Date: Tue, 18 Apr 2006 23:24:50 +0000
Reply-To: Zhi Song <song_zhi@HOTMAIL.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Zhi Song <song_zhi@HOTMAIL.COM>
Subject: Re: Looking for advise
Content-Type: text/plain; format=flowed
Actually, to measure how long a customer stays in one program is the same
idea as doing survival analysis, although logistic regression is more
popular in marketing research.
One point not clear in your question message is why you need to do random
sampling, just to reduce your computing time or to need to collete more
information. Please remember ramdomization at this point or at this type
study isn't really to control confounding, just trying to let your sample
representing your target population. So the point is what is your target
population, or what kind of your target population you believe. Based on
that, you can set up your sampling strategy.
>From: David L Cassell <davidlcassell@MSN.COM>
>Reply-To: David L Cassell <davidlcassell@MSN.COM>
>Subject: Re: Looking for advise
>Date: Tue, 18 Apr 2006 15:55:41 -0700
>> I have a group of people who were offered and accepted a subscription
>>discount. I also have a group of people who subscribed at the same time
>>but were not offered the discount.
>> I think I want to take a random sample from those who subscribed but
>>were not offered the discount.
>> The figure out how long subscribers stayed. Then summarize this so
>>I have how many subscribers stayed 1 month, 2 months, etc. for both
>> Then I can do min, max, mode, mean, variance, stddev.
>> Am I heading in the right direction ? I can code this. I am NOT
>>looking for code. I am hoping someone can say "yes you are heading in the
>>right direction' or 'no, you are not'
>I wish I could say 'yes' or 'no', but I can't even figure out what you
>want to do with the data. Sorry.
>Am I correct that you have WAY more non-discount than discount entries,
>so you're looking for a way to get a reasonable data set to work with? If
>so, I'd want to know the data set sizes you have.
>If all you want is min/max/mode/... then you can do those without
>sampling from the data. (Unless your data sets are so large that you
>need to use PROC STDIZE instead of PROC UNIVARIATE to get quantiles
>in a reasonable amount of time.)
>If your long-range goal is some manner of logistic regression, you may
>need PROC SURVEYLOGISTIC and careful sampling of the original data
>sets to get down to a workable study size.
>David L. Cassell
>3115 NW Norwood Pl.
>Corvallis OR 97330
>FREE pop-up blocking with the new MSN Toolbar ?get it now!