Notes 9/30/2004
More on trend analysis, contrasts
EPSY
6301
30 Sep 2004
If you use existing data, your grade will be weighted accordingly
- you will get more credit if you collect data yourself
A more sophisticated methodology will be expected for those who start with existing data
We have a couple more weeks to locate a data set on our area of interest
- then all my critiques for this class can revolve around that area of interest / topic
we can also simulate data, but that is no fun because nothing gets rejected
to do one way ANOVA you can go three ways
-
- Comparing Means
- GLM
look at this list of SASS syntax
- SASS is so simple it is scary
- SASS does not have point and click like SPSS
- Data step is first, entering it and specifying labels
- Dr O likes to use titles to keep different tasks separate
- PROC step is where you request different statistical procedures
- since 1999 I have been trying to switch to SPSS because many people do not like SASS
when we are testing for normality or the homogeneity of variance, we do NOT want to reject the null hypothesis
Most research tier 1 institutions prefer SASS to SPSS
SASS is NOT forgiving: 1 typo = it does not work
I am just showing you now how it was done in the past
- SASS has a semicolon as a delimiter
- miss one semicolon, the entire program doesn't work
90% of computer output is ancillary, only about 10% is really needed
Dr O does not know how to do this in SPSS, he owes us this
My thought: everything tonight up to this point (7:40 pm) has been a waste of time, none of us have access to SASS and are going to use it....
Orthagonal contrasts
- in contrast analysis you are really dissecting questions in ways that give you a lot of power
you are going to see the beauty in this because SPSS cannot show you this
omnibus F considers everything in the experiment: all the comparisons that you can think of
we can increase our power by many degrees when we do these
when you add the significance of all the pairwise comparisons, you will get more than the significance indicated by the omnibus F
no one cares if you know Scheffe or Tukey
- these types of pairwise comparisons are what are most powerful and most interesting
- we are doing this by futher partitioning what we are looking at
- the varaiability in what we are looking at is limited to fewer factors, this gives us more power
Dr O has seen dissertation reviews where the person could not identify the IV and DV, and levels
big question/hypothesis: is there a trend in terms of the effect the treatment had on the outcome variable?
monotonically increasing = increasing consistently
can do linear tests or quadratic tests
Different comparisons will yield different results
- don't just go with pairwise comparisons
"they don't print 'significance' on your diploma" - Pat
can you do trend analysis on a variable like ethnicity?
- no, it cannot be categorical, it must be continuous or semi-continuous
- any time your levels have to do with "kinds" (categories) - don't use trend analysis
my question: how do you select the comparisons that will test for these different types of patterns?
trend analysis deals with orthagonal contrasts
- when you have 4 levels, you can only do 3
there is a table that tells you the coefficients to use for linearity, quadratic, and cubic
- these numbers are weights attached to the group means
- examples are 0, 1, -3, etc
- when you add up these weights they must equal zero for you to have orthagonality
2nd step for checking for orthagonality
- must have pairs to test for this
- multiply the coefficients for each level, then sum them: this must also equal zero!
uncorrelated contrasts = independent contrasts
Is this data showing a linear pattern?
we will not get into how we get these coefficients in these tables
- this is the last night we are going to talk about this (trend analysis) unless someone asks Dr O about this
if your contrasts are orthagonal, then the sum of your found signficance should never exceed the omnibus F
30 Sep 2004
If you use existing data, your grade will be weighted accordingly
- you will get more credit if you collect data yourself
A more sophisticated methodology will be expected for those who start with existing data
We have a couple more weeks to locate a data set on our area of interest
- then all my critiques for this class can revolve around that area of interest / topic
we can also simulate data, but that is no fun because nothing gets rejected
to do one way ANOVA you can go three ways
-
- Comparing Means
- GLM
look at this list of SASS syntax
- SASS is so simple it is scary
- SASS does not have point and click like SPSS
- Data step is first, entering it and specifying labels
- Dr O likes to use titles to keep different tasks separate
- PROC step is where you request different statistical procedures
- since 1999 I have been trying to switch to SPSS because many people do not like SASS
when we are testing for normality or the homogeneity of variance, we do NOT want to reject the null hypothesis
Most research tier 1 institutions prefer SASS to SPSS
SASS is NOT forgiving: 1 typo = it does not work
I am just showing you now how it was done in the past
- SASS has a semicolon as a delimiter
- miss one semicolon, the entire program doesn't work
90% of computer output is ancillary, only about 10% is really needed
Dr O does not know how to do this in SPSS, he owes us this
My thought: everything tonight up to this point (7:40 pm) has been a waste of time, none of us have access to SASS and are going to use it....
Orthagonal contrasts
- in contrast analysis you are really dissecting questions in ways that give you a lot of power
you are going to see the beauty in this because SPSS cannot show you this
omnibus F considers everything in the experiment: all the comparisons that you can think of
we can increase our power by many degrees when we do these
when you add the significance of all the pairwise comparisons, you will get more than the significance indicated by the omnibus F
no one cares if you know Scheffe or Tukey
- these types of pairwise comparisons are what are most powerful and most interesting
- we are doing this by futher partitioning what we are looking at
- the varaiability in what we are looking at is limited to fewer factors, this gives us more power
Dr O has seen dissertation reviews where the person could not identify the IV and DV, and levels
big question/hypothesis: is there a trend in terms of the effect the treatment had on the outcome variable?
monotonically increasing = increasing consistently
can do linear tests or quadratic tests
Different comparisons will yield different results
- don't just go with pairwise comparisons
"they don't print 'significance' on your diploma" - Pat
can you do trend analysis on a variable like ethnicity?
- no, it cannot be categorical, it must be continuous or semi-continuous
- any time your levels have to do with "kinds" (categories) - don't use trend analysis
my question: how do you select the comparisons that will test for these different types of patterns?
trend analysis deals with orthagonal contrasts
- when you have 4 levels, you can only do 3
there is a table that tells you the coefficients to use for linearity, quadratic, and cubic
- these numbers are weights attached to the group means
- examples are 0, 1, -3, etc
- when you add up these weights they must equal zero for you to have orthagonality
2nd step for checking for orthagonality
- must have pairs to test for this
- multiply the coefficients for each level, then sum them: this must also equal zero!
uncorrelated contrasts = independent contrasts
Is this data showing a linear pattern?
we will not get into how we get these coefficients in these tables
- this is the last night we are going to talk about this (trend analysis) unless someone asks Dr O about this
if your contrasts are orthagonal, then the sum of your found signficance should never exceed the omnibus F
Posted: Thu - September 30, 2004 at 08:45 PM
Quick Links
Calendar
Sun | Mon | Tue | Wed | Thu | Fri | Sat |
Categories
Archives
XML/RSS Feed
Statistics
Total entries in this blog:
Total entries in this category: 4
Published On: Sep 30, 2004 08:45 PM
Total entries in this category: 4
Published On: Sep 30, 2004 08:45 PM
