基于已拟合2次多项式模型通过Y值反推caliber(X)值的方法
数据集与问题说明
已知校准数据集dataset1结构如下:
dataset1 = data.frame( caliber = c("5000", "2500", "1250", "625", "312.5", "156", "80", "40", "20", "0"), var1 = c(NA, NA, NA, 30458, 13740,11261, 9729, 5039, 3343, 367), var2 = c(463000, 271903, 154611,87204, 47228, 28082, 14842, 8474, 5121, 1308), var3 = c(308385, 184863, 89719, 48986, 27968, 18557, 9191, 5248, 3210, 703), var4 = c(290159, 149061, 64045, 36864, 19092, 12515, 6805, 3933, 2339, 574), var5 = c(270801, 163657, 51642, 48197, 23582, 14544, 7877, 4389, 2663, 482), var6 = c(NA, NA, NA, 37316, 21305, 11823, 5692, 3070, 1781, 363) )
caliber与各观测变量的定量关系为二次多项式形式:var = poly(caliber, 2, raw=T)。
拟合效果示意图
核心需求
待推算数据集dataset2中所有样本的caliber列均为缺失值,仅已知var1-var6的观测值,需要通过校准集的多项式关系反推caliber取值。dataset2结构如下:
dataset2 = data.frame( caliber = c(NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA), var1 = c(1120, 1296, 1132, 1280, 1096, 1124, 1004, 8384, 1072, 1104, 1568, 1044, 1108, 1012), var2 = c(5044, 4924, 5088, 4804, 4824, 4844, 4964, 4788, 4804, 4964, 4824, 4788, 4844, 4944), var3 = c(2836, 2744, 2744, 2668, 2688, 2940, 2756, 2720, 2668, 2892, 2636, 2700, 2836, 2668), var4 = c(8872, 61580, 3036, 4468, 12132, 3000, 7920, 6868, 6896, 9392, 4728, 6896, 21076, 3228), var5 = c(2312, 4236, 1928, 4448, 2388, 2108, 3644, 3060, 2168, 1912, 1812, 3528, 4100, 2176), var6 = c(1156, 1228, 1224, 1364, 1128, 1176, 1184, 1640, 1188, 1300, 1332, 1176, 1176, 1152) )
已查阅的参考内容
- predict x values from simple fitting and annoting it in the plot
- Predict X value from Y value with a fitted model
- get x-value given y-value: general root finding for linear / non-linear interpolation function
- aproxfun function from binsmooth package, find x from y value
遇到的具体报错
问题1:approx插值返回NA
执行以下代码时,formula$caliber不存在,导致插值结果异常:
formula <- lm(var2~poly(caliber,2,raw=T), dataset1) approx(x = formula$fitted, y = formula$caliber, xout = 0)$y
问题2:predict预测报错
执行以下代码时触发对象缺失错误:
mod<-lm(var2~poly(caliber, 2, raw=T), data=dataset1) summary(mod) newdata=data.frame("var2"=dataset2[1:24,c("var2")]) pred<-predict(mod,newdata, type = 'response')
Error in poly(caliber, 2, coefs = list(alpha = c(998.35, 3691.21383929929 :object 'caliber' not found
同时存在predict预测失败、数据集行数不匹配、插值结果错误的问题。
可运行解决方案
步骤1:修正数据格式
校准集中caliber为字符型,需转为数值型避免拟合异常:
dataset1$caliber <- as.numeric(dataset1$caliber)
步骤2:批量拟合多项式模型
对var1-var6分别拟合raw形式的二次多项式,存储模型系数:
var_names <- paste0("var", 1:6) # 剔除NA值后拟合模型 model_list <- lapply(var_names, function(v) { lm(as.formula(paste0(v, " ~ poly(caliber, 2, raw = TRUE)")), data = dataset1, na.action = na.omit) }) names(model_list) <- var_names
步骤3:编写反向求解函数
二次多项式的标准形式为y = a + b*x + c*x²,整理为一元二次方程c*x² + b*x + (a - y) = 0,代入求根公式求解caliber(即x),并筛选在合理取值范围(0-5000)内的有效解:
reverse_caliber <- function(y, fit_mod, calib_min = 0, calib_max = 5000) { mod_coef <- coef(fit_mod) intercept <- mod_coef[1] coef_linear <- mod_coef[2] coef_quad <- mod_coef[3] # 计算判别式 discriminant <- coef_linear^2 - 4*coef_quad*(intercept - y) if(discriminant < 0) return(NA) # 计算两个实根 root1 <- (-coef_linear + sqrt(discriminant))/(2*coef_quad) root2 <- (-coef_linear - sqrt(discriminant))/(2*coef_quad) # 筛选有效范围内的根 valid_roots <- c(root1, root2)[c(root1, root2) >= calib_min & c(root1, root2) <= calib_max] if(length(valid_roots) == 0) return(NA) return(valid_roots[1]) }
步骤4:批量计算dataset2的caliber值
对每个样本遍历所有可用变量的观测值,分别求解caliber后取均值,得到最终推算结果:
dataset2$caliber <- apply(dataset2, 1, function(single_row) { calib_vals <- c() for(v in var_names) { obs_y <- as.numeric(single_row[v]) if(!is.na(obs_y)) { calib_vals <- c(calib_vals, reverse_caliber(obs_y, model_list[[v]])) } } mean(calib_vals, na.rm = TRUE) })
报错根因说明
approx返回NA的原因:lm对象不会自动存储原始数据集的caliber字段,且拟合过程中因NA值剔除了部分行,直接引用formula$caliber会出现对象不存在、长度不匹配的问题。- predict报错的原因:模型的自变量是caliber,传入的newdata仅包含var2列,没有模型要求的caliber字段,poly函数匹配预计算系数时找不到caliber对象,因此触发报错。
内容的提问来源于stack exchange,提问作者Henrique
相关产品推荐
相关产品推荐

