90 likes | 223 Views
University of Warwick, Department of Sociology, 2012/13 SO 201: SSAASS (Surveys and Statistics) (Richard Lampard) Index Construction (Week 13). Scale/Index Construction.
E N D
University of Warwick, Department of Sociology, 2012/13SO 201: SSAASS (Surveys and Statistics) (Richard Lampard)Index Construction(Week 13)
Scale/Index Construction • Researchers sometimes want to measure some latent characteristic of their respondents, e.g. whether they have a ‘traditional’ or a ‘modern’ viewpoint on couple relationships. • This is often done by asking a number of questions which each tap that characteristic and, when aggregated, collectively do so in more reliable way.
Some issues... • How does one know that a measure constructed by aggregating various items to give a scale is measuring the ‘right’ quantity, i.e. is a valid measure? • How does one ensure that what a measure is measuring is unidimensional, i.e. that it is not a composite of measures of two or more underlying concepts? • How does one assess which items need to be included to maximise the reliability of the measure? • And how does one assess the overall reliability of the scale?
Some answers... • For a discussion of assessing various forms of validity, see Oppenheim (1992). • Unidimensionality can be assessed using a technique called factor analysis, see DeVellis (2003). • Reliability can be assessed using a measure called Cronbach’s alpha (see De Vaus, 2001; DeVellis, 2003).
Factor analysis Factor analysis generates a set of underlying factors which successively maximise the amount of (remaining) variation in the items that they can explain. If a scale is working properly unidimensionally, then the first factor will explain a high proportion of the variation, and the subsequent factors similar, small amounts.
Cronbach’s alpha • According to DeVellis (2003: 95), “Alpha is an indication of the proportion of variance in the scale scores that is attributable to the true score”. • Items are chosen for inclusion so as to maximise that proportion, and if they have relatively high correlations with the rest of the items within the scale (viewed collectively). • De Vaus (2001) suggests that a value of at least 0.7 is preferable.
Some additional issues... • Should all the items in the scale be treated as of equal importance? Or should their values be added in such a way as to increase/ decrease the relative importance of some items? • Are the gaps between the values that a variable can take uniform in meaning?
Some examples... • Masculinity/Femininity: The Bem Sex-Role Inventory • Psycho-social well-being: The General Health Questionnaire (GHQ) • The Warwick-Edinburgh Mental Well-being Scale (WEMWBS): see http://www.hqlo.com/content/5/1/63
Another example... • BSA left-right scale (used since 1986) • See Park et al. (2012) British Social Attitudes 28. London: Sage/NatCen. http://ir2.flife.de/data/natcen-social-research/igb_html/pdf/1000001_e.pdf • Can be found in the BSA 2006 dataset...