Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: algorithmic question : running sum and computations


From   Nick Cox <njcoxstata@gmail.com>
To   Francesco <k7br@gmx.fr>
Subject   Re: st: algorithmic question : running sum and computations
Date   Fri, 17 Aug 2012 13:55:23 +0100

I don't have easy advice on this. As I understand it sorting on

id product (date)

can't distinguish between

id 1  product A date 42  quantity 12
id 1  product A date 42  quantity -12
id 1  product A date 42  quantity 21
id 1  product A date 42  quantity -21

and

id 1  product A date 42  quantity 12
id 1  product A date 42  quantity -21
id 1  product A date 42  quantity 21
id 1  product A date 42  quantity -12

In the first case you have two spells to 0, and in the second one
spell to 0. Your example shows that spells need not be two
observations long, so I don't know what to suggest.

Nick

On Fri, Aug 17, 2012 at 1:45 PM, Francesco <k7br@gmx.fr> wrote:
> Actually Nick there is only a slight problem : dates could be repeated
> for the same individual AND the same product  : for example there
> could be several round trips during the same day for the same
> product... In that case I would consider that there are as many
> delta_Date equal to zero as different round trips during the day for a
> particular product... My apologies I did not think of this particular
> and important case...
>
> Could the trick  egen panelid = group(id product) be adapted in that case ?
>
> Many thanks
> Best Regards
>
> On 17 August 2012 13:58, Francesco <k7br@gmx.fr> wrote:
>> Many, Many thanks Nick and Scott for your kind and very precise
>> answers! Spells is indeed what I needed ;-)
>>
>>
>> On 17 August 2012 13:43, Nick Cox <njcoxstata@gmail.com> wrote:
>>> Using your data as a sandpit
>>>
>>> .  clear
>>>
>>> .  input id    date str1 product quantity
>>>
>>>             id       date    product   quantity
>>>   1.  1       1           A           10
>>>   2.  1       2           A           -10
>>>   3.  1       1           B            100
>>>   4.  1       2           B            -50
>>>   5.  1       4           C            15
>>>   6.  1       8           C            100
>>>   7.  1       9           C            -115
>>>   8.  1      10          C            10
>>>   9.  1      11          C            -10
>>>  10.  end
>>>
>>> it seems that we are interested in the length of time it takes for
>>> cumulative quantity to return to 0. -sum()- is there for cumulative
>>> sums:
>>>
>>> .  bysort id product (date) : gen cumq = sum(q)
>>>
>>> In one jargon, we are interested in "spells" defined by the fact that
>>> they end in 0s for cumulative quantity. In Stata it is easiest to work
>>> with initial conditions defining spells, so we negate the date
>>> variable to reverse time:
>>>
>>> .  gen negdate = -date
>>>
>>> As dates can be repeated for the same individual, treating data as
>>> panel data requires another fiction, that panels are defined by
>>> individuals and products:
>>>
>>> .  egen panelid = group(id product)
>>>
>>> Now we can -tsset- the data:
>>>
>>> .  tsset panelid negdate
>>>        panel variable:  panelid (unbalanced)
>>>         time variable:  negdate, -11 to -1, but with a gap
>>>                 delta:  1 unit
>>>
>>> -tsspell- from SSC, which you must install, is a tool for handling
>>> spells. It requires -tsset- data; the great benefit of that is that it
>>> handles panels automatically. (In fact almost all the credit belongs
>>> to StataCorp.) Here the criterion is that a spell is defined by
>>> starting with -cumq == 0-
>>>
>>> .  tsspell, fcond(cumq == 0)
>>>
>>> -tsspell- creates three variables with names by default _spell _seq
>>> _end. _end is especially useful: it is an indicator variable for end
>>> of spells (beginning of spells when time is reversed). You can read
>>> more in the help for -tsspell-.
>>>
>>> .  sort id product date
>>>
>>> .  l id product date cumq _*
