簡體   English   中英

plm:使用 fixef() 手動計算固定效應雙向模型的擬合值

[英]plm: using fixef() to manually calculate fitted values for a fixed effects twoways model

請注意:我試圖讓代碼同時處理時間和個人固定效應以及不平衡的數據集。 下面的示例代碼適用於平衡數據集。

也請參閱下面的編輯

我正在嘗試使用plm包手動計算固定效應模型(具有個體效應和時間效應)的擬合值。 這更像是一個練習,以確認我了解模型和包的機制,我知道我可以從plm對象,從兩個相關問題( 此處此處)中獲取擬合值本身。

plm小插圖 (p.2),底層模型是:

y _it = alpha + beta _transposed * x _it + ( mu _i + lambda _t + epsilon _it)

其中 mu_i 是誤差項的單個分量(又名“個體效應”),而 lambda_t 是“時間效應”。

可以通過使用fixef()提取固定效應,我想我可以使用它們(與自變量一起)來計算模型的擬合值,使用(使用兩個自變量)以這種方式:

擬合_it = alpha + beta _1 * x1 + beta _2 * x2 + mu _i + lambda _t

這就是我失敗的地方——我得到的值遠不及擬合值(我得到的是模型對象中實際值和殘差之間的差異)。 一方面,我在任何地方都看不到alpha 我嘗試將固定效果顯示為與第一個、平均值等的差異,但沒有成功。

我缺少什么? 這很可能是對模型的誤解,或者是代碼中的錯誤,恐怕……提前致謝。

PS:其中一個相關問題暗示pmodel.response()應該與我的問題有關(以及沒有plm.fit函數的原因),但它的幫助頁面並不能幫助我理解該函數的實際作用,我找不到任何示例來解釋它產生的結果。

謝謝!

我所做的示例代碼:

library(data.table); library(plm)

set.seed(100)
DT <- data.table(CJ(id=c("a","b","c","d"), time=c(1:10)))
DT[, x1:=rnorm(40)]
DT[, x2:=rnorm(40)]
DT[, y:=x1 + 2*x2 + rnorm(40)/10]
DT <- DT[!(id=="a" & time==4)] # just to make it an unbalanced panel
setkey(DT, id, time)

summary(plmFEit <- plm(data=DT, id=c("id","time"), formula=y ~ x1 + x2, model="within", effect="twoways"))

# Extract the fitted values from the plm object
FV <- data.table(plmFEit$model, residuals=as.numeric(plmFEit$residuals))
FV[, y := as.numeric(y)]
FV[, x1 := as.numeric(x1)]
FV[, x2 := as.numeric(x2)]

DT <- merge(x=DT, y=FV, by=c("y","x1","x2"), all=TRUE)
DT[, fitted.plm := as.numeric(y) - as.numeric(residuals)]

FEI <- data.table(as.matrix(fixef(object=plmFEit, effect="individual", type="level")), keep.rownames=TRUE) # as.matrix needed to preserve the names?
setnames(FEI, c("id","fei"))
setkey(FEI, id)
setkey(DT, id)
DT <- DT[FEI] # merge the fei into the data, each id gets a single number for every row

FET <- data.table(as.matrix(fixef(object=plmFEit, effect="time", type="level")), keep.rownames=TRUE) # as.matrix needed to preserve the names?
setnames(FET, c("time","fet"))
FET[, time := as.integer(time)] # fixef returns time as character
setkey(FET, time)
setkey(DT, time)
DT <- DT[FET] # merge the fet into the data, each time gets a single number for every row

# calculate the fitted values (called calc to distinguish from those from plm)
DT[, fitted.calc := as.numeric(coef(plmFEit)[1] * x1 + coef(plmFEit)[2]*x2 + fei + fet)]
DT[, diff := as.numeric(fitted.plm - fitted.calc)]

all.equal(DT$fitted.plm, DT$fitted.calc)

我的會話如下:

