How to derive SinglePredictionBands from a NonlinearModelFit model

Question

I have found NonlinearModelFit to be extremely useful in my work. Here I have fit 4 points to a function and am being asked to explain how the 95% SinglePredictionBands associated with the plot below were derived/computed. Can someone show how these bands can be computed from “first principles” ? In other words, I’d need to use Mathematica to independently produce the functions labeled lowerband and upperband. My code is below. Many thanks.

x = {1, 2, 3, 4};
y = {0.891, 0.885, 0.844, 0.836};
data1 = Transpose[{x, y}];
plot1 = ListPlot[data1, Frame -> True, 
FrameTicks -> {Range[0, 5, 1], Range[0.8, 0.9, .05], None, None}, 
FrameLabel -> {"X", "Y"}, PlotRange -> {{0, 5}, {0.8, 0.9}}, 
PlotStyle -> Red, ImageSize -> 500, 
LabelStyle -> Directive[Black, FontSize -> 14]];
Remove[a, b];
model1 = NonlinearModelFit[data1, a + b t^2, {a, b}, t] ;
Normal[model1];
params = model1["BestFitParameters"];
{a = params[[1, 2]], b = params[[2, 2]]};
predictionbands = model1["SinglePredictionBands"];
lower[t_] := predictionbands[[1]];
upper[t_] := predictionbands[[2]];
plot1a = Plot[model1[t], {t, 1, 4}, Frame -> True, PlotStyle -> Blue, 
PlotRange -> {{0, 4}, All}];
plot1b = Plot[model1[t], {t, 4, 9}, Frame -> True, 
PlotStyle -> {Blue, Dashed}, PlotRange -> All];
plot2a = Plot[lower[t], {t, 5, 9}, Frame -> True, 
FrameTicks -> {Range[0, 9, 1], Range[0, 0.9, .1], None, None},
PlotRange -> All, PlotStyle -> {Green, Dashed}];
plot2b = Plot[upper[t], {t, 5, 9}, Frame -> True, 
FrameTicks -> {Range[0, 9, 1], Range[0.5, 1, .1], None, None},
PlotRange -> {0.5, 1}, PlotStyle -> {Green, Dashed}];
text1 = Text["95% Upper Prediction Band", {7.5, .92}];
text2 = Text["95% Lower Prediction Band", {3.8, .68}];
text3 = Text["Nominal Prediction", {7, .75}];
plot3 = Show[plot1b, plot1a, plot2a, plot2b, plot1, 
AxesOrigin -> {0, 0}, PlotRange -> {.6, 1}, 
FrameTicks -> {Range[0, 9, 1], Range[.6, 1, .05], None, None}, 
FrameLabel -> {"X", "Y"}, 
Epilog -> {Line[{{-1, .65}, {9, .65}}], text1, text2, text3}, 
ImageSize -> 500, ImageSize -> 500, 
LabelStyle -> Directive[Black, FontSize -> 14]]

enter image description here

Here are the SinglePredictionBands that I need to independently produce.

lowerband = model1["SinglePredictionBands"][[1]]

0.894058 - 0.00400775 t^2 - 4.30265 Sqrt[0.000237726 - 0.0000163949 t^2 + 1.09299*10^-6 t^4]

upperband = model1["SinglePredictionBands"][[2]]

0.894058 - 0.00400775 t^2 + 4.30265 Sqrt[0.000237726 - 0.0000163949 t^2 + 1.09299*10^-6 t^4]

http://en.wikipedia.org/wiki/Confidence_and_prediction_bands#Prediction_bands — Chris Degnen, Oct 25 '14 at 16:03
Thanks for the pointers. The wikipedia article seems to indicate that the key to my question is w(x) but I don't see a definition for w(x) for the context of a prediction band there. — Steve, Oct 25 '14 at 16:15
The detail is in the prediction interval link. You need to calculate interval points to form the band line. — Chris Degnen, Oct 25 '14 at 16:23
Also something here: http://www.itl.nist.gov/div898/handbook/pmd/section5/pmd512.htm — Chris Degnen, Oct 25 '14 at 16:37

Chris Degnen · Accepted Answer · 2015-01-14T13:35:43.307

The fit given in this question, a + b t^2, is actually linear in a and b, so the regression line can be found with a linear fit, i.e.

x = {1, 2, 3, 4};
y = {0.891, 0.885, 0.844, 0.836};
data1 = Transpose[{x, y}];

n1 = Normal@NonlinearModelFit[data1, a + b t^2, {a, b}, t];

n2 = Normal@LinearModelFit[data1, {1, t^2}, t];

n1 == n2

True

The prediction intervals can be found using the two-variable least-squares model. First the data is linearised. Note this does not affect the regression coefficients:

data2 = Transpose[{x^2, y}];
n3 = Normal@LinearModelFit[data2, {1, s}, s]; 
CoefficientList[n2, t^2] == CoefficientList[n3, s]

True

Using the linearised data various quantities are computed:

MapIndexed[(X@#2 [[1]] = #1) &, x^2]; 
MapIndexed[(Y@#2 [[1]] = #1) &, y];

{n = Length[data2],
 ΣX = Sum[X[i], {i, n}],
 ΣY = Sum[Y[i], {i, n}],
 ΣXY = Sum[X[i] Y[i], {i, n}],
 ΣX2 = Sum[X[i]^2, {i, n}]}

{4, 30, 3.456, 25.403, 354}

{{a, b}} = {a, b} /. Solve[{
     (* Normal equations for straight line *)
     ΣY == n a + b ΣX,
     ΣXY == a ΣX + b ΣX2}, {a, b}];

Show[Plot[a + b x, {x, 0, 20}], ListPlot[data2]]

enter image description here

(* Least-squares regression of Y on X *)
Array[(Yhat[#] = a + b X[#]) &, n];

Array[(e[#] = Y[#] - Yhat[#]) &, n];
(* Residual or unexplained sum of squares *)
RSS = Sum[e[i]^2, {i, n}];

(* Estimate of disturbance variance, s^2 *)
s2 = RSS/(n - 2);
s = Sqrt[s2];

Xmean = ΣX/n;
(* Derivation from sample x mean *)
Clear[x]; Array[(x[#] = X[#] - Xmean) &, n];
Σx2 = Sum[x[i]^2, {i, n}];

(* Confidence limit for +/-2 S.D. of the mean *)
σ = 2; cl = 2 (CDF[NormalDistribution[0, 1], σ] - 0.5)

0.9545

nvars = Length[{a, b}];
(* t-statistic *)
t = Quantile[StudentTDistribution[n - nvars], (1 + cl)/2]

4.52654

(* Estimated variance of a new observation *)
ev[X0_] := t s Sqrt[1 + 1/n + (X0 - Xmean)^2/Σx2]

interval[X0_] := {a + b X0 + ev[X0], a + b X0 - ev[X0]}

In accordance with the data linearisation the input to interval is squared.

lines[x_] := interval[x^2]

xvals = Table[i, {i, 5, 9, 0.04}];
{line1, line2} = Transpose[lines /@ xvals];

Show[Plot[a + b x^2, {x, 0, 10}], ListPlot[data1, PlotStyle -> Red],
 ListPlot[{Transpose[{xvals, line1}], Transpose[{xvals, line2}]},
  PlotStyle -> Directive[Green, Dashed],
  Joined -> True], AxesOrigin -> {0, 0}, ImageSize -> 480, 
 Frame -> True, PlotRange -> {{0, 10}, {0.5, 1}}]

enter image description here

The derivation of the estimated variance can be read in Econometric Methods (Ed.3) by J.Johnston, (chapter 2, The Two-Variable Linear Model), page 43.

How to derive SinglePredictionBands from a NonlinearModelFit model

1 Answers1

Linked