MTest: A Procedure for Multicollinearity Testing using Bootstrap.

About

Functions for detecting multicollinearity. This test gives statistical support to two of the most famous methods for detecting multicollinearity in applied work: Klein’s rule and Variance Inflation Factor (VIF). See the URL for the paper associated with this package, Morales-Oñate and Morales-Oñate (2023) doi:10.33333/rp.vol51n2.05

Consider the regression model

$$ Yi = \beta{0}X{0i} + \beta{1}X{1i} + \cdots+ \beta{p}X{pi} +ui $$

where $i = 1,\ldots,n$, $X{j,i}$ are the predictors with $j = 1,\ldots,p$, $X0 = 1$ for all $i$ and $u_i$ is the gaussian error term.

In order to describe Klein's rule and VIF methods, we need to define auxiliary regressions associated to the abobe model (global). An example of an auxiliary regressions is:

$$ X{2i} = \gamma{1}X{1i} + \gamma{3}X{3i} + \cdots+ \gamma{p}X{pi} +ui. $$

In general, there are $p$ auxiliary regressions and the dependent variable is omitted in each auxiliary regression. Let $R{g}^{2}$ be the coefficient of determination of the global model and $R{j}^{2}$ the $j\text{th}$ coefficient of determination of the $j\text{th}$ auxiliary regression.

MTest

Given a regression model, Mtest is based on computing estimates of $R{g}^{2}$ and $R{j}^{2}$ from $n{boot}$ bootstrap samples obtained from the dataset, $R{g{boot}}^{2}$ and $R{j_{boot}}^{2}$ respectively.

Therefore, in the context of MTest, the VIF rule translates into:

$$H0:\mu{R{j{boot}}^{2}}\geq 0.90,$$

and

$$Ha:\mu{R{j{boot}}^{2}}<0.90.$$

We seek an achieved significance level (ASL)

$$ASL = Prob{H0}{\mu{R{j_{boot}}^{2}}\geq 0.90}$$

estimated by

$$ASL{n{boot}} = Card(\mu{R{j{boot}}^{2}}\geq 0.90)/n{boot}$$

In a similar manner, the Klein's rule translates into:

$$H0:\mu{R{j{boot}}^{2}}\geq \mu{R{g_{boot}}^{2}},$$

and

$$ Ha:\mu{R{j{boot}}^{2}}<\mu{R{g_{boot}}^{2}}. $$

We seek an achieved significance level

$$ASL = Prob{H0}{\mu{R{j{boot}}^{2}}\geq \mu{R{g{boot}}^{2}}}$$

estimated by

$$ASL{n{boot}} =Card(\mu{R{j{boot}}^{2}}\geq\mu{R{g{boot}}^{2}})/{n_{boot}}.$$

It should be noted that this set up let us formulate VIF and Klein's rules in terms of statistical hypothesis testing.

MTest: the algorithm

$R{g{boot}}^{2}$ and $R{j{boot}}^{2}$ are the distributions of $R{g}^{2}$ and $R{j}^{2}$ induced by applying the bootstrap procedure to the dataset. Achieved significance level is computed for the VIF and Klein's rule. In the following we describe the procedure step by step:

• Create $n{boot}$ samples from original data with replacement of a given size ($n{sam}$).
• Compute $R{g{boot}}^{2}$ and $R{j{boot}}^{2}$ from each $n{boot}$ samples. This outputs a $B{n_{boot}\times (p+1)}$ matrix.
• Compute $ASL{n{boot}}$ for the VIF and Klein's rule.

Note that the matrix $B{n{boot}\times (p+1)}$ allow us to inspect results in detail and make further tests such as boxplots, pariwise Kolmogorov-Smirnov (KS) of the predictors and so on.