For preparation I'm afraid the Blog is all there is. I will comment on questions. Kyla informs me that many of you will still meet at U3 today. I will be available to answer questions at that time.
Mike G
UPdate.
Here is the memo for the first two catchup questions. I will add as I go and answer items on the blog. Check back as I will update this file each time I add a new memo. Will note this in the blog post.
New revised file here
Hi Mike! Is there an e-mail address that we could perhaps contact you on?
ReplyDeletemike@dmsa.co.za
ReplyDeleteHi Mike!
ReplyDeleteWhat do you do if there is more than one covariate in an ANCOVA? Do you do the same as if there is one covariate but say after controlling for the effectsof all of the covariates? We havn't done it in class or in the tuts so I'm not sure what to do?
Thanks
Hi Mike!
ReplyDeleteHow do you interpret the sample regression equation for significant results in multiple regression? What does > 0 mean? They all look greater than 0 in the class example, is it meant to be from 1 and up?
hi mike
ReplyDeletehope you feeling better.
for the post hoc analysis, for two way anova, is it necessary to comment on every pair. pls refer to 2009 question 1. there seem to be approx 30 pairs. is this correct?
thank you
Saleha and Fatima
Some Reponses.
ReplyDeleteFOR ANCOVA. From the 2007 Paper
DV : Contract Diff - The degree to which transactional elements predominate over relational elements of the psychological contract.
IV: Tenure
Levels
0-6
7-18
19-30
31-42
43+
Covariates
Supervisor Support
Organisation Support
Contract Violation
Hence the H0: is
No Tenure effect on Contract Diff after accounting for Supervisor Support, Organisation Support and Contract Violation.
etc
I.e. your solution is correct. You only look at the p-value for Tenure here. Th pvalues for the covariates are ignored.
SAMPLE REGRESSION COEFFICIENT
Simply >0 implies a postive relationship.
<0 implies a negative relationship.
The strength of the relationship is examined by the standardised regression coefficient.
POST HOCS for ANOVA.
ReplyDeleteIn an interaction you need only interpret a subset. (on those of either across the rows or down the columns)Nevertheless it may well be a substantial subset. Try to find a way to summarize the description. I will send an example later
Thanks Mike! Get well soon.
ReplyDeleteHi Mike
ReplyDeleteHope you get well soon.
There is some confusion in calculating partial eta squared η2 ( eta-squared ) - in my tutorial notes partial eta squared is calculated the same way as the Global effect size - but in my Lecture notes I have the following formula:
1. Partial Eta Squared = SS(treatment)/(SSerror + SStreatment).
2. Global Eta Squared =SStreatment/SScorrected total .
My gut goes with the lecture notes - please advise if this is the correct formula?
Hi Mike,
ReplyDeleteIm a bit confused on what steps to follow in regression analysis where a condition index is greater than 30 (multicollinearity is present). I think this is the case in Question 1 of the 2010 paper.
Get well soon :)
Thanks
Hi Mike, sorry you are not well and hope you get better soon, or otherwise enjoy the time off.
ReplyDeleteRegarding 2009 paper, Q3, I'm not sure if my interpretations are correct.
The eta2 for the LCL, mean and UCL are 0.3, 0.58 and 0.85. Is the correct interpretation then that at best the computer based assessment score is 85% better than the paper based score, and at worst 30% better, after controlling for the covariates?
and Qc, is the correct answer Yes it is a good measure as respondents are allowed to choose which test they take and the two versions are equivalent, and that the results suggest a correlation between computer literacy and non-verbal reasoning?
Latly in regression, is the sample regression coefficient b0? If b0 is negative does that mean that the predictors's are negatively related to the DV?
Hope this all makes sense.
Thanks
. Partial Eta Squared = SS(treatment)/(SSerror + SStreatment).
ReplyDeleteis the correct form.
Im a bit confused on what steps to follow in regression analysis where a condition index is greater than 30 (multicollinearity is present). I think this is the case in Question 1 of the 2010 pa.
ReplyDeleteper.
See outline I've just posted.
