Shift in sign

[Serena Menny]

Dear members,

This table presents the marginal effects derived from probit regressions, where the dependent variable, sustainable_investment, is a binary indicator set to 1 if the investor holds a sustainable fund, and 0 otherwise. The key explanatory variables are defined as follows:

Sustainable_concerns: An average score derived from three items, each measured on a 7-point Likert scale, reflecting the level of concern individuals have regarding sustainability issues.
Sustainable_behavior: An average score derived from two items, each measured on a 7-point Likert scale, indicating the extent to which individuals engage in recycling and other environmentally protective behaviors.
Sustainable_vote: A single item measured on a 7-point Likert scale, assessing whether individuals vote for political parties with sustainable programs.

What could possibly explain the shift in the sign for sustainable_concerns?

The three variables—sustainable_concerns, sustainable_behavior, and sustainable_vote—reflect general attitudes towards sustainability issues and are not directly related to sustainable funds. Additionally, three other variables specifically related to sustainable funds are included, each measured on a 7-point Likert scale (sustainable_attitude, confidence, and impact).

After including the new variables related to sustainable funds, sustainable_concerns becomes significant while sustainable_behavior loses its significance.

[George Ford]

not surprising. everything is positively correlated (and the effects will be influenced not by var2var correlation but multivariate correlation). This is a good example of why univariate regressions are iffy.

I'd do some research about treating the Likert variables as continuous. It is done, but might make more sense to treat them as ordinal (i.x). That may soften up the correlation (or might not).

also run it with concerns & behavior, then with concerns & vote to see which one is eating your concern variable, or if it's both.

[Mike Lacy]

I share much of George's views, but I'd add that a "suppressor" effect such as you see often arises from your variable of interest having direct and indirect effects that go in opposite directions. However, in that case, it would be likely (certain??) that at least some of the bivariate correlations you show would be negative and some would be positive, but that's apparently not true. Are you sure that no minus signs were accidentally removed from your correlation matrix?

[Serena Menny]

"and the effects will be influenced not by var2var correlation but multivariate correlation" what does this mean if I might ask?

"also run it with concerns & behavior, then with concerns & vote to see which one is eating your concern variable, or if it's both." When I analyze concerns and behavior alone or concerns and voting alone, concerns stop being significant but retain a positive sign. However, when I include behavior or voting as the third variable in the analysis, the sign of concerns flips to negative. So I guess it is both of them.

"Are you sure that no minus signs were accidentally removed from your correlation matrix?" Yes, I checked again just now, and the correlations are correct. This means I have no suppressor effect?

Is it possible to keep the model, or are these significant flaws? Because I don't know how to explain that.

[Mike Lacy]

I think you'll get a better idea of what's going with your data if you explicitly analyze it for mediation, i.e., direct and indirect effects. Stata's built-in -mediate- command, and its associated documentation (-help mediate-) should help you think about your situation. Even if you don't do a formal mediation analysis, thinking in those terms should be useful.

[George Ford]

might try factor analysis (tetrachoric) to see if all these sustainable variables load to one factor. I suspect they will. Looks like many variables all being driven by some "sustainability' latent factor.

mediation is an interesting possibility.

as for #4, my point is that the correlation coefficient between two variables does not tell you everything about the correlation structure when you have multivariate regression. you see this in the way the coefficients change as you add variables.

I think the biggest problem here relates to timing, as I've mentioned before. this problem is not cross sectional. investment takes time, and intent may be reflected years later in investments.

[Serena Menny]

Many thanks, Mike and George, for your insights. Your responses are greatly appreciated!