Re: st: Re: rolling and ts operators

[Nick Cox <njcoxstata@gmail.com>]

That's what I thought too, although I'd guess at the problem being a
missing value rather than something being empty. But how come the use
of f. doesn't exclude the last observation in the window too? This is
a lazy post, because I have not looked at the code or done any
experimentation.

Nick

On Tue, Oct 25, 2011 at 11:19 AM, Richard Herron
<richard.c.herron@gmail.com> wrote:
> -rolling- passes seven observations to -regress-, but when -regress-
> creates the regressors, the one with the lag operator in the first
> observation for each individual is empty. -regress- still runs the
> regression, but can only use six observations since the first one is
> empty.
>
> It doesn't have to do with the number or linearity of regressors.
> Maybe it's illustrative to look at that first seven observation window
> and see that the first lagged value is missing and that -regress- has
> N = 6.
>
> * begin code
> webuse grunfeld, clear
> xtset company year
> list company year invest mvalue l.mvalue in 1/7
> regress invest mvalue l.mvalue in 1/7
> * end code
>
> On Tue, Oct 25, 2011 at 05:12, Nuno Soares <ndsoares@gmail.com> wrote:
>> Dear Richard,
>>
>> Thank you so much! I didn't know - rolling - was panel data aware!
>> It's much more efficient than the code I have.
>>
>> However, although your insights solve my problem, the main question
>> still remains: why does Stata use 6 observations and not 7....
>>
>> If I just tweak your code a litle bit (to include 5 indep variables as
>> in my case I squared mvalue);
>>
>> * start code *
>>
>> webuse grunfeld, clear
>> rename company id
>> gen mvaluesq=mvalue^2
>> xtset id year, year
>> gen end = year
>> tempfile stats
>> rolling _b N = e(N), window(7) recursive saving(`stats', replace) ///
>>    : reg invest mvalue l.mvalue f.mvalue kstock mvaluesq
>> merge 1:1 id end using `stats', nogenerate
>> xtset id year, year
>>
>> * calculating error term *
>> gen e = invest - (_b_cons + _b_mvalue*mvalue + _stat_2*l.mvalue ///
>>    + _stat_3*f.mvalue + _b_kstock*kstock+_b_mvaluesq*mvaluesq)
>>
>> * end code *
>>
>> If we just check the data we'll notice that N starts with 6
>> observations, which doesn't make sense... I must be missing
>> something...
>>
>>
>> Best wishes,
>>
>> Nuno
>>
>>
>> Re: st: rolling and ts operators
>>
>> From      Richard Herron <richard.c.herron@gmail.com>
>> To        statalist@hsphsun2.harvard.edu
>> Subject   Re: st: rolling and ts operators
>> Date      Mon, 24 Oct 2011 09:50:16 -0400
>>
>>  It may be easiest to think of -rolling- and -regress- in two steps.
>> First -rolling- grabs a window of 7 observations and hands off these
>> data to -regress-. Then -regress- regresses, but is only able to use 6
>> of these observations because in the first window the lagged value
>> isn't available.
>>
>> If you add -N = e(N)- to your -rolling- command you can see how many
>> observations -regress- used.
>>
>> Also, there is no need to use loops with -rolling-; it is panel aware
>> (at least in 11.2). The following should help with the -N = e(N)- bit.
>>
>> webuse grunfeld, clear
>> rename company id
>> keep if id == 2 // just a test firm
>> xtset id year
>> gen end = year
>> tempfile stats
>> rolling _b N = e(N), window(7) recursive saving(`stats', replace) ///
>>    : reg invest mvalue l.mvalue f.mvalue kstock
>> merge 1:1 id end using `stats', nogenerate
>> xtset id year
>>
>> * calculating error term *
>> gen e = invest - (_b_cons + _b_mvalue*mvalue + _stat_2*l.mvalue ///
>>    + _stat_3*f.mvalue + _b_kstock*kstock)
>>
>>
>>
>> On 24 October 2011 13:52, Nuno Soares <ndsoares@gmail.com> wrote:
>>> Hi everyone,
>>>
>>> I'm having a problem with the - rolling - command the the use of time
>>> series operators. I need to estimate a model for several firms (id)
>>> and several years in a rolling regression with a window of at least 7
>>> years and the recursive option.
>>>
>>> The code I'm using is the following:
>>>
>>> use "test.dta", clear
>>> keep if id==2 * just a test firm
>>> xtset id year
>>> gen end=year
>>> tempfile stats
>>> levelsof id, local(ids)
>>> foreach id of local ids {
>>>        cap rolling , window(7) recursive saving(`stats', replace): reg
>>> depvar indepvar1 l.indepvar1 f.indepvar1 indepvar2 indepvar3 if
>>> id==`id'
>>>        cap merge 1:1 id end using "`stats'", update replace
>>>        cap drop _merge
>>> }
>>> sort id year
>>> xtset id year
>>> * calculating error term *
>>> gen e=depvar-(_b_cons+_b_indepvar1*indepvar1+_b_stat2*l.indepvar1+_b_stat3*f.indepvar1+_b_indepvar2*indepvar2+_b_indepvar3*indepvar3)
>>>
>>>
>>> Given the l. operator I was expecting that Stata would use 8
>>> observations given that the first l.indepvar1 would be missing an thus
>>> it wouldn't have enough observations to estimate the first model with
>>> only six observations (while the window option is set at 7,
>>> restricting the sample to the first 7 observations would lead to have
>>> only 6 eligible observations, and thus not enough data to estimate the
>>> regression). This however isn't happening with Stata reporting
>>> estimated coefficient starting from the first 7 observations. Any
>>> clues on why this is happening?
>>>

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/