Increase Accuracy of Program: How Many Subjects Needed?

  • Thread starter Thread starter jt.harperjr
  • Start date Start date
  • Tags Tags
    Analysis
AI Thread Summary
To achieve an accuracy rating above 80% in classifying individuals based on two variables, the model must be refined to account for variations and exceptions among subjects. Testing a larger sample size can improve the model's representativeness of the general population, but the definition of "accuracy" is crucial. The discussion emphasizes the need to transition from a binary classification to a scale that captures a continuum of characteristics. Incorporating more subjects will likely enhance accuracy, but careful consideration of sample selection is necessary. Ultimately, a well-defined model and a comprehensive understanding of accuracy are essential for reliable outcomes.
jt.harperjr
Messages
3
Reaction score
0
I'm a programmer, but I know very little about statistics and am not even sure where or how to ask this. Let's say you have 2 variables about people in general, var A and var B, that are tangible characterists of these people. People either possesses A or B.

I then take 11 different measurements about the person and use those to determine if they are actually A or B without looking at them. The program successfully determines if someone is A or B in a group of 10 people. But as I test more and more people, I find that some people have slight differences or exceptions in their variables that I have to account for.

Example: All A people have the first variable in a range of 12 to 13, the 2nd variable in a range of 5 to 6, but then I find an A person who has a range of 1 for the 2nd variable. So I add to the formula that if the 2nd variable = 1, then the person has A.

My question - How many people would I have to test out to get an accuracy rating above 80% of the program, or is that even possible. As I add more and more subjects that fit that equation, does that translate into an increase in accuracy of the program when used on the general population?
 
Physics news on Phys.org
jt.harperjr said:
I'm a programmer, but I know very little about statistics and am not even sure where or how to ask this. Let's say you have 2 variables about people in general, var A and var B, that are tangible characterists of these people. People either possesses A or B.
So to start with your model assumes that someone is one or the other ... i.e. maybe A=gay and B=straight ... and you cannot be bisexual.

I then take 11 different measurements about the person and use those to determine if they are actually A or B without looking at them. The program successfully determines if someone is A or B in a group of 10 people. But as I test more and more people, I find that some people have slight differences or exceptions in their variables that I have to account for.

Example: All A people have the first variable in a range of 12 to 13, the 2nd variable in a range of 5 to 6, but then I find an A person who has a range of 1 for the 2nd variable. So I add to the formula that if the 2nd variable = 1, then the person has A.
So you discover with testing that it the initial model needs to be refined to account for those mostly straight people who have experimented with same-sex relationships in college or something?

Maybe you noticed that some people with mostly blue eyes have a bit of brown flecks in them or some people who are basically dark-skinned are, yet, not exactly black either.

My question - How many people would I have to test out to get an accuracy rating above 80% of the program, or is that even possible. As I add more and more subjects that fit that equation, does that translate into an increase in accuracy of the program when used on the general population?
You need to change your model. Perhaps you need to set A and B as opposite ends of a scale.
You also need to define what you mean by "accuracy rating".
But your question is too general.
 
Simon Bridge said:
So to start with your model assumes that someone is one or the other ... i.e. maybe A=gay and B=straight ... and you cannot be bisexual.

So you discover with testing that it the initial model needs to be refined to account for those mostly straight people who have experimented with same-sex relationships in college or something?

Maybe you noticed that some people with mostly blue eyes have a bit of brown flecks in them or some people who are basically dark-skinned are, yet, not exactly black either.


You need to change your model. Perhaps you need to set A and B as opposite ends of a scale.
You also need to define what you mean by "accuracy rating".
But your question is too general.

To give more detail, in the example above, you could be bisexual. I was just trying to keep it simple. There is actually A, B, C, D, and E, they are on a scale going from A to E, placing people somwhere in between or on the ends.

By accuracy rating, I wonder if I incorporate enough subjects, would it translate to being more accurate for the worlds population as a whole?
 
The bigger your sample, the more representative the results will be of the population responses - if you were to test the entire population. Does that mean it is "accurate"? Depends what you mean by "accurate". Depends how you choose your sample.

You also need a model that accounts for the continuum between the types.
Like height is pretty continuous, but someone who is 185cm tall is actually between 184.5cm and 185.4cm
You could be more granulated than that - using wider ranges for "small", "medium", "tall", "giant" etc.
 
Thanks for your help Simon, after some work on the dry erase board, I've found an answer.
 
Well done. I use a windowpane myself ;)
 

Similar threads

Replies
14
Views
2K
Replies
43
Views
2K
Replies
1
Views
1K
Replies
7
Views
2K
Replies
16
Views
3K
Back
Top