Interpreting Coeficients in Standard Deviations

[Edgar Sanchez]

Hello Everyone,

I am new in Statalist and I have a question related to intrepreting betas in standard deviations in a 2SLS with id and year fixed effects, and clusterized at id level. In that sense I am using the command "reghdfe".

My regression is:

reghdfe outcome $controls (mainx=instrument) if ano>2008 & ano<2014, absorb(id year) vce(cluster id) ivsuite(ivreg2) stages(first)

I know that for interpreting coeficients in standard deviations I have to z standarize all my variables (outcome mainx controls instrument) by removing the mean and dividing by the standard deviation; however, some of my variables must be non_negative (population, homicides per 100,000 inhabitants, military bases interacted to US GDP, etc.) and when I standarize them I get negative values (below mean) and I cannot generate variables like log(variable).

Besides, since the command "reghdfe" doesn't have an option like "beta" for reg or "listcoef" for logit, I have no idea how to express the coefficient in terms of standard deviations.

Thanks to everyone,

Best Regards


Édgar Hernando Sánchez Cuevas
Research Assistant
Faculty of Economics
Universidad de los Andes
Bogota, Colombia

[Doug Hemken]

If you *first* standardize your variables, your coefficients are *already* in units of SD. But, as you have noted, fully standardizing everyything in your model might not actually give you a result with a clear interpretation. Mathematically,what you have is a model expressed with a change of basis - this is just a linear transformation of your original model.

[Edgar Sanchez]

Thanks for your reply professor Hemken,

I have found that an alternative way of interpreting the coefficients in standard deviations is standarizing the directly the beta by multiplying it by the relative standard deviation:

beta*[std. dev (X_i)/std. dev (Y)]

I was wondering whether the beta obtained by doing that is identical to the one obtained by z_standarizing all the variables and running the same regression.
Again, thanks for your response.

Best regards,

Édgar Hernando Sánchez Cuevas
Research Assistant
Faculty of Economics
Universidad de los Andes
Bogota, Colombia

[Doug Hemken]

Try running “stdBeta” (you’ll have to install the package) as a post-estimation command after reghdfe. It will perform the z_standardization for you.

[Edgar Sanchez]

Thanks a lot professor Hemken. It works perfectly and the coefficient is identical to the one calculated by doing:

beta*[std. dev (X_i)/std. dev (Y)]

Once again, thank you very much,

Best regards,

Édgar Hernando Sánchez Cuevas
Research Assistant
Faculty of Economics
Universidad de los Andes
Bogota, Colombia

[Xiaoke Ye]

Try running “stdBeta” (you’ll have to install the package) as a post-estimation command after reghdfe. It will perform the z_standardization for you.

Hi Prof Hemken
When I tried to use "stdBeta", I get the following statement:
"Failure to specify nodepvar where needed can produce meaningless results.", I have no idea what is meants

[Federico Tedeschi]

"Failure to specify nodepvar where needed can produce meaningless results.", I have no idea what is meants

It means: "Please check that you really want to standardize also the dependent variable, instead of the independent variables only". By reading the ".ado" file, I would expect such warning to be received only in case of binary outcomes, but it's not the case (for example, I always get it if I use "stdbeta" after the "mixed" command: I don't know whether it depends by some options I always use, or it's always given after running "mixed" followed by "stdBeta").