Mediation testing with -medeff-: levels for continuous treatment

[Florian Zapkau]

Dear Statalist members,

I am using Stata 14.2 on MacOS to test a mediator model where the treatment (IV) is continuous, the mediator (M) is binary and the outcome (DV) is continuous. I use the user-written command -medeff- to estimate whether the indirect effect (IV on DV via M) is statistically significant. -medeff- is part of the Mediation package from SSC and documented in the following article in the Stata Journal:

Hicks, R., & Tingley, D. (2011). Causal mediation analysis. Stata Journal, 11(4), 605-619.

However, I am not sure (and sincerely hope that I haven't missed it somewhere) whether the IV needs to be a binary variable. The -medeff- command requires users to set "levels" of the treatment variable (i.e., control and treatment condition). If nothing is specified, the command assumes that the levels of the treatment are 0 and 1. However, my treatment variable (IV) can take values from 1 to 7 and everything in between, as the variable was measured with multi-items each on a Likert scale from 1 to 7.

I have estimated three different sets of models (x1-3 are control variables):

1) Using my original treatment variable (without specifying the levels for the treatment, thus, the command assumes [0 1]

medeff (logit M IV x1 x2 x3) (regress DV IV M x1 x2 x3), treat(IV) mediate(M)

[output omitted]

------------------------------------------------------------------------------------
        Effect                 |  Mean           [95% Conf. Interval]
-------------------------------+----------------------------------------------------
        ACME                   |  .0002402     -.0203698      .0240611
        Direct Effect          |  .1552511     -.0003194      .3100036
        Total Effect           |  .1554913      .0033906      .3079827
        % of Tot Eff mediated  |  .0014808      .0006634      .0110117
------------------------------------------------------------------------------------

2) Using a normalized treatment variable that is in the range between 0 and 1

medeff (logit M IV_normalized x1 x2 x3) (regress DV IV_normalized M x1 x2 x3), treat(IV_normalized) mediate(M)

[output omitted]

------------------------------------------------------------------------------------
        Effect                 |  Mean           [95% Conf. Interval]
-------------------------------+----------------------------------------------------
        ACME                   | -.0940051     -.2580492      .0304379
        Direct Effect          |  .8625058     -.0017743      1.722242
        Total Effect           |  .7685007     -.0944182      1.656022
        % of Tot Eff mediated  | -.1130681     -.9799373      .6520454
------------------------------------------------------------------------------------

3) Original treatment variable, but I specified treatment levels 1 to 7 (and tried other levels as well) leading to an error message

medeff (logit M IV x1 x2 x3) (regress DV IV M x1 x2 x3), treat(IV [1 7]) mediate(M)

r(3000)
nothing found where subexp expected

Interestingly, the regressions in all three models (which -medeff- estimate) yield the exact same results in terms of Chi2-values (logit) and F-values (OLS) and the indirect effect (ACME) seems to be non-signifikant (CI include 0) in 1) and 2), while 3) yields only an error message.

Long story short: Can I use -medeff- for my model after all and if not, what would be alternatives (-paramed- requires also levels for the treatment and -gsem- cannot distinguish the direct/indirect effect)?

My apologies for the long text and thank you in advance for any suggestions on this matter!

Kind regards,
Florian

[Steve Samuels]

Added: The square brackets are not to be typed: they indicate that what's inside is optional. See, for example, the help for regress.

The program expects one or three terms in the treat() option; the required term is the treatment variable; the optional ones are the "to" and "from" values. So specify

treat(IV 1 7)

[Florian Zapkau]

Added: The square brackets are not to be typed: they indicate that what's inside is optional.

So specify

treat(IV 1 7)

Thank you for the hint, Steve - that has worked!

The results of my set of models 2 (normalized treatment variable) and 3 (treat 1 7) are rather similar (which makes sense, as in both cases the difference in levels comprises the full range of the treatment scale).

I will still ask the authors of -medeff- whether their command can be used for models with a continuous treatment to be on the safe side before I report the results somewhere.

[Lisa Spoelstra]

Dear Florian, I came to this topic because I also work with a continuous treatment variable. I was wondering, did you ask the authors if you can use the model for a continuous treatment variable and what was their response?

[Florian Zapkau]

Dear Lisa, sorry for the belated response. Unfortunately, I haven't contacted them, as the research project was halted for other reasons in the meantime. If you find a solution, I would be pleased to hear about it!

[Karin Freiberg]

Dear Lisa, did you by any chance find support for having a continuous treatment variable in medeff?

[xiao xiao]

Dear Lisa, did you by any chance find support for having a continuous treatment variable in medeff?

Dear Karin, did you find any support for having a continuous or discrete treatment used in medeff?