Dear STATS,

I received the following comment from a reviewer.

Could anyone give me more details on how to perform such analysis.
(Can be done with SPSS?)

The question of directionality could have been possibly explored
through further statistical analysis, e.g. mediation analysis (see
Baron &.Kenny (1986).The Moderator-Mediator Variable Distinction in
Social Psychology Research: Conceptual, Strategic, and Statistical
Considerations. Journal of Personality and Social Psychology 51 (6):
1173 - 1182

Thank you in Advance!

Chrys

On Apr 29, 11:06 am, Chrys <despinaa_de...@yahoo.com> wrote:
Dear STATS,

I received the following comment from a reviewer.

Could anyone give me more details on how to perform such analysis.
(Can be done with SPSS?)

The question of directionality could have been possibly explored
through further statistical analysis, e.g. mediation analysis (see
Baron &.Kenny (1986).The Moderator-Mediator Variable Distinction in
Social Psychology Research: Conceptual, Strategic, and Statistical
Considerations. Journal of Personality and Social Psychology 51 (6):
1173 - 1182

Thank you in Advance!

Chrys

I’m surprised no one has responded. I’ll try to help…

Here’s a straightforward example:

*Interaction term will tell you if there is a significant moderating
relationship.
------------------------------------------------------------------

If you think a variable is a mediator, you can run two models: one
with the mediator and one without. See if the relationship between the
IV and DV essentially disappears in the full model as compared to the
reduced model. If yes, then you may have mediation. This is an
oversimplification b/c there are levels of mediation.

BTW, bidirectionality makes me think of a nonrecursive model (feedback
loop). I don't believe that's what you are talking about.

Ryan

Ryan wrote:
On Apr 29, 11:06 am, Chrys <despinaa_de...@yahoo.com> wrote:
Dear STATS,

I received the following comment from a reviewer.

Could anyone give me more details on how to perform such analysis.
(Can be done with SPSS?)

The question of directionality could have been possibly explored
through further statistical analysis, e.g. mediation analysis (see
Baron &.Kenny (1986).The Moderator-Mediator Variable Distinction in
Social Psychology Research: Conceptual, Strategic, and Statistical
Considerations. Journal of Personality and Social Psychology 51 (6):
1173 - 1182

Thank you in Advance!

Chrys

I’m surprised no one has responded. I’ll try to help…

Perhaps no-one responded because they don't buy B&K's  "mediation"
analysis.  For example, see Bob Ling's review of Kenny's book
"Correlation and Causation" in JASA, 1982, Vol. 77, pp. 489-491.
(It's available through JSTOR.)



Here’s a straightforward example:

--- snip example of interaction, which B&K call "moderation" ---

*Interaction term will tell you if there is a significant moderating
relationship.
------------------------------------------------------------------

If you think a variable is a mediator, you can run two models: one
with the mediator and one without. See if the relationship between the
IV and DV essentially disappears in the full model as compared to the
reduced model. If yes, then you may have mediation. This is an
oversimplification b/c there are levels of mediation.

That sounds like "confounding" to me.

When I did some reading on mediation versus confounding a few
years ago, I concluded that the distinction between them is that a
mediating variable falls on a *suspected* causal pathway between X
and Y.  The webpage notes I was reading at the time (which are no
longer available) actually said "on the causal pathway". IMO, that
is far too strong a statement, because it suggests that there is
just one cause, and that it is known.  But many of the things
studied in the social & medical sciences have multiple causes.
Therefore, I changed the wording to "on A SUSPECTED causal
pathway":  "A" instead of "the", to signify that there may be more
than one cause; and "suspected", because in observational studies,
one cannot establish causation conclusively.

BTW, bidirectionality makes me think of a nonrecursive model (feedback
loop). I don't believe that's what you are talking about.

Ryan

--
Bruce Weaver
bwea...@lakeheadu.cawww.angelfire.com/wv/bwhomedir
"When all else fails, RTFM."- Hide quoted text -

- Show quoted text -

What you state above is consistent with my conceptual understanding.
If you believe that a mediating variable Z falls on "a suspected
causal path" between X1 and Y, how would you explore/test this in
SPSS? Would it not be similar to the previous mehtods mentioned
above?

My main problems with this are as follows:

1. It is my understanding that observational data are often used for
mediation analyses. Therefore, I would say "associated with", not
"caused by" in the descriptions of those regression equations.

2. If the association between X and Y changes when M is included in
the model, I would take that (as noted in my earlier post) as evidence
of confounding, not mediation.

Here's another question that occurred to me: What would mediation
analysts say if the absolute value of coefficient for X got *larger*
with M in the model? In terms of confounding, this would be an
example of "negative confounding", meaning that the crude (confounded)
effect underestimates the true effect.

