Missing values in exploratory factor analysis: A ‘best of all possible
worlds’ approach to imputation for incomplete survey data
Abstract
In the social sciences, multi-item scales and factor analyses are
standard tools in survey research. In the social sciences, such tools
are omnipresent, as are, unavoidably, nonresponses. The question is how
to handle missing values when an exploratory factor analysis is
intended. Deletion methods will result in — oftentimes substantial and
damaging — reduction of power. The seemingly obvious alternative is to
keep all respondents and apply imputation to missing values. However,
with the true factor structure unknown, theoretically recommendable
multiple imputation methods cannot simply be applied. Instead of
declaring an entire method unsuitable for exploratory analysis, we
propose an approach that keeps the relevant aspects of various methods
and combines these by sacrificing less relevant aspects. Doing so, we
keep understanding and ease of use in mind, aiming for an approach that
is more rigorous and ‘correct’ than what is commonly used in practice,
whilst still being straightforward enough to actually be used.