R version 3.2.2 (2015-08-14)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 8 x64 (build 9200)

locale:
[1] LC_COLLATE=English_United States.1252  LC_CTYPE=English_United States.1252    LC_MONETARY=English_United States.1252 LC_NUMERIC=C                          
[5] LC_TIME=English_United States.1252    

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] plm_1.4-0           Formula_1.2-1       RJSONIO_1.3-0       jsonlite_0.9.17     readxl_0.1.0.9000   data.table_1.9.7    bit64_0.9-5         bit_1.1-12          RevoUtilsMath_3.2.2

loaded via a namespace (and not attached):
 [1] bdsmatrix_1.3-2  Rcpp_0.12.1      lattice_0.20-33  zoo_1.7-12       MASS_7.3-44      grid_3.2.2       chron_2.3-47     nlme_3.1-122     curl_0.9.3       rstudioapi_0.3.1 sandwich_2.3-4  
[12] tools_3.2.2  

編輯:(2015-02-22)由於這引起了一些興趣,我將嘗試進一步澄清。 我試圖擬合一個“固定效應”模型(又名“內部”或“最小二乘虛擬變量”,正如plm 包小插圖在第 3 頁上的段落中所稱的那樣)——相同的斜率,不同的截距。

這與在為timeid添加虛擬變量后運行普通 OLS 回歸相同。 使用下面的代碼,我可以使用 base lm()plm包中復制擬合值。 對於假人,很明顯 id 和 time 的第一個元素是要比較的組。 我仍然不能做的是如何使用plm包的功能來做同樣的事情,我可以使用lm()輕松完成。

# fit the same with lm() and match the fitted values to those from plm()
lmF <- lm(data = DT, formula = y ~ x1 + x2 + factor(time) + factor(id))
time.lm <- coef(lmF)[grep(x = names(coef(lmF)), pattern = "time", fixed = TRUE)]
time.lm <- c(0, unname(time.lm)) # no need for names, the position index corresponds to time

id.lm <- coef(lmF)[grep(x = names(coef(lmF)), pattern = "id", fixed = TRUE)]
id.lm <- c(0, unname(id.lm))
names(id.lm) <- c("a","b","c","d") # set names so that individual values can be looked up below when generating the fit

DT[, by=list(id, time), fitted.lm := coef(lmF)[["(Intercept)"]]  +  coef(lmF)[["x1"]] * x1  +  coef(lmF)[["x2"]] * x2  +  time.lm[[time]]  +  id.lm[[id]]]
all.equal(DT$fitted.plm, DT$fitted.lm)

希望這對其他可能感興趣的人有用。 問題可能與plmfixef如何處理我有意創建的缺失值有關。 我嘗試使用fixeftype=參數,但沒有效果。

這適用於具有effect="individual"和時間假人y ~ x +factor(year)的不平衡數據:

fitted <- pmodel.response(plm.model)-residuals(plm.model)

我發現這可以幫助你,因為 lm() 解決方案在我的情況下不起作用(與 plm 包相比,我得到了不同的系數)

因此,這只是應用 plm 包作者的建議http://r.789695.n4.nabble.com/fitted-from-plm-td3003924.html

所以我所做的只是申請

plm.object <- plm(y ~ lag(y, 1) + z +z2, data = mdt, model= "within", effect="twoways")
fitted <- as.numeric(plm.object$model[[1]] - plm.object$residuals) 

我需要 as.numeric 函數的地方,因為我需要將它用作向量來插入以進行進一步的操作。 我還想指出,如果您的模型在右側有一個滯后的因變量,那么上面的 as.numeric 解決方案提供了一個向量,由於滯后,已經包含缺失值的 NET。 對我來說,這正是我需要的。

我非常接近 Helix123 的建議,即減去within_intercept (它包含在兩個固定效果中的每一個中,因此您需要對此進行更正)。

我的重建錯誤中有一個非常具有啟發性的模式:個體a總是偏離 -0.004858712(對於每個時間段)。 個體b, c, d在每個時間段總是偏離 0.002839703,除了在第 4 期(沒有觀察a ),它們偏離 -0.010981192。

