Cross-sectional independence tests returning p-value NA

Question

I'm a Master student researching for his thesis.

In a panel data analysis, I have a sample of 141 observations related to airports. First lines are:

   Airport       Date Year PH.tot.seats Seat.year Seasonality
1      OLB 22/08/2014 2014         2534   2751080      2.6000
2      OLB 23/08/2015 2015         2449   2798846      2.5700
3      MAH 03/08/2014 2014         2460   2987482      2.5500
4      OLB 02/08/2013 2013         1740   2516710      2.5300
5      MAH 17/08/2015 2015         3686   3431666      2.5000
6      MAH 20/08/2012 2012         2681   2758799      2.4600
7      WMI 04/11/2012 2012         1863   1320165      2.4400
8      OLB 20/08/2012 2012         1932   2326278      2.4200
9      MAH 01/08/2011 2011         2344   2860084      2.3800
10     IBZ 03/08/2014 2014         3801   7170858      2.3700 
11     IBZ 06/08/2015 2015         3956   7465915      2.3500 
12     IBZ 13/08/2012 2012         3561   6345849      2.3400 
13     IBZ 03/08/2011 2011         3297   6616744      2.3100 
14     RHO 25/08/2015 2015         2703   3850181      2.3000 
15     IBZ 02/08/2013 2013         3438   6711774      2.2800 

The panel is unbalanced, with every unit between 1 and 7 times in the panel. After I find out that the fixed effect model is the most suitable and I run it, I test the usual assumptions. I followed this document as a guideline: http://www.princeton.edu/~otorres/Panel101R.pdf This is the code of the bit that does not work as I expected:

#panel regression-fixed effects n entity specific intercepts
fixed = plm(log(PH.tot.seats)~log(Seat.year)+log(Seasonality),data=db, 
index=c("Airport", "Year"),model="within")
summary(fixed)
fixef(fixed)
#tests for cross sectional dependence in panels
pcdtest(fixed,test=c("lm"))
pcdtest(fixed,test=c("cd"))

And this is the output for both the Pesaran and the BP test.

Is there any problems with the dataset or the code? Any suggestion would be highly appreciated.

Answer 1

It seems that pcdtest is no implemented for unbalanced panels. Weird enough it works with tiny unbalancedness but not with larger ones...

Here a reproducible example:

library(plm)

#in a data.frame like this:
testData <- data.frame(id=c(rep(1:200,each=20)),time=rep(1991:2010,200),
                   var1=rnorm(4000,10)/1000,var2=rnorm(4000,123))
#pcdtest will work:
pcdtest(var1 ~ var2, data=testData,
    model="within")

#even with slight unbalancedness:
testData1 <- testData[-(3:4),] 
pcdtest(var1 ~ var2, data=testData1,
    model="within")

#but not with slightly larger ones:
testData2 <- testData[-(3:30),] 
pcdtest(var1 ~ var2, data=testData2,
    model="within")

I do not know where the treshold is exactly or why it does not work.

To be honest, I'm starting to loose my faith in real world applications of plm which is a pity because it implements exactly what is usually needed. However, it performs quite badly with many fixed effects and often times even breaks down (just compare a large panel dataset where you do the within transformations yourself followed by lm() with an model=within,effects=twoways plm attempt). It also has troubles with unbalanced panels many times from my experience... Maybe pglm is better?