Essentially
1.you note the multicollinearity
2. Make relvant suggestions on how to improve it based on what you know about the problem.
These could include.
a) Removing of similar or duplicate variables.
b) creation or use of an index.
c) stepwise regression of some kind.
Hi Mike
ReplyDelete2 quick questions
1) STANDARDISED ESTIMATES:
Is there any criteria to know whether it is
weak/strong?
2) 2-WAY ANOVA/ANCOVA:
If the Ivs are significant and the
interaction is significant, do you still do
direction and effect sizes for the main
effects?
Thanks!
In response to ...
ReplyDeleteRegarding 2009 paper, Q3, I'm not sure if my interpretations are correct.
The eta2 for the LCL, mean and UCL are 0.3, 0.58 and 0.85. Is the correct interpretation then that at best the computer based assessment score is 85% better than the paper based score, and at worst 30% better, after controlling for the covariates?
Mikes Response
Cohen's D is not a Percentage. It means that (in this case because ideally we would like them to be the same) at best the difference is small to moderate and at worst he difference is large.
...and Qc,
is the correct answer Yes it is a good measure as respondents are allowed to choose which test they take and the two versions are equivalent, and that the results suggest a correlation between computer literacy and non-verbal reasoning?
Mikes Response
This would be one of several possible correct answers. Importantly your reasoning suggests a valid argument.
An equally valid argument might say that a better design (for example a within subjects design) would be needed to assess equivalence so no.
You could also argue that a test that correlates with computer knowledge is problematic from a validity perspective.
In smple terms I was looking for a sensible argument linking the data to common sense psychometrics.
..Lastly in regression, is the sample regression coefficient b0? If b0 is negative does that mean that the predictors's are negatively related to the DV?
Mikes response
I'm a bit confused by the question. A typical regression equation looks like
Dv = B0 + B1 IV1 + B2 IV2 + ...+e
b0 is thus in general the sample intercept which we don't interpret at all in this course.
or b1, b2 etc.
If a given beta value is less than zero then yes we understand this to indicate a negative relationship so that if b1<0 this implies that large values of IV1 are associated with small values of Y.
to
ReplyDelete...1) STANDARDISED ESTIMATES:
Is there any criteria to know whether it is
weak/strong?
a standardise regression coefficient is treated like a correlation.
2) 2-WAY ANOVA/ANCOVA:
If the Ivs are significant and the
interaction is significant, do you still do
direction and effect sizes for the main
effects?
No if the interaction is significant you only interpret the interaction
From an email
ReplyDeleteWhen do we use a covariance matrix?
It is possible to use a covariance matrix as the basis for factor analysis when two conditions are present.
a) All the measured items must be on the same scale.
b) There must be some link between the importances of the variable to the factor analysis and the size of its variance. In class I suggested that item difficulty/discrimination provides an example of where this might be true.
In practice start by identifying if items are on the same scale. If not then correlation matrix. If so then suggest that covariance may be possible and ry to see if any reason for the second criteria suggests itself.
In factor analysis- what do we do if the constructs are broken down from the original question?
No Idea what you mean
How to read an orthogonal transformation matrix and orthomax graphs?
You don't look at the transformation matrix
How to read the graphs that cluster items together visually for factor analysis in 2010 paper?
The are effectively just plots of the factor loadings. If they don't help you just interpret the loadings themselves
What would you say if the confidence interval had a high effect size but the result was not significant?
It sugests that your sample size is to small to rule out the possibility that there is an effect. Hence try to get a bigger sample.
One of the t-tests allowed us to choose from matched or independent but suggested that we may have to use information from the independent test in order to answer the question. Is there anything we might have to use?
To calculate cohens D you need the average Std Deviation from the 2 independent table.
What are values the show high/low skewness?
Skewness >1 or less <-1
On the R studentized graphs for regression, would points that lie above/below 2 and -2 be important, or are points that lie to the left also important? If so, how?
Generally we look at outside -3 or +3. We would expect 5% of the sample to be outside -2 and +2
The graph is not importantly directional. A point far to the left is an influential point. The same as a point far to the right.