In the end, I'm still left wondering what real evidence there is for a
causal pathway from X -> M -> Y when the data is purely observational,
and all you have is a difference between c and c' in the equations
described above.

Hi Bruce--see below--

In sci.stat.consult Bruce Weaver <bweaver@lakeheadu.ca> wrote:
My main problems with this are as follows:

1. It is my understanding that observational data are often used for
mediation analyses. Therefore, I would say "associated with", not
"caused by" in the descriptions of those regression equations.


And I think this is still generally the norm in mediational tests (using
association not 'caused.')

2. If the association between X and Y changes when M is included in
the model, I would take that (as noted in my earlier post) as evidence
of confounding, not mediation.

In terms of the underlying algebra, there's absolutely no difference
between mediation and confounding. The difference lies in your
hypothesized 'causal chain.' We hopefully know enough about what we're
modeling to distinguish the two on substantive grounds.


Here's another question that occurred to me: What would mediation
analysts say if the absolute value of coefficient for X got *larger*
with M in the model? In terms of confounding, this would be an
example of "negative confounding", meaning that the crude (confounded)
effect underestimates the true effect.

In the end, I'm still left wondering what real evidence there is for a
causal pathway from X -> M -> Y when the data is purely observational,
and all you have is a difference between c and c' in the equations
described above.


None, other than your own theory. You have to bring extra-statistical
arguments to bear, such as plausibility given what we know about the
phenomenon under study. That said, the argument is usually made that the
test is really one of disconfirmation, that is, finding NO relation among
the variables in question, or NO change in the x->y relation after
introducing m (the putative mediator), is taken as evidence of no
mediation. Finding that the variables are related and then that there is
some sort of substantial diminution of x->y after introducing b is only
interpreted as something like consistent with the possiblity of mediation.
Not airtight, by any stretch (for example, always potential for dozens of
lurking unmeasured variables, but isn't that almost always true?), but my
experience with the whole mediation/path analysis literature is that very
few few would venture to make any firm firm positive causal statement from
such analyses.

My own view is that it's not such an egregious a thing--I think that when
the vast majority of us put forth any sort of regression model (other than
the purely blind empirical predictive application) we're making a sort of
soft causal hypothesis.

Mike Babyak

Hi Bruce--see below--

In sci.stat.consult Bruce Weaver <bweaver@lakeheadu.ca> wrote:
My main problems with this are as follows:

1. It is my understanding that observational data are often used for
mediation analyses. Therefore, I would say "associated with", not
"caused by" in the descriptions of those regression equations.


And I think this is still generally the norm in mediational tests (using
association not 'caused.')


2. If the association between X and Y changes when M is included in
the model, I would take that (as noted in my earlier post) as evidence
of confounding, not mediation.

In terms of the underlying algebra, there's absolutely no difference
between mediation and confounding. The difference lies in your
hypothesized 'causal chain.' We hopefully know enough about what we're
modeling to distinguish the two on substantive grounds.


Here's another question that occurred to me: What would mediation
analysts say if the absolute value of coefficient for X got *larger*
with M in the model? In terms of confounding, this would be an
example of "negative confounding", meaning that the crude (confounded)
effect underestimates the true effect.

In the end, I'm still left wondering what real evidence there is for a
causal pathway from X -> M -> Y when the data is purely observational,
and all you have is a difference between c and c' in the equations
described above.


None, other than your own theory. You have to bring extra-statistical
arguments to bear, such as plausibility given what we know about the
phenomenon under study.

That said, the argument is usually made that the
test is really one of disconfirmation, that is, finding NO relation among
the variables in question, or NO change in the x->y relation after
introducing m (the putative mediator), is taken as evidence of no
mediation. Finding that the variables are related and then that there is
some sort of substantial diminution of x->y after introducing b is only
interpreted as something like consistent with the possiblity of mediation.
Not airtight, by any stretch (for example, always potential for dozens of
lurking unmeasured variables, but isn't that almost always true?), but my
experience with the whole mediation/path analysis literature is that very
few few would venture to make any firm firm positive causal statement from
such analyses.

My own view is that it's not such an egregious a thing--I think that when
the vast majority of us put forth any sort of regression model (other than
the purely blind empirical predictive application) we're making a sort of
soft causal hypothesis.

Mike Babyak

There is one thing that you guy have not mentioned,
which is a point that always struck me as important --
It is much easier to argue for "mediation", etc., when there
is a distinctive ordering in time.

Under "extra-statistical arguments", that is the most convincing.
Of course, when the direction is demonstrably wrong, there
is no "mediation" argument left at all.