iSixSigma

Analyzing Discrete Data

Six Sigma – iSixSigma Forums General Forums General Analyzing Discrete Data

Viewing 13 posts - 1 through 13 (of 13 total)
  • Author
    Posts
  • #54493

    Liam Collinge
    Participant

    I have a data set for Turnaround Times. It is measured in days so I am having problems analysing it for normality. Is there a way dealing with this data that I should be using? Any help would be appreciated as I’m new to the LSS world.

    Cheers,

    Liam.

    0
    #195610

    Bill McNeese
    Member

    It seems to me that this got complicated very fast. You have several customers with different delivery requirements. Let’s assume you are just looking at whether the shipment was on-time for a customer. I would start by making sure that you have a good operational definition of that requirement. Say customer A wants the product this Friday. You won’t get it done until Monday. So, you call the customer and ask if Monday is OK. He says yes. Did you still meet the delivery requirement? Many companies would say yes – after all, it makes there on-time delivery look better. Make sure you have good operational definitions in place for your delivery requirements or your analysis will not be valid.

    From there I would keep it simple. If you are looking at on-time delivery, I would simply plot the % of on-time deliveries (per day, per week or month depending on the number of shipments)using an individuals control chart. Your capability is whatever your average % on-time is once you have enough data (only need 4 to 5 points to start the chart; use historical data if operational definitions were OK) and is stable. Then use Pareto diagrams to look at reasons for late deliveries, by customer etc. That should help guide your problem solving efforts.

    The different delivery requirements will most likely have different means/medians. I would not worry about normality. You may well not have enough points with each different requirement to test that and the objective is to improve performance versus the delivery requirements.

    0
    #195635

    Charles Kimball
    Guest

    I agree with Bill, but would like to add one additional comment. If you have a order that is available to ship on an agreed date and the customer does not release it or otherwise pushes the ship date out, it is not late. Some systems track the actual ship date compared to the promise ship date and do not pick up customer extensions of release dates and calculations count as “late” when in fact it is not. If this were to occur frequently, your data would be skewed unfavorably.

    0
    #195638

    Ken Feldman
    Participant

    Why doesn’t somebody tell the original poster that, as a rule, time data will NOT be normally distributed? It is expected to be skewed because there is no upper bound to the time but certainly a lower bound of zero unless you have a time machine. Non parametrics and non normal capability analysis are both easily done. Normality assumptions in many cases are robust. Transforming data is the LAST thing you do!! Understand your data first and then go from there.

    0
    #195656

    Neo Bourne
    Guest

    Darth, It will be incorrect to generalize that all time related data will be non-normal. Why should the lower end always be close to zero. If the centering of the data is far away from zero, you can always have a perfect normal distribution even in case of time relate data.

    0
    #195657

    Anonymous
    Guest

    @Darth I’ve often wondered about this. It’s my understanding events follow a Poisson distribution – a fact that forms the basis of Spatial Yield Analysis within the semiconductor industry. One example of this approach in Ed Bluestein was to compare the yield of various families of devices – we found very similar random defect, but considerable variation in the ‘functional entitlement’ from one reticle-set to another. One cause of ‘systematic defects’ was edge dislocations caused from pushing and pulling wafers into furnaces too quickly.

    0
    #195662

    Ken Feldman
    Participant

    I supposed anything can happen but in the case of the original poster who was “having problems analysing it for normality” I assume that this was not the case. Please present “real” where the actual distribution was a “perfect normal distribution”. By that I assume you mean a p-value of 1.0 on the normal distribution plot. Come on, no such thing as a “perfect normal distribution”.
    @andy-u WTF are you talking about???? Love the babble though. Made for a good laugh, thanks. Now back to the Don Julio.

    0
    #195663

    MBBinWI
    Participant

    @Darth – I’m sure that Andy has some useful points, but for this question, the issue lies in how close the distribution is to the zero point. As Neo identifies, if the actual measured value of the time variable is displaced away from zero, then normality can be a valid assumption.

    0
    #195667

    Ken Feldman
    Participant

    MBBinWI, sorry I didn’t recognize your continued contribution to isicksigma along with Andy. Since you and the guys were going down memory lane the other night, here is one for the trip. This exact issue was my first posting back in the good old days and my first interaction with Stan. We got into it pretty heavily over whether there was any instance where time might be normally distributed. I took your position from a theoretical standpoint but frankly haven’t seen a real application. As I stated, it is not the case for the original poster since he was ready to transform the data…..that will start another uproar on here.

    0
    #195671

    Anonymous
    Guest

    @Darth My posts aren’t for everyone. This particular post was for those working in the semiconductor industry. You say you haven’t seen a real application, yet you criticize someone when they describe a real application. As I alluded earlier, the method I referred to enabled MOS 8 to break yields records month on month. Perhaps if we attracted a few more people into the forum, it would be more successful and a more enjoyable community!

    Since you’ve referred to ‘isicksigma’ I can only assume you’re losing money; you’ve had some poor investments .. bad luck!

    0
    #195673

    Ken Feldman
    Participant

    @andy-u If you are referring to the post by Neo, he said “what if” and did not provide an actual application. He also mentioned a “perfect normal distribution” so his credibility is a bit lame.

    0
    #195676

    Mike Carnell
    Participant

    @liamocol I am not a big fan of getting into these long drawn out stats strings but going back to the original post as whats his name suggested – I think it was Darth. What does that mean “It is measured in days so I am having problems analysing it for normality.” Data is data. It isn’t more or less difficult to run. You run the test on your data and it either is or it isn’t normally distributed. What you are having difficulty with is that it isn’t turning out the way you want it to. That is mindset that will take a lot of integrity out of any analysis you do.

    There are tools for analyzing non normal data so if it isn’t normal put a little effort into this and figure it out. Just because some half *ssed instructor taught you Six Sigma and wasn’t smart enough to understand that you might run into non normal data doesn’t mean you need to run around as ignorant as they are.

    Just my opinion

    0
    #195920

    P L Narasimhan
    Guest

    My opinion is this.
    Normality can be checked with in a group and not for hetrogeneous groups with different set of operating condition. But we can analyze fulfilment of the delivery time.
    Suppose we get a data ie delivery time for 100% satisfaction and 0%. We can work out the percentage in between. When we deliver we wiil find out what percentage of satisfaction we have given. Analyze this for normality and find your mean satisfaction level. We can use Pareto chart for the causes of failure.
    narasimhan

    0
Viewing 13 posts - 1 through 13 (of 13 total)

You must be logged in to reply to this topic.