1. ## Statistical analysis

Hi there

I'm a student doing a dissertation about the influence of reward satisfaction on turnover intention. At the moment I'm wondering how to statistically analyse my results using SPSS, I'm not sure which kind of analysis to use.

My hypotheses are:
1. The relation between reward satisfaction and turnover intention increases if the employee is educated.
2. The relation between psychological satisfaction and turnover intention increases if the employee is educated.
3. Psychological satisfaction has a greater influence on turnover intention than reward satisfaction.

For the first and second I was wondering to use a correlation analysis.
For the 3rd one I thought a regression analysis is the way to go, is this correct?

Kind regards

2. ## Re: Statistical analysis

Originally Posted by Hypnoz
Hi there

I'm a student doing a dissertation about the influence of reward satisfaction on turnover intention. At the moment I'm wondering how to statistically analyse my results using SPSS, I'm not sure which kind of analysis to use.

My hypotheses are:
1. The relation between reward satisfaction and turnover intention increases if the employee is educated.
2. The relation between psychological satisfaction and turnover intention increases if the employee is educated.
3. Psychological satisfaction has a greater influence on turnover intention than reward satisfaction.

For the first and second I was wondering to use a correlation analysis.
For the 3rd one I thought a regression analysis is the way to go, is this correct?

Kind regards
Can you give us some more background. I have an idea what your variables are, but more clarity is better.

What are all your variables and how are they measured?
What's your sample size (total and per group)?
Indicate the dependent variable.

If you believe all of these things influence the outcome (seems to me like you're trying to model turnover intention, maybe it's a score?), you can probably run a least squares regression (provided some things work in your favor, including the assumptions) or a rank regression (one model). Questions 1 & 2 sound like classic interactions, and by coding your variables right, an upper tailed test of hypothesis would work for each of these (or looking at a [1-alpha]*100% CI). The third question is a little trickier to determine which has a "greater influence", because you'd have trouble rationalizing the units of of measure are equivalent for the two independent variables (sounds like you're trying to compare slopes, basically). I would address 3 last because you may have significant interactions from steps 1 & 2.

I'll wait for you to get back regarding my specific questions, and feel free to ask me if anything isn't clear!

3. ## Re: Statistical analysis

Originally Posted by ondansetron
Can you give us some more background. I have an idea what your variables are, but more clarity is better.

What are all your variables and how are they measured?
What's your sample size (total and per group)?
Indicate the dependent variable.

If you believe all of these things influence the outcome (seems to me like you're trying to model turnover intention, maybe it's a score?), you can probably run a least squares regression (provided some things work in your favor, including the assumptions) or a rank regression (one model). Questions 1 & 2 sound like classic interactions, and by coding your variables right, an upper tailed test of hypothesis would work for each of these (or looking at a [1-alpha]*100% CI). The third question is a little trickier to determine which has a "greater influence", because you'd have trouble rationalizing the units of of measure are equivalent for the two independent variables (sounds like you're trying to compare slopes, basically). I would address 3 last because you may have significant interactions from steps 1 & 2.

I'll wait for you to get back regarding my specific questions, and feel free to ask me if anything isn't clear!
My variables are psychological satisfaction measured with the PreSS (questionnaire using likert scales), reward satisfaction divided in pay structure, pay level, pay raises and advantages, measured with the PSQ (questionnaire using likert scales), turnover intention using ATS (questionnaire using a seven-point-scale). However I also have information about age (ratio scale), sex (standard m or f question), education degrees (multiple choice), management position or not (yes/no question) and net pay (nominal scale).

The sample size is 86 at the present moment. However, I'm studying 4 different companies and 2 still need to participate. (So I now have 2 groups at 74 and 12).

Like you already assumed I'm trying to model turnover intention which results in a score. So I could do square regression or rank regression (one model)? And for questions 1 & 2 an upper tail test?

4. ## Re: Statistical analysis

Questions 1 and 2 concern interactions (the relationship between two variable changes if the value of a thrid variable changes). You could use a multiple regression approach for them. One important question will be how "education" is defined. Maybe you have something like higher/medium/lower education, in such a case with k categories you can add k-1 dummy variables for "education", and k-1 interaction variables to your model.

