non-parametric one-sample test?
Locked
non-parametric one-sample test?
 
|
Hi We want to hear all your funny, exciting and crazy Hotmail stories. Tell us now |
Re: non-parametric one-sample test?
|
|
Hi Chris,
Would a sign test be the one ?
Best Wishes,
Martin Holt
|
Re: non-parametric one-sample test?
|
|
In reply to this post by Chris Smith-49
Chris,
Consider using the one-sample wilcoxon signed rank test, which tests whether the median value in the population is equal to zero. It is available starting with SPSS 18. HTH, Steve Brand www.StatisticsDoc.com On Thu, Mar 18, 2010 at 6:28 AM, Chris Smith wrote: > Hi > > I have a data set of roughly 50 companies - each were asked to state > the percentage of their work that they would classify as 'green' in > nature. I want to run a test to show that over the sample, this > differs from 50%, i.e. on average there is not an even balance between > green and non-green work being done. > > > If this was a normally distributed variable, I'd run a standard > one-sample t-test - however the distribution is fairly positively > skewed (most companies seem to report between 10-30% - a few report > higher), but with a small spike at 100%. Therefore at the very least > trying a non-paramtric equivalent alongside a t-test seems appropriate > - however going through the SPSS menus I can't find the non-parametric > equivalent to the one-sample t-test. Any ideas/refernces - I'm sure > I'm missing something obvious here! > > Thanks in advance for any help > > Chris > > _________________________________________________________________ > Tell us your greatest, weirdest and funniest Hotmail stories > http://clk.atdmt.com/UKM/go/195013117/direct/01/ ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
Re: non-parametric one-sample test?
|
|
In reply to this post by Chris Smith-49
Dear Chris, You might try a data transformation such as 1/x or log(x) to see
if that makes the distribution closer to normal, and then use a t-test. Or, f
you have the stomach for it, bootsrapping. Garry Gelade Business Analytic Ltd. From: SPSSX(r) Discussion
[mailto:[hidden email]] On Behalf Of Chris Smith Hi We
want to hear all your funny, exciting and crazy Hotmail stories. Tell us
now __________ Information from ESET NOD32 Antivirus, version of virus signature database 4954 (20100318) __________ The message was checked by ESET NOD32 Antivirus. http://www.eset.com |
Re: non-parametric one-sample test?
|
|
In reply to this post by Chris Smith-49
Chris Smith wrote:
> I have a data set of roughly 50 companies - each were asked to state the > percentage of their work that they would classify as 'green' in nature. > I want to run a test to show that over the sample, this differs from > 50%, i.e. on average there is not an even balance between green and > non-green work being done. > > > If this was a normally distributed variable, I'd run a standard > one-sample t-test - however the distribution is fairly positively skewed > (most companies seem to report between 10-30% - a few report higher), > but with a small spike at 100%. Therefore at the very least trying a > non-paramtric equivalent alongside a t-test seems appropriate - however > going through the SPSS menus I can't find the non-parametric equivalent > to the one-sample t-test. Any ideas/refernces - I'm sure I'm missing > something obvious here! > Hi Chris: You have already received a lot of responses, but I think that none focused on the problem that SPSS can't (won't?) do a one sample non-parametric test, like Wilcoxon signed ranks test or signs test. At the end of this message I will give you a workaround to get both. Concerning which one is better for your data, I would discard Wilcoxon test, since you have mentioned that your data are positively skewed, and symmetry (but not normality) is a condition for Wilcoxon test. You can either try what's already been suggested to you: log-transform your data and check symmetry again, or use sign test. Workaround with SPSS: you have to compute a new variable with the reference value (50% in your case), and then run a two related samples non-parametric test: COMPUTE RefValue=50. NPAR TEST /WILCOXON= RefValue WITH PercentGreen /SIGN = RefValue WITH PercentGreen. HTH, Marta GG -- For miscellaneous SPSS related statistical stuff, visit: http://gjyp.nl/marta/ ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
Re: non-parametric one-sample test?