有什么想法嗎? 看起來個體的固定效應被不平衡所拋棄。 重新運行它平衡工作正常。

完整代碼:

DT <- data.table(CJ(id=c("a","b","c","d"), time=c(1:10)))
DT[, x1:=rnorm(40)]
DT[, x2:=rnorm(40)]
DT[, y:= x1 + 2*x2 + rnorm(40)/10]
DT <- DT[!(id=="a" & time==4)] # just to make it an unbalanced panel
setkey(DT, id, time)

plmFEit <- plm(formula=y ~ x1 + x2,
               data=DT,
               index=c("id","time"),
               effect="twoways",
               model="within")

summary(plmFEit)

DT[, resids := residuals(plmFEit)]

FEI <- data.table(as.matrix(fixef(plmFEit, effect="individual", type="level")), keep.rownames=TRUE) # as.matrix needed to preserve the names?
setnames(FEI, c("id","fei"))
setkey(FEI, id)
setkey(DT, id)
DT <- DT[FEI] # merge the fei into the data, each id gets a single number for every row

FET <- data.table(as.matrix(fixef(plmFEit, effect="time", type="level")), keep.rownames=TRUE) # as.matrix needed to preserve the names?
setnames(FET, c("time","fet"))
FET[, time := as.integer(time)] # fixef returns time as character
setkey(FET, time)
setkey(DT, time)
DT <- DT[FET] # merge the fet into the data, each time gets a single number for every row

DT[, fitted.calc := plmFEit$coefficients[[1]] * x1 + plmFEit$coefficients[[2]] * x2 +
     fei + fet - within_intercept(plmFEit)]

DT[, myresids := y - fitted.calc]
DT[, myerr := resids - myresids]

編輯:適應雙向不平衡模型,需要plm版本>= 2.4-0

這是你想要的嗎? 通過fixef提取固定效應。 以下是非平衡雙向模型上的 Grunfeld 數據示例(對平衡雙向模型的工作方式相同):

gtw_u <- plm(inv ~ value + capital, data = Grunfeld[-200, ], effect = "twoways")
yhat <- as.numeric(gtw_u$model[ , 1] - gtw_u$residuals) # reference
pred_beta <- as.numeric(tcrossprod(coef(gtw_u), as.matrix(gtw_u$model[ , -1])))
pred_effs <- as.numeric(fixef(gtw_u, "twoways")) # sum of ind and time effects

all.equal(pred_effs + pred_beta, yhat) # TRUE -> matches fitted values (yhat)

如果您想在其組件中拆分個體和時間效果的總和(由effect = "twoways" ),您需要選擇一個參考,並且很自然地想到兩個,如下所示:

# Splits of summed up individual and time effects:
# use one "level" and one "dfirst"
ii <- index(gtw_u)[[1L]]; it <- index(gtw_u)[[2L]]
eff_id_dfirst <- c(0, as.numeric(fixef(gtw_u, "individual", "dfirst")))[ii]
eff_ti_dfirst <- c(0, as.numeric(fixef(gtw_u, "time",       "dfirst")))[it]
eff_id_level <- as.numeric(fixef(gtw_u, "individual"))[ii]
eff_ti_level <- as.numeric(fixef(gtw_u, "time"))[it]

all.equal(pred_effs, eff_id_level  + eff_ti_dfirst) # TRUE
all.equal(pred_effs, eff_id_dfirst + eff_ti_level)  # TRUE

(這是基於 fixef 的手冊頁?fixef 。那里還展示了如何處理(平衡和非平衡)單向模型)。

暫無
暫無

聲明:本站的技術帖子網頁,遵循CC BY-SA 4.0協議,如果您需要轉載,請注明本站網址或者原文地址。任何問題請咨詢:yoyou2525@163.com.

 
粵ICP備18138465號  © 2020-2024 STACKOOM.COM