Mediation in health research: A statistics workshop using SPSS Dr. Sean P. . An introduction to mediation analysis using SPSS software (specifically, model by adding a third variable (i.e., the “mediator”) • In mediation, the. Question: Is PROCESS available for any program other than SPSS or SAS? .. Does this mean there is no way that M could mediate the relationship between X . The coefficient for X*Z tells you how much the coefficient for the relationship . kinenbicounter.info While testing moderating effect, is it necessary to center the variables? the models and included the corresponding annotated syntax of both programs in this chapter.
No, unless you consider equal weighting a use of sampling weights. I address some fairly unsophisticated but easy to implement means of entertaining questions of causal order in the book. For latent variable models, I recommend Mplus, for it has the ability to estimate latent variable models and parameters that are functions of model coefficients while producing bootstrap confidence intervals for these parameters without having to jump through all the hoops many other covariance structure modeling programs require.
If your "latent" variable is a average of indicators and available in your data as such, then technically it isn't a latent variable; it is observed. There is a vocal minority that takes the position that you should not or, even worse, cannot do mediation analysis with correlational data, and no doubt you will encounter a critic now and then who takes this perspective. In my opinion, this position confuses the roles of data analysis, research design, and theory in causal inference.
My position on the role of data analysis in causal inference is discussed in books and journal articles I have written, and it is more relaxed, empowering, and trusting of the intelligence of the scientist than the extreme "manipulationist" position. Other places where I discuss this include Hayes and Rockwood The position I take--that inference is a product of our minds and not our mathematics or our software! All this said, it remains your responsibility to keep your brain attuned to the inferential task at hand and not be lulled into complacency when interpreting your output from PROCESS or elsewhere.
A statistically significant indirect effect is in no way a proof of causality. Make your argument, if an argument is the best you can do given the nature of your research design, but don't overstate or convey overconfidence in what your analysis is telling you about cause-effect.
PROCESS uses ordinary least squares OLS regression to estimate variables on the left sides of model equations, except for the model of outcome variable Y, which can be estimated with logistic regression if it is dichotomous. PROCESS uses ordinary least squares OLS regression to estimate variables on the left sides of model equations, except when outcome variable Y is dichotomous, in which case the model of Y is estimated with logistic regression. Note that in version 3.
Models are moderated serial mediation models, and you can program your own if none of the preprogrammed models correspond to what you want to do. I am interested in estimating a moderation model PROCESS model 1 but my independent variable X or moderator is categorical with more than two categories.
This topic is also discussed in Chapter 10 of the 2nd edition of Introduction to Mediation, Moderation, and Conditional Process Analysis. I would like to estimate a mediation model model 4 but my X is a multicategorical variable rather than dichotomous or continuous.
For a discussion of mediation analysis with a multicategorical independent variable, see Chapter 6 of the second edition of Introduction to Mediation, Moderation, and Conditional Process Analysis. You can also read about this in Hayes, A.
Statistical mediation analysis with a multicategorical independent variable.
FAQ - The PROCESS macro for SPSS and SAS
British Journal of Mathematical and Statistical Psychology, 67, My confidence intervals for indirect effects change each time I do a mediation analysis. There must be something wrong with your code. There is nothing wrong with the code.
The end points of a the confidence interval are determined by percentiles in the distribution of bootstrap estimates of the indirect effect. See Appendix A for instructions. Alternatively, set the number of bootstrap samples to a very large number in order to minimize sampling error in the estimation of the end points of the confidence interval.
In PROCESS version 2, bias corrected bootstrap confidence intervals was the default, with the percentile method available as an option. In version 3, percentile bootstrap confidence intervals are the default.
Bias correction is not available in version 3. So even if you use the same seed, you will find discrepancies in the end points of confidence intervals in output between versions 2 and version 3. I've been told that it is wrong to control for a mediator when estimating the effect of X on Y. If you are interested in only the total effect of X on Y, then you would not want to control for a mediator M when estimating Y from X. If you control for M, then you are estimating only the direct effect of X, meaning that your estimate of X's effect does not include the component of X's effect that operates through the mediator.
But when doing a mediation analysis, of interest is the indirect effect of X and, perhaps, the direct effect.
To estimate these, you have to include mediator M in the model of Y. The total effect carries no information about mediation, and whether the total effect is significant or not is irrelevant to whether you can or should conduct a mediation analysis to examine the indirect effect of X on Y.
It appears that I have evidence of an indirect effect of X on Y through a proposed mediator, but there is no evidence of an association between X and Y. What should I do? This is not only possible, but it is probably much more common than people realize.
