Whenever you use a multi-item scale to measure a construct, a key step is to create a score for each subject in the data set.
This score is an estimate of the value of the latent construct (factor) the scale is measuring for each subject. In fact, calculating this score is the final step of running a Confirmatory Factor Analysis.
The simplest way to create a score is to add up (or average) the values of each variable on the scale (indicator) for each subject. This is called a factor-based score. It’s based on the factor in the factor analysis, but is not technically a factor score since it doesn’t use the factor weights.
Is this an acceptable method? The chances of this being acceptable is between slim and none, and Slim left town.
Why can’t we do this?
A true factor score is the best estimate of the subject’s value on the latent construct you’re measuring with your observed indicator variables. The factor loadings, as calculated by the analysis, determines the optimal weighting for each indicator.
When the EFA/CFA and structural equation models “predict” the factor scores it uses a linear regression, incorporating the factor loadings into the model.
To compare the difference between the “add up the scores” and linear regression approach we will use an example. In this example, five indicators together model the latent construct of Assertiveness:
AS3 Automatically take charge
AS4 Know how to convince others
AS5 Am the first to act
AS6 Take control of things
AS7 Wait for others to lead the way
The table below gives the coefficients generated by the linear regression approach. For the addition method the mean of the linear regression coefficients was used to evenly weight the variables.
The two approaches will not generate similar scores.
AS3’s weighting is 52% greater and AS4 is -47% less using linear regression compared to addition.
There is one situation where it makes sense to use addition. If the factor loadings are all equal, we can add up the scores for each indicator to generate the factor scores. But we cannot assume they are equal without testing them.
Let’s visually compare the factor scores generated by a structural equation model and adding the indicator scores together. The factor scores generated by the structural equation model are standardized, with a mean of zero and a standard deviation of 1. To compare the addition approach scores we will standardize them as well.
Below is a scatterplot of the SEM and addition generated scores. The diagonal line represents the line where the scores generated by the two methods are equal.
There are very few points on the diagonal line. The generated scores are not the same.
There are many situations in statistics where the best approach isn’t clear. This isn’t one of them. Running a confirmatory factor analysis and computing accurate scores is important for getting accurate measurements.