Assessing Partial Association Between Ordinal Variables: Quantification, Visualization, and Hypothesis Testing
Dungang Liu,Shaobo Li,Yan Yu,Irini Moustaki
DOI: https://doi.org/10.1080/01621459.2020.1796394
IF: 4.369
2020-08-26
Journal of the American Statistical Association
Abstract:Partial association refers to the relationship between variables <span class="NLM_disp-formula inline-formula"><math>Y1,Y2,…,YK</math></span> while adjusting for a set of covariates <span class="NLM_disp-formula inline-formula"><math>X={X1,…,Xp}</math></span>. To assess such an association when <i>Y<sub>k</sub></i>'s are recorded on ordinal scales, a classical approach is to use partial correlation between the latent continuous variables. This so-called polychoric correlation is inadequate, as it requires multivariate normality and it only reflects a linear association. We propose a new framework for studying ordinal-ordinal partial association by using Liu-Zhang's surrogate residuals. We justify that conditional on <span class="NLM_disp-formula inline-formula"><math>X</math></span>, <i>Y<sub>k</sub></i>, and <i>Y<sub>l</sub></i> are independent if and only if their corresponding surrogate residual variables are independent. Based on this result, we develop a general measure <span class="NLM_disp-formula inline-formula"><math>ϕ</math></span> to quantify association strength. As opposed to polychoric correlation, <span class="NLM_disp-formula inline-formula"><math>ϕ</math></span> does not rely on normality or models with the probit link, but instead it broadly applies to models with any link functions. It can capture a nonlinear or even nonmonotonic association. Moreover, the measure <span class="NLM_disp-formula inline-formula"><math>ϕ</math></span> gives rise to a general procedure for testing the hypothesis of partial independence. Our framework also permits visualization tools, such as partial regression plots and three-dimensional P-P plots, to examine the association structure, which is otherwise unfeasible for ordinal data. We stress that the whole set of tools (measures, <i>p</i>-values, and graphics) is developed within a single unified framework, which allows a coherent inference. The analyses of the National Election Study (<i>K</i> = 5) and Big Five Personality Traits (<i>K</i> = 50) demonstrate that our framework leads to a much fuller assessment of partial association and yields deeper insights for domain researchers. <a class="ext-link" href="https://doi.org/10.1080/01621459.2020.1796394">Supplementary materials</a> for this article are available online.
statistics & probability