st: Fwd: tscollap query

[Christopher F Baum <baum@bc.edu>]

Begin forwarded message:

> From: Restrepo J <J.Restrepo@rhul.ac.uk>
> Date: January 27, 2004 6:57:53 AM EST
> To: Christopher F Baum <baum@bc.edu>
> Subject: RE: tscollap query
>
> This is great! Thank you very much for your help and directions.
> Jorge 
>
> -----Original Message-----
> From: Christopher F Baum [mailto:baum@bc.edu] 
> Sent: 27 January 2004 02:09
> To: Restrepo J
> Subject: Re: tscollap query
>
> Dear Jorge
>
> Thanks for the clarifying information. I don't think there's anything fancy
> needed here. Say that your 19,000 records of raw data are coded with the
> calendar day of the event and the location. Then create three new Stata date
> variables: the week in which the event took place, the month, the year.
> Those are all simple applications of Stata's date functions. Then you can
> readily collapse the raw data on any one of those indicators, as you say,
> independently of the location (for the whole country). That can be done with
> just the collapse command. If you wanted to generate a similar
> lower-frequency time series for each region, you could collapse on both the
> date variable and the region indicator. So it is not really an issue of
> reducing the frequency of a variable (which is what tscollap does--changing
> monthly inflation to quarterly or annual, let's say) but rather grouping the
> raw data and averaging them (or summing them, or counting them) over a lower
> data frequency -- really a matter of "binning", but binning by calendar date
> intervals. Similar to the kind of work that people do with hospital-stay
> data, which can be aggregated over time, regions, etc.
>
> Best wishes
> Kit
>
> On Jan 26, 2004, at 3:52 PM, Restrepo J wrote:
>
> > Thank you very much for your prompt reply! Maybe a short explanation 
> > of my dataset would clarify my question.
> >
> > This is a crime dataset that I collected with other colleagues over 
> > the last year. It includes more than 19 thousand events for which we 
> > have around 20 variables, including the day in which the event took 
> > place and the location.
> > I would need to associate each event with a time period and a spatial 
> > code, which I plan to do with -tabulate-. What I really want then is 
> > to generate a panel from those events for n regions and t months (or, 
> > say, days).
> >
> > My problem is the degree of "aggregation" for each one of these 
> > dimensions.
> > Regarding time, I need to reduce the frequency of variables using sum 
> > and count in order to obtain, dayly, weekly, monthly, and yearly 
> > variables independently of the location (for the whole country). Later 
> > I can generate panels by region and month, say.
> >
> > Would then it be possible to use tscollap or a modification of it to 
> > generate daily and monthly variables from the initial time code in 
> > each event? Is there another command to "generate" panels from event data?
> >
> > Many thanks.
> >
> > Jorge A Restrepo
> >
> > -----Original Message-----
> > From: Kit Baum [mailto:baum@bc.edu]
> > Sent: 26 January 2004 17:42
> > To: Restrepo J
> > Cc: statalist@hsphsun2.harvard.edu
> > Subject: Re: tscollap query
> >
> > Dear Jorge
> >
> > tscollap is just a convenience, "hard wiring" various features of
> > -collapse- to take advantage of what we know about calendar-time data.
> > Since -collapse- can generate arbitrary subsets of a dataset, where 
> > you specify exactly how the data are to be collapsed, I should think 
> > that you can use -collapse- to deal with your event-oriented data.  
> > I'm not sure how, though, you plan to "aggregate the variables on each 
> > event according to different time
> > > > periodicity (days, weeks, months, quarter, years) and/or by spatial
> > > clusters
> > > > (both defined by the user and statistically generated)". It seems to 
> > > > me that an observation in the result data
> > would either be identified by a time period, or by an event, but could 
> > not very well be associated with both; and how would you deal with a 
> > result dataset in which the observations either belong to a time 
> > period or to a specific event?
> >
> > Best wishes
> > Kit
> >
> > On Jan 26, 2004, at 12:24 PM, Restrepo J wrote:
> >
> > > Dear Professor Baum:
> > >
> > > I am taking the liberty of writing you regarding the program you 
> > > wrote in Stata for collapsing time series. Nick Cox pointed to it in 
> > > statalist after my query in Statalist:
> > >
> > > > I am working with a relatively large data set organised by events 
> > > > (19.000
> > > in
> > > > total). Each event has time and spatial descriptors and several 
> > > > variables.
> > > >
> > > > My query is: Is there a command-routine-program for Stata to 
> > > > aggregate the variables on each event according to different time 
> > > > periodicity (days, weeks, months, quarter, years) and/or by spatial
> > > clusters
> > > > (both defined by the user and statistically generated)? I would 
> > > > appreciate any kind of guidance on this question.
> > >
> > > It seems that your program in combination with the command Collapse 
> > > would allow me to do most of what I need, as I would be able to 
> > > aggregate from monthly to lower frequencies. The problem is that I 
> > > have my original data in "event" form, i.e. a collection of daily 
> > > events that I would need also to collapse to daily and monthly 
> > > frequencies. Do you know of a way of doing this? Is it possible to do 
> > > this with your programme or with a variation of it?
> > >
> > > Best regards,
> > >
> > >
> > > Jorge Alberto Restrepo
> > > ____________________________
> > > Department of Economics
> > > Royal Holloway-University of London
> > > Egham Hill, Egham, Surrey
> > > TW20 0EX, United Kingdom