Polynomial Regression nonsense Predictions

Question

Suppose I want to fit a linear regression model with degree two (orthogonal) polynomial and then predict the response. Here are the codes for the first model (m1)

x=1:100
y=-2+3*x-5*x^2+rnorm(100)
m1=lm(y~poly(x,2))
prd.1=predict(m1,newdata=data.frame(x=105:110))

Now let's try the same model but instead of using $poly(x,2)$, I will use its columns like:

m2=lm(y~poly(x,2)[,1]+poly(x,2)[,2])
prd.2=predict(m2,newdata=data.frame(x=105:110))

Let's look at the summaries of m1 and m2.

> summary(m1)

Call:
lm(formula = y ~ poly(x, 2))

Residuals:
     Min       1Q   Median       3Q      Max 
-2.50347 -0.48752 -0.07085  0.53624  2.96516 

Coefficients:
              Estimate Std. Error t value Pr(>|t|)    
(Intercept) -1.677e+04  9.912e-02 -169168   <2e-16 ***
poly(x, 2)1 -1.449e+05  9.912e-01 -146195   <2e-16 ***
poly(x, 2)2 -3.726e+04  9.912e-01  -37588   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 

Residual standard error: 0.9912 on 97 degrees of freedom
Multiple R-squared:     1,      Adjusted R-squared:     1 
F-statistic: 1.139e+10 on 2 and 97 DF,  p-value: < 2.2e-16 

> summary(m2)

Call:
lm(formula = y ~ poly(x, 2)[, 1] + poly(x, 2)[, 2])

Residuals:
     Min       1Q   Median       3Q      Max 
-2.50347 -0.48752 -0.07085  0.53624  2.96516 

Coefficients:
                  Estimate Std. Error t value Pr(>|t|)    
(Intercept)     -1.677e+04  9.912e-02 -169168   <2e-16 ***
poly(x, 2)[, 1] -1.449e+05  9.912e-01 -146195   <2e-16 ***
poly(x, 2)[, 2] -3.726e+04  9.912e-01  -37588   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 

Residual standard error: 0.9912 on 97 degrees of freedom
Multiple R-squared:     1,      Adjusted R-squared:     1 
F-statistic: 1.139e+10 on 2 and 97 DF,  p-value: < 2.2e-16 

So m1 and m2 are basically the same. Now let's look at the predictions prd.1 and prd.2

> prd.1
        1         2         3         4         5         6 
-54811.60 -55863.58 -56925.56 -57997.54 -59079.52 -60171.50 

> prd.2
         1          2          3          4          5          6 
  49505.92   39256.72   16812.28  -17827.42  -64662.35 -123692.53 

Q1: Why prd.2 is significantly different from prd.1?

Q2: How can I obtain prd.1 using the model m2?

Answer 1

m1 is the right way to do this. m2 is entering a whole world of pain...

To do predictions from m2, the model needs to know it was fitted to an orthogonal set of basis functions, so that it uses the same basis functions for the extrapolated new data values. Compare: poly(1:10,2)[,2] with poly(1:12,2)[,2] - the first ten values are not the same. If you fit the model explicitly with poly(x,2) then predict understands all that and does the right thing.

What you have to do is make sure your predicted locations are transformed using the same set of basis functions as used to create the model in the first place. You can use predict.poly for this (note I call my explanatory variables x1 and x2 so that its easy to match the names up):

px = poly(x,2)
x1 = px[,1]
x2 = px[,2]

m3 = lm(y~x1+x2)

newx = 90:110
pnew = predict(px,newx) # px is the previous poly object, so this calls predict.poly

prd.3 = predict(m3, newdata=data.frame(x1=pnew[,1],x2=pnew[,2]))