# Hypothesis test

Six Sigma – iSixSigma › Forums › Old Forums › General › Hypothesis test

- This topic has 15 replies, 9 voices, and was last updated 13 years, 2 months ago by AJ.

- AuthorPosts
- November 30, 2007 at 1:50 pm #48803

itl190370Participant@itl190370**Include @itl190370 in your post and this person will**

be notified via email.Can anyone help. I am currently monitoring 2 machines. We are recording defects per unit produced on both machines (blow holes in a weld). Which hypothesis test would I use to campare. ANOVA?

0November 30, 2007 at 2:00 pm #165529F test and two sample t.

There are much better ways if you are looking to quantify goodness of welds. In the time it will take to get enough data to make DPU a meaningful stat, you could have both machines working better.0November 30, 2007 at 2:16 pm #165530

itl190370Participant@itl190370**Include @itl190370 in your post and this person will**

be notified via email.Thanks for the response. What are your recommendations for assessing the differing performance in machines?

0November 30, 2007 at 2:43 pm #165532You have to keep in mind that if you use ANOVA or a 2 sample t-test or any other test requiring assumptions of normality, you have to make sure that your data is independent and normal at the least. If your data is not normal, then you may have to transform the data or use a non parametric test. Sampling the data from your population with n size greater than 25 will also most likely get you normal data to work with.

0November 30, 2007 at 3:21 pm #165535AJ,

With all due respect, your advice is just plain ignorant.

This guy is using DPU. If his DPU is 1, your sample size advice will give him an average of 25 defects per sample – no problem. If his DPU is .01, most samples of 25 will have no defects, some will have 1 – big problem. You need to know what defect level you are dealing with to give any advice on sample size. It is clear you are parroting some bad training and not talking from experience.

The other thing that is wrong with what is being done is who cares if he can show a difference in two welders? Take a big enough sample over time, one will emerge statistically better than the other one. Big deal. If his parts are designed right, welding is a critical operation. Welding is known to be one of those processes that can be taken to six sigma capability. Instead of wasting time showing a difference, the time should be spent improving both welders. We should be talking process maps with inputs and outputs, measurement capability including the ability to do cross sections and pull tests, and DOE. DOE’s for welding come down to how to hold, how clean the surfaces are, time and temperature. Done – go make the welders better.0November 30, 2007 at 6:39 pm #165552You need continuous data to use anova calcualtions. You have attribute data

0November 30, 2007 at 11:38 pm #165561Ron,DPU when given proper sample size can be treated as continuous.

0December 6, 2007 at 4:26 pm #165793Since you have a discrete data For comparison of two machines for defect you use two sample propotion proportion test.

0December 6, 2007 at 6:11 pm #165803I agree with Amin. The 2P test should give you what you want and is easy to do in Minitab, if that’s what you’re using. Good luck!

0December 7, 2007 at 7:56 am #165835

itl190370Participant@itl190370**Include @itl190370 in your post and this person will**

be notified via email.I have tried using the 2 proportion test but on occasion the defect rate ie blow holes in the weld can be greater than the number of units welded, ie 1 unit can have more than one defect. In this case the failures are greater than the number of units produced and in this case how can I use a proportion failure as the defects is greater than the number of units produced?

0December 7, 2007 at 10:24 am #165838

Krishnam Raju PVParticipant@Krishnam-Raju-PV**Include @Krishnam-Raju-PV in your post and this person will**

be notified via email.Can you send data to me (Excel file) so that I will try to analize and get back to you.

[email protected]

Regards,

Krishnam Raju PV

0December 7, 2007 at 1:00 pm #165842As you are testing number of defects, it is a discrete data. Hence, you should be using 2 proportion test or Chi Square test. 2 Sample t test can be used when you have Y as a continuous variable but that is not the case here. Your ‘X’ is a discrete variable and so is your ‘Y’.

0December 7, 2007 at 1:03 pm #165843I had posted my earlier message without reading the entire thread, first you need to be sure if you are concerned about defects or defectives. In my opinion you should go by the defectives and your defectives will never be greater than the number of units processed and thereby you can use 2 Proportion test.

0December 7, 2007 at 1:25 pm #165846

Harish GoyalParticipant@Harish-Goyal**Include @Harish-Goyal in your post and this person will**

be notified via email.Hi

To make comparison of defects on two machines you can make use of Chi Square Testing.

Regards

Harish0December 7, 2007 at 8:55 pm #165882I am going to attempt to add something here even though Stan thinks I’m ignorant and inexperienced.

I agree with harish, I think since that this is considered analyzing two sets or proportions (% defectives) p1 vs p2, a chi squared test would be the best test for significance.0December 7, 2007 at 9:11 pm #165883You can also use a two sample Z test (not t-test) using the two proportions

(p1-p2-0)/(SE)

Where SE is the (pbar*qbar/n1 + pbarqbar/n2)^.5

where pbar is (x1+x2)/(n1+n2)

You may need to pool the SE if you have equal variance. To pool the variance you will need to:

Then look this up in the Z-table to see if the P-value is below the desired setting which is normally p<.05 for significance. But you need to make sure if the alpha value of .05 is the right alpha risk for your application.

0 - AuthorPosts

The forum ‘General’ is closed to new topics and replies.