check_collinearity() does not work with orthogonal polynomials

[jmgirard]

Related to https://github.com/easystats/see/issues/346

I know that orthogonal polynomials won't be collinear (that's their point). But I want to be able to show that while teaching.

library(easystats)
#> # Attaching packages: easystats 0.7.2 (red = needs update)
#> ✔ bayestestR  0.13.2   ✔ correlation 0.8.4 
#> ✔ datawizard  0.11.0   ✔ effectsize  0.8.8 
#> ✖ insight     0.20.0   ✖ modelbased  0.8.7 
#> ✔ performance 0.12.0   ✔ parameters  0.21.7
#> ✔ report      0.5.8    ✔ see         0.8.4 
#> 
#> Restart the R-Session and update packages with `easystats::easystats_update()`.

x <- runif(n = 100, min = 100, max = 110)
y <- 0.2 + 1.2 * x + rnorm(100, 0, 1)
df <- data.frame(x, y)

fit <- lm(y ~ poly(x, degree = 2), data = df)
check_collinearity(fit)
#> Not enough model terms in the conditional part of the model to check for
#>   multicollinearity.
#> NULL

^{Created on 2024-06-12 with reprex v2.1.0}

When other variables are included, there is no error but the orthogonal polynomials are considered a single term. Is that correct?

library(easystats)

x <- runif(n = 100, min = 100, max = 110)
g <- sample(c(0, 1), size = 100, replace = TRUE)
y <- 0.2 + 1.2 * x + 0.4 * g + rnorm(100, 0, 1)

df <- data.frame(x, y, g)

fit <- lm(y ~ poly(x, degree = 2) + g, data = df)
check_collinearity(fit)
#> # Check for Multicollinearity
#> 
#> Low Correlation
#> 
#>                 Term  VIF   VIF 95% CI Increased SE Tolerance Tolerance 95% CI
#>  poly(x, degree = 2) 1.04 [1.00, 4.96]         1.02      0.96     [0.20, 1.00]
#>                    g 1.04 [1.00, 4.96]         1.02      0.96     [0.20, 1.00]

^{Created on 2024-06-12 with reprex v2.1.0}

[strengejacke]

Have you checked with car::VIF()?

[jmgirard]

car returns an error:

x <- runif(n = 100, min = 100, max = 110)
y <- 0.2 + 1.2 * x + rnorm(100, 0, 1)
df <- data.frame(x, y)
fit <- lm(y ~ poly(x, degree = 2), data = df)

performance::check_collinearity(fit)
#> Not enough model terms in the conditional part of the model to check for
#>   multicollinearity.
#> NULL

car::vif(fit)
#> Error in vif.default(fit): model contains fewer than 2 terms

^{Created on 2024-06-12 with reprex v2.1.0}

[mattansb]

When other variables are included, there is no error but the orthogonal polynomials are considered a single term. Is that correct?

I think that makes sense - there is only 1 underlying variable, and adding orthogonal functions of the variable won't have any collinearity between them.

[jmgirard]

When other variables are included, there is no error but the orthogonal polynomials are considered a single term. Is that correct?

I think that makes sense - there is only 1 underlying variable, and adding orthogonal functions of the variable won't have any collinearity between them.

They won't have collinearity and they are powered by a single variable, but technically won't there be one term per degree? And in some weird world, couldn't it be possible for the different terms to have different VIFs in a model with additional variables?

[jmgirard]

Also, note that this same behavior occurs even when using raw polynomials, which definitely could be collinear:

x <- runif(n = 100, min = 100, max = 110)
y <- 0.2 + 1.2 * x + rnorm(100, 0, 1)
df <- data.frame(x, y)
fit <- lm(y ~ poly(x, degree = 2, raw = TRUE), data = df)

performance::check_collinearity(fit)
#> Not enough model terms in the conditional part of the model to check for
#>   multicollinearity.
#> NULL

car::vif(fit)
#> Error in vif.default(fit): model contains fewer than 2 terms

^{Created on 2024-06-12 with reprex v2.1.0}

[mattansb]

poly is considered one term - like factors or matrices:

data("mtcars")

mod <- lm(mpg ~ cbind(disp, hp) + poly(qsec, 2) + factor(cyl),
          data = mtcars)

insight::find_terms(mod)
#> $response
#> [1] "mpg"
#> 
#> $conditional
#> [1] "cbind(disp, hp)" "poly(qsec, 2)"   "factor(cyl)"

