Hi, All,
I have one question about "SmepMontreal08rPruzek.pdf".
For the 2nd graph "2 x 2 factorial design".
Is the disign like


A 


Yes 
No 
B 
Yes 
1 
3 
NO 
2 
4 
So we can have the contrast as
$contrasts (standardized)
A B A*B
[1,] 1/2 1/2 1/2
[2,] 1/2 1/2 1/2
[3,] 1/2 1/2 1/2
[4,] 1/2 1/2 1/2
If I am right, can any one explain the result (also calculation) as
$means.pos.neg.coeff
neg pos diff stEftSze
A 10.58 10.0 0.54 0.18
B 10.38 10.2 0.14 0.05
A*B 8.88 11.7 2.86 0.94
ANSWER:
Please go to the pdfs I gave you 2 weeks ago on your flash drives. You will see there
is a document I wrote w/ the approximate title CSDAtalk08... In it you will find more information
about the same data (possibly with one value changed, so about the same). Use the contrast coefficient weights above to find the averages of the FOUR MEANS in the 2 x 2 layout, as per the particular contrasts. You should find that the neg and pos columns shown here contain the means of the two negative coefficient means, same for positive, for each column (as rows A, B and A*B). Use the granoova function itself with your own (small?) data set (say n =3 in each group, w/ four groups, inputing as con the three columns here. Then compute the means, as per my suggestion here, and then see what you get. You should be able to reproduce the means as printed when you run the function.
I might generate the initial vector using, say, sample(2:7,12, repl=T) to keep it simple. Let me know what you get. BP
ps. The talk was very well received, and many people said the really liked the graphics.
See third panel above
Thank you. (Yi Sun)
Comments (8)
Yi Sun said
at 8:01 am on Sep 26, 2008
Thank you. (Yi Sun)
harryxkn@... said
at 8:52 pm on Oct 1, 2008
I have got a maybe very silly question, please bear with me to read it through, thanks!
We have seen a lot of papers using PSA to infer causal effects, I am wondering if we could use the inferred causal effects to help choose treatment for new observations.
Here, I want to use the following graph from passa's ppt presentation (CharacterEducationPassa07ppt, page 34) as an example. The two outcome lines for conventional education and character education cross at propensity score around 0.5.
If we simply look at the graph and take the two lines as treatment 1 and treatment 2
(rather than the original meaning of this graph by Passa), we would find treatment 1 better than treatment 2 when PS score is fewer than around 0.5, and treatment 2 better than treatment 1 when PS score is more than around 0.5. Now, if we have a new observation(N) which has got all the covariate information but N is not sure about which treatment (1 or 2) to choose to have a better outcome, can we just simply use the same model to caculate the PS score for N, if the PS score for N is smaller than 0.5, we then suggest N to choose treatment 1, otherwise treatment 2.
If this is not OK, why?
If it is OK, then we have to be sure the model we used to get the PS score is very good and reliable, and as comprehensive as possible. In this case, if the model to get PS score is reliable, it is better if N has a PS score closer to 0 or 1, so that we could be sure about selecting treatment 1 or 2 for N, otherwise if N has a PS score close to 0.5, it is hard to say which treatment to suggest for N.
bob pruzek said
at 4:42 am on Oct 2, 2008
Harry, This is an interesting issue, one that deserves discussion. Let's do that next week. It is not a silly question, but disentangling the covariate score or category combinations could become so complicated that doing as you suggest might
be so difficult as to be impractical. But not always. And that's where it becomes interesting. Lot's to say on this. Soon. BP
And Yi,
Your premise about not being able (?) to include (estimated) propensity scores in regression models is not clear to me.
Where did this idea come from?
In fact, PScrores are often used in regression applications, and it can make sense. But there can be problems too, and they need to be spelled out. Let's talk about this soon too.
Keep those comments coming!
BP
Yi Sun said
at 10:46 am on Oct 2, 2008
Hi, Harry, I do believe that we can use the suggested casual relationship between the variates and outcome of new observation (patient) to refer him/her to the better treatment. And it's the object of why we compare different treatments. But as Dr. Pruzek said, the complexity of real world in disentangling covariates could make this idea impractical. And hidden (or missed) covariate could totally changed the expectation we want. However, the more we understand the "casual effects", the more safer we can provide the suggest the patients better treatment. Am I right, Dr. Pruzke, correct me if I make mistakes. Thanks.
Yi
bob pruzek said
at 4:19 pm on Oct 2, 2008
Yi, You are correct in all that you say here. Note that there is a function called parallels (maybe not w/ 's'); use help.search
to find details. It can be invaluable for searches that reveal the PROFILES of the individuals on selected (co)variates, say, within narrow bands of PS's. Try it. Get familiar with it. And don't forget you can retrieve the individual id info for, say, the PS vector using the [ ] system. Get familiar w/ that too. Look at the code inside granova.1w that illustrates different ways this can work. But TRY IT, several ways, as that is the only way to assure a working understanding. The better you prepare for a discussion, with examples (!), next week the more we can learn about a key way to use PSA for incisive analysis of observational data.
Finally, I want to note that BOTH the PSAgraphics and the granova packages have now been updated and should be downloaded from CRAN in their latest versions. Please try to do that before next Wed too. Best, BP
Yi Sun said
at 12:55 pm on Oct 3, 2008
I checked on CRAN and found that both of packages are in version 1.2. Are these newest versions?
Yi
bob pruzek said
at 3:21 pm on Oct 4, 2008
Yes, for both. BP
Yi Sun said
at 1:33 pm on Oct 7, 2008
Thx, Porf Pruzek.
Do we have the regular class on Wednesday (Oct 8, 08) at 10:30?
Yi
You don't have permission to comment on this page.