In the Project EAT study (Neumark-Sztainer) there was an interest in the adolescents' self-esteem. The following 6 questions taken from the Rosenberg Self-Esteem Scale (Society and the adolescent self-image. Princeton, NJ: Princeton University Press) were asked on the questionnaire. A random sample of 267 of the students from the total surveyed group can be found in slfesteem.xls (Excel file) or slfesteem.csv (a comma delimited file), the columns correspond to the following: Column 1 - Student ID number Column 2 - Gender (1 is boy, 2 is girl) Column 3 - satisfed "On the whole, I am satisfied with myself" Column 4 - goodqual "I feel that I have a number of good qualities" Column 5 - nogood "At times I think I am no good at all" Column 6 - dothings "I am able to do things as well as most other people" Column 7 - morespec "I wish I could have more respect for myself" Column 8 - useless "I certainly feel useless at times" The responses for the self esteem questions (Columns 3-8) are on a 4 point Likert scale, with 1 anchored at "Strongly disagree", 2 at "Disagree", 3 at "Agree" and 4 anchored at "Strongly agree" ------------------------------------------------------------------------ A self esteem scale can be created by summing these 6 Likert-Type items (after appropriately reversing the order of variables that have contrary meaning. 1. Using SAS (or some other statistical package) calculate the Cronbach alpha for the Self-esteem scale. Report the alpha for the raw variables and the standardized variables. 2. Give the output for the Deleted Variable Cronbach alpha values. Is there any item that can be thrown out of the scale in order to increase the reliability? 3. Create a slfesteem scale and compare the mean value for this scale across boys and girls. ------------------------------------------------------------------- Some interesting comments related to the positivity and negativity of the way questions are asked Hello, > As part of my predoctoral research project, I am analyzing the >Rosenberg Self Esteem Inventory using confirmatory factor analysis. >Some of the items on this scale are "reverse coded". That is, on some >of them, a higher number indicates more self esteem, and on others it >represents less. Should I rereverse the scoring so that all the items >(or nearly all) correlate positively? Will this make any difference? > Any opinions or references welcome. > >Peter >flom@murray.fordham.edu > > Date: Fri, 13 Jan 1995 13:10:35 -0600 Reply-To: SEMNET Discussion List Sender: SEMNET Discussion List From: deborah bandalos Subject: Re: Factor analysis with items reverse scored In-Reply-To: <01HLSKB8ENAK0000UF@crcvms.unl.edu> from "Ronald Goldsmith" at Jan 11, 95 01:40:15 pm > Ron Goldstein replied to Peter Flom's question about reverse coding items as follows: > Peter, the only difference it will make is that the signs on the factor > loadings will reflect the direction of wording of the items. It is > worthwhile to not reverse the scoring and do the factor analysis to see > dramatically what effect direction of item wording has on the factor > structure. You might, for instance, see all the postive items load on one > factor and all the negative items load on another. Be sure to reverse the > scoring when you evaluate reliability, however, because no doing so will > give you some negative correlations and a low coefficient alpha. See, R. E. > Goldsmith, Dimensionaliyt of the Rosenberg Self-Esteem Scale," Journal of > Social Behavior and Personality, 1 (2), 1986, 253-264. > > Good luck, Ron G. > On Fri, 13 Jan 1995 13:10:35 -0600 deborah bandalos said: > It is well known that positively an negatively oriented items will >tend to form separate factors. However, Nunnally & Bernstein (p. 318) >point out that this is due to the fact that items with similar >distributions will tend to correlate more highly than dissimilarly >distributed items. If this is the case, then it would seem that the >"positive" and "negative" factors are really just statistical artifacts >rather than a manifestation of some aspect of self-esteem. I would >therefore argue for reverse coding to avoid that artifact. The problem is that reverse-scoring does not avoid the possible artifact--it only reverses the tails of the distribution. Those items will still share distributional characteristics they don't share with positively worded items for reasons I described in an earlier e-mail. The problem, as I see it, is allocating that commonality, be it artifactual or substantively interesting. A bi-factor model with one general factor (global self-esteem) and two group factors (positively worded, negatively worded) would deal with it. Otherwise, one will either end up with a poor fit and, using some modification strategies, a bunch of correlated errors, or the appearance of a two-factor model.