anova(mod)
#> Analysis of Variance Table
#> 
#> Response: mpg
#>                 Df Sum Sq Mean Sq F value  Pr(>F)    
#> cbind(disp, hp)  2 842.55  421.28 51.7367 1.3e-09 ***
#> poly(qsec, 2)    2  11.95    5.97  0.7337 0.49020    
#> factor(cyl)      2  67.98   33.99  4.1742 0.02728 *  
#> Residuals       25 203.57    8.14                    
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

performance::check_collinearity(mod)
#> # Check for Multicollinearity
#> 
#> Low Correlation
#> 
#>           Term  VIF     VIF 95% CI Increased SE Tolerance Tolerance 95% CI
#>  poly(qsec, 2) 4.39 [ 2.91,  6.99]         2.09      0.23     [0.14, 0.34]
#> 
#> Moderate Correlation
#> 
#>         Term  VIF     VIF 95% CI Increased SE Tolerance Tolerance 95% CI
#>  factor(cyl) 9.45 [ 5.99, 15.31]         3.07      0.11     [0.07, 0.17]
#> 
#> High Correlation
#> 
#>             Term   VIF     VIF 95% CI Increased SE Tolerance Tolerance 95% CI
#>  cbind(disp, hp) 23.03 [14.25, 37.63]         4.80      0.04     [0.03, 0.07]

^{Created on 2024-06-13 with reprex v2.1.0}

If you want each feature to be a term, you will need to hand code them.

mod <- lm(mpg ~ disp + hp + qsec + I(qsec^2) + I(cyl==6) + I(cyl==8),
          data = mtcars)

insight::find_terms(mod)
#> $response
#> [1] "mpg"
#> 
#> $conditional
#> [1] "disp"        "hp"          "qsec"        "I(qsec^2)"   "I(cyl == 6)"
#> [6] "I(cyl == 8)"

anova(mod)
#> Analysis of Variance Table
#> 
#> Response: mpg
#>             Df Sum Sq Mean Sq F value    Pr(>F)    
#> disp         1 808.89  808.89 99.3390 3.428e-10 ***
#> hp           1  33.67   33.67  4.1344   0.05277 .  
#> qsec         1   6.71    6.71  0.8235   0.37282    
#> I(qsec^2)    1   5.24    5.24  0.6438   0.42990    
#> I(cyl == 6)  1  59.13   59.13  7.2616   0.01241 *  
#> I(cyl == 8)  1   8.85    8.85  1.0867   0.30718    
#> Residuals   25 203.57    8.14                      
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

performance::check_collinearity(mod)
#> # Check for Multicollinearity
#> 
#> Low Correlation
#> 
#>         Term  VIF       VIF 95% CI Increased SE Tolerance Tolerance 95% CI
#>  I(cyl == 6) 2.09 [  1.53,   3.25]         1.45      0.48     [0.31, 0.65]
#> 
#> Moderate Correlation
#> 
#>  Term  VIF       VIF 95% CI Increased SE Tolerance Tolerance 95% CI
#>  disp 7.88 [  5.03,  12.72]         2.81      0.13     [0.08, 0.20]
#>    hp 9.03 [  5.74,  14.62]         3.01      0.11     [0.07, 0.17]
#> 
#> High Correlation
#> 
#>         Term    VIF       VIF 95% CI Increased SE Tolerance Tolerance 95% CI
#>         qsec 348.67 [212.31, 573.00]        18.67  2.87e-03     [0.00, 0.00]
#>    I(qsec^2) 320.53 [195.20, 526.75]        17.90  3.12e-03     [0.00, 0.01]
#>  I(cyl == 8)  11.38 [  7.17,  18.48]         3.37      0.09     [0.05, 0.14]

^{Created on 2024-06-13 with reprex v2.1.0}

[jmgirard]

If I modified the function to do the following, would you accept the PR?

• Check if there is only one term
• If so, check if it is a poly term
• If so, check if raw is true or false
• If true, refit with I() and output collinearity
• If false, output a note saying there is no collinearity

[bwiernik]

I don't think it makes sense to treat polynomials as being more than one variable/term. The collinearity of polynomial terms is non-essential, so VIF diagnostics don't make much sense for them. Collinearity isn't diagnostic of model issues for polynomial terms (raw or not)--the use of orthogonal polynomials is for computational reasons, not model validity reasons.

[jmgirard]

That makes sense to me. What about just a message when giving it only a poly term that explains this?

For example:

check_collinearity(lm(y ~ poly(x, 2), data = df))
#> Message: Polynomial transformations are considered a single term and thus VIFs are not calculated between them.

[DominiqueMakowski]

I think a note in the details section of the function documentation should suffice to avoid a very verbose messaging