>>>
>>>      +---------------------------------------------------+
>>>      | id   product   date   cumq   _spell   _seq   _end |
>>>      |---------------------------------------------------|
>>>   1. |  1         A      1     10        1      2      1 |
>>>   2. |  1         A      2      0        1      1      0 |
>>>   3. |  1         B      1    100        0      0      0 |
>>>   4. |  1         B      2     50        0      0      0 |
>>>   5. |  1         C      4     15        2      3      1 |
>>>      |---------------------------------------------------|
>>>   6. |  1         C      8    115        2      2      0 |
>>>   7. |  1         C      9      0        2      1      0 |
>>>   8. |  1         C     10     10        1      2      1 |
>>>   9. |  1         C     11      0        1      1      0 |
>>>      +---------------------------------------------------+
>>>
>>> You want the mean length of completed spells. Completed spells are
>>> tagged by _end == 1 or  cumq == 0
>>>
>>> .  egen meanlength = mean(_seq/ _end), by(id)
>>>
>>> This is my favourite division trick: _seq / _end is _seq if _end is 1
>>> and missing if _end is 0; missings are ignored by -egen-'s -mean()-
>>> function, so you get the mean length for each individual. It is
>>> repeated for each observation for each individual so you could go
>>>
>>> . egen tag = tag(id)
>>> . l id meanlength if tag
>>>
>>> I wrote a tutorial on spells.
>>>
>>>  SJ-7-2  dm0029  . . . . . . . . . . . . . . Speaking Stata: Identifying spells
>>>         . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  N. J. Cox
>>>         Q2/07   SJ 7(2):249--265                                 (no commands)
>>>         shows how to handle spells with complete control over
>>>         spell specification
>>>
>>> which is accessible at
>>> http://www.stata-journal.com/sjpdf.html?articlenum=dm0029
>>>
>>> Its principles underlie -tsspell-, but -tsspell- is not even
>>> mentioned, for which there is a mundane explanation. Explaining some
>>> basics as clearly and carefully as I could produced a paper that was
>>> already long and detailed, and adding detail on -tsspell- would just
>>> have made that worse.
>>>
>>> For more on spells, see Rowling (1997, 1998, 1999, etc.).
>>>
>>> Nick
>>>
>>> On Fri, Aug 17, 2012 at 11:30 AM, Francesco <cariboupad@gmx.fr> wrote:
>>>> Dear Statalist,
>>>>
>>>> I am stuck with a little algorithmic problem and I cannot find an
>>>> simple (or elegant) solution...
>>>>
>>>> I have a panel dataset as (date in days) :
>>>>
>>>> ID    DATE    PRODUCT QUANTITY
>>>> 1       1           A           10
>>>> 1       2           A           -10
>>>>
>>>> 1       1           B            100
>>>> 1       2           B            -50
>>>>
>>>> 1       4           C            15
>>>> 1       8           C            100
>>>> 1       9           C            -115
>>>>
>>>> 1      10          C            10
>>>> 1      11          C            -10
>>>>
>>>>
>>>>
>>>> and I would like to know the average time (in days) it takes for an
>>>> individual in order to complete a full round trip (the variation in
>>>> quantity is zero)
>>>> For example, for the first id we can see that there we have
>>>>
>>>> ID PRODUCT delta_DATE delta_QUANTITY
>>>> 1         A               1=2-1                  0=10-10
>>>> 1         C               5=4-9                  0=15+100-115
>>>> 1         C               1=11-10               0=10-10
>>>>
>>>> so on average individual 1 takes (1+5+1)/3=2.3 days to complete a full
>>>> round trip. Indeed I can discard product B because there is no round
>>>> trip, that is 100-50 is not equal to zero.
>>>>
>>>> My question is therefore ... do you have an idea obtain this simply in
>>>> Stata ? I have to average across thousands of individuals... :)
*
*   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/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index