# Significant Increase or Decrease

Six Sigma – iSixSigma Forums General Forums Tools & Templates Significant Increase or Decrease

Viewing 14 posts - 1 through 14 (of 14 total)
• Author
Posts
• #53362

J
Participant

Hi Six Sigma Guru’s,
I gather a weekly population data of several metrics in our organization. How would I determine if the increase or decrease in percentage is significant or not? Is there a better way to show a graphical representation of population data aside from a Trend Chart that would give a statistical conclusion? Hoping for a friendly response.

GBU

0
#189782

Participant

Sounds like you need to do something that gives you a p-value. How about a regression, checking to see if the slope is significant? A Fitted-Line plot can do the same thing. (If your change isn’t linear, then it gets a little more complicated, but you could still use multiple regression).

0
#189797

Szentannai
Member

Hi,
maybe a quick and not really tidy way would be to present the confidence intervals together with the values. As long as they overlap you have a rather low (though not zero) probability of missing a significant increase.
After you see a signal, in the sense of a confidence interval NOT overlapping with the others, you should re-sample and do a hypothesis test.

This is not perfect but I guess it should work in practice.

regards
Sandor

0
#189798

J
Participant

Delenn,
Appreciate the input. Let me give additional context on the nature of the data. At the end of the week, I want to compare the current week’s percentage against the previous week’s percentage leaving me only 2 data points(both in percentage). Take note, the data is a population data(our data system is capable of gathering all the data in our production). I want to convince the management team that a certain increase or decrease in percentage is “not alarming” or “alarming”.

Thanks,
GBU

0
#189810

Participant

Well, if it is population data, then you have a hard number of increase or decrease — you don’t need statistics to infer to the population. The “alarming or not alarming” then becomes a business decision.

With only two data points, I don’t think you can make a call. You need a greater period of time to see if the increase (or decrease) between the weeks is due to “normal” variation or a special cause (your “alarming” scenario). I think you’re back to collecting more data over multiple weeks before you can give an analysis.

So,

1. Test your measurement system. Make sure your measurement system can actually differentiate between these values, and you’re not reporting noise. (Gage R&R, for example).
2. Run a control chart over multiple weeks. Can you get a good chunk of historic data so you have a baseline?
3. If your process is measurable and in control, then you can start to say whether individual weeks are outliers or within normal range.

If there is a shortcut to this, I would be glad to here it, but that’s my best advice.

0
#189812

Andrew Banks
Participant

GBU,

I’d like to ask some clarifying questions:

-At the end of the week, I want to compare the current week’s percentage against the previous week’s percentage leaving me only 2 data points(both in percentage).

Do you have a history of previous weeks’ data? Why stop at just last week and this week? What about week before last? And the week before that? Does the data only exist for two weeks?

-Take note, the data is a population data(our data system is capable of gathering all the data in our production).

I’m not sure that 100% sampling is a population. Establishing the population distribution of your process would likely require data collection over a much longer time frame than a week.

-I want to convince the management team that a certain increase or decrease in percentage is “not alarming” or “alarming”.

It sounds like you would like to construct a p-chart (attribute data (# defective) and you know the total # of units). However, I agree with Delenn – you’ll need more data. Is weekly the only possible subset? How about daily?

Regards,

0
#189817

J
Participant

Delenn,
In this case, where I’ve got the population data in percentage format calculated on a weekly basis, what type of control chart best represents the data? As an an additional context to this, the population data is taken from 5 factors(our suppliers) but the nature and the method of calculation of the data are the same.

0
#189818

J
Participant

BBinNC,
The history of up to 54 previous week’s data is available. Most of the time, the audiences are more interested in the current week’s and previous week’s performance. Given that the data is reported to the management only once a week, sometimes the previous weeks’ data are taken for granted and ignored, but I got your point.
I thought a P-chart is the option but that would give additional groundworks because I don’t have visibility of the attributes and the number of units. The data comes out as a percentage from the system. If there’s no other alternative, I have no choice but to go deeper and collect the number of attributes.

0
#189824

Participant

Can you tell us if the data is for defectives (good parts vs bad parts) or defectives (number of errors in a lot)? Also, could you make up a number that shows the general idea of % (i.e., typically in a week you see 5 out of 1,000 defects, or is is 200 / 1,000 defects)? All these things have an effect.

0
#189825

J
Participant

Delenn,
The data is for defectives (number of errors in a week of production). Typically in a week of around 300,000 units produced, 50,000 is defective. Like I said before, the 300,000 units are coming from 5 sources(typically 5 independent suppliers), with unequal number of produced units and defectives.

0
#189831

Andrew Banks
Participant

GBU,

Thanks for clarifying. If you have 5 independent suppliers – I don’t think combining them is a good idea. You probably need 5 charts. Can you stratify the data? It seems like your data system is pretty capable – it may require some work by the IT dept, but I bet it is possible.

Here are some other thoughts that popped into my head…

1. Why aren’t the suppliers managing their own quality and supplying you the control charts if you want them?

2. Can you determine the time order that the parts were mfg? Control charts are time-ordered. If all you can track by is receive date, you may have trouble with creating meaningful charts.

3. Wow! 166,667 DPMO is pretty high. Are some suppliers substantially better than others?

4. Given the size of your data sets (large – which is good), if you can get the data stratified by supplier you could calculate the attribute capability and then ask any suppliers not meeting requirements for an improvement strategy.

There are a lot of very experienced folks here on the Forum (of which I am the probably the least). There may be something I am missing.

0
#189833

Participant

Just for the heck of it, why don’t you try an I-MR chart? Your counts are so high, and your percentage is so high, that you may be able to treat the data as continuous. Try plotting each week’s percentage in an I-mR chart and see what happens.

0
#189834

J
Participant

Delenn,
Appreciate all the input. Looks like I am leaning towards creating an IMR chart for my future presentation.

0
#189836

J
Participant

BBinNC,
You’re right, the data system is pretty capable but requires huge amount of time. These 5 individual suppliers are managing their own quality. They do control charts on their own, they do their own analysis on the performance of their production lines. Being the customer, who is assigned to do a management-level presentation of the suppliers’ performance, I am looking for a quick but statistically-valid style of doing the presentation.
Unfortunately, 16% failure rate is acceptable in the industry where I’m coming from, to your point, yes! some suppliers are substantially better than others(a huge variance in performance. One supplier is 4%, other 25%, 15%, etc.
Thank you for the learning experience. Expect more questions coming from me in the future.

GBU

0
Viewing 14 posts - 1 through 14 (of 14 total)

You must be logged in to reply to this topic.