Date: Thu, 12 Mar 2009 17:48:35 -0700
Reply-To: Bminer <b_miner@LIVE.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Bminer <b_miner@LIVE.COM>
Subject: Re: Proper Test for Data
Content-Type: text/plain; charset=ISO-8859-1
On Mar 12, 5:37 pm, HERMA...@WESTAT.COM (Sigurd Hermansen) wrote:
> I would take a close look first thing at the distribution of data within the six month interval (and around the interval if available). To illustrate the potential for estimation errors, consider these different monthly distributions:
> sales by month
> 0 1 2 3 4 5 6 7
> 1) 5 5 5 5 7 7 7 7
> 2) 3 4 5 6 6 7 8 9
> 3) 5 4 5 6 8 7 6 5
> 4) 4 6 5 4 8 7 6 7
> Each of the first three month intervals (1-3) has the same sales total, as does each of second three month intervals. Case 1) indeed appears to be a result of something happening at the month 4 mark (although somewhat suspiciously since interventions often take time to take effect), but 2) seems more likely to be the result of an upward trend. 3) doesn't look much like the result of an intervention at month 4, and 4) looks more like seasonal lows at months 0,3,5, ...
> If you are looking only at 15 vs. 21 in these cases, all will look the same in aggregate, but distributions over time may tell you much about other influences that have conditioned the values that you are observing. And distributions over time don't tell the whole story either. A change in demand for a product line could happen about the same time as the intervention. If you can't rule out or control for confounding variables, your experiment can give you misleading answers. In these situations diagnostic statistics can uncover very subtle sources of bias, but statistics don't supercede an initial review of alternative explanations for what you observe
> -----Original Message-----
> From: SAS(r) Discussion [mailto:SA...@LISTSERV.UGA.EDU] On Behalf Of Bminer
> Sent: Thursday, March 12, 2009 2:50 PM
> To: SA...@LISTSERV.UGA.EDU
> Subject: Proper Test for Data
> Hi All-
> I have data for two groups: Group A and Group B. They are created as being as close to each other as possible except for the fact that Group A had an intervention.
> I have two pieces of data on each individual in each group:
> Sales in 3 months before intervention
> Sales in 3 months after intervention
> There are zeroes in the data sets.
> Should I subtract before and after and then compare this difference as an independent two group test (non parametrically because of the >=0 nature of the data)
> Or is there a better way? Ultimately I want to know if the intervention worked and use the control (group B) to help validate.
Would you suggest plotting means, medians for each group by month
leading up to intervention. Is it then a visual judgment call on if
trend explains the *change* after intervention or is there a test (is
it repeated measures)?
If I had many months prior I could accomplish an intervention analysis
with a time series and I think protect against what you are