Oh, and how to red the variance explained graph for the 2010 factor analysis question
ReplyDeleteJust a plot of cumulative variance explained. Just use the table and ignore the graph.
See my answers to Q10 on the blog.
ReplyDeleteim hving sum trouble with multiple regression. In d 2010 paper, Q1 we r given 2 IV's & the second the IV is divided into 3 measures. So do i list the 2 IV's or the IV's in the parameter estimated table?
The description to my mind is clear that the three separate ivs are used but it i clear from the analysis that the ivs are
Locus Of Control
Optimism
Utilization
Appraisal
Also for the assumptions how do we comment on thngs like the R student- do we jst mention that there r outliers bcoz some points lie below 2 & -2?
Points -3 or above +3.
Suggest a remedy e.g. remove outliers and refit the model.
&if theres no scatterplot how do we then test for linearity?
If there a plot you need which isn't there. Sy something like.
To test for linearity we would need the scatterpolt of studentized residuals against predicted values.
Hi Mike!
ReplyDeleteI'm a bit unsure as to how to go about selecting factors with particular reference to the cumulative proportion in factor analysis
Hi Mike
ReplyDeleteIf we are just given a Type III table and no Type I, how do we comment if it is a balanced/ unbalanced design? Do we just say we using a Type III table?
Thanks!
Hi Mike
ReplyDeleteAm I correct in saying that for an unbalanced design the r-square is not equal to the partial global effect size? The partial global effect size for an unbalanced design is SSeff/SSeff+SSerror
Thanks.
Hi Mike,
ReplyDeleteI hope that you are feeling better! I just want to ask 2 quick questions:
1) for question 3 in 2010, why have you used the 95% CL Std Dev value difference (0.9449) in the Cohen's D equation instead of the normal Std Dev value diff (1-2) (1.1032)?
2) why is it that because fading had a bigger mean than systematic desensitization, this shows that the fading group is more anxious? I thought that if there was a bigger mean (and therefore a bigger time delay) then that means the group was less phobic/anxious?
thank you so much and I hope that you get well soon!
When can you post the memo for the other questions? Esp the 2 way ANOVA.
ReplyDeleteHi Mike
ReplyDeleteHow do we interpret a negative effect?
Just re uploaded some more solutions to file.
ReplyDeleteI'm a bit unsure as to how to go about selecting factors with particular reference to the cumulative proportion in factor analysis
ReplyDeleteSee example in tut. But generally just see how many factors do you need to get cum prop > 40%
If there <10 items push this up to 50%
If there are more than 100 items drop this to around 35%
Hi Mike
ReplyDeleteIf we are just given a Type III table and no Type I, how do we comment if it is a balanced/ unbalanced design? Do we just say we using a Type III table?
Then you can't
Unless of course you are told about the design in the description
ReplyDeleteHi Mike
ReplyDeleteAm I correct in saying that for an unbalanced design the r-square is not equal to the partial global effect size? The partial global effect size for an unbalanced design is SSeff/SSeff+SSerror
R Sqare is used in reression. Eta squared in Anova.
Use partial eta squared for unbalanced designs. See note above
Hi Mike,
ReplyDeleteI hope that you are feeling better! I just want to ask 2 quick questions:
1) for question 3 in 2010, why have you used the 95% CL Std Dev value difference (0.9449) in the Cohen's D equation instead of the normal Std Dev value diff (1-2) (1.1032)?
I might have got this wrong. Head is not all there. The tables have also swapped around in the last couple of versions. It should be the SD not eith of CLs for the SD
2) why is it that because fading had a bigger mean than systematic desensitization, this shows that the fading group is more anxious? I thought that if there was a bigger mean (and therefore a bigger time delay) then that means the group was less phobic/anxious?
No delay is time for the spider to come closer. So short delays = less anxious
thank you so much and I hope that you get well soon!
Hi Mike
ReplyDeleteHow do we interpret a negative effect?
In what?
That's me for today good luck for tomorrow
ReplyDeleteSorry I meant a negative effect size.
ReplyDelete