I need some help. I need to determine the sample size to validate the extraction of medication instances from free text reports. I have something like (I'm making up the numbers since another person is counting the reports, etc).
I have 200 patients divided in four groups:
group 1: 50 patients exposed to a medication who did not have a side effect
group 2: 50 patients exposed to medication who had a side effect
group 3: 50 patients not exposed to medication who did not have a side effect
group 4: 50 patients not exposed to medication who had a side effect (caused by something else)
for these 200 patients I have (I'm making this number up) 15432 reports that have been processed with an Natural Language Processing tool. Since we need to validate the extraction process against a gold standard (actual person checking the reports to see if extraction was correct). The extraction process consists of basically identifying all instances of a given medication inside a report. So, for example, if I want to know if there's a mention of aspirin in the report, the tool will extract it. There could be no mention, or the medication (in this case, aspirin) could appear one or more times in the document/report.
I think I can merge groups 1 and 2 since I'm interested in detecting medication exposure and not the presence of a side effect. So, this will give me 2 groups:
Group A: pts exposed to medication and
Group B: patients not exposed to medication - each group with 100 pts and their corresponding number of reports adding up to 15432.
In order to do this, I need to:
1. determine whether I need to calculate my sample for validation at a medication instance level or at a document/report level so I can calculate specificity, sensitivity, accuracy, precision. However, I do not know the actual number of instances -since they are being picked by the tool, although I know the actual number of reports.
2. determine the number of reports or instances I need so my results are statistically significant.
I think I need the following to determine the sample size:
confidence level - I think that 95% is customary??
confidence interval <- I DO NOT KNOW HOW TO DETERMINE THIS
Population size <_ SHOULD THIS BE MY 15,432 reports or should it be another number?
What are the underlying assumptions? Is this a binomial distribution?
Any ideas??
Thanks so much!
This Question has been solved and asker verified All Experts Exchange premium technology solutions are available to subscription members.
Experts Exchange has been collecting answers to technology questions since 1996…3 million and counting! If you have a question, chances are we already have your answer.
If you can't find the exact answer you're looking for, ask our exclusive community of 50,000 experts. You’ll get a personalized answer from a trusted professional.
Thousands of free tech tips, tricks, how-to’s and tutorials are available in our peer reviewed articles section. See for yourself how smart our experts are, no login required.
Save solutions to your questions, answers you’ve discovered through searching plus helpful articles in your personal knowledgebase for easy future access.
Access the answers to your technology questions today.
30-day free trial. Register in 60 seconds.
Members of the expert community talk about why the experience at Experts Exchange is different than what you will find anywhere else.
Try it out and discover for yourself.
30-day free trial. Register in 60 seconds.
Join the community of experts here and help other tech pros by answering question in your area of expertise. You can earn FREE access to all Experts Exchange's premium features and resources.
Hi richdiesal,
Thanks for your reply. I'm trying to validate the output of the NLP processing, so I need to know how many reports I need to pick and manually review to see if the output of the NLP tool is correct or not without going through all of them, but a number large enough so I can say the NLP works correctly.
Maybe my question was too long. I was trying to explain the problem.
If you're just interested in double-checking the output of an automated tool (I'm not familiar with NLP in particular), there isn't really any sort of test to use. It would just be a matter of what seems reasonable. I know in my field, by-hand recodes of 25% of the data to check accuracy aren't uncommon. But I imagine that varies depending on who is checking and what sounds reasonable to them.
Business Accounts
Answer for Membership
by: richdiesalPosted on 2009-02-11 at 17:28:24ID: 23618447
There are a lot of questions here...
You seem to be talking about power analysis, but power analysis is unnecessary if you already have data. If you don't already have statistical significance, you need more people. If you do, you don't. If you want to know the specific number more cases you should get to have a specific power level (for example, an 80% chance to detect your effect), then that is a more appropriate quesiton - is this what you want?
Confidence level is the probability that you are willing to accept of a Type I Error occurring. So, if you are willing for there to be a 5% chance that if there was really no difference between groups and there really was one, 95% confidence would be the result. In medicine, 99% or 99.9% are more typical. Social sciences usually use 95%.
Confidence intervals are computed differently depending on the test, but generally are of the form:
computed statistic +/- (standard error * test statistic)
We refer to the population as the count of whatever our base level of comparison is, which is usually a theoretical value. If you can collect all data from the entire population, there is no need for the use of inferential statistics (i.e. statistical significance testing).