Two points:
1) Are you aware of new work on mediation by, e.g., MacKinnon? e.g Annual Review Pschology, this web page and this book?
Eseentially, MacKinnon treats mediation as existing on a continuum rather than being either present or absent (which is what Baron and Kenny's old article does).
2) You can include the covariates in the regression equations and proceed with either approach. There is no need to split the data for the categorical variable.
I can't help with SPSS, but it should be straightforward. If you still have SPSS problems with this, however, the better site is StackOverflow.
The general framework for mediation analysis is to run three models:
- Model 1: a * IV -> DV (to establish some kind of total effect)
- Model 2: b * IV -> MV (to establish some effect on the mediator)
- Model 3: c * IV + d * MV -> DV
where a to d stand for regression coefficients.
The first thing to check for in mediation is evidence of an indirect effect of A on C. The path for that is b (from IV to MV) multiplied by d (from MV to DV with IV in the model). If there is no indirect effect, there is no mediation. If b * d is not statistically significant, then there is not sufficient evidence to detect the mediation effect. I will call b * d, ie for indirect effect: $ie = b \times d$.
Because ie comes from two models, models 2 and 3, we need some method to test whether it is statistically significant. The more traditional approach is the Sobel test, which is just a specialized form of the delta method, which is a first order Taylor Series expansion. The assumption built into this method is the normality of ie. You will often need a large sample size for this assumption to be satisfied. So these days, it is more common to perform bootstrapping. Within each resample, retrieve b and d from Models 2 and 3 respectively, multiply them to get ie, and obtain the confidence interval for ie at the end. Since you use SPSS, there is a macro called Process that implements bootstrapping for the indirect effect.
So if your indirect effect is not statistically significant, you can end your talk of mediation.
If it is statistically significant, it is also reasonable to expect that a in Model 1 would be statistically significant, suggesting evidence of a total effect. The same goes for b suggesting that a at least influences your mediator. Also d suggesting that your mediator influences the outcome. The question then becomes: which type of mediation do you have: partial or complete?
Here's the standard approach:
- If c is statistically significant in the presence of an indirect effect, then the relationship from the IV to the DV persists in the presence of the mediation effect, so we only have partial mediation.
- If c is not statistically significant in the presence of an indirect effect, then the relationship from the IV to the DV vanishes in the presence of the mediation effect, so we have complete mediation.
In your particular situation, I do not see the indirect effect, so it is difficult to begin the conversation about mediation. I will assume that your Steps 3 and 4 are my Model 3. If that is true, then $b=.435$ and $d=.231$, so it is possible to conduct the Sobel test.
With the Sobel test, the formula for the standard error of the indirect effect is: $$\sqrt{b^2 \times se_{d}^2 + d^2 \times se_{b}^2}$$ In your case, $se_d = d/t_d= .231/2.484=0.092995$ and $se_b = b/t_b= .435/4.4=0.09886$. So the standard error of your indirect effect is: $$\sqrt{.435^2 \times 0.092995^2 + .231^2 \times 0.09886^2}=0.04645$$ Your indirect effect will be: $.435\times .231=0.100$. So your $t$-statistic for the indirect effect will be $.1/.04645=2.15$.
I hope this helps. Personally, I doubt the results of mediation analyses. I think it is a causal analysis, so when estimated using ordinary least squares (as you are doing), it is plagued by omitted variable bias. So I do not believe any of these results generally.
Best Answer
No, this is not possible.
An ANOVA aims to test differences between means of different groups/conditions on one independent variable. E.g. DV: number of strawberries/strawberry plant; IV1: three watering groups (100ml/day, 200ml/day, 300ml/day); IV2: three fertilizer groups (5g/day, 10g/day, 15g/day). Thus, you'd have a 3x3 factorial ANOVA without repeated measures, which you could turn into a 3x3 repeated measures ANOVA by measuring plant height on several points in time (e.g. after 7, 14, and 21 days). An ANCOVA aims to do the same, but in this case you introduce at least one continuous variable, a so-called covariate, for which you do not specify an interaction with the other factors. This reduces error variance in your IV (plant height) due to this covariate and thereby increases statistical power for the effects of your IVs. Yet, you could as well set up AN(C)OVAs as regression models instead (for which you'd have to code your factorial IVs differently), although in some statistics courses ANOVAs are taught differently, i.e. by formulas which actually compare different parts of the total variance in your DV (variance within treatment groups and variance between treatment groups), in order to receive an F-value, which lets you determine a corresponding p-value from the F-distribution. Well I am going into details. The important part is: AN(C)OVAs compare means!
Yet, a mediation analysis has a totally different aim. Here, you test whether your IV (e.g. water/day) predicts your DV (number of strawberries/plant). This, so far, isn's any different from the proceeding layed out above and you might predict mean number of berries/plant from the watering conditions. All of this you could analyse in an ANOVA But now comes the critical part: Here you are also interested in whether the effect of your IV changes the value of your DV mediated by some other value. E.g. you could reason that water/day affects the height of your strawberry plants, thereby leading to variation in sun exposure of the plants, and therefore to variation in berries/plant (For the simplicity of the example let's assume plant height and sun exposure to be one combined variable). This effect (water/day -> plant height|sunexposure -> berries/plant) is called the indirect effect, while the effect water/day -> berries/plant is called the direct effect. Both the direct and indirect effect should be tested for significance with specialized models.
I guess that your intention to test three ANOVAs could be inspired by the model by Baron&Kenny, which examplifies the logic of mediation analyses. Here, you test three mediations (IV->DV, IV->M, and M->DV). Yet, there are more appropriate procedures based on bootstrapping and I would suggest you to use the PROCESS-Macro by Andrew F. Hayes, which is available for both SPSS and SAS and furthermore includes a nice GUI for SPSS (I do not know about SAS). Furthermore, there is a nice template pdf available for many different mediation (and moderation) analyses, which allows you to easily figure out which model is appropriate for you. I hope this helps. Oh, and Andrew Hayes has also published a book on mediation and moderation. Check whether you have access to this book via an institution with which you are associated. It explains quite nicely what happens in mediation and moderation.