WEBVTT WEBVTT Kind: captions Language: en 00:00:00.030 --> 00:00:06.300 Convergent and discriminant validity are two  terms that you often see after a factor analysis. 00:00:06.300 --> 00:00:10.860 What exactly is convergent validity and  discriminant validity is a bit difficult   00:00:10.860 --> 00:00:15.420 to understand because these concepts were  introduced without proper definitions. 00:00:15.420 --> 00:00:20.520 The concept of discriminant validity and  convergent validity date back to a paper   00:00:20.520 --> 00:00:27.990 by Campbell and Fiske from 1955 and in that  article the authors presented a concept of   00:00:27.990 --> 00:00:33.240 multi trait multi-method matrix which is  a set of correlations when you measure   00:00:33.240 --> 00:00:40.110 at least two traits each with at least  two methods. And they argued that some   00:00:40.110 --> 00:00:44.670 correlations in the table and lack of  correlations are evidence of convergent   00:00:44.670 --> 00:00:50.580 validity and discriminant validity without  really explaining what exactly convergent   00:00:50.580 --> 00:00:56.280 validity means beyond the empirical criteria  and the same with discriminant validity. 00:00:56.280 --> 00:01:05.430 So the current meaning of these terms has somehow  evolved from the original meaning because we don't   00:01:05.430 --> 00:01:13.410 really do multi trait multi method matrices it is  a very difficult to get two independent measures   00:01:13.410 --> 00:01:18.960 of the same construct. For example measure  to construct with paper and a pen first and   00:01:18.960 --> 00:01:24.390 then with online form next - so that's not a  practical way of measuring for most people. 00:01:24.390 --> 00:01:30.210 So the current practice in discriminant  unconverted validity is that discriminant   00:01:30.210 --> 00:01:37.800 validity refers to whether two scales  measure distinct constructs. If two   00:01:37.800 --> 00:01:42.390 scales measure the same construct they  don't have discriminant validity. Then   00:01:42.390 --> 00:01:47.490 convergent validity refers to whether  indicators measure the same construct.   00:01:47.490 --> 00:01:52.920 So it's kind of like used as a  synonym for internal consistency. 00:01:52.920 --> 00:02:02.610 These are used also very inconsistently.  So the definitions are a bit difficult   00:02:02.610 --> 00:02:07.290 to understand. People use them without  the definition. They are defined in a   00:02:07.290 --> 00:02:11.190 different ways in the literature and  they are also applied differently. 00:02:11.190 --> 00:02:22.150 This is a good example from an article by  Carlson and they say that some authors argue   00:02:22.150 --> 00:02:28.870 that a correlation as low as 0.28 of two  indicators indicate convergent validity.   00:02:28.870 --> 00:02:35.320 So if two indicators are correlated at  0.28 then they measure the same thing. 00:02:35.320 --> 00:02:47.920 They also point out that some indicators - some  authors argue that correlation as high as 0.75   00:02:47.920 --> 00:02:55.180 between two indicators is still not high enough  to claim that these indicators are not distinct. 00:02:55.180 --> 00:03:01.690 So the current practice with these terms is  that you have a number and then you make a   00:03:01.690 --> 00:03:09.880 claim and there's really no connection of how  high the number is and because the discriminant   00:03:09.880 --> 00:03:15.130 validity and convergent validity concepts are  not well defined then it is very difficult to   00:03:15.130 --> 00:03:20.080 interpret what that correlation means. So you  have to have an interpretation for the statistic. 00:03:20.080 --> 00:03:28.360 For composite reliability for alpha you have  an interpretation that it's a reliability of   00:03:28.360 --> 00:03:35.020 a sum of indicators. For these correlations  between indicators it's very difficult to   00:03:35.020 --> 00:03:42.760 interpret them directly. Therefore my  take on this is that it's important to   00:03:42.760 --> 00:03:48.910 assess whether your scales measure distinct  thing things. If you do a factor analysis   00:03:48.910 --> 00:03:54.070 each scale should load on what - each  indicator should load on one factor   00:03:54.070 --> 00:03:58.900 and two scales that are supposed to measure  two distinct things should be two factors. 00:03:58.900 --> 00:04:05.380 If that is the case then you have evidence of  both a discriminant validity and convergent   00:04:05.380 --> 00:04:09.010 validity and you don't really have  to look at any of these correlations.