Author Topic: Histogram Interpretations  (Read 48 times)

admin

  • Administrator
  • Full Member
  • *****
  • Posts: 124
Histogram Interpretations
« on: November 04, 2015, 11:48:16 am »
This is a question that gets posted from may of my students, clients and industry professionals.

I want to use Histogram to arrive on standard time of a given process. This will be easy if the data is normally distributed (symmetric). I have some cases where it is flat, bi-modal, skewed etc.

Can someone help me in getting guideline in this regard. Pl. note that this standard is going to be base for me to arrive upon headcount required to perform a particular activity.

Here is my response..
It is quite likely in situations to encounter non-normal histograms. You may want to consider the following options.

a. Check if there are "special/assignable/sporadic" causes that result in extremely high or extremely low values. Verify if they can be traced to "Special" cause occurances, if so, waitlist / keep them out and with the rest of the data check for normality.

b. Stratify - Check for diversity of streams / sources from which data may be generated for e.g,. If the data can be segregated source wise and viewed again, check if it is showing normality. In such a case, you may have to evolve different standards (if applicable/agreeable) (or) establish a target what is "business / customer wise" acceptable and improve the "non-compliant" process stream to catch up.

c. If the data is of long term nature, check if the data is mixed up of any major change-events that has impacted your result or process measure. In such a case, it is a good idea to separate the data pre and post change event so that the baseline status is representative of "true current status" (instead of mix-up of data arising out of numerous change events over a period of time). Run chart that chronologically plot the data and control charts that show if the process has "shifted" to a new control zone can help you in doing this

d. Use control charts and Box-whisker plots to help in identifying sporadic and stratification approach to take an informed judgement on the actual performance

e. If the data is short term (stable process and collected considering process "in control" then normality should be more likely. If it is long term data that is the result of numerous process influences, then the above approaches can be tried out to make better sense of the process behaviour.

At times when looking visually, the distribution may not look normal, but if you do a probability plot and p-value computations (via Anderson darling analytical tests), you may be able to arrive at a right conclusion.

A more sensible strategy to arrive at "realistic" head count (even if the distribution is not-normal), is to use "simulation" techniques that factor in "demand flow", "task time" (and many other factors for e.g., availability, skill level, productivity, defect rate etc. based on current data) to obtain a close to "real world" estimate. It will also provide the opportunity to "revise" and re-compute resourcing requirements in future, given changes in operational conditions over a period of time.

Hope this helps!!

Share on Facebook Share on Twitter