Hi All,
I was trying to run Factor analysis for a dataset with around 150 variables but only have around 80 observations.
I tried the "factanal" function in R and R reported error:
Error in solve.default(cv) :
system is computationally singular: reciprocal condition number = 3.0804e-20
Any suggestions on alternative methods / packages?
Here is an example on when would the error occurr:
# This will work (dataset with 80 obs and 15 predictors)
set.seed(1234)
fake_df = as.data.frame(matrix(rnorm(80*15), nrow = 80))
factanal(fake_df, factors = 2, rotation = "varimax")
# This will not (dataset with 80 obs and 150 predictors)
set.seed(1234)
fake_df = as.data.frame(matrix(rnorm(80*150), nrow = 80))
factanal(fake_df, factors = 2, rotation = "varimax")
So far I've replaced the solve function in the "factanal" source code with a numerical solving function one that I created in the attached R-code file using svd, but it did not resolve the issue:
Instead it created new error:
Error in factanal.fit.mle(cv, factors, start[, i], max(cn$lower, 0), cn$opt) :
could not find function "factanal.fit.mle"
You are trying to solve for 150 parameters with 80 data points, so a factor solution is vastly under-determined. This would be like trying to fit a straight line to a single data point. Either reduce the number of variables or increase the number of data points (or both).
Basically Dwight is right.
If you transpose the data matrix (getting 150 rows and 80 columns) SVD will do.
But interpretation will be ... tricky.
Good luck.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R