MANOVA allows us to test hypotheses regarding the effect of one or more independent variables on two or more dependent variables.

Perhaps our paper on life-satisfaction gives you an idea. I would do regression analyses.

Hi Jack,

can you elaborate more how you proceeded. Having "identified persons with high turnover rates" suggests that you focused especially on this group and you surveyed them (and potentially those with low levels?).

You should conduct a comprehensive survey and then investigate relationships. Selecting persons with an certain value on an focused outcome is called endogenous selection bias and probably will give your nonsensical results.

For more info, see

Elwert, F., & Winship, C. (2014). Endogenous selection bias: The problem of conditioning on a collider variable. Annual Review of Sociology, 40, 31-53. doi:10.1146/annurev-soc-071913-043455

In addition: What is your goal when doing this analysis? You do not need another study to show that low job satisfaction causes turnover (there are already a trillion).

Best,

Holger

Hi, Jack!

Today multiple regression is a ubiquitous choice for the analysis you describe if one's objective is to maximize explained variation in an ordered turnover intention dependent measure, and logistic regression is the choice if the turnover intention measure is binary and one's objective is to maximize likelihood.

No regression model is formulated to identify the most accurate model, and neither variance or likelihood are synonymous with accuracy. Not surprisingly, regression models have difficulty making accurate predictions of data values which deviate from the mean (ordered data) or the prevalent (categorical data) response.

• https://odajournal.com/2013/09/20/maximizing-the-accuracy-of-multiple-regression-models-using-unioda-regression-away-from-the-mean/
• https://odajournal.com/2013/10/21/assessing-technician-nurse-and-doctor-ratings-as-predictors-of-overall-satisfaction-of-emergency-room-patients-a-maximum-accuracy-multiple-regression-analysis/
• Yarnold, P.R., Hart, L.A., & Soltysik, R.C. (1994). Optimizing the classification performance of logistic regression and Fisher’s discriminant analyses. Educational and Psychological Measurement, 54, 73-85.

There is a modern non-parametric alternative which identifies the most accurate (weighted) model(s) which exist for your data and hypotheses.

• https://odajournal.com/2014/04/06/a-statistical-guide-for-the-ethically-perplexed-chapter-4-panter-sterba-handbook-of-ethics-in-quantitative-methodology-routledge-2011-clarifying-disorientation-regarding/
• https://odajournal.com/2020/06/28/what-is-novometric-data-analysis/
• https://odajournal.com/2016/09/19/novometric-analysis-with-ordered-class-variables-the-optimal-alternative-to-linear-regression-analysis/

Holger Steinmetz Yes I can elaborate, thanks for the response. These groups that I am currently focusing on, as I said in my post, are the 3 medical staff groups with the highest rates of turnover intention. Those 3 groups also happen to have the lowest overall job satisfaction. This conclusion was arrived at after a comprehensive survey was done for all medical staff, and after analysing the results through ANOVA and correlation coefficient.

Hi Jack,

as I said, you shouldn't do that as it creates a collider bias. In addition, the problems how to proceed have to with this hole you dug and in which you are now. Isn't there a possibility to do a full-blown survey with a broad sample (including all, high, middle, and low satisfaction folks.

Apart from that, Jack gave you some tips but remember we talking about appropriate methods for certain goals. Pure prediction has purely practical goals, hence, if this is with you want, he gave you some sound advise. If you do want to learn or understand something behind the data than there is not shortcut to proper sampling, definition of a population, understanding causal processes and their links to the data, and measurement models (linking attributes to measures of these attributes). It procedure you choose depends on the task.

HTH

Holger

Yup you can proceed. You'll have to identify the possible confunders. Good luck

Holger Steinmetz

I'm confused by what you mean with a full-blown survey. As I said, I did a survey for all medical staff, in other words, the entire population. There are 7 different types of medical staff. Out of those 7, I have chosen 3 to focus on AFTER I completed my survey about satisfaction & turnover intention. There was no prediction involved at all.

Also, could you expand on the collider bias and how it relates to me please?

