17 replies to this topic

[bubbagump]

Hello,

I really hope someone can help me (I can't find any info about this issue anywhere on the web)...My son is doing regression in his high school math class and as you might expect, most of the students (including the teacher) use TIs (usually 84) and my son is only one with a Casio. I believe the Casio is superior, and it sure is much easier to use than the TI, and my son really likes it.

BUT...When doing regression analysis (and some other stats functions), the Casio gives a DIFFERENT answer than the TI. Sometimes it looks like only rounding differences, other times answers are dramatically different.

Needless to say, this is causing a serious issue for my son in his class. He asked his teacher about it and he thought maybe the Casio was using a different (or even wrong) algorithm. Of course I guess it is also possible that the TI is the one with wrong answer, but how to know?

Have any of you heard about this issue before? I would really appreciate it if you could assist, otherwise I am going to have to fork over big bucks for a new TI calc and my son will have to learn how to use a brand-new calc...Not a good thing in the middle of the school year.

Please help!

Thank you very much for your time.

Sincerely,
Mike

[flyingfisch]

I don't think you should have to buy a ti calc and this problem is probably fixable...

Do you have any example problems where the casio and ti differ? It might help to check in shift+menu (setup) and see if the rounding mode is the same as the ti calc's. Otherwise, elaborating on the problems it doesn't work with would help.

[MicroPro]

Could you give examples that the result of calculation is different? We can see what is wrong and what is right then. Maybe it's just different settings in the calcs?

By the way, welcome to UCF!

[flyingfisch]

Was just thinking, another possibility could be that the problems involve a random seed, and so casio and ti calcs would not produce the same numbers.

[bubbagump]

Hello all...Thanks for fast responses. I was also thinking it must be some sort of "numerical analysis" difference between the two cals.

Yes, indeed I will try to provide some examples. My boys (actually twins, and they each have a Prizm ) will discuss with their math teacher today and do same problems on Casio vs. TI and I will report numbers back to you (they didn't write down the differences yesterday).

Note that we have not just seen this for regression, but one of my boys had prob/stats class last semester and he was getting different answers for standard deviation than others in his class. At the time, we just brushed it off and made the best of it, but my son did get some answers marked WRONG on his exam just because his calc was giving different answer, and he certainly did everything correctly on the calc. It was frustrating, especially since everyone else in class had TIs and obviously it must be the Prizm that has wrong answer (being sarcastic of course)

Thanks

Sorry, forgot to ask...Exactly which menu setting is for "rounding mode"?

Thanks

[flyingfisch]

Oh that's cool, I have twin sisters

About rounding mode, you can access it by entering run-mat mode, pressing shift+menu, and scrolling down to "Display: ". By default it is Norm1, but there are other options including engineering, scientific, and fixed.

Besides that, until I can actually see the problem, I can't think of any other things.

Hope that helps.

[bubbagump]

Ok, below are the examples. One is way off, others are close, but different enough to where their teacher might mark answers wrong...


Problem 1
1 2.6
11 3
21 3.7
31 4.5
41 5.3
51 6.1
54 6.3
Using exponential regression (a*b^x):
Casio Answers
a=2.5764
b=1.0171
r=0.9973
r2=0.9946

TI Answers
a=2.5574
b=1.0173
r=0.9977
r2=0.9954

Same data set using logistic regression (c/(1+a*e^(-bx)))…This one is WAY off:
Casio Answers
a=10.7221
b=0.0197
c=30.0806

TI Answers
a=4.2134
b=0.0261
c=12.8547

Problem 2
1 22.6
6 46.2
11 59
16 65.7
23 68.9
Using logarithmic regression (a+b*lnx):
Casio Answers
a=21.2058
b=15.4639
r=0.9933
r2=0.9866

TI Answers
a=21.7785
b=15.2021
r=0.0.9955
r2=0.9911


Thanks for help.

[hayzel]

Just checked for curiosity, your data with Mathematica:

Problem1 Using exponential regression (a*b^x):
{a -> 2.59978, b -> 1.01688}


Problem1 Same data set using logistic regression (c/(1+a*e^(-bx)))
{a -> 4.21338, b -> 0.0261056, c -> 12.8547}



As you can see on the graphs for the above results and your results (TI,Casio) all are really close to the real data. (red data, blue Mathematica, Dots Brown TI, Dashes Green Casio).
Regression is a numeric procedure, so the differences are rounding errors, different algorithms for minimizing error, maybe different norma checking, etc....
I feel more secure with the results of Mathematica, but there is no real error and on the other solutions.
It is ridiculus that a lesson, and a grade is based on numeric results of a task such the above. Your teacher could search it more before naming an error or correct solution.

