I am trying to decide how best to use STATA to assess whether the quality of recorded music has declined since the advent of the internet/streaming services. To this end, I have used 'best of' lists of albums (i.e. 100 best albums of the 21st century, 500 greatest albums of all time) to collect data on the number of high-quality albums released each year, according to critics' opinion of whether albums surpass a certain threshold for inclusion in lists. I have so far collected data on 15 lists, ranging from a 10-year to 56-year span and from 100 items to 500 items. I plan to collect similar data for songs in the near future. I have had an attempt at deriving meaning from my data (see details below) but am quite adolescent to STATA and so would appreciate guidance on how best to go about this.
What I have done so far:
- Created a panel dataset: Yit = number of albums in given year (t) featured on given critics list (i) (i.e. number of albums released in 2014 on Billboards top 100 albums of the 2010s list). I have divided this raw value (i.e. 9) by the number of items on the list (i.e. 100) to derive the portion of the list allocated to said year. In order to make fractions from different years comparable (i.e. 2014 might have 9/100 albums on billboards list, for which each year has E(X)=10, and 4/100 on 'the guardians' list which covers 20 years with 100 items so E(X)=5) I have multiplied said fraction by the number of years the list spans (a uniform transformation? Am I allowed to do this?). Let's call my standardised variable Wis.
- Imported dataset into STATA. Assigned each list dummy variable an id number. Regressed Wis by year (OLS regression). Also created additional variable LnWis and regressed again. Am I right in thinking the coefficient on year in this second regression reflects the % change in Wis across years?
- Performed FEM regression (once by LSDV and once within transformation) on Wis and year. Coefficients/F-test imply individual effects do exist. Interestingly, F-test does not imply this when similar regression is performed with LnWis.
- Performed REM regression. Again, Breusch-Pagan implies individual effects exist for Wis but not for LnWis.
- Hausman test to assess FEM vs REM returned negative chi2 value and error that model fails to meet asymptotic assumptions, but xtoverid suggested FEM should be used.
On the above, is any of what I've done valid? Is there a better way to do this? Am I right in thinking FEM/panel data is the best approach to this?
Thanks for any help that you can give and best wishes to all!
0 Response to Regression for Assessing Quality of Recorded Music
Post a Comment