With kind regards

K.

5. ## Re: Statistical analysis

Originally Posted by Karabiner
Questions 1 and 2 concern interactions (the relationship between two variable changes if the value of a thrid variable changes). You could use a multiple regression approach for them. One important question will be how "education" is defined. Maybe you have something like higher/medium/lower education, in such a case with k categories you can add k-1 dummy variables for "education", and k-1 interaction variables to your model.

With kind regards

K.
Education is defined by using 3 groups. professional bachelors are 1, academic bachelors are 2 and masters are 3. It is still possible to use K categories here, right? Because I have tried it and it resulted in some strange graphs... (i.e. turnover intention increases when psychological satisfaction increased with academic bachelors and master's degrees.)

Kind Regards

6. ## Re: Statistical analysis

Originally Posted by Hypnoz
My variables are psychological satisfaction measured with the PreSS (questionnaire using likert scales), reward satisfaction divided in pay structure, pay level, pay raises and advantages, measured with the PSQ (questionnaire using likert scales), turnover intention using ATS (questionnaire using a seven-point-scale). However I also have information about age (ratio scale), sex (standard m or f question), education degrees (multiple choice), management position or not (yes/no question) and net pay (nominal scale).

The sample size is 86 at the present moment. However, I'm studying 4 different companies and 2 still need to participate. (So I now have 2 groups at 74 and 12).

Like you already assumed I'm trying to model turnover intention which results in a score. So I could do square regression or rank regression (one model)? And for questions 1 & 2 an upper tail test?
I would probably wait for the rest of your data to come in to avoid maybe influencing how you proceed with analysis (i.e. if you see something in a prelim analysis you might (consciously or subconsciously) alter the analysis plan when the rest of the data come in later.

Would you mind clarifying the actual number of independent variables that you'd plan to use in trying to model intention to turnover? I'm trying to make sure I understood what you wrote above. Depending on this number, it may call for some creativity or careful selection of what terms to include in the model.

As for my comment on using a multiple linear regression (same as ordinary least squares): if you fit a preliminary model, you can investigate the model assumptions. If they're reasonably satisfied, you might be able to use this method, provided it checks out after you go through some of the modeling steps. If it doesn't work, or if you want to check the robustness of your conclusions, you could use a rank regression (predicting the rank of the intention to turnover score, think of it as a nonparametric analysis). Either of these would allow you to test for interactions as you want.

In terms of the "tailedness" of the test, you can disregard my comment before, as it wasn't very clear. Realistically, you can just test the coefficient at that time to see if the interaction coefficient meshes with your hypothesis.

Originally Posted by Hypnoz
Education is defined by using 3 groups. professional bachelors are 1, academic bachelors are 2 and masters are 3. It is still possible to use K categories here, right? Because I have tried it and it resulted in some strange graphs... (i.e. turnover intention increases when psychological satisfaction increased with academic bachelors and master's degrees.)

Kind Regards
In general, it's much easier to interpret model coefficients (for linear regression, for example) when you fit the intercept and use k-1 dummy variables for a variable with k levels. Using 1/0 coding will create a unique permutation of 0/1 for each of the k-levels.

7. ## Re: Statistical analysis

Originally Posted by ondansetron
I would probably wait for the rest of your data to come in to avoid maybe influencing how you proceed with analysis (i.e. if you see something in a prelim analysis you might (consciously or subconsciously) alter the analysis plan when the rest of the data come in later.

Would you mind clarifying the actual number of independent variables that you'd plan to use in trying to model intention to turnover? I'm trying to make sure I understood what you wrote above. Depending on this number, it may call for some creativity or careful selection of what terms to include in the model.

As for my comment on using a multiple linear regression (same as ordinary least squares): if you fit a preliminary model, you can investigate the model assumptions. If they're reasonably satisfied, you might be able to use this method, provided it checks out after you go through some of the modeling steps. If it doesn't work, or if you want to check the robustness of your conclusions, you could use a rank regression (predicting the rank of the intention to turnover score, think of it as a nonparametric analysis). Either of these would allow you to test for interactions as you want.

