arnold on Sep 16, 2008
Spector et al (1997) wrote an article to show that artifactual items can be produced by factor analysis when researchers use negative items on a survey because people will only endorse (agree with) items that are close to their true level of the construct, and will disagree with items that are far away in either direction. When assessing a scale that contains oppositely worded items that form separate factors, it is important to ascertain whether the scale measures one or two constructs. The seeming independence of oppositely worded items can be caused, not by underlying constructs, but by the way in which people respond to items.
Reasons for artifactual factors:
1) Lack of ability to understand negatively worded items
2) Carelessness of reading
3) The way in which people respond to items that vary in direction can produce item direction factors.
How to distinguish an artifact from a construct:
Factors that come from normally distributed item data that have perfectly homoscedastistic, linear joint distributions are unlikely to have arisen from the artifact (Spector et al, 1997, p. 674).
1) A procedure where responses are converted to a two choice format (yes or no).
2) A simple test which requires a scaled set of items that vary on the underlying construct of interest. Place items as sorted from low to high on the construct. A correlation matrix of the items is computed with the items in the sorted order. Those items that are closest together in their scale value should be most highly intercorrelated. This produces a simplex pattern in the matrix whereby the magnitude of correlations decreases with distance from the main diagonal (as you go down and toward the left).
3) Be cautious in making a case for two substantive factors when constructs are confounded with wording direction. Since the artifact of item wording direction can produce factors, one must provide strong evidence that the factors are substantive.
4)Required are studies that go beyond scales and use alternative means of validation. One might also relate scores on the scales to appropriate physiological measures. The establishment of separate nomological networks for each construct should serve as the basis for their acceptance. The factor analysis results suggest what might be rather than what is concerning the existence of constructs.
Spector et al (1997) suggests using more moderate items rather than extreme items. For example it is better to measure “I enjoy my job” rather than “I love my job” because the latter is more extreme.
Spector, P.E., VanKatwyk, P.T., Brannick, M.T. & Chen, P.Y. (1997), “When two factors don’t reflect two constructs: How item characteristics can produce artifactual factors. Journal of Management, 23, 659 – 678.
Anyone have anything to add?