Dear Statalists,
before getting into detail, here‘s a short version of my question:
How can I demonstrate ib Stata that an independent variable, that is suspected to be correlated with unobservables, i.e. endogenous, is actually not correlated with the error term, i.e. not endogenous?
The common suggestion is to find an instrument and test the executed model via Hausman-Wu test to demonstrate that using an instrument is justified. However, in my case it’s very hard to get ex-post data that could build an instrumental variable for the respective variable.
My data comprises weekly sales of movie new releases that are released over several years and sales are recorded parallel for the first twelve weeks for each movie on three different sales channels. Most independent variables are fixed per movie and do not change per weekly data point (e.g. Production Budget, Genre, Critics Rating, etc.), only the weekly sales on channel C (= dependent variable) and the average price during that week (= independent variable, suspected to be endogenous).
I’ve set the data as panel (panel variable = movieID+channel; time variable: sales week) and calculated my base model as RE including fixed effects that control the sales channel, sales week and release year. Then I stored residuals and standard errors and tested the correlation of price with them:
xtreg ln_sales ln_price ln_budget ln_critics i.genre i.channel i.salesweek i.year, re
predict std_Error, stdp
predict resid, ue
pwcorr ln_price std_Error resid, sig
Price correlates with the SE -.0301 (p <.000) and Residual .0168 (p <.000)
As the correlation is significant I#d conclude that price is indeed endogenous. I also tried eregress in the following test constellation
eregress ln_sales ln_budget ln_critics i.genre i.channel i.salesweek i.year, endogenous (ln_price = ln_budget i.sweek i.yearc i.channel)
The output displays the following concerning the relation of price and sales: corr(e.ln_price, e.ln_sales) = 4.44e-17 (p = 1.000)
However I’m not sure whether it’s applicable in my case.
Thanks for any advice for my problem 😊
Nicolas
before getting into detail, here‘s a short version of my question:
How can I demonstrate ib Stata that an independent variable, that is suspected to be correlated with unobservables, i.e. endogenous, is actually not correlated with the error term, i.e. not endogenous?
The common suggestion is to find an instrument and test the executed model via Hausman-Wu test to demonstrate that using an instrument is justified. However, in my case it’s very hard to get ex-post data that could build an instrumental variable for the respective variable.
My data comprises weekly sales of movie new releases that are released over several years and sales are recorded parallel for the first twelve weeks for each movie on three different sales channels. Most independent variables are fixed per movie and do not change per weekly data point (e.g. Production Budget, Genre, Critics Rating, etc.), only the weekly sales on channel C (= dependent variable) and the average price during that week (= independent variable, suspected to be endogenous).
I’ve set the data as panel (panel variable = movieID+channel; time variable: sales week) and calculated my base model as RE including fixed effects that control the sales channel, sales week and release year. Then I stored residuals and standard errors and tested the correlation of price with them:
xtreg ln_sales ln_price ln_budget ln_critics i.genre i.channel i.salesweek i.year, re
predict std_Error, stdp
predict resid, ue
pwcorr ln_price std_Error resid, sig
Price correlates with the SE -.0301 (p <.000) and Residual .0168 (p <.000)
As the correlation is significant I#d conclude that price is indeed endogenous. I also tried eregress in the following test constellation
eregress ln_sales ln_budget ln_critics i.genre i.channel i.salesweek i.year, endogenous (ln_price = ln_budget i.sweek i.yearc i.channel)
The output displays the following concerning the relation of price and sales: corr(e.ln_price, e.ln_sales) = 4.44e-17 (p = 1.000)
However I’m not sure whether it’s applicable in my case.
Thanks for any advice for my problem 😊
Nicolas