In terms of the "tailedness" of the test, you can disregard my comment before, as it wasn't very clear. Realistically, you can just test the coefficient at that time to see if the interaction coefficient meshes with your hypothesis.
The actual numbers of independent variables I'm going to use is 2 namely, reward satisfaction and psychological satisfaction. I'm also planning to use education as the moderating variable and ofcourse turnover intention as the dependent variable.
Reward satisfaction = (pay level + pay raises + pay structure)/3 however is is possible to use them each as a independant variable as well? (then I'm using 4 independant variables)

Kind regards

8. ## Re: Statistical analysis

Originally Posted by Hypnoz
The actual numbers of independent variables I'm going to use is 2 namely, reward satisfaction and psychological satisfaction. I'm also planning to use education as the moderating variable and ofcourse turnover intention as the dependent variable.
Reward satisfaction = (pay level + pay raises + pay structure)/3 however is is possible to use them each as a independant variable as well? (then I'm using 4 independant variables)

Kind regards
So:
Y= turnover intention (measured in points)
X1= Reward satisfaction (points)
X2= psych sat. (points)
Education: X3= 1 if masters; 0 else AND X4= 1 if academic bachelors; 0 else
(this would implicitly make professional bachelors the base/reference category coded as 0,0 for an X3,X4 pair. You could change the coding however you like such as making masters the base, but 0/1 coding is the easiest to interpret).

Interaction of : X1*X3, X1*X4; X2*X3, X2*X4; any plans to test for interaction between X2 and X1? I would probably do so for modeling purposes (I.e. if you say "No, I don't believe there exists interaction of X1 and X2" then you'd actually want to test that to make sure there isn't evidence for a significant interaction. If nonsignificant, this doesn't prove lack of interaction, but it does provide you with the ability to say no significant interaction of X1 and X2 was detected).

Let's stop here and address the following (before building the model):

For reward satisfaction-- what is the reason for combining all of those measures? I would think that these don't necessarily indicate how satisfied the person is with pay (think of people who earn a lot but hate their jobs--I can think of quite a few). Have you looked at the principle components to see if how similar the information is between these three variables (to justify combining at least 2)? If you want to combine all 3, maybe the variable is better interpreted as "financial incentive/benefit"-- what do you think? If you want to use each separately, you'd have to make separate interactions for those (are they score variables?).

9. ## Re: Statistical analysis

Originally Posted by ondansetron
Interaction of : X1*X3, X1*X4; X2*X3, X2*X4; any plans to test for interaction between X2 and X1? I would probably do so for modeling purposes (I.e. if you say "No, I don't believe there exists interaction of X1 and X2" then you'd actually want to test that to make sure there isn't evidence for a significant interaction. If nonsignificant, this doesn't prove lack of interaction, but it does provide you with the ability to say no significant interaction of X1 and X2 was detected).
To analyse the interactions I need to multipy an independent variable with the moderator? Is this correct? If so, you made it a bit clearer to me!
Testing for an interaction between X1 and X2 seems helpful to me even if it is solely for modeling purposes. If there might be an interaction between the 2 variables, could it be a help with hypothesis 3?

Originally Posted by ondansetron
Let's stop here and address the following (before building the model):

For reward satisfaction-- what is the reason for combining all of those measures? I would think that these don't necessarily indicate how satisfied the person is with pay (think of people who earn a lot but hate their jobs--I can think of quite a few). Have you looked at the principle components to see if how similar the information is between these three variables (to justify combining at least 2)? If you want to combine all 3, maybe the variable is better interpreted as "financial incentive/benefit"-- what do you think? If you want to use each separately, you'd have to make separate interactions for those (are they score variables?).
I combined all 3 because I'm indeed interested in financial reward satisfaction. Each one seperatly is scored by a likert scale, so it ranged from 1 to 5. Is this what you mean with score variables?