Hi Jack,

ok, these are good news. Now let's come to what your goals are. Analyses are means to an end to reach a goal / answer a question. What are your questions?

With regard to collider bias, please read the paper by Elwert & Winship.

In a nutshell, a collider is a variable that is affected by your IV and DV. If you control for that collider (or select a subsample) you get biased results. The same happens if you have a structure like X --> Y --> S and you control for S or select a subsample of S as well as if you select a subsample of Y and estimate the X-Y effect. All of this biases (and sometimes reverses) the estimated effect of X on Y.

In this thread, you find a nice example of a researcher who found seemingly strange effects which supposedly were caused by collider bias (and a misspecification of his model):

https://www.researchgate.net/post/How_to_solve_the_collinearity_problem_in_mediation#view=5e75f55c5d34f82a212c23dd

Best,

Holger

Holger Steinmetz

Thank you for your explanation. That was very helpful.

To respond to your question about what my questions are, I have this main research question:

"To what extent does employee satisfaction have an impact on medical employee intended voluntary turnover rates at SKMH (hospital)?"

and these sub-research questions:

1. Which group(s) of medical employees have the highest intended voluntary turnover rate, and why?

2. Which group(s) of medical employees have the lowest job satisfaction rates, and why?

3. How do job satisfaction levels relate to medical staff intended voluntary turnover rates?

4. Which identifiable strategies will impact medical employee satisfaction levels?

5. Which identifiable strategies will impact medical staff intended voluntary turnover rates?

The first 3 I plan to answer using quantitative analysis, followed by interviews + qualitative analysis of these. The final 2 will be using relevant literature, depending on results.

Paul Yarnold, Béatrice Marianne Ewalds-Kvist

I have read this:

"Multiple regression" refers to situations in which you have more than one predictor / explanatory variable (𝑋X).

"Multivariate regression" refers to situations in which you have more than one response / outcome / dependent variable (𝑌Y).

I have two outcome variables: Overall job satisfaction + turnover intention, and I have around 25 predictor variables that are regarding various specifics of job satisfaction.

As this is the case, would you agree that I should use multivariate regression instead of multiple regression?

You have already done a lot as establishing relationship between, low job satisfaction and increased turnover intention, now run the both multiple regression and multivariate regression, and see how both tests are interpreting results, it will give more clear idea what test is displaying more logical picture of relationship.

Hi, Jack,

Several different legacy linear methods are used to analyze data in the multivariate (more than one dependent measure) design you describe--which differ to some extent in terms of their data and assumption requirements: examples are multivariate regression, canonical correlation, path analysis, and structural equation modeling. They have several things in common:

1) None have ever found relationships which yielded a Nobel prize.

2) None are capable of making accurate point-predictions (perhaps most clearly explaining point 1).

3) All require substantial time and cognitive investment to "learn".

4) They all make various untenable, never true assumptions about things like underlying distributions, and nature of error terms--which call into question the validity of their findings.

5) They all assume the universe truly is a linear phenomenon!

6) They all find strong effects in random data.

7) None of them can guarantee finding the most accurate answer which exists for a given hypothesis and data set.

The modern method I described earlier corrects all of the previous errors except for number 1. Number 1 can't be solved until problems 2-7 are fixed. The problems other than number 1 ARE fixed. Care to guess what is coming next, for the new methodology, a bit down the road?

Should you decide to stay within the Matrix for now, it is never too late to take the red pill...

PS: I've been at this a long time, more than half a century. It took me a long time to learn all the legacy methods (2 decades), and more than 3 decades subsequently to replace them with a new paradigm which literally fixes all legacy problems (points 1-7 are only a subset). After I learned the legacy paradigm, I produced two introductory books for APA Books with my friend Larry Grimm, PhD. These books introduce all the legacy methods I mentioned above--and the books are available in many academic libraries. They are called Reading and Understanding Multivariate Statistics, and Reading and Understanding More Multivariate Statistics.

Subsequently, APA Books published a book starting the solution to the failed past: https://www.apa.org/pubs/books/4316000