Tau, Kappa, Eta?
Locked
 
|
Dear all,
This is a basic question. My background is economics. So I am not very familiar with analysing data from surveys with short lickert scales. My dependent variable (Y) is overall satisfaction with service, measured with a 5 items scale: Very satisfied, satisfied, indifferent, unsatisfied, very unsatisfied. My independent variables (x(1), x(2), ..., x(K)) are satisfaction with particular aspects of the service, measured with a 4 items scale: veru satisfied, satisfied, unsatisfied, very unsatisfied (in my independent variables I do not have middle category). Don't ask me why, but data is coded like that. Fortunately I only have few observations who answer indifferent in the dependent variable (as the person who interviewed people did not read "indifferent". The degree of correlation between k the independent variable is really high, as people tend to say either satisfied or insatisfied to most questions on the particular aspects of the service. Also, most X(k) show a bivariate Person r square really high with Y (most people tend to answer equally to most X(k) and to Y). I doubt which measure for correlation I should use for bivariate correlations. Pearson's r is not very suitable as I have only 5 integer values in my independent variable and 4 in my dependent variable. I might use Eta, trying to explain the reduction of the error prediction in Y, once I incorporate one X(k). Nevertheless, I am not very satisfied with the results, as this measure does not take into account that very satisfied in Y, is closer to satisfied than to satisfied in X(k). I have several questions: 1) I dunno if any weighted form of Kappa will be useful here. I don't have data of two judges on the same topic, but of many people in several topics. 2) Can I obtain a weighted Kappa in SPSS? 3) If I try to run an ordinal logistic model I violate the proportional Hazards Assumption. So, what should I do? Should I run a multinomial logistic instead? 4) As all my predictors (X(k)) show a very high degree of correlation. What should I do to avoid multicollineality if I run a logistic regression? Thanks J. Pulido |
|
|
Sorry I made a mistake, where I said Eta, I was meaning Goodman and Kruskall's Tau...
|
|
|
In reply to this post by jmdpulido
> Date: Wed, 16 Mar 2011 04:23:23 -0700
> From: [hidden email] > Subject: Tau, Kappa, Eta? > To: [hidden email] >[snip] > > I doubt which measure for correlation I should use for bivariate > correlations. Pearson's r is not very suitable as I have only 5 integer > values in my independent variable and 4 in my dependent variable. [... snip] Use Pearson's r. Yes, it shows effects of "attenuation", especially between variables that are dichotomies. Other than for dichotomies, the attenuation is pretty negligible. And you don't have dichotomies. People are accustomed to Pearson's r. Its size has implications that are better known to everyone, statisticians or other readers, than any other correlations. Where do you get that notion, that the Pearson's is not suitable? - A Pearson is a good estimate of a weighted kappa. - An intra-class correlation is the useful generalization for multiple raters, if you can't be better satisfied by looking at pair-wise Pearson's. One proper way to avoid problems with high inter-correlations among a set of predictors is to compute one or more composite scores to use as predictors instead. -- Rich Ulrich ===================== 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 |
|
|
Pearson r for dichotomies is the same as phi. On the other hand, percentage difference in a 2x2 table is equivalent to the regression coefficient b, which is itself a function of Pearson r (multiplied by the ratio of STDDEVs). Furthermore, there is no chance of nonlinear effects if you have just dichotomies, because only one interval exists for each variable (i.e. the interval between the two values). For these and other reasons I prefer using measures linked to the generalized linear model, a.k.a. linear regression, of which Pearson’s r is a key index. Hector De: SPSSX(r) Discussion [mailto:[hidden email]] En nombre de Rich Ulrich > Date: Wed, 16 Mar 2011 04:23:23 -0700 No virus found in this message. |
|
|
In reply to this post by jmdpulido
Thanks for your comments...
I'll just R square then in a regression. However, Do you advise to use OLS? or as my values are integer (going from 1 to 4) should I prefer to run a logistic regression? Thanks again, you are being very usefull. |
| Free forum by Nabble | Edit this page |