It sounds like you want to study coping strategies and organizational change. There have been a good number of studies done on this topic:

https://scholar.google.com/scholar?q=coping+strategies+organizational+change

You should read a bunch of these studies and see if there has been some relationship that has been measured repeatedly, for example, use of emotion-focused strategies and change in stress during periods of organizational change. If the same relationship has been studied several times, you can do a meta-analysis.

Celeste,

I think to the extent that those studies do not capture the exact same variables, you may very quickly run into an "apples and oranges" problem. Combining studies investigating different variables into a cumulative estimate (that is supposed to approximate a population value) would give you meaningless results.

For example, if you had one set of studies looking at coping strategy A (say, versus not coping at all, whatever that means) and another set of studies looking at coping strategy B, it would make little sense to combine them via meta-analysis. What would the results mean? Some combination of A and B? Of course you could keep A and B separate by testing moderation. But the omnibus effect would be meaningless.

By the way, I have found this paper helpful in understanding meta-analysis:

Article If It Were Only That Easy: The Use of Meta-Analytic Research...

I agree with Heiko. Meta-analysis is a method to cope with questions, so the variables of interest are not necessarily the same. But you need to have a focus, a specific perspective in terms of your research question (your example question is rather ... I would say, ambiguous for meta-analysis). You can analyze a group of variables relevant to a specific topic but from different contexts. For example, you may want to study employees' reaction to organizational changes from the perspective of, say, innovation. In this case you can analyze the variables related to employee's innovation in response to your research question. And it is not advised to analyze variables across topics (or even without a specific topic/theme ), which is meaningless, like Heiko said.

The meta-analysis provides an effect size (with confidence interval) from the results of several previous studies, on the same variable or relationship. As Heiko points out, it does not make sense to mix different variables in the meta-analysis. In case there are several variables, a meta-analysis can be made for each of them separately.

The homogeneity test allows us to know if the meta-analyzed studies differ significantly from each other. In the affirmative case, analysis of the moderating variables can be made.

As a fundamental reference about meta-analysis, I recommend the book by Lipsey, M. W., & Wilson, D. B. (2001). Practical meta-analysis. London: Sage.

David Dunaetz , Heiko Breitsohl, Wenhong Chen and José García-Arroyo many thanks for your clarifications and suggestions for additional readings!

Someone is always going to find something to critique, but if you truly feel the study is relevant, you could consider converting any standardized effects based upon the mean, to a nonparametric generalized odds ratio, which is related to the more general hypothesis:

"To what extent is a random score in one group, greater than a random score in another?"

You would have to do this for all of the data, obviously.

Some potential criticisms:

1. Someone is destined to complain about the loss of "power" when using the nonparametric POV. My rebuttal would be: the generalized odds ratio is closely related to the Mann-Whitney-Wilcoxon (MWM) test, and the Hodge-Lehmann estimator.

The MWM has at least 0.86 efficiency in the worst case (data from a distribution with tails thinner than normal -- an unlikely scenario), and infinite efficiency in cases of skew or excessive variance. In the case where the data is precisely normal, you lose only about 5% efficiency compared to the T test.

2. The "apples-oranges" critique: The conversion of everything to generalized odds ratios somewhat mitigates that argument. If the study is relevant, then the treated group should have higher (lower) scores vs. control, regardless of the measure.

You might want to look up the rebuttal by Gene V. Glass et. al in his pioneering work: Meta-Analysis in Social Science Research (p.22-23).

"The claim that the only studies which are the same *in all respects* (emphasis in original) that can be compared is self-contradictory ... The only studies which need to be integrated are *different studies* (again emphasis in original). Generalization will necessarily entail ignoring some distinctions that can be made among studies."

3. I would avoid the NHST point of view, and emphasize the estimation perspective; that will also help with critique mentioned in 1. You will need to decide on the best method of aggregating the data. I would consider reporting robust variants of the measure of central tendency and dispersion.

I would also do a sensitivity analysis -- ie. which studies have the most impact on a naive reporting of central tendency and dispersion. A statistical consultant would be invaluable to give guidance on data analysis methods. I'm partial to the likelihood school of thought FWIW.

Some References:

Article The Missing Medians: Exclusion of Ordinal Data from Meta-Analyses

Article Generalized Odds Ratios for Ordinal Data

A Conversation With R. Clifford Blair On The Occasion Of His Retirement https://digitalcommons.wayne.edu/jmasm/vol3/iss2/22/

*A thorough discussion of Monte Carlo simulations of the MWM test vs the conventional T test.

Nice discussion by academic statisticians on the methodological issues related to this nonparametric estimate.

When to use the Wilcoxon Rank sum instead of unpaired T test.

https://stats.stackexchange.com/questions/19681/when-to-use-the-wilcoxon-rank-sum-test-instead-of-the-unpaired-t-test

Hi Robert Ryley , many thanks for your feedback and the useful references. Really appreciate the help!

I've collected a host of resources on meta-analysis as well as non-parametric effect measures on Frank Harrell's data methods website. That is a great place to discuss the statistical aspects of research, particularly in the health care context.

Nonparametric Effect Size estimation, Likelihood Methods, and Meta Analysis

https://discourse.datamethods.org/t/nonparametric-effect-size-estimation-likelihood-methods-and-meta-analysis/1525/3

Added: June 28, 2019:

Another alternative, should the studies be too heterogeneous for effect size aggregation, you can try methods based upon P value combination. There are a number "admissible" p value combination techniques, but I'm partial to the Stouffer-Lipak method based upon the inverse transform of the p value.

These authors adapt it into a useful Neyman-Pearson-Fisher theory of evidence, closely related to the hypothesis testing and estimation taught in every intro stats class.

Meta analysis : a guide to calibrating and combining statistical evidence

https://www.worldcat.org/title/meta-analysis-a-guide-to-calibrating-and-combining-statistical-evidence/oclc/180982658

The widely cited Hedges and Olkin text Statistical Methods for Meta-Analysis (p. 28) states:

"The nonparametric nature of combined significance tests gives them great flexibility in applications. Such tests can be used to combine any independent tests of hypotheses even though the individual tests examine somewhat different hypotheses."

The limitation -- you will only be able to claim there is evidence for a positive (negative) effect. You will sacrifice the ability to claim any magnitude for that effect.

The following is a good discussion of the merits/drawbacks of the p value approach compared to effect size aggregation:

Article Reply To: The Liptak-Stouffer Test for Meta-Analyses