# Survey Question

Six Sigma – iSixSigma › Forums › Old Forums › General › Survey Question

This topic contains 9 replies, has 9 voices, and was last updated by George Chynoweth 12 years, 7 months ago.

- AuthorPosts
- April 5, 2007 at 12:04 am #46633
Hello,

I have a survey where internal customers rate us on a scale of 1 to 5, with 5 being the best. Can I calculate a standard devaition and use confidence limits with this type of data. It is not really continuous data. If not, what is the best to analyze the results.

Thanks,0April 5, 2007 at 7:58 am #154443

reshma kalutayParticipant@reshma-kalutay**Include @reshma-kalutay in your post and this person will**

be notified via email.hi dear,

am sorry, nt really replying ur question!!! instead, am adding up to ur troubles….

i think we can calculate std dev for any sets of numbers. am i right???0April 5, 2007 at 10:25 am #154451Hey Bill

Since the specification limit has been set in your case. You can use the DMAIC method & measure,analyse,improve & control as per the std. deviation.

Cheers!!!

0April 9, 2007 at 11:21 am #154541A simple visual display can be used box plot with X axis 1 -5

0April 9, 2007 at 11:23 am #154542

karthikeyanParticipant@karthikeyan**Include @karthikeyan in your post and this person will**

be notified via email.Bill,

The data from a survey is considered to be attribute data (discrete). Attribute data are analyzed with the help of Atrribute Agreement Analysis.

The data that you are dealing with ordered (1,2,3,4,5) as compared to (red, blue white0 which is norminal.

I woud use ordinal data method in Atrribute Agreemeent Analysis.

Thanks

Karthik0April 9, 2007 at 12:03 pm #154543Sorry to disagree with my everyone but the Anova is the best way to determine the info for which you are looking.

0April 9, 2007 at 12:27 pm #154545

karthikeyanParticipant@karthikeyan**Include @karthikeyan in your post and this person will**

be notified via email.Hello,

But before that, why would someone want to determine standard deviation for a discrete (attribute) data.

If I am not wrong, standard deviation or dispersion is used to find the spread of the data.

For discrete data concordance / agreement within each other / is analogous to dipersion used for continuoous data.

But before that, one needs to understand the end goal / objective of the survey.

Thanks

Karthik0April 9, 2007 at 1:27 pm #154547Tippically when you look at survey information, if you are trying to measure or improve satisfaction you will be looking at percentage of top box or top two boxes. If looking at reducing dissatisfaction you are looking at percentage o bottom box or bottom two boxes. If you were to do a Control Chart it would be a P chart.

It is important to know the size of the universe and of course the size of your sample to be able to calculate measurement error and discern statistically meaningful improvements from just noise.

Survey practicioners usually look at error rates of 2-3% with 90% confidence level, which for large populations means over 200 surveys made0April 9, 2007 at 7:44 pm #154575I’m not the swiftest boat in the ocean, but this is what I do:

1. I do not bother with the Standard Deviation. True, it is just a formula and any set of numbers will produce Std Dev value. However, the value ONLY has the intpretation of encompasing 68% of the data if the data is normally distributed. I have no reason to suppose that survey results are normally distributed (example: you intenty to teach that 2+2= 4. At the end of the eaching you survey if you met your objective. Hopefully all of the participants would agree that you did — thus the resonses are not normally distributed).

2. I use surveys on the classes I teach. I run a Chi-squared test to compare the survey results from two different sessions. Normally I look for at leat an 80% confidence level that the survey results are not the same. If that trigger is met, then I re-examine the activities to determine what I did that increased the survey results ( If I modified the instructional material, is that the cause? Did someone make an insightful observation that helped the class? etc. Then make sure that factor is retained for the next class.). On the other end of the spectrum is the decrease in the survey response (Did I forget a topic? Dis a change, intended for the better, make the link between topics less clear? Dis the background of the participant change in an unexpected way?, etc.)

I do not have a good basis for the 80% confidence level I use, other than I can usually figure out what changed that may have created the different survey response — i.e., the 80% level was empirically determined by me and allows me to focus on a few course changes at a time.

Eugene0April 13, 2007 at 6:19 pm #154777

George ChynowethParticipant@george-chynoweth**Include @george-chynoweth in your post and this person will**

be notified via email.You must understand the nature of the data before you select an analysis. The problem here is one of scaling and trying to measure an intangible concept. (e.g., a survey scale does not yield the same type or level of data that a pressure gauge does). When testing people, as opposed to testing physical objects, you enter the field of Psychometrics, which is a different measurement realm than most six sigma practitioners are used to. Data are categorized into 4 types: nominal, ordinal, interval, and ratio. Each data type requires a specific type of analysis, or analytical family (parametric vs. non-parametric). Interval and ratio level data contain more information than nominal and ordinal, and hence, the more powerful parametric analyses are used. For this reason, folks who know what they are doing regarding survey design and scaling will try to construct a scale that yields interval level data. An interval level scale has equal distances between each of the scale points: the numerical distance between 1 and 2 is the same as between 9 and 10. An example is the Fahrenheit thermometer. A true Likert scale has equal appearing distances between each of the scale points, and a mid-point or neutral option. Psychologists have argued for decades that the true Likert scale yields interval level data and have used parametric statistics (t-Test, ANOVA, Pearson correlation, regression, etc.) to analyze the data. However, there are many Likert-type scales which do not have equal distances or a mid-point – they are very unbalanced, and clearly yield ordinal level data. Analyzing ordinal level data with parametrics is not a sound practice as it introduces measurement error – something to be avoided whether surveying for satisfaction or determining tolerance.So, the survey in question is clearly ordinal (suggesting non-parametric analyses such as Chi-Square, median tests, etc.), but can it be considered interval? I’d have to see the exact scale used in order to offer an opinion. The standard deviation calculated from ordinal level data is misleading, and depending on the scale, can provide dis-information. And here is the real caution: if decisions regarding resouce allocation and expenditure will result from the analyses and interpretation of the survey results, you better get it right. :)

0 - AuthorPosts

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