[bubbagump]

Hello Hayzel,

Thank you very much for checking this on Mathematica. Indeed it appears issue is numerical analysis differences, which is what I guessed was what was happening.

BUT...What about logistic regression?

Again, this is what I get on the Casio:
a=10.7221
b=0.0197
c=30.0806

And on the TI:
a=4.2134
b=0.0261
c=12.8547

Mathematica:
a=4.21338
b=0.0261056
c=12.8547

It seems you do not see this huge discrepancy?

Thanks

[hayzel]

You 're right, it is a big difference, in the results between TI-Mathematica and Casio but the interesting thing is that it really fits the data in both cases.

So apparently the iterative solution of a such non-linear regression is done, with different algorithms. I cannot say if casio uses a correct way or not, but I can really say that the estimation of fitting of Casio is also too close to the data.

[weirleader]

Hi all,

Thanks for looking into this! I'm trying to pin down the cause of the discrepancy and my Prizm is producing ExpReg and LogReg results identical to the TI results mentioned above. The only discrepancy I've been able to confirm is with the LogisticReg results. If anybody else observes the non-matching ExpReg or LogReg results (NOT Logistic), please share! And let me know what OS version number you're working with.

Thanks!

[bubbagump]

All,

In an act of desperation, I performed a full system reset of both calculators and regression results are now matching the TI (except for logistic, no idea what is going on there, different algorithm I guess?). Success!

Very strange and not sure what setting was wrong, or maybe something was corrupted, no idea. But it is working now and that is all that matters. My boys are very happy :-)

Thanks to everyone for all of your help.

[flyingfisch]

It sounds like your sons may have had something in setup (possibly a different rounding mode) that was setting them apart from the TI's and that by resetting the calcs you reset the rounding mode to system default.

Anyway, I'm glad you were able to fix the problem.

Feel free to come back if you have any more questions.

[weirleader]

I really don't think it was a rounding setup. I tried each mode (Eng/Fix/Norm1/Norm2, etc.) and couldn't get it to affect the regression results. And yet I've no idea what else could've caused it. I guess this'll give me a puzzle to stew on for a while! :-)

[flyingfisch]

Yeah I suppose so... the only other thing I can think of would be the mode (rad, deg) but it wasn't a trig problem so that is highly unlikely

[MicroPro]

You have noticed that the curves from the answers of the Prizm and the TI both almost fit the data, despite the constants being far different? (i.e y=c/(1+a*e^(-bx)) )

So the "problem" here is that they use different algorithms (or the same algorithm with different constants, starting vars, iterations, etc but I can't tell because I don't know how logistic regs are calculated).

If the teacher really insists on the answers being the same, maybe we can find out which algorithm TI and Mathematica are using and make a Basic program to do that Prizm. Can somebody find a good tutorial about logistic regression calculation methods?

[weirleader]

I have been looking for such an algorithm for that exact reason. There's a lot of obfuscation due to conflicting usage of the term "Logistic Regression".

A couple of sites I've found helpful, so far:
http://www.hpmuseum....cgi?read=204621
http://math.fullerto...quationMod.html

If anybody does find a good algorithm described, please do share it!

[ZweiLynx]

This post is the perfect example of why I hate high school math teachers. They're uncapable of using their heads! (The problem described here also makes me seriously question this particular math teacher's capabilities as a mathematics professor - even at high school level)

Also the business agreement TI has with high school math textbooks has been hurting students for far too long, it has to be stopped.

Hayzel's analysis above making a graph in Mathematica should give you the answer as to what's going on here.

In regression analysis (statistics), and by extension mathematics, you can very well have two equations with different coeficients and get the same results. Not going to get into the details of it, I'll leave that to you.

It is the teacher's job to figure out why a student is getting a wrong (or different) result. It's not very difficult to just plug in numbers into the different equations and compare results. An advanced software like Mathematica is not even needed, it can be done in Excel. Couldn't this teacher have thought of that?

But no, this math "teacher" had to take the easy road, and as a result put a student under emotional stress and his parents under financial stress.

I hope you didn't buy a new calculator bubbagump, and I applaud your willingness to go the extra mile (searching the Internet, coming to this forum, etc.) to help your kid.