On Tue, 05 Mar 2002 18:03:47 GMT, Doug Hoy <[EMAIL PROTECTED]> wrote: [...] > I am considering hiring a research analyst who has submitted some examples > of work performed. It involves reporting on market survey data. > > The work seems mostly OK, but a few oddities make me wonder... > > For example, when reporting on the statistic accuracy of a sample of 500, a > 'confidence' of 95% is given with a 'Std. Error' +/- 2.0%.
If I said that the mean score was 115, with a std.err. of 2.0, that would imply that the 95% Confidence interval for the average was about 111-119. A Std. err. of 2% could be used to describe the variation in the mean response on a dichotomous item, in a survey. That would be approximately right for N=500 -- a little larger, at a 50% split, a little smaller for extreme values. > > I'm not used to seeing a Standard Error given in this context. Usually it's > a margin of error, which a non-statistically-inclined client can > intuitively apply to their data for confidence limits (which at 95% > confidence level, would be larger than +/- 2%). The impression given is > that one can trust the various results to be within 2% of the population > value. > > Am I misunderstanding something here? Personally, I'm used to seeing 'margin of error' written and described very poorly or ambiguously, mainly in newspaper articles, where it is impossible to tell what it refers to. Unless the numbers are there so I can do the math. You are thinking (I think) of the 95% confidence interval, which will be plus-and-minus twice the std. err. I think you should not fault your applicant for using more precision. I hope he should be able to explain the difference. It is (usually, I hope) trivial for a decent statistician to adapt the conventions of any client. It is not bad to have on hand an assistant who is wise to the conventions, but your serious work is almost bound to be faulty, if there is any serious work, if there is not a statistician involved. > > Later, the sample is split into two sub-groups of n= 350 and 150 > respectively. Results for a particular question are 56% and 68%. It seems > to me that the sample is not big enough to declare these subgroups > different, at 95%. No? I construct a 2x2 table with (102+48, 196+154); that yields an uncorrected chisquared value of 6.27, p= 0.012. I declare them "apparently different." -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
