Did I find a bug in the tseries or urca packages?

Question

I'm trying the functions to check the cointegration of a matrix.

I'm using Phillips & Ouliaris Cointegration Test

The function in tseries package is po.test and ca.po in urca

The results with urca are:

> ca.po(prices, demean='none')

######################################## 
# Phillips and Ouliaris Unit Root Test # 
######################################## 

Test of type Pu 
detrending of series none 


Call:
lm(formula = z[, 1] ~ z[, -1] - 1)

Residuals:
    Min      1Q  Median      3Q     Max 
-7.4960 -0.2912  0.7116  1.4530  3.3962 

Coefficients:
        Estimate Std. Error t value Pr(>|t|)    
z[, -1] 0.559705   0.004678   119.6   <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 

Residual standard error: 1.73 on 749 degrees of freedom
Multiple R-squared: 0.9503, Adjusted R-squared: 0.9502 
F-statistic: 1.431e+04 on 1 and 749 DF,  p-value: < 2.2e-16 


Value of test-statistic is: 12.9648 

Critical values of Pu are:
                  10pct    5pct    1pct
critical values 20.3933 25.9711 38.3413

The result with tseries are:

> po.test(prices, demean=FALSE)

    Phillips-Ouliaris Cointegration Test

data:  prices 
Phillips-Ouliaris standard = -25.6421, Truncation lag parameter = 7,
p-value = 0.01

Warning message:
In po.test(prices, demean = FALSE) : p-value smaller than printed p-value

As you can see I'm testing the same matrix (prices). How is it possible that urca tells there is NO cointegration and tseries: YES?

Prices max it's a simple matrix with two columns (stock1 - stock2), take a look to an extract of that.

1  3.065448  5.244870
2  3.094924  5.806821
3  2.873858  5.647601
4  3.205457  6.190820
5  3.315990  6.453064
6  3.168612  6.865161
7  3.271777  7.230428

Thank you

Answer 1

The devil is in the details. The help page po.test, you would have found this:

If lshort is TRUE, then the truncation lag parameter is set to trunc(n/100), otherwise trunc(n/30) is used.

And in help page of ca.po:

Usage

ca.po(z, demean = c("none", "constant", "trend"), lag = c("short", "long"), type = c("Pu", "Pz"), tol = NULL)

...

lag Either a short or long lag number used for variance/covariance correction.

So you can guess that the number of lags is chosen differently. The code from the functions justify this hypothesis. The code from po.test:

if (lshort) 
        l <- trunc(n/100)
    else l <- trunc(n/30)

From the ca.po:

if (lag == "short") {
        lmax <- trunc(4 * (nobs/100)^0.25)
    }
    else if (lag == "long") {
        lmax <- trunc(12 * (nobs/100)^0.25)
    }

Hence the statistics are actually different and so are the results.

This is not uncommon situation in testing for unit-roots and cointegration. If different statistics give different results, this usually means that something is missing. Also note that in general these statistics do not deal well with structural breaks, so if there are events which might of introduced structural breaks it would be prudent to take them into account.