0
$\begingroup$

I would like to perform factor analysis to identify latent variables and create indices from my data set that has 144 variables (mostly ordinal) all together. Before doing so I would like impute my missing data however I cannot do this for the entire data set without reducing the number of variables in it first (error message). My questions are:

1) How should I handle this? Can I run a factor analysis to see which variables need imputing, then go back perform multiple imputation and re-do the factor analysis this time without missing values...is this silly?

2)Can I group my binary variables together without using factor analysis? This would partially reduce the nr of variables in my data set.

error message: The imputation model for inCL_care contains more than 100 parameters. No missing values will be imputed. Reducing the number of effects in the imputation model, by merging sparse categories of categorical variables, chane gin the measurement level of ordinal variables to scale, removing two-way interactions, or specyfying constraints on the roles of some variables, may resolve the problem. Alternatively increase the max.number of parameters allowed on the MAXMODELPARAM keyword of the IMPUTE sub-command.

$\endgroup$
1
$\begingroup$
  1. It might help to tell us what the error message is.
  2. You can group variables together for any reason. Imagine the question: "Do you have a dog / cat / fish / tortoise?" That's 4 variables - you can sum them to get the number of types of animal a person has. No need for factor analysis.
$\endgroup$
1
$\begingroup$
  1. There is no need for factor analysis for the purpose of seeing which variables need imputing. You can simply obtain descriptive statistics for that purpose.

  2. It probably would be very helpful to combine some binary variables as suggested by @Jeremy Miles, thereby reducing the number that would be used in factor analysis. Other ways to eliminate variables would no doubt be helpful too. It would be a rare study indeed for which latent constructs required 144 measured variables in order to show up clearly. The use of so many variables might strain the ability of your software program, first to impute values and then to estimate a factor structure. At any rate you will want to adjust your commands so as to eliminate interactions as part of the basis for imputation.

Remember that factor analysis was designed for use with continuous variables, and that results you obtain should be treated as rough estimates. I assume that all along you are referring to exploratory factor analysis; the confirmatory type, which is used to formally test hypothesized structures and which rests on some restrictive assumptions, would hardly apply in the case you describe.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.