Simple means to improve the interpretability of regression coefficients
2010 (English)In: Methods in Ecology and Evolution, ISSN 2041-210X, Vol. 1, no 2, 103-113 p.Article in journal (Refereed) Published
1. Linear regression models are an important statistical tool in evolutionary and ecological studies. Unfortunately, these models often yield some uninterpretable estimates and hypothesis tests, especially when models contain interactions or polynomial terms. Furthermore, the standard errors for treatment groups, although often of interest for including in a publication, are not directly available in a standard linear model. 2. Centring and standardization of input variables are simple means to improve the interpretability of regression coefficients. Further, refitting the model with a slightly modified model structure allows extracting the appropriate standard errors for treatment groups directly from the model. 3. Centring will make main effects biologically interpretable even when involved in interactions and thus avoids the potential misinterpretation of main effects. This also applies to the estimation of linear effects in the presence of polynomials. Categorical input variables can also be centred and this sometimes assists interpretation. 4. Standardization (z-transformation) of input variables results in the estimation of standardized slopes or standardized partial regression coefficients. Standardized slopes are comparable in magnitude within models as well as between studies. They have some advantages over partial correlation coefficients and are often the more interesting standardized effect size. 5. The thoughtful removal of intercepts or main effects allows extracting treatment means or treatment slopes and their appropriate standard errors directly from a linear model. This provides a simple alternative to the more complicated calculation of standard errors from contrasts and main effects. 6. The simple methods presented here put the focus on parameter estimation (point estimates as well as confidence intervals) rather than on significance thresholds. They allow fitting complex, but meaningful models that can be concisely presented and interpreted. The presented methods can also be applied to generalised linear models (GLM) and linear mixed models.
Place, publisher, year, edition, pages
2010. Vol. 1, no 2, 103-113 p.
confidence intervals, generalized linear models, interaction terms, null hypothesis testing, partial correlation coefficients, partial regression coefficients, standard errors, standardized effects sizes
IdentifiersURN: urn:nbn:se:uu:diva-153234DOI: 10.1111/j.2041-210X.2010.00012.xISI: 000288914100001OAI: oai:DiVA.org:uu-153234DiVA: diva2:415704