Why do we care about df?
Six Sigma – iSixSigma › Forums › Old Forums › General › Why do we care about df?
 This topic has 9 replies, 7 voices, and was last updated 18 years, 6 months ago by abasu.

AuthorPosts

May 19, 2004 at 4:55 am #35582
Dev ShresthaParticipant@DevShrestha Include @DevShrestha in your post and this person will
be notified via email.Some one still need to explain, whay it is important to consider so called degree of freedom? Say in ANOVA?
0May 19, 2004 at 6:43 am #100462If you were thristy and you wanted something to quench your thirst, and you had water, juice or beer in ur fridge, the outcome of your behaviour would be quite different than if all you had was some water…. hope you’r with me till now.
Whats your degree of freedom in the latter scenario…none you drink water. But in the former scenario.. you v got more options, your df is higher.
So??
Well same is the outcome of all this other things like MSS,SS, etc. that you see in say Anova, when this df thing changes.
Lets take a simple case, statistics is mostly about prediction, if it told you that in (a+b+c)/2=d we had an idea about a and b and c, that tells you all u need about d, but if if were to tell you we had no idea what say c was , now you have got like d misbehaving all over again, and so on and so forth, the greater the degree of freedom of things, the different are the implications of the outcome… that shall do for now… if you dont bother about it, well, the results wont bother to help you either. I suggest u read Statistics for Managers by Levin et al.
All the best and do take care of the df.
GGBB
0May 19, 2004 at 7:03 am #100464
Hemant GhamParticipant@HemantGham Include @HemantGham in your post and this person will
be notified via email.Why Degrees of Freedom or n1?
The use of DF or n1 can be considered as a mathematical device used for the purpose of deriving an unbiased estimator of the population variance.
n1 is referred to as degrees of freedom.
When the total sumsofsquared deviations is given and the pairwise deviation contrasts are made for n observations, the last contrast is fixed; hence, there are n1 degrees of freedom from which to accumulate the total.
More specifically, degrees of freedom can be defined as (n1) independent contrasts out of n observations. For example, in a sample with n = 5, measurements X1, X2, X3, X4, and X5 are made. The additional contrast, X1X5, is not independent because its value is known from
(X1X2) + (X2X3) + (X3X4) + (X4X5) = (X1X5)
Therefore, for a sample of n = 5, there are four (n1) independent contrasts of degrees of freedom. In this instance, all but one of the contrasts are free to vary in magnitude, given that the total is fixed. But when n is large, the degree of bias is small; therefore, there is little need for such a corrective device.
Example: An experiment was run on two different days, Monday and Thursday. The company involved three operators O1, O2, O3. The Wear testing process was run at three settings A, B, C. Each operator made two determinations of wear at each day and setting (Mixed Model ANOVA).
So we have following DF.
Source DF
Day 1 (21) Here we have only one independent contrast (MonThu)
Operator 2 (31) Here we have two independent contrasts (O1O2) + (O2O3)
Setting 2 (31) Two independent contrasts (AB) + (BC)
0May 19, 2004 at 7:11 am #100465
Hemant GhamParticipant@HemantGham Include @HemantGham in your post and this person will
be notified via email.Please read the check sign in the previous post as ‘‘ (minus) since the characters are not converted in the post correctly.
For e.g. n1 = n – 1
Also read n as ‘ n ‘ and
(X1X2) + (X2X3) + (X3X4) + (X4X5) = (X1X5) as (X1X2) + (X2X3) + (X3X4) + (X4X5) = (X1X5)
Regards,
Hemant0May 19, 2004 at 1:14 pm #100478
Ken FeldmanParticipant@Darth Include @Darth in your post and this person will
be notified via email.The two responses so far have described what df is. I believe the original poster asked why does he/she care about it. What actions are taken one way or the other based on the df?
0May 19, 2004 at 3:43 pm #100486
Hemant GhamParticipant@HemantGham Include @HemantGham in your post and this person will
be notified via email.Darth,
What and Why both are answered.
Regards,
Hemant0May 19, 2004 at 4:10 pm #100490In many uses, such as ANOVA, the degrees of freedom are essentially an “adjusted” measure of sample size for the sample used to estimate the respective variance(s).
If you divide the squared deviations from the mean by n instead of (n1), you will get a biased estimate of the population variance. That means that as the sample size gets larger and larger, the estimate will tend toward SQRT(n1)/n)*sigma, rather than sigma. By dividing by (n1) instead of just n, we create an unbiased estimate of the population variance – one that will tend toward the true value of sigma.
In the ANOVA, the Fstatistic is simply a ratio of variances, so each of these variances needs to be calculated using an appropriate degrees of freedom – so they are not biased.
Keep in mind that the idea behind most statistical tests is that we find a test statistic whose distribution can be predicted under the null hypothesis assumption. If the test statistic calculated from your data shows sufficient evidence of not coming from the hypothesized distribution, then we may decide that the null hypothesis is not true.
In the case of ANOVA, the test statistic is the ratio of two varainces. The degrees of freedom are a necessary bit of information when trying to predict the distribution of the Fstatistic under the null hypothesis assumption that there are no significant model terms (e.g. no differences between the means in a oneway ANOVA).
The same thing can be said about predition of the tstatistic under the null hypothesis for a ttest.0May 19, 2004 at 5:33 pm #100494You have clearly explained the “what”, but not the “who cares”. You state that it is used for the “purpose of deriving an unbiased estimator of the population variance”, but the questions “so what?” and “who cares?” still stand. Please explain further if you could…
Thanks in advance,
SP0May 20, 2004 at 2:43 pm #100535If the estimation of sigma is biased, then the calculated Fstatistic will may not follow the theorized F distribution, even if the null hypothesis is true. The result could be a much greater tendency to reject the null hypothesis than expected. You’ll see differences where differences don’t exist.
0May 20, 2004 at 3:48 pm #100541Think of df as statistical currency. You have to spend a df to get information such as 1 df to calculate sample mean and so on.
One instance you care about df is while analyzing single replicate DOEs. Minitab (or other software) will not generate pvalues for the different terms. So one trades df associated with higher order interations to get an estimate of experimental error and P values.0 
AuthorPosts
The forum ‘General’ is closed to new topics and replies.