Premium
Statistical Myths About Log‐Transformed Dependent Variables and How to Better Estimate Exponential Models
Author(s) -
Villadsen Anders R.,
Wulff Jesper N.
Publication year - 2021
Publication title -
british journal of management
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 2.407
H-Index - 108
eISSN - 1467-8551
pISSN - 1045-3172
DOI - 10.1111/1467-8551.12431
Subject(s) - mythology , generalized linear model , econometrics , exponential family , exponential function , statistical model , statistics , maximum likelihood , mathematics , computer science , history , classics , mathematical analysis
We review 10 years of research published in the Strategic Management Journal ( SMJ ) and find the wide use of log‐transformed dependent variables (LTDVs) to be based on statistical myths, with possible detrimental effects for the validity of research findings. We find that many researchers use LTDVs for the wrong reasons, and very often in a way that is misaligned with the hypothesis they intend to examine. Researchers also appear unaware of the severe shortcomings of LTDVs. Using LTDVs implies estimating an exponential model, which represents a non‐linear relationship. We identify three myths that are widely followed by researchers: (1) LTDVs should be used to make distributions more normal; (2) linear hypotheses can be tested with LTDVs; and (3) LTDVs are the best way to estimate an exponential model. We call on researchers to exhibit caution when planning to use LTDVs and recommend instead the use of generalized linear models (GLMs) with quasi‐maximum likelihood estimation. The superiority of GLMs is demonstrated by two empirical examples from recently published studies.