10. ## Re: Statistical analysis

Originally Posted by Hypnoz
To analyse the interactions I need to multipy an independent variable with the moderator?
Right. Create a new variable from the cross product of the suspected terms involved in the interaction. If you suspect X2 moderates X1, a new variableV X1_X2 should be created as the product of X1 and X2. The coefficient for this will allow you to test for the interaction.

Originally Posted by Hypnoz
Is this correct? If so, you made it a bit clearer to me!

Originally Posted by Hypnoz
Testing for an interaction between X1 and X2 seems helpful to me even if it is solely for modeling purposes. If there might be an interaction between the 2 variables, could it be a help with hypothesis 3?
I don't think it will necessarily help with #3. As mentioned earlier #3 is going to be tricky to test. I would first construct a careful definition of what it means to be "more important" than another variable (explained variance, for example might be one approach).

Originally Posted by Hypnoz
I combined all 3 because I'm indeed interested in financial reward satisfaction.
I guess my question is this: how does summing these variables measure satisfaction with the financial rewards and reward potential? You want to make a convincing, logical argument for why this is the case (because it may not be measuring satisfcation-- you're simply asking factual questions about pay and potential, not how satisfied someone is with respect to these things).

Originally Posted by Hypnoz
Each one seperatly is scored by a likert scale, so it ranged from 1 to 5. Is this what you mean with score variables?
Yes, that's what I meant.

11. ## Re: Statistical analysis

Originally Posted by ondansetron
I guess my question is this: how does summing these variables measure satisfaction with the financial rewards and reward potential? You want to make a convincing, logical argument for why this is the case (because it may not be measuring satisfcation-- you're simply asking factual questions about pay and potential, not how satisfied someone is with respect to these things).
Well the questions relate to what they receive each month (i.e. take home pay), their most recent raise (what they have received) and how the company manages their pay. It all relates to the money they already earned or still will receive, that's why the 3 are grouped under financial reward satisfaction. Hopefully this will make it a bit clearer for you?

Attached you can find an analysis I already did by using interactions like you explained before. If I can interpret it correctly I see that the relation between X1 and Y, and X2 and Y declines when the employee has a higher degree. What are your thoughts on this?

12. ## Re: Statistical analysis

Originally Posted by Hypnoz
Well the questions relate to what they receive each month (i.e. take home pay), their most recent raise (what they have received) and how the company manages their pay. It all relates to the money they already earned or still will receive, that's why the 3 are grouped under financial reward satisfaction. Hopefully this will make it a bit clearer for you?
Right, I understand this is what you're doing. What I'm saying is that this seems to involve an element of psychology-- you're trying to use some kind of satisfaction (or lack thereof) to determine a form of employee turnover. What still hasn't been explained or justified is that this measurement shows any "satisfaction" with the rewards received by an employee. The objective reporting of monetary benefits doesn't allow you to assess how satisfied the employees are with their financial rewards. In other words, the variable you created doesn't indicate how happy the employees are with their financial rewards, unless you make the assumption that a higher value of the variable means more employee satisfaction. I don't necessarily think that assumption is reasonable, overall. Speculating, I would say there are many employees who have reasonably high values of the financial reward variable, but aren't necessarily any happier than someone with a lower value (for whatever reason). This is something I would consider as a limitation if you're planning to submit it for peer review (unless you have a variable measured where you explicitly asked employees how satisfied they are with their financial rewards).

If you want the short version of that: I'm not convinced enough (without other data) that the variable you label "financial reward satisfaction" reasonably measures that satisfaction. If I were a peer-reviewer, I'd ask how you know or why you believe this measures satisfaction, and whether you actually asked the employees explicitly about their satisfaction.

Hopefully I was a little more clear with my point this time

Originally Posted by Hypnoz
Attached you can find an analysis I already did by using interactions like you explained before. If I can interpret it correctly I see that the relation between X1 and Y, and X2 and Y declines when the employee has a higher degree. What are your thoughts on this?
The picture, for some reason or another, is very small. When I try to zoom in, it gets too blurry to read.

