2

I am modelling a outcome in gam were two variables (x1 and x2) are continuous and the other four are factors. I have a suspicion that x1 and x2 might be collinear so I want to check that. As I understand I need to check with concurvity. The figures of approx. 0.37 which I expect to be OK, or?

   A1 <- gam(y~ s(x1) + s(x2), data= df, method = "REML")

concurvity(A1)
                 para     s(x1)     s(x2)
worst    1.604547e-25 0.3765562 0.3765562
observed 1.604547e-25 0.3637563 0.2497401
estimate 1.604547e-25 0.2926522 0.2767589


 > concurvity(A1, full = FALSE)
    $worst
                  para        s(x1)        s(x2)
    para  1.000000e+00 9.035414e-26 7.145648e-26
    s(x1) 9.194812e-26 1.000000e+00 3.765562e-01
    s(x2) 7.033065e-26 3.765562e-01 1.000000e+00

    $observed
                  para        s(x1)        s(x2)
    para  1.000000e+00 4.660693e-34 4.259220e-30
    s(x1) 9.194812e-26 1.000000e+00 2.497401e-01
    s(x2) 7.033065e-26 3.637563e-01 1.000000e+00

    $estimate
                  para        s(x1)        s(x2)
    para  1.000000e+00 3.875674e-28 3.293079e-28
    s(x1) 9.194812e-26 1.000000e+00 2.767589e-01
    s(x2) 7.033065e-26 2.926522e-01 1.000000e+00

The problem is that when I include my factors I get high concurvity with the parametric terms which I cannot understand. To check, I have made x3-x6 as completely random variables (0 and 1) and I still get the high concurvity.

df$x3 <- sample(c(0,1), replace = TRUE, size=368)
df$x4 <- sample(c(0,1), replace = TRUE, size=368)
df$x5 <- sample(c(0,1), replace = TRUE, size=368)
df$x6 <- sample(c(0,1), replace = TRUE, size=368)

df$x3 <- as.factor(df$x3)
df$x4 <- as.factor(df$x4)
df$x5 <- as.factor(df$x5)
df$x6 <- as.factor(df$x6)


A2 <- gam(y~ s(x1) + s(x2)  +x3 +x4+ x5 + x6, data= df, method = "REML")


> concurvity(A2)
              para     s(x1)     s(x2)
worst    0.8005751 0.3817906 0.3921831
observed 0.8005751 0.3701165 0.3551481
estimate 0.8005751 0.2994385 0.2887118


> concurvity(A2, full = FALSE)
$worst
              para        s(x1)        s(x2)
para  1.000000e+00 9.035414e-26 7.145648e-26
s(x1) 9.038433e-26 1.000000e+00 3.765562e-01
s(x2) 7.125601e-26 3.765562e-01 1.000000e+00

$observed
              para        s(x1)        s(x2)
para  1.000000e+00 3.565108e-34 1.338613e-31
s(x1) 9.038433e-26 1.000000e+00 3.393421e-01
s(x2) 7.125601e-26 3.637036e-01 1.000000e+00

$estimate
              para        s(x1)        s(x2)
para  1.000000e+00 3.875674e-28 3.293079e-28
s(x1) 9.038433e-26 1.000000e+00 2.767589e-01
s(x2) 7.125601e-26 2.926522e-01 1.000000e+00

Can someone please explain why and what I am doing wrong?

curious
  • 21
  • 4

1 Answers1

0

I'm not an expert on GAMs by any stretch of the imagination, but I've been wrestling with them for a few months now. My understanding of the issue of high concurvity in the parametric terms is this: since the parametric terms are estimated with a very simple function (linear) it is very easy to explain that function in terms of the more complex smooths in the model. There is another function in the package mgcv.helper called vif.gam() that can calculate the variance inflation factors for the parametric terms... Unfortunately, it doesn't work for factor variables that require a GVIF (generalized variance inflation factor). I found that it wasn't to hard to modify the vif() function in the car package to work for gam objects, though. You just have to make it sure it recognizes the difference between parametric terms and smooth terms.

Take everything I said with a grain of salt. I'd love for a true expert to weigh in, as I would like to feel more confident about this as well.

Edit:

You will probably find this answer very helpful: vif.gam()

dante
  • 121
  • 6