🚀 go-pugleaf

In article <19930722.053951.626@almaden.ibm.com>
DaveRoberts@vnet.IBM.COM (Dave Roberts) writes:
> Aco-worker asks the following:
> Someone did the following survey:
>     - Population = 600
>     - Surveyed   = 600   (ie. non-random)
>     - Question posed is binary: happy - yes or no
>     - Return: 225 answered, 80% yes, 20% no
>
> One should be able to calculate the deviation on a random survey. How
> about a non-random one with a serious non-response bias like the above
> example? What deviation should be used? Can it be calculated? These
> kinds of surveys are common in the corporate world but I always raise
> the issue of it being non-random and having serious non-response. Now
> I'd like some expert feedback and some numbers.
>
> smw
>+---------------------------------------------------------------------+
>| Dave Roberts - OOUIs R US                                           |
>+---------------------------------------------------------------------+
>.
HAVE YOU EVER NON-RESPONSE.  SURELY THE "NUMBERS" ARE
SIMPLE, WE JUST STATE WHAT WE KNOW AND DON'T KNOW.

WE KNOW:
180 PEOPLE SAID THE WERE HAPPY.  [WHAT THEY THOUGHT WE DO NOT KNOW.]
45 SAID THEY WERE NOT.   375 AREN'T TELLING.

WE DO NOT KNOW:
THAT THE 375 WOULD HAVE ANSWERED IN EVEN APPROXIMATLEY THE SAME
PROPORTIONS IF ....

THIS PUTS THE ESTIMATED FRACTION OF STATED HAPPINESS SOMEWHERE BETWEEN
                   0.3    AND   0.925
                [180/600  AND   (180+375)/600].
WITH THAT KIND OF SPREAD, WHAT GOOD ARE CONFIDENCE INTERVALS, STANDARD
DEVIATIONS AND ALL THAT STUFF?

STATISTICS IS NO USE.  OUR ONLY HOPE IS INTUTITION.

IF INTUITION IS ALL WE HAVE, USE IT, BUT BE HONEST:
DON'T DISGUISE INTUITION AS (FRAUDULENT) CALCULATION, AS TOO
OFTEN HAPPENS IN SUCH CASES.

                     Bill Knight / University of New Brunswick / Canada