13. ## Re: Statistical analysis

Originally Posted by ondansetron
Right, I understand this is what you're doing. What I'm saying is that this seems to involve an element of psychology-- you're trying to use some kind of satisfaction (or lack thereof) to determine a form of employee turnover. What still hasn't been explained or justified is that this measurement shows any "satisfaction" with the rewards received by an employee. The objective reporting of monetary benefits doesn't allow you to assess how satisfied the employees are with their financial rewards. In other words, the variable you created doesn't indicate how happy the employees are with their financial rewards, unless you make the assumption that a higher value of the variable means more employee satisfaction. I don't necessarily think that assumption is reasonable, overall. Speculating, I would say there are many employees who have reasonably high values of the financial reward variable, but aren't necessarily any happier than someone with a lower value (for whatever reason). This is something I would consider as a limitation if you're planning to submit it for peer review (unless you have a variable measured where you explicitly asked employees how satisfied they are with their financial rewards).

If you want the short version of that: I'm not convinced enough (without other data) that the variable you label "financial reward satisfaction" reasonably measures that satisfaction. If I were a peer-reviewer, I'd ask how you know or why you believe this measures satisfaction, and whether you actually asked the employees explicitly about their satisfaction.

Hopefully I was a little more clear with my point this time
Ok I understood it the first time as well, but I wasn't very clear, sorry for this. I'll certainly keep this in mind! I based myself on an article of Heneman & Schwab where they created the PSQ to measure pay satisfaction.

Originally Posted by ondansetron
The picture, for some reason or another, is very small. When I try to zoom in, it gets too blurry to read.
Again sorry for this, I wasn't aware.
Hopefully this is not blurry and understandable:

14. ## Re: Statistical analysis

Originally Posted by Hypnoz
Ok I understood it the first time as well, but I wasn't very clear, sorry for this. I'll certainly keep this in mind! I based myself on an article of Heneman & Schwab where they created the PSQ to measure pay satisfaction.
I see now. If that's the case, they may have already checked its validity as a measure of satisfaction. I was under the impression you just decided to go ahead with it without another reference (my mistake).

Originally Posted by Hypnoz
Again sorry for this, I wasn't aware.
Hopefully this is not blurry and understandable:
No problem at all! It looks like you're using SPSS?

For testing, I would enter all the interactions for one variable into a new "block" in the model dialog box. In other words, enter everything in block 1 except for the interaction terms relating to PsySat . Then in the next block, put those interactions in and request that SPSS also shows the change in R-squared statistics (so put PsySat_Abach & Psy sat_master in block 2, for example). When you hit enter, you'll get results for the block 1 model and if the change is significant when adding the block 2 terms to the model.

If significant, leave in the model and do follow up testing later (after we do this for other terms). Put a different set of interaction terms in the second block, and move the significant terms to block 1. If not significant, drop the terms from the model, and put a different set of interaction terms in the 2nd block. You can do this so you've tested all interaction terms in joint subset tests first, before specifying which should be dropped on an individual basis before interpreting the final model.

15. ## Re: Statistical analysis

Originally Posted by ondansetron
No problem at all! It looks like you're using SPSS?
Yes, SPSS is the program I'm using however I also have acces to R and Excel. I know little about R and lots about Excel.

Originally Posted by ondansetron
For testing, I would enter all the interactions for one variable into a new "block" in the model dialog box. In other words, enter everything in block 1 except for the interaction terms relating to PsySat . Then in the next block, put those interactions in and request that SPSS also shows the change in R-squared statistics (so put PsySat_Abach & Psy sat_master in block 2, for example). When you hit enter, you'll get results for the block 1 model and if the change is significant when adding the block 2 terms to the model.

If significant, leave in the model and do follow up testing later (after we do this for other terms). Put a different set of interaction terms in the second block, and move the significant terms to block 1. If not significant, drop the terms from the model, and put a different set of interaction terms in the 2nd block. You can do this so you've tested all interaction terms in joint subset tests first, before specifying which should be dropped on an individual basis before interpreting the final model.
Attached you can find the results of this if I understood it correctly. I have added soms variables (all except pay itself (I mean the numbers of pay, not the satisfaction)). But nothing is significant...