Merry Christmas!

...and see you in ANOVA in the New Year.

Reporting Significance Levels

John asked,

When reporting significance level at the .05 level, if my value is .051 can I report as p = .051 (using three decimal places to show approximation)?

Missing Values

Mari said earlier that

Absolutely you should drop any participant who is missing both predictors and outcome.

I have 2 predictors and 1 outcome.
In the event that the outcome and one predictor is missing values do I delete the case even if the other predictor has an value listed?

Assumptions

Can you test assumptions for a categorical variable? This doesn't seem to make sense, because looking at residuals for a contrast-coded variable doesn't seem to provide useful information. So maybe rather than can you test the assumptions, my question is am I supposed to?
Also, for a moderated model, do you check assumptions for the moderator variable? I know there was an earlier post that said you don't have to check assumptions of the predictor for each level of the moderator, but I'm unclear about checking the moderator as its own variable.

reflecting part 2

I noticed that after I had reflected my data and transformed it, the relationship became a positive one, where betas went from negative to positive. I do conceptually understand why it turned out this way, but in my results write up, do I indicate that after i have reflected and transformed my data, there was a significant negative relationship in transformed units, and then report the positive betas that I got from the regression?

mirroring

The distribution of my residuals were slightly negatively skewed, and when I mirrored and transformed my outcome, the histogram still showed a slight negative skew. Can this be normal, or did I overlook a step in my transformation process?

Thank you for all the lightning fast responses!

Violation of homoskedasticity?

Sorry if it is not okay to ask this outright, but does this look okay?  I went ahead and tried the transformations and it got progressively worse. Should I include in my write-up that I tried to transform but it got worse, or should I just say this is "good enough" and only include my first tests of assumptions?

Vanishing Post?

I was responding to a post, and blogger wigged out on me. Suddenly my response and the original post were both gone. If your post vanished, and you still need the answer, please repost!

How Much Overlap is Too Much?

There have been a couple of blog and backchannel questions about the acceptable degree of overlap between previous homework assignments and the final project.

The general rule is that as long as 51% or more of the model has changed, you are OK. That is,

1. if you are using some of the same variables, but in a different model (i.e., the homework model was an additive hierarchical model and now you are looking at the same variables in a mediated or moderated model), that's OK. (Note that the converse is not true--a mediated or moderated homework model cannot revert to an additive hierarchical model, because in that case, you would have no need to run the regressions, because they were already done in the original homework assignment. That is, regression 3 from a mediated model is exactly what you would run to test an additive model with those variables. Similarly, step 1 of a moderated model is the additive multiple regression model.)

1. if the outcome is the same, but predictor and mediator are different, that’s OK.

2. if one predictor is the same, but a moderator, mediator, or second predictor have been added, and the outcome is different, that's OK.

3. if a mediator is the same, but the predictor and outcome are different, that's OK.

If you aren't sure if your model fits those parameters, feel free to ask, but have patience for the reply. I start jury duty tomorrow and will not have internet access...

Transformations

I know that you just do a transformation when both assumptions (homoskedasticity and normality are violated), but what do you need to do when the two violated assumptions are linearity and homoskedasticity (a different pair)and normarlity is perfect. I found in my notes about polinomial regression or adding a squared term to the regression. Do I need to do that? If yes, HOW? If not, I will have 2 assumptions violated. Does that mean that this model does not work?

transformed results write up

would it be possible for us to see an example of an APA results section where the data was transformed? or is it just a matter of just indicating that the results were transformed and discussing the numbers in transformed units?

comparing negative Betas

Hello, I have another question regarding negative betas:

For my mediated model, if the B from the first regression (predictor->outcome) was, let's say -0.55, and the B of predictor->outcome on the third regression was -0.02. Would I still treat this as a situation where the B from the first regression is "greater" than the B from the third? Since mathematically, -0.55 is a smaller value than -0.02, but in terms of relationships, -0.55 still represents a greater amount of change per unit of X?

A question regarding my model

Will be acceptable to do a moderated model with sex as moderator, but using different predictor and outcome varibles not used in the last assigment?

Mediated Model

I have a question regarding my mediated model.

My unstand. B in R3 (-0.015) is not lower than B in R1 (-0.17), but the 3 regressions showed that each is significant (p-values of R1, R2, and R3 are less than .001). This shows that there is no mediation? What's going on?

And when I put this into the Sobel's test (for negative betas), the test statistic shows to be -8.48, but the p-value is 0. So there's a significant level of mediation, right?

These two results are not supporting one another.

Another question: for the Sobel's test statistic, do I regress Mediator (M) to Outcome (O) separately to find the B and SE? Or do I just use the M and O figures from my R3, which is (P and M on O)?

Thanks!

Transformed Variables and the Sobel Test

To add to Chris's question,
If we need to do a transformation in a mediated model, do we have to do anything to the unstandardized beta's before we use them for the Sobel test or can we use them as is?
Thanks,
Andrew

Exam 3 at Front Desk

Feel free to drop by and take a look at your exam, and where you currently stand in the class. Note that your final article reviews are NOT included in that grade estimate (just because the TAs are working on final exams and projects too, not because of any special thing about these assignments).

