Running regressions by decade when observations are by year

[Chinmay Sharma]

Hi All,

I currently have data that looks like the following:


* Example generated by -dataex-. To install: ssc install dataex
clear
input float(y x year)
23 2143 1980
23 431 1980
3 34 1980
23 134 1980
123 13 1981
12 4 1981
213 314 1981
123 31 1982
123 31 1982
12 4 1982
123 4 1982
123 4 1983
end
[/CODE]
Here, there is information on regressand y, and regressor x, by year. I wish to conduct regressions by decade. I have a panel of observations starting from 1850 onwards, so I want to conduct the regressions for 1850-1860, 1860-1870 and so on. The ultimate aim is to obtain predicted values for y, but based on coefficients that have been estimated by decade. The code I used in the past was to do this by year:

gen fitted=.
levelsof year, local(levels)
foreach i of local levels{
regress y x
predict temp
replace fitted=temp if year==`i'
drop temp
}

This however, estimates coefficients by year, which I need done by decade. One option is to group different decades by using the egen command with the group function, but I am at a loss at how to group by decade. Any suggestions are most welcome.

[Marcos Almeida]

Perhaps you wish to take a look at rolling window regression. Just type - help rolling - to check it out. This won't make it decade by decade, but each 10-yr period. Otherwise, creating groups by decade - and using it as predictors, instead of years - will do fine.

[Chinmay Sharma]

Thank you Marcos. I think I will first generate a decade variable:

gen decade = 10 * floor(year/10)

and thereafter, estimate using OLS by groups. Thanks!

[Nick Cox]

1850-1860 looks like 11 years to me, not a decade, and similarly for the other examples.

But consider some variant -- using rangestat (SSC) -- on

gen decade = 10 * floor(year / 10) 
rangestat (reg) y x, int(decade 0 0) 

On the whole, rolling regressions make more sense to me.

 
rangestat (reg) y x, int(year -9 0) 

For other examples of rangestat, look at its help file or search archives here.

An overview of rounding and binning will appear in Stata Journal 18(3).