turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- Analytics
- /
- Stat Procs
- /
- Strange output with weight option

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

08-12-2010 01:52 PM

Hi,

I'm working on patient satisfaction with first level health care in Mexico. The Data base contains a variable called 'factores de expansion final' to allow the sample to be representative of the whole population. According to what I know, this variable should be included in the weight option of any procedure used in SAS.

The data base contains a lot of potential independant variables ( variables of partial satisfaction) which are all correlated to each others. Thus I decided to run a factor analisis. As the variables are defined on lickert scale, I choose to transform them with proc prinqual and then run the factor analisis with the transformed variables.

This solution was really meaningfull and logical as long as I didn't use the weight option. With the weight option, factors and transformations had no sense, even with rotations.

I dont understand why the inclusion of this option is changing so much the correlations between variables.

Then I used the factors from the solution without weight in a binary logistic regression to modelize global satisfaction and this time, the solution with the weight option in proc logit is very very good (R2=0.9, too good?) and very very bad without it ( R2 =0.1).

I stilll dont understand why does the solution changes so much with or without the option.

Which solutin should I use? Should I never use the option weight, always use it or only use it when the results are coherent?

Please help me, this analisis is suppose to help improve the mexican health car system,

Thanks for your attention and help,

Chompy

I'm working on patient satisfaction with first level health care in Mexico. The Data base contains a variable called 'factores de expansion final' to allow the sample to be representative of the whole population. According to what I know, this variable should be included in the weight option of any procedure used in SAS.

The data base contains a lot of potential independant variables ( variables of partial satisfaction) which are all correlated to each others. Thus I decided to run a factor analisis. As the variables are defined on lickert scale, I choose to transform them with proc prinqual and then run the factor analisis with the transformed variables.

This solution was really meaningfull and logical as long as I didn't use the weight option. With the weight option, factors and transformations had no sense, even with rotations.

I dont understand why the inclusion of this option is changing so much the correlations between variables.

Then I used the factors from the solution without weight in a binary logistic regression to modelize global satisfaction and this time, the solution with the weight option in proc logit is very very good (R2=0.9, too good?) and very very bad without it ( R2 =0.1).

I stilll dont understand why does the solution changes so much with or without the option.

Which solutin should I use? Should I never use the option weight, always use it or only use it when the results are coherent?

Please help me, this analisis is suppose to help improve the mexican health car system,

Thanks for your attention and help,

Chompy

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

08-14-2010 11:14 AM

Chompy,

I'm not a statistician, but I'd think it would depend upon whether your measure is is still meaningful using weight. Realize that the weight option is adding records with precisely the same values, thus all of the actual within unit variance is lost.

Consider, for example, the equivalent data sets:

using actual values:

score group

2 0

3 0

3 0

3 0

4 0

4 0

5 0

8 0

10 0

5 1

correlation=-0.03155

using non-weighted aggregated values:

score group

5.3 0

5.0 1

correlation=-1.0

using weighted values:

score group

5.3 0

5.3 0

5.3 0

5.3 0

5.3 0

5.3 0

5.3 0

5.3 0

5.3 0

5 1

correlation=-1.0

In the case shown above I would think that only the non-aggregated, non-weighted correlation actually reflects reality. Yes, there are instances where the use of weighted, aggregate values makes sense, but that would all depend on your data.

Art

p.s. FWIW, I disagree with the previous comment that cross posting doesn't sometimes help. I never would have seen your post if it had only been posted here in the statistics' section.

I'm not a statistician, but I'd think it would depend upon whether your measure is is still meaningful using weight. Realize that the weight option is adding records with precisely the same values, thus all of the actual within unit variance is lost.

Consider, for example, the equivalent data sets:

using actual values:

score group

2 0

3 0

3 0

3 0

4 0

4 0

5 0

8 0

10 0

5 1

correlation=-0.03155

using non-weighted aggregated values:

score group

5.3 0

5.0 1

correlation=-1.0

using weighted values:

score group

5.3 0

5.3 0

5.3 0

5.3 0

5.3 0

5.3 0

5.3 0

5.3 0

5.3 0

5 1

correlation=-1.0

In the case shown above I would think that only the non-aggregated, non-weighted correlation actually reflects reality. Yes, there are instances where the use of weighted, aggregate values makes sense, but that would all depend on your data.

Art

p.s. FWIW, I disagree with the previous comment that cross posting doesn't sometimes help. I never would have seen your post if it had only been posted here in the statistics' section.

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

08-20-2010 05:49 AM

Hi Chompy.

First i must ask. What kind of data are you looking at? Is it questionares, on a scale mabye?

Before using weights, you must ask yourself: "What is, and how does my bias look?" - weighted regression is not meaningful unless you weight AGAINST your bias, that you have to argument for intuitively.

Also, what are you estimateing with? And what rotation method do you use?

We need more exact info, to be able to help you.

First i must ask. What kind of data are you looking at? Is it questionares, on a scale mabye?

Before using weights, you must ask yourself: "What is, and how does my bias look?" - weighted regression is not meaningful unless you weight AGAINST your bias, that you have to argument for intuitively.

Also, what are you estimateing with? And what rotation method do you use?

We need more exact info, to be able to help you.