Re: Doubt in Regression

Sometimes the simplest possible solution works fine, sometimes it
don't.
Here are a couple of very simple solutions:

First problem:
Analyse the scale as nominal, not ordinal.
It probably will be necessary to combine levels, for example:
Category 1 = 1,2,3 - Frustrating
Category 2 = 4 - Neutral
Category 3 = 5,6,7 = Good
Category 4 = Unsure/na

You define a dichotomous (0/1) variable for each category, then select
one of them as the reference category - the Neutral group would be a
natural choice for this.
Then calculate a regression model with the three remaining category
variables as independent factors.
This model will give you a total test if there there are important
differences between the categories you have defined (H0: All category
effects are equal). And also separate tests of significant difference
between each of the remaining categories and the reference category.

Second problem:
Calculate three models:
Dependent <- Group 1
Dependent <- Group 2
Dependent <- Group 3
and note the over-all measures of modell fit.

Then calculate the total model
Dependent <- Group 1 + Group 2 + Group 3
and note the over-all measure of fit.

For example if you use linear regression models, you may use the model
r-square.
And if that gave you 0.37 0.43 0.39 for the three separate models and
0.68 for the total model, you would know that not one of the groups
explains all that is possible to explain - you should combine two of
them or perhaps all three to find the optimal model.
But if the total model gave r-square = 0.45, you could assume that the
group with best explanatory power is sufficient, and that the other two
groups may be discarded in the analysis without loss of predictive
power.

You have yourself to decide if those simple solutions are
excellent/interesting/meaningless. It depends on the type of questions
you want to answer from your data.

Valter Sundh
Dep of Community Medicine and Public Health
Sahlgrenska academy
Göteborg University
Sweden





Lawrence wrotew:

Warm Greeting to all!

I have some doubts in my Regression model. I explained below. Please
give me a conclusion on my problem and show me the right way to get the
solution.

Thanks,

Lawrence.

First Problem:

1-Extremely Frustrating
2-Very Frustrating
3-Somewhat Frustrating
4-Neutral
5-Somewhat Good
6-Very Good
7-Extremely Good
8-Unsure/Not applicable

This is my independent variable scale structure. I want to see the
contribution of this kind of independent variables with dependent
variable(5 point scale -Satisfaction). Before running multiple linear
regression analysis I created new set of variables for independent
variables without Unsure/Not applicable option. After done this action
I am getting 40% of system missing in my data files. But I don't want
to loss this much information from my analysis and I want use all the
cases. I know some of the options like Series Mean, Replacing with near
mean or median and interploation methods are available for replacing
system missing. Can I use which method for my problem and it can be
effective? explain.....

Second Problem:
All my independent variables are significantly
correlated together with 99% significance level. What can I do? This
model R-Square value is .35 and I am used Backward method. I lossed
some variables when I seen last model. I classified my independent
variable into three groups. I want to see how the first group
contribute the dependent variable and in the same way for remaining 2
groups. And also how this 15 attributes contribute dependent variable.
Here is the flow diagram.


|<------Group 1<------- 5 attributes
|
|
Dependent<----- <------Group 2<------- 5 attributes
|
|
|<------Group 3<------- 5 attributes

.

Relevant Pages

  • Re: JSH: The Math
    ... > Not surprisingly, since mathematics IS mathematics, my position can be ... I have picked m and f, a and b are dependent on them, though ... Therefore by the Harris Independence Principle (HIP), ... > The independent variables are just independent. ...
    (sci.math)
  • survey analysis -- ordinal scales
    ... My friend asked for some statistical help on a finance department ... 14 scale questions for management ... independent variables, and it will turn out to be a linear ... be put into a multivariate logistic regression. ...
    (sci.stat.math)
  • relationship investigation with large dataset
    ... I want to investigate the relationship between 6 independent variables ... variables because I cant argue that they are of interval scale.). ... The p-values of 30 multivariate ANOVA tests are used in a T-test to ... investigate wether the p-values of ANOVA's are less than 0.05 or not. ...
    (sci.math)
  • Re: Univariate Analysis
    ... regression" should be used when we have more than one dependent ... variable and one or more independent variables (in the latter case, ... should probably speak of "multivariate multiple regression"). ... multiple independent variables, then we may want to call it ...
    (sci.stat.math)
  • Re: multiple regression model
    ... > I have multiple regression model consist of one dependent and 7 ... determine the thresholds on t. ... > between the dependant and an independent variables. ...
    (sci.stat.consult)