Divergent/discriminant validity

With convergent validity, this is part of concurrent validity, it’s whether values on a measure in a dataset do not show a statistical association with measures on another measure where theory and design says they should not show an association.

Details #

That’s fine and logical. In the physical sciences this is nice and clear: measurements of the mass of an object should not change if the temperature of the object is changed. (Now of course, it all gets a bit more complicated when one gets up near the speed of light and since relativity was discovered: I confess I haven’t a clue whether temperature and mass get to be related at near light speed!)

Whatever the complexities of relativity, in terrestrial life divergent validity holds usefully when assessing measures: if the recorded mass (actually, it’ll be weight) change for a set of objects when the air temperature changes then the scales have a bias: they are affected by temperature when they shouldn’t be. However, this is not so simple in our realm as the variables that shouldn’t affect others are either so silly to postulate, current wind direction and well-being perhaps, or perhaps not so clear. For example, there has been a real tradition for testing measures to see if they show divergent validity against gender and age but how many psychological state or trait variables are really completely independent of gender or age? There are measures of depression say that which were constructed selecting items until the measure showed no statistically significant gender differences. I remain rather sceptical about whether “depression” really is independent of gender in many cultures and countries.

So for me divergent validity is not really about demonstrating, with good statistical power to have detected relationships, that those relationships were not detected but it should be about exploring relationships with demographic variables and, where that seems reasonable, finding that they are reasonably precisely estimated in datasets from which we can reasonably generalise to the populations in which we want to use the measure and that the relationships may be statistically significant but are very small compared with the convergent validity relationships.

A common naming error #

One sometimes sees a convergent validity exploration that depends on finding a difference between two groups referred to as discriminant/divergent validity. Typically that’s where a study has compared a set of scores from people seeking help with scores from people not feeling a need to seek help. This is convergent validity if the measure is anything that might be expected to relate to help seeking: i.e. for all our well-being, quality of life, functioning or “psychopathology” measures

Try also #

Convergent validity
Predictive validity

Chapters #

Mostly in Chapter 3, mentioned in Chapter 7.

Online resources #

This largely boils down to test for correlations so various of my shiny apps relating to correlation are pertinent I guess.

Dates #

First created 21.iv.24.

Powered by BetterDocs