Modern thinking about mediation analysis does not impose the requirement that there be evidence of a simple association between X and Y in order to estimate and test hypotheses about indirect effects. Beyond Baron and Kenny: Statistical mediation analysis in the new millennium. Communication Monographs, 76, Regression-based mediation and moderation analysis in clinical research: Observations, recommendations, and implementation.
Behaviour Research and Therapy, 98, When the same model is estimated using the same data with the same output options, the results will be the same as what you get with SPSS or SAS's regression procedures. There are many sources of discrepancies you may notice when discrepancies exist, and they are all generated by the user, not by PROCESS.
When you do, standard errors, t-values, p-values, and confidence intervals are different than what SPSS and SAS's internal regression procedures produce, as they should be.
Most other sources of discrepancies are due to the user not acknowledging the existence of missing data. For example, if you mean center or standardize "univariately" i. I don't recommend doing centering or standardization computations manually.
If you do, do them to a high degree of precision three or four decimal places generally is not sufficient and only after purging the data of cases missing on variables that will end up in the analysis. Read about the problems with manual mean centering and standardization as well as my debunking of the mean centering myth in the 2nd edition of Introduction to Mediation, Moderation, and Conditional Process Analysis.
In a mediation analysis, another common mistake I see users make is estimating the effect of X on M and the effect of M on Y controlling for X in separate regressions without acknowledging the existence of missing data.
Moderator Analysis with a Dichotomous Moderator using SPSS Statistics
Suppose, for example, some cases are missing on Y. Although we can debate the merits and faults of listwise deletion, it is generally not good practice to piece together a mediation analysis using different subsets of the data for the estimation of different parts of the model. If there is a difference between these, you have missing data you are not properly acknowledging somewhere.
If there is no difference, then the source of the discrepancy is something else you have done differently compared to what PROCESS is doing.
Two-condition within-participant statistical mediation analysis: In this paper we discuss the estimation of the indirect effect and inference using bootstrapping and Monte Carlo confidence intervals. This paper also discusses parallel and serial multiple mediator versions of this model not originally addressed by Judd et al.
In my mediation analysis examining the direct and indirect effects of X on Y through M, the path from X to M or the path from M to Y is not statistically significant. Does this mean there is no way that M could mediate the relationship between X and Y. According to Baron and Kennyit cannot. Should I bother estimating the indirect effect in this case? The "criteria to establish mediation" approach popularized by Baron and Kenny is historically important but not consistent with modern practice and advice.
These days, we don't rely on statistical significance criteria as described in Baron and Kenny for the individual paths in a mediation model in order to assess whether a variable M functions as a mediator of the relationship between X and Y.
The pattern of significance or nonsignificance for individual paths in a mediation model is not pertinent to whether the indirect effect is significant.
You absolutely should estimate the indirect effect. I am interested in mediated moderation rather than moderated mediation. Do you have a macro for that?
As I discuss in my book on mediation analysis, in my opinion, mediated moderation is rarely very interesting or substantively interpretable. The same model can be conceptualized in terms of moderated mediation, and the results usually are more meaningful when you change your interpretative focus from the indirect effect of a product to the conditional indirect effects.
I recommend avoiding use of the term "mediated moderation" or any attempt to muster support for such a process. Although PROCESS can be used to construct the indirect effect of a product in a "mediated moderation" model, it turns out that this is equivalent to the index of moderated mediation, and moderation of mediation is much more interesting and substantively meaningful.
A discussion of the index of moderated mediation can be found in HayesMultivariate Behavioral Research. I don't see a discussion of this in your book. How can I tell if an indirect effect is moderated? The index of moderated mediation is available when the indirect effect is a linear function of a single moderator. If your model has more than one moderator, an indirect effect may be a function of two moderators simultaneously, in which case no index is provided.
For a discussion of various tests of moderated mediation in models with more than one moderator, see HayesCommunication Monographs. These are outdated concepts with little place in modern mediation analysis. They are based on the size and significance of the total and direct effects. All this information is in the output, but I recommend you avoid the use of these terms or interpreting your analysis based on the significance of the total and direct effects and whether the effect of X becomes nonsignificant after adding the mediator to the model.
For a discussion, see Hayes or Hayes and Rockwoodwhich you can download from here. I have missing data. It has no internal procedure for dealing with missing data other than listwise deletion. If the data file you are analyzing is tagged as derived from the multiple imputation routine, it will not analyze it and an error is likely to result The problem in SPSS is that the MATRIX language does not honor split file designations.