I am working in a moderated model were I need to run separate regressions for each level of the moderator variable. I want to check my assumptions, but I don't know if I need to use the unstandarized predicted values and residuals of the separate regressions or I need to run a single regression for all cases and use those unstandarized predicted values and residuals. In other words, do I need to check assumptions separate for each level of the moderator (2 graphics per assumption)?

negative betas and sobel's test

I was using the sobel's test website to test a mediated model, but i noticed that when one of the B's i put in the equation was negative, the significance values became what looks to me like 1-p? (i.e., if i had a 2-tailed p value of .95 with the negative B, it would turn into .05 if i changed that B to positive.) How would significance be determined/reported for this situation? And would I report the sobel's test statistic as negative as well?

thanks!

Assumptions for a Mediated Model

Hello friendly and helpful TAs (and/or Mari),

I have a question related to GLM assumptions for a mediated model. When we utilize Baron and Kenny's three model approach for testing mediation, there are three separate regressions performed to test the mediation; during these steps the 'predictor' and 'outcome' variables change according to what you want to analyze. (The relationship of the true predictor to true outcome, relationship of true predictor to mediator, and relationship of both true predictor and mediator to true outcome).

So, my questions:
1) Do we have to test each seperate relationship for GLM assumptions? I assume that at least the first two models do need to be tested, as the relationship of the predictor to outcome and the relationship of the predictor to mediator would have to conform to the assumptions to be valid, but I am unsure of the last model.

2) If assumptions are not met and a transormation needs to be tested on the outcome variable, should I transform for each step of B & K's mediation approach? If I do, this makes interpretation sort of hairy in the second step of the model, as the "outcome" variable to be transformed is no longer the outcome of true interest. Or should I just transform my true outcome on the first regression? In the case that I need to transform for all steps of B & K's approach, how do I interpret the relationship at each step?

Gracias,

Chris

clear-up on small detail

I know that we aren't allowed to use any model previously used in homework assignments or lecture, but we can still use a variable that was used in those assignments, right?

For example, would I be able to hypothesize that perceived humor and morality would predict global self worth? (Not actually using this model for my assignment, but global self worth is a variable we used in other assignments, but since the overall model uses different variables to accompany global self worth, I can go ahead with it?)

thanks!

Failed Assumptions

If when you test your assumptions and cannot pass no matter which transformation applied, can you still use that regression include in your write up that it didn't pass assumptions so that the data cannot be interpreted reliably? Do you want us to start the complete experiement over again with new variables until we find something that works?

Which residuals to use?

I am trying to test the assumptions for a hierarchical multiple regression. After I create the new variables (in this case, square root transformed), do I re-run my assumptions by plotting the new square root variables with the residuals that were generated from my first regression, or do I save new residuals by running the same hierarchical mutliple regression with the new transformed variables and then re-run the assumptions using those new residuals that were generated from the second regression?

Testing Assumption of Homoskedasticity

After creating my scatterplot, i would like to test the ratio of highest to lowest variance. I am not sure how to split the file- the options within this drop down menu do not seem to include the ability to separate the predictor (or is it the predicted values i use?) into meaningful levels. Is there another step i am missing? I messed up my data file once already trying to accomplish this and had to red0 everything all over again :-(

Also, is there a generally accepted cut-off point when you have missing data (i.e. if the missing data represents < 5% of the total sample) that is applied when deciding to omit participants? Predicting values has already proven to be difficult due to more missing data in the alternative variables used.

Final Project: Missing Values

My outcome variable is missing 8 values. I used a multiple regression with my outcome and two other variables that are significantly related to my outcome but aren’t my variables of interest. Yet, my outcome variable is still missing 8 values because those 8 cases were also missing values for the two significantly related variables. Next, I used the regression equation (Transform à Computer Variable) and formed a “predictedoutcome” variable a follows:

Predictedoutcome – B + (B * variable) + (B * variable)

My new predictedoutcome variable is still missing 8 values.

Where shall I go from here? Am I supposed to continue to look for additional variables that are significantly related to my outcome variable? Should I drop a case (subject) if there are missing values on all three of my variables of interest (i.e., my two predictors and outcome)?



Also, can you please specifically explain how to merge two variables on SPSS? I tried merging two variables in “Compute Variable” under the “Function group” but received an error messsage.

Thank you.

levels of moderation

Hello, I hope it's not too late to ask this question...

For categorical moderators, what does it mean exactly when you regress the outcome on the predictor for each level of the moderator? Would they be separate simultaneous multiple regressions where you regress predictor and level 0 of the moderator on outcome, and regress predictor and level 1 of the moderator on outcome?

note discrepancy

I found a discrepancy between the typed out notes and the slides, and could only speculate as to which one was correct. In the notes (for "Prediction and Validation", 3rd page, last sentence above Validation of models) it says
"As sample size increases and number of predictors INCREASE, these three types of R^2 grow closer..." (emphasis added).
The slides however (slide 17 i believe) says these three types "merge as sample size N increases and number of predictors (k) DECREASES"
It makes sense to me that k should decrease because an increase will take away from degrees of freedom, decreasing power. But i don't want to jump to conclusions...which is right? Thanks!!!

Review Sheet and Final Project Posted

The review sheet for exam 3 and the guidelines for the final project have now been posted to Portico. The TA's will also have copies of the review sheet at this evening's review session.

For those of you who would appreciate individual review sessions, please feel free to sign up for times with the TA's.

Article Review

I am working on an article review and I came across an article with an author listed his/her name as A. H. Y. Ho. Should I cite his/her name as "Ho, A. H. Y." or "Ho, A. H." in the reference section of the article review? Thanks.