5
$\begingroup$

I am trying to find out how to conduct run tests on one sample but with a non 0.5-0.5 underlying occurrence pattern of the two dichotomous values.

For instance +++---++++, but assume that +'s had an underlying probability of 0.8 of happening whereas -'s had 0.2.

Is there a runs test which can address this?

(If so, is there an R function for it?)

$\endgroup$
4
  • $\begingroup$ Asking for code / packages is off-topic here. Did you try Googling it? $\endgroup$ – gung - Reinstate Monica Apr 2 '15 at 18:19
  • 1
    $\begingroup$ Yes of course. You can alternatively suggest a test that can conduct this analysis. $\endgroup$ – appletree Apr 2 '15 at 18:21
  • 1
    $\begingroup$ There's many different "runs test" test-statistics. What test statistic are you interested in using? The total number of runs of +'s and -'s? The number of runs of one sign? the length of the longest run? ... Note that the widely used Wald-Wolfowitz* runs test conditions on the total +/- counts, so doesn't assume p(+)=0.5. $\:\:$ *(incorrectly attributed to them, but that's Stigler's law for you) $\endgroup$ – Glen_b Apr 3 '15 at 0:35
  • $\begingroup$ in R, randtests::runs.test (on CRAN) can do it (code - as -1 and + as 1, and set the threshold to 0) $\endgroup$ – Glen_b Apr 3 '15 at 0:56
5
$\begingroup$

If you're interested in a runs test of randomness (which I suppose from your tags, but your text doesn't indicate), the Wald-Wolfowitz runs test conditions on the totals of the +/- counts, so it doesn't assume $p(+)=\frac{1}{2}$.

The Wald-Wolfowitz test (e.g. see Stephens, 1939) is a permutation test based on the total number of runs of both kinds, conditional on the total number of symbols of each kind.

I presume you're interested in too few runs (the usual case), which indicates "clumping of signs".

Your example data has 3 runs, with $n_+=7$ and $n_-=3$. There are 120 arrangements of the symbols, of which 2 have 2 runs (all + followed by all - and vice versa) and of the 3-run cases there are 2 with 7 "+"s in the center and 6 with 3 "-"s in the center, for a total of 10 cases with 3 or fewer runs:

2 runs:
--- +++++++
+++++++ ---

3 runs:
- +++++++ --
-- +++++++ -

+ --- ++++++
++ --- +++++
+++ --- ++++
++++ --- +++
+++++ --- ++
++++++ --- +

so the exact p-value of that case is 10/120.

It's a widely used test, most stats packages offer it in some way, though some only offer the asymptotic approximation (which works better with larger $n_+$ and $n_-$).

If you're interested in testing something else or using some different statistic, you'll need to say more.

(In R, randtests::runs.test (on CRAN) can do the Wald-Wolfowitz test (code - as -1 and + as 1, and set the threshold to 0); if your sample size is large enough to use the asymptotic approximation, it's easy to code by hand, in any case)

[1] Stevens W.L. (1939),
"Distribution of groups in a sequence of alternatives"
Annals of Eugenics 9:10-17

$\endgroup$
2
  • $\begingroup$ Thank you very much. Great to confirm this. Do you by any chance have any suggestions on a runs test where the underlying distribution can be given and a comparison is made to it? $\endgroup$ – appletree Apr 3 '15 at 20:11
  • $\begingroup$ What do you intend by 'the distribution can be given'? The pmf for the Wald-Wolfowitz statistic is given in many places, not least the reference in my answer. That same reference has a 'runs of one kind' test which has a pmf with a 'standard' distribution (hypergeometric), but I wouldn't normally recommend it over the Wald-Wolfowitz. $\endgroup$ – Glen_b Apr 4 '15 at 0:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.