You can read about the bootstrap with multiple imputation in mediation analysis here. The mathematics for the derivation of the regions of significance are quite complicated and even impossible with more than a few groups. But check out a macro written by Amanda Montoya called OGRS that will find the boundary points for regions of significance using an iterative approach rather than a purely analytical one.
The derivations for the JN regions of significance for a conditional indirect effect discussed in Preacher, Rucker, and HayesMultivariate Behavioral Research assume the sampling distribution of the conditional indirect effect is normal. This is a faulty assumption, and the reason bootstrapping or some other method is preferred for inference about an indirect effect. In version 3, PROCESS produces information relevant to probing an interaction only when the corresponding interaction has a p-value of 0.
This is the default but, it can be changed to a larger or smaller value using the intprobe option. Use the mcx, mcw, or mcz options to specify one of these variables as multicategorical. How do I respond? There are advantages to using SEM, but some disadvantages as well.
The hard line position your reviewer or editor is taking is probably not consistent with his or her own behavior. Any OLS regression analysis is subject to the weaknesses discussed in Chapter 5 and Hayes, Montoya, and Rockwoodincluding bias in estimation of effects due to ignoring measurement error. Yet no doubt your critics have probably used OLS regression and have probably published their own work using it, with all its flaws.
And the editor has probably accepted papers with regression analyses in them. I articulate my position on this in Hayes, Montoya, and Rockwood I have been told I am supposed to mean center or standardize focal predictor and moderator prior to estimating an interaction. You can mean center if you want to, but doing so is not required.
For a discussion of why mean centering is a choice you can make rather than a requirement, the dangers of manually centering and standardizing, and some other myths about centering and standardization, see Introduction to Mediation, Moderation, and Conditional Process Analysis. As discussed in the documentation, PROCESS treats the group with the numerically smallest number on the multicategorical variable coding groups as the reference.
If you want a different group as the reference, recode your multicategorical variable so that your desired reference group has the numerically smallest code prior to running PROCESS. I believe X's effect on Y is moderated by two variables. How do I choose between model 2 and model 3?
Both model 2 and model 3 allow X's effect to depend on both W and Z. Furthermore, for both models, it is possible for X's effect on Y to be statistically significant only for some combinations of W and Z.
But there is an important constraint built into model 2 that does not exist in model 3. You would use model 2 if you want or predict or hypothesize the moderation of the effect of X on Y by W to be independent of Z.
That is, model 2 would be appropriate if you want the amount by which the effect of X on Y changes as W changes to be the same across values of Z. But if you feel that the moderation of X's effect on Y by W would or should depend on Z, then model 3 is appropriate.
Moderation by Z of the moderation by W of the effect of X on Y is "moderated moderation" or "three-way interaction," and this is set up and tested using model 3, not model 2. For a discussion of the mathematical distinction between these two models, see chapter 9 of the 2nd edition of Introduction to Mediation, Moderation, and Conditional Process Analysis.
My advisor tells me I should use the Baron and Kenny strategy for assessing mediation. MEMORE also provides an option that conducts pairwise contrasts between specific indirect effects in models with multiple mediators. Regression analysis and linear models: Concepts, application, and implementation. It can be downloaded from the book's web page and is documented in Appendix A of the book.
In addition to the usual regression program output, it has options for heteroscedasticity-consistent inference using either the HC0, HC1, HC2, HC3, or HC4 variance-covariance matrixautomatic coding of a multicategorical categorical regressor, options for estimating and probing interactions involving a multicategorical regressor, all subsets regression, spline regression, crossvalidation indices for the multiple correlation, an implementation of a limited form of dominance analysis, contrasts between regression coefficients, and a few other features.
An SPSS macro for multilevel mediation and conditional process analysis. Stay tuned for a paper or tutorial describing its use and application in A tutorial on estimating, visualizing and probing an interaction involving a multicategorical independent variable in linear regression analysis.
It was written by Amanda Montoya and is described, tested, and documented in Montoya The paper referenced above illustrates its use.
Here is a direct link to the OGRS files. The archive contains five folders, one for each macro. I no longer support or respond to questions about these macros. Quantifying and testing indirect effects in simple mediation models when the constituent paths are nonlinear. Multivariate Behavioral Research, 45, There is an error in the equation for Y-hat at the bottom of page of Preacher and Hayes This does not affect any of the computations anywhere in the manuscript.
Mediation and the estimation of indirect effects in political communication research. Lance Holbert EdsSourcebook for political communication research: