问题
This is what I would like to do:
library("lmtest")
library("dynlm")
test$Date = as.Date(test$Date, format = "%d.%m.%Y")
zooX = zoo(test[, -1], order.by = test$Date)
f <- d(Euribor3) ~ d(Ois3) + d(CDS) + d(Vstoxx) + d(log(omo)) + d(L(Euribor3))
m1 <- dynlm(f, data = zooX, start = as.Date("2005-01-05"),end = as.Date("2005-01-24"))
m2 <- dynlm(f, data = zooX, start = as.Date("2005-01-25"), end=as.Date("2005-02-14"))
summary(m1)
summary(m2)
coeftest(m1, vcov=NeweyWest)
coeftest(m2, vcov=NeweyWest)
There is no problem with the function summary(m1)
However if I want to use HAC by NeweyWest i.e.coeftest(m1, vcov=NeweyWest)
I get the following error message and I do not know why: Error in na.fail.default(as.ts(x)) : missing values in object
What do I have to change on my code in order to get the result using coeftest()? Note: there are no missing values as you can see from the sample data. Many thanks!
Sample Data:
Date Euribor3 Ois3 Vstoxx CDS omo
1 03.01.2005 2.154 2.089 14.47 17.938 344999
2 04.01.2005 2.151 2.084 14.51 17.886 344999
3 05.01.2005 2.151 2.087 14.42 17.950 333998
4 06.01.2005 2.150 2.085 13.80 17.950 333998
5 07.01.2005 2.146 2.086 13.57 17.913 333998
6 10.01.2005 2.146 2.087 12.92 17.958 333998
7 11.01.2005 2.146 2.089 13.68 17.962 333998
8 12.01.2005 2.145 2.085 14.05 17.886 339999
9 13.01.2005 2.144 2.084 13.64 17.568 339999
10 14.01.2005 2.144 2.085 13.57 17.471 339999
11 17.01.2005 2.143 2.085 13.20 17.365 339999
12 18.01.2005 2.144 2.085 13.17 17.214 347999
13 19.01.2005 2.143 2.086 13.63 17.143 354499
14 20.01.2005 2.144 2.087 14.17 17.125 354499
15 21.01.2005 2.143 2.087 13.96 17.193 354499
16 24.01.2005 2.143 2.086 14.11 17.283 354499
17 25.01.2005 2.144 2.086 13.63 17.083 354499
18 26.01.2005 2.143 2.086 13.32 17.348 347999
19 27.01.2005 2.144 2.085 12.46 17.295 352998
20 28.01.2005 2.144 2.084 12.81 17.219 352998
21 31.01.2005 2.142 2.084 12.72 17.143 352998
22 01.02.2005 2.142 2.083 12.36 17.125 352998
23 02.02.2005 2.141 2.083 12.25 17.000 357499
24 03.02.2005 2.144 2.088 12.38 16.808 357499
25 04.02.2005 2.142 2.084 11.60 16.817 357499
26 07.02.2005 2.142 2.084 11.99 16.798 359999
27 08.02.2005 2.141 2.083 11.92 16.804 355500
28 09.02.2005 2.142 2.080 12.19 16.589 355500
29 10.02.2005 2.140 2.080 12.04 16.500 355500
30 11.02.2005 2.140 2.078 11.99 16.429 355500
31 14.02.2005 2.139 2.078 12.52 16.042 355500
EditI think the problem is in this command: zooX = zoo(test[, -1], order.by = test$Date), namely the function order.by(). If you delete this part, all else equal you can compute HAC by NeweyWest.( Of course you also need to change start = as.Date("2005-01-25")to the index number, e.gstart=50.) But by deleting it you lose the start and end Date in the regression output, which was very useful. So if anyone knows a way around it, please let me know!
回答1:
NeweyWest calculates the 'lag' with this code:
lag <- floor(bwNeweyWest(x, order.by = order.by, prewhite = prewhite,
ar.method = ar.method, data = data))
... and when called with the default arguments it replicates your (and my replication of it) error:
>bwNeweyWest(m2,lag = NULL, order.by = NULL, prewhite = TRUE, adjust = FALSE,
+ diagnostics = FALSE, sandwich = TRUE, ar.method = "ols",
+ data = list(), verbose = FALSE)
Error in na.fail.default(as.ts(x)) : missing values in object
The example on the ?NeweyWest page suggests that pre-specifying a lag was the original strategy. I've run your coeftest calls with lags of 2,3, and 4 and do not see much sensitivity to the choice of lags:
coeftest(m1, vcov=NeweyWest(m1, lag = 2, prewhite = FALSE) )
#-------------
t test of coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.00088591 0.00024838 -3.5667 0.007329 **
d(Ois3) -0.01256761 0.20243315 -0.0621 0.952020
d(CDS) 0.00010732 0.00097169 0.1104 0.914774
d(Vstoxx) 0.00121163 0.00051398 2.3573 0.046150 *
d(log(omo)) 0.01245017 0.01916762 0.6495 0.534190
d(L(Euribor3)) -0.42173541 0.11765274 -3.5846 0.007141 **
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
#----------
coeftest(m2, vcov=NeweyWest(m2 , lag = 2, prewhite = FALSE ) )
#------------
t test of coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -0.00055562 0.00029246 -1.8998 0.08992 .
d(Ois3) 0.25659641 0.17004507 1.5090 0.16558
d(CDS) -0.00276703 0.00197776 -1.3991 0.19530
d(Vstoxx) -0.00091397 0.00063662 -1.4357 0.18493
d(log(omo)) -0.01524269 0.02810579 -0.5423 0.60076
d(L(Euribor3)) -0.51430803 0.17335182 -2.9668 0.01578 *
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
I'm not sure whether this is an oversight on the part of the dynlm authors or the sandwich authors. You might send them an email to get more authoritative comment on statistical validity matters.
来源：https://stackoverflow.com/questions/30003047/r-hac-by-neweywest-using-dynlm