Since this message comes from GAO, I had assumed we were talking something
more than 30 people. It is true that there are problems of independence in
trend analysis; that is why there is a lot of sophistication in time series
techniques. As to fitting a regression, time series analysts generally (not
uniformly) reject the idea of time indexed regression, but if you are going
to do it, you can correct for autocorrelation with well known techniques
which overcomes the independence issue. Those techniques are found in
standard time series packages.
The t-test that Gary is proposing assumes that only the beginning and end
are of any interest. If this is true, then it makes things a lot simpler,
but it may not be true. If you really do have only an increment of 30, then
perhaps a t-test is all that is possible. But, in this case, you must also
ask whether significance testing is relevant at all, you are looking at
population data in two instances, the facts are as determined.
As to calculating weekly observations, there are very few time series that
are analyzed at the weekly level, but you can do it if you want.
Dan Williams
-----Original Message-----
From: Political Methodology Society [mailto:[log in to unmask]] On
Behalf Of Gary Klass
Sent: Thursday, June 22, 2006 8:37 AM
To: [log in to unmask]
Subject: Re: [POLMETH] addendum - trend data
Dan's solution poses more problems, I think.
Yes, If you have monthly data for the same trend, you will have 60 cases
rather than 5.
Or, with weekly data you will have 260 cases.
And if you actually record staffing down to the minute, you will have
thousands of cases and
you are guaranteed to have a significant trend, unless there has been no
change at all.
This problem occurs in all time series analysis that uses statistical
significance as its measure of significant change.
Measures of significance assume random independent sampling.
Time series data is never random and the time points are never selected
independently
(each year you select determines the next year you will select).
The whole idea of a statistically significant trend is bogus, I think.
In this case, you probably have a situation where you've added 20 or 30
people to the staff over five years.
A significance test would allow you to test the hypothesis that you have
added nobody.
Does that make sense?
Dan Williams wrote:
> Not if they know what they are doing. Is the problem that budget
practices
> bump up staffing once a year? Otherwise, why don't you/they look at
monthly
> staffing practices?
>
> Dan Williams
>
> -----Original Message-----
> From: Political Methodology Society [mailto:[log in to unmask]] On
> Behalf Of Anna Maria Ortiz
> Sent: Wednesday, June 21, 2006 10:40 AM
> To: [log in to unmask]
> Subject: [POLMETH] addendum - trend data
>
> I should have also asked expressly whether anybody would buy the simplest
> significance test, OLS with time as the independent variable, for a five
> data-point trend.
>
> ___________________________
>
> Good morning -
>
> A pair of economists has approached me wanting to use a Logest regression
> (the kind available as an Excel function, not logistic) to determine
whether
> growth in staff levels at an organization is "significant" over a 5 year
> period. Although I can articulate concerns with the method and their
> execution in particular, I'm having difficulty offering alternative firm
> standards for "significant" growth.
>
> Other than setting arbitrary standards (e.g. "a change of more than x% is
> significant"), does anyone have thoughts on how to establish criteria for
> whether a growth trend involving no independent variables is
"significant"?
> Any recommendations for models, standard levels, minimum data points in
time
> necessary, or references to peer-reviewed articles using Logest would be
> greatly appreciated.
>
> Thank you,
>
> Anna Maria
>
> Anna Maria Ortiz, Ph.D.
> Senior Statistician
> Applied Research and Methods
> U.S. Government Accountability Office
> (202) 512-2788
> [log in to unmask]
>
> **********************************************************
> Political Methodology E-Mail List
> Editor: Karen Long Jusko <[log in to unmask]>
> **********************************************************
> Send messages to [log in to unmask]
> To join the list, cancel your subscription, or modify
> your subscription settings visit:
>
> http://polmeth.wustl.edu/polmeth.php
>
> **********************************************************
>
> **********************************************************
> Political Methodology E-Mail List
> Editor: Karen Long Jusko <[log in to unmask]>
> **********************************************************
> Send messages to [log in to unmask]
> To join the list, cancel your subscription, or modify
> your subscription settings visit:
>
> http://polmeth.wustl.edu/polmeth.php
>
> **********************************************************
>
>
--
Gary M. Klass
Associate Professor
Department of Politics and Government
Illinois State University
Normal, Illinois 61790
(309)438-7852
ISU\IWU Habitat Collegiate Home
<http://lilt.ilstu.edu/habitat/>
SECA # 903 - 0235
**********************************************************
Political Methodology E-Mail List
Editor: Karen Long Jusko <[log in to unmask]>
**********************************************************
Send messages to [log in to unmask]
To join the list, cancel your subscription, or modify
your subscription settings visit:
http://polmeth.wustl.edu/polmeth.php
**********************************************************
**********************************************************
Political Methodology E-Mail List
Editor: Karen Long Jusko <[log in to unmask]>
**********************************************************
Send messages to [log in to unmask]
To join the list, cancel your subscription, or modify
your subscription settings visit:
http://polmeth.wustl.edu/polmeth.php
**********************************************************
|
