want to avoid making a type I error.

Type I Error: = releasing a harmful drug

From my previous discussion on hypothesis testing, we know that to decrease the probability of committing a type I error, we choose an alpha level that is lower. In a t-test, if we are really afraid of making a type I error, we might set alpha ( the significance level) at .05, .01, or to go overboard.001 etc.

As previously discussed, setting alpha lower and lower increases beta, or the probability of committing a type II error. Recall, a type II error is accepting a false Ho. In this case that would be equivalent to falsely concluding that the ‘drug is harmful’ when it could actually be released and improve the lives of millions.

Type II error bias = over precaution; setting alpha so low, or setting the standard of proof so high as to almost always reject Ho, and biasing the decision in such a way that the probability of a type two error (beta) is greatly increased.

As a result of type II error bias, many life improving drugs never make it to the market.


Anonymous said...

your Ho and Ha as stated are
not realistic.
[where has the FDA stated an alternative hypothesis
'drug is safe' ?]

but the general issue you are trying to get at _is_ very relevant-> see eg Piaggio et al.
JAMA. 2006 Mar 8;295(10):1152-60.

agEconomist said...

Thank you for pointing me to the JAMA article. Perhaps after reading it my comments will need further revision.

But, you are correct. In real experiments with pharmaceuticals, I'm sure an actual hypothesis is much more technical and specific than 'drug is safe' etc.

The purpose of Ho and Ha in this example is not to be realistic , but to serve as a behavioral model that abstracts from reality to focus on the importance of incentives and how they affect decision making. This example where a hypothesis is specified as 'drug is safe' is a generalization used proliferantly throughout economic literature.

However, I think your comment brings up a good point. Maybe, to be somewhat more precise, and still maintain the models explanatory power conveyed by abstraction, we should clarify what we mean by 'drug is safe'.

Perhaps it should be noted that this implies that given the pelthora of biochemical and statistical tests associated with drug approval, a decision maker claims that the 'drug is candidate for release' or something like that.

Sometimes it may be the case that economists, as modeling fanatics, and going for big picture applications, get carried away and aren't explicit enough about what their assumptions are. More congnizance about that issue, and the fact that their audience may often include non-economists would go far in winning acceptance for models they present for policy applications.

Joseph said...

Dear Matt

Thank you for explanation

In essence the FDA practises defensive medicine in order not to be sued

If potentially useful medicines are not reaching patients because of fear of litigation then there neeeds to be a change to the system to protect the FDA from lawsuits but at the same time still holding them accountable for their decisions

Kind regards


Joseph said...

Dear Matt

Thank you for your explanation
It appears that the FDA is practising defensive medicine because of fear of litigation.

If the end result is that potentially useful drugs are not reaching patients for this reason the system in which the FDA operates has to change

For example maintain accountability but explore other methods of risk assessment.

Kind regards