|
In reply to this post by Chris Smith-49
At 06:28 AM 3/18/2010, Chris Smith wrote:
I have a data set of roughly 50 companies - each were asked to state the percentage of their work that they would classify as 'green' in nature. I want to run a test to show that over the sample, this differs from 50%, i.e. on average there is not an even balance between green and non-green work being done. This is one of those problems that raise at least as many problems of interpretation as of methodology. You want to show that "on average there is not an even balance between green and non-green work". First, there's your measure: "the percentage of their work that they would classify as 'green'" (emphasis added). I dare say your actual question was more rigorous than this sounds; but as phrase, the terms 'green' and 'percentage' are both free-floating, with respondents using what definitions they choose. 'Green' is a notoriously flexible word, but 'percentage' isn't much better: Percent by total revenue? By number of product lines? By labor-hours applied? By square feet of factory floor? Do you know what definition each respondent used, of either 'green' or 'percentage'? On top of that, are the percentages measured, or estimated? By common experience (and, I believe, psychological experiments), humans don't estimate proportions well; particularly, low-proportion categories are commonly either inflated or ignored in estimates. And this lays aside the question of bias, for example the highly likely desire to report one's own company's work as 'green' as possible. I lay it aside, but you'll have to discuss it carefully when you report the results. Then, if you can argue you have a meaningful measure of the "percentage of their work that [is] 'green'", and want to test whether "on average there is an even balance" on that measure, then what does "on average" mean? Is there any statistic on your data that you can defend as a measure that may legitimately be called an 'average'? Among other things, a data transformation that doesn't have theoretical backup makes any 'average' hard to defend. (What is the meaning of the geometric mean of a set of measures of proportion? But if you log-transform, you're saying that that is the meaningful measure.) I guess you could classify each company as "<50%" and ">50%", and test whether the counts showed a significant difference from equal proportion (see BINOMIAL or CHISQUARE in NPAR TESTS), but what does the answer tell you? OK, let's see what you might do with your data. You say that "most companies seem to report between 10-30% - a few report higher, but with a small spike at 100%". That's a 'cloud and outlier' distribution -- the "10-30%" is the 'cloud', the 100% the 'outlier'. In the absence of a good theoretical model, it's best to analyze that in two steps: outlier vs. cloud, and variations within cloud. Unfortunately, that takes a goodly sample size; 50 companies is probably far too few. But it sounds like remotely reasonable summary statistics will give a value within the 'cloud', i.e. "between 10-30%". That solves your problem whether the value is above or below 50%. The only question is whether the "small spike at 100%" can be used to argue against that conclusion. Either you say no, that spike doesn't change the picture; or, you address two more questions: First, do you believe those "100%" reports? Are they making up self-congratulatory numbers? It they're companies with specifically 'green' lines of work like wind energy, do you agree that makes them 100% 'green'? (What about pollution from a wind-turbine factory?) If you do believe them, you can look for a summary statistic that you can claim is reasonable, and that assigns enough weight to the 'green' work by the companies reporting high numbers, to offset the work of the many that report much below 50%. -Best of luck to you, Richard Ristow ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
Re: non-parametric one-sample test?
predicting winners/losers and ranks?
|
|
In reply to this post by Richard Ristow
Hi all
I have a couple of situations where one respondent's response on an outcome variable is systematically related to other respondents, and am not sure quite how to model them: i) I have 20 teams of 5 individuals i.e. 100 people nested within 20 teams. Within each team there is a competition to find the best employee (winner coded 1, 4 losers coded 0). I want to (attempt to) predict the winners from their demographic characteristics. At the moment I have the data at the employee level i.e. ID TEAM WINNER AGE SEX 1 1 0 25 0 2 1 0 27 0 3 1 1 21 1 4 1 0 45 0 5 1 0 27 1 6 2 0 35 0 7 2 1 55 0 8 2 0 17 1 etc My initial though was to use logistic regression, but within each team the outcome var, WINNER, is systematically related between employees, in that if one employee = 1, the rest must equal 0. Any ideas how to proceed? ii) is very similar, in that the data is identical, but instead of winner I have a ranking of employees performance in the competition from 1 to 5 i.e. ID TEAM CPRANK AGE SEX 1 1 2 25 0 2 1 3 27 0 3 1 1 21 1 4 1 4 45 0 5 1 5 27 1 6 2 4 35 0 7 2 1 55 0 8 2 3 17 1 etc I could use ordinal or standard multiple regression, but again the outcomes are related (can only have 1 '1', 1 '2', 1 '3' etc within each team) Any help or suggestions for references much appreciated ;-) Get a free e-mail account with Hotmail. Sign-up now. <a href="http://"> Get a new e-mail account with Hotmail - Free. Sign-up now. |
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |