第三章多元线性回归模型第六节预测应变量的点预测将解释变量预测值代入估计的方程便可:0102030k0012300...kYXXXXYX或者=一、的置信区间)XY(E0000YX=对于模型给定样本以外的解释变量的观测值X0=(1,X10,X20,…,Xk0),可以得到被解释变量的预测值:它可以是总体均值E(Y0)或个值Y0的预测。但严格地说,这只是被解释变量的预测值的估计值,而不是预测值。为了进行科学预测,还需求出预测值的置信区间,包括E(Y0)和Y0的置信区间。00YX=易知0000022000002100E(Y)E(X)XE(YX)Var(Y)Var(X)E[XE(X)]E[X()]X(X'X)X'容易证明)1(~ˆˆknt)E(YY00010XX)X(X于是,得到(1-)的置信水平下E(Y0)的置信区间:010000100)(ˆˆ)()(ˆˆ22XXXXXXXXtYYEtY其中,t/2为(1-)的置信水平下的临界值。2100000Y~N[E(YX),X(X'X)X']二、的置信区间如果已经知道实际的预测值Y0,那么预测误差为:0Y000ˆYYe容易证明0)YY(E)e(E0002210000Var(e)Var(Y)[1+X(X'X)X']e0服从正态分布,即)))(1(,0(~01020XXXXNe)))(1(ˆˆ010220XXXXe构造t统计量)1(~ˆˆ000kntYYte可得给定(1-)的置信水平下Y0的置信区间:010000100)(1ˆˆ)(1ˆˆ22XXXXXXXXtYYtYe0服从正态分布,即)))(1(,0(~01020XXXXNe)))(1(ˆˆ010220XXXXe构造t统计量)1(~ˆˆ000kntYYte可得给定(1-)的置信水平下Y0的置信区间:010000100)(1ˆˆ)(1ˆˆ22XXXXXXXXtYYtY中国居民人均收入-消费支出二元模型例中:2001年人均GDP:4033.1元,于是人均居民消费的预测值为Ŷ2001=120.7+0.2213×4033.1+0.4515×1690.8=1776.8(元)实测值(90年价)=1782.2元,相对误差:-0.31%预测的置信区间:00004.000001.000828.000001.000001.000285.000828.000285.088952.1)(1XX3938.0010XX)X(X于是E(Ŷ2001)的95%的置信区间为:3938.05.705093.28.1776或(1741.8,1811.7)3938.15.705093.28.1776或(1711.1,1842.4)同样,易得Ŷ2001的95%的置信区间为第四章非线性回归模型的线性化变量间的非线性关系线性化方法案例分析非线性回归模型非线性关系线性化的几种情况:①对于指数曲线,令,可以将其转化为直线形式:,其中,;②对于对数曲线,令,,可以将其转化为直线形式:;③对于幂函数曲线,令,,可以将其转化为直线形式:其中,;bxdeyxbayxbaylnxbaybdxyxbayyylnxxdalnyyxxlnyylnxxlndaln④对于双曲线,令,转化为直线形式:;⑤对于S型曲线,可转化为直线形式:;⑥对于幂乘积:,只要令,就可以将其转化为线性形式:其中,;xbay1xbayxxexyybeay,1,1令xbaykkxxdxy2121kkxxxy22110xxyy1,1,ln,,ln,ln,ln2211kkxxxxxxyydln0⑦对于对数函数和只要令,就可以将其化为线性形式:例:下表给出了某地区林地景观斑块面积(Area)与周长(Perimeter)的数据。下面我们建立林地景观斑块面积A与周长P之间的非线性回归模型。kkxxxylnlnln22110kkxxxy22110kkxxxxxxyyln,,ln,ln,2211序号面积A周长P序号面积A周长P110447.370625.39242232844.3004282.043215974.730612.286434054.660289.307330976.770775.7124430833.840895.98049442.902530.202451823.355205.131510858.9201906.1034626270.300968.060621532.9101297.9624713573.9601045.07276891.680417.0584865590.0802250.43583695.195243.90749157270.4002407.54992260.180197.239502086.426266.54110334.33299.729513109.070261.8181111749.080558.921522038.617320.396122372.105199.667533432.137253.335138390.633592.893541600.391230.030146003.719459.467553867.586419.406表某地区各个林地景观斑块面积(m2)与周长(m)15527620.2006545.291561946.184198.66116179686.2002960.4755777.30556.9021714196.460597.993587977.719715.7521822809.1801103.0705919271.8201011.1271971195.9401154.118608263.480680.710203064.242245.049614697.1301234.1142469416.7008226.0091624519.867326.3171225738.953498.6566313157.6601172.916238359.465415.151646617.270609.801246205.016414.790654064.137437.3552560619.0201549.871665645.820432.3552614517.740791.943676993.355503.7842731020.1001700.965684304.281267.9512826447.1601246.977696336.383347.136297985.926918.312702651.414292.235解:(1)作变量替换,令:,,将上表中的原始数据进行对数变换,变换后得到的各新变量对应的观测数据如下表所示。AylnPxln序号y=lnAx=LnP序号y=lnAx=LnP19.2541066.4383794212.358138.36218629.6787636.4172438.3076225.667487310.340996.6537824410.336376.79791849.1530196.273258457.5084335.3236559.2927427.5528164610.176196.87529469.9773387.168551479.5159096.95184178.838076.0332264811.091187.71887988.2147895.4967894911.965727.78636497.72325.284414507.6432085.585528105.8121354.602457518.0420795.567651119.371536.326008527.6200275.769558表经对数变换后的数据127.7715335.296653538.1409385.534711139.0348716.385013547.3780035.438211148.7001346.130066558.2603866.0388391513.176138.786501567.5736265.2915971612.098977.993105574.3477554.041328179.5607486.393579588.9844086.5733341810.034927.005852599.8663996.9188211911.173197.051092609.0196016.523136208.0275565.501457619.5954087.1181092113.059259.015056628.4162385.787871228.6550326.211917639.4847597.067248239.031156.028643648.7974386.413133248.7331136.027773658.3099576.0807442511.012367.345927668.6386716.069247269.5831276.67449678.8527166.222147(2)以x为横坐标、y为纵坐标,在平面直角坐标系中作出散点图。很明显,y与x呈线性关系。456789101112131445678910LnPlnA图林地景观斑块面积(A)与周长(P)之间的双对数关系(3)根据所得表中的数据,运用建立线性回归模型的方法,建立y与x之间的线性回归模型,得到:x与y的相关系数高达=0.9665。(4)将还原成双对数曲线,即5057.0505.1xy5057.0ln505.1lnPAxyr不可线性化的非线性回归模型的线性化估计方法直接搜索法直接优化法迭代线性化法设定一般的回归模型yi=f(xi;)+uiOLS估计不能得到其明确的解的形式例:2222220110110011101212(;)()2[(;)]0()2[]0()2[]0()2[]0iiiiixniiiinxiinxxiinxxiiiyeufxSyfxSyeSyeeSyexeβββββββ泰勒级数展开与迭代估计公式:将模型yt=f(x,)+u在参数(0,0,1,0,…)展开:00,01,0,00,0002,0,00000,00000:(,,)(,)(,)()1()()2!(,)kktttiiiikkiijjijijkkttiitiiiiInitialestimatorsfyfffffyfuβxβxβxβ0,010,11,1,10000(,)(,,)kkttiitkiiiiffyfuxββ,1,,:ijijijStopRule几种对数模型的比较对数模型比较模型形式含义经济解释yt=β0+β1xt+utdy=β1dxx变化一个单位,y变化β1个单位yt=β0+β1Ln(xt)+utdy=β1dx/xx变化1%,y变化β1/100个单位Ln(yt)=β0+β1xt+utdy/y=β1dxx变化一个单位,y变化(100β1)%Ln(yt)=β0+β1Ln(xt)+utdy/y=β1dx