On 8 May 2010 12:40, Timothy Green <[email protected]> wrote: > I think that's error in the mean, and since the Lib Dems had the most > responses it's the lowest (error = sigma/sqrt(n), right?). The variance in > the answers is stddev. > > Someone who has done stats more recently than myself can probably correct > me. >
I'd bet that I did stats much longer ago, but... Yes "(standard) error" is a name for the sample standard deviation divided by the square root of the sample size. It can be used as an estimate of the standard deviation of the mean of samples of the same size. Whether its at all useful or helpful to give it is another matter - its really a religious discussion. "Religious" in the sense that, while some aspects of statistics are simply matters of maths or logic, or in some cases philosophy (eg whether its meaningful to talk about posterior probabilities or not), descriptive statistics is about data presentation, which is a matter on which people argue a great deal. For my money, standard deviation is a *pants* measure. Even for nice (normally distributed) data, most people will read a measure like: mean=50 +/- 10 (standard deviation) as meaning that the range 40-60 covers pretty much all the cases of interest, whereas all it means is that about 68% of the possible values are in that range. Even 30-70 only gets you 95% of the possibilities (i.e. 1/20 fall outside, and 1/20 is quite a lot of things sometimes - when I commuted to work 1/20 of my journeys took place each fortnight, or many times a year). The only virtue of s.d. is its what everyone uses - so you can just put it there without much more explanation. A more useful comment: the MDS looks fine (its a neat rough and ready way to reduce data to 2D) but the "distance" is interesting. Simply taking the difference between two % agreements has an odd effect. Consider: [1] Party A and party B have 0% and 50% agreement with a particular policy - they score 50% difference. The probability that any pair of people from either party would agree on that topic is 50%. [2] The figures are now 25% and 75%, and the probability of pairwise agreement falls to about 38% but the score difference is 50%. Well you see what I mean: its not a linear measure - just something worth bearing in mind. -- Francis Davey _______________________________________________ Mailing list [email protected] Archive, settings, or unsubscribe: https://secure.mysociety.org/admin/lists/mailman/listinfo/developers-public
