I am just beginng an investigation involving characterizing the dependence between two 1D stochastic point processes $x$, $y$. The natural approach seems to involve Ripley's K-function:

$$ K(t) = \frac{T}{n_xn_y} \sum_{i=1}^{n_x} \sum_{j=1}^{n_y} w(x_i,y_j) I[d(x_i,y_j)<t] $$

where $n_x$ is the number of observations in $x$ and $n_y$ is the number of observations in $y$ across the interval $T$. Deviation from $K(t)=t $ is an indication of correlation between the two point processes.

However, it's not clear to me how to estimate the edge correction $w(x_i,y_j)$ the 1D case. Some papers refer to Hani Doss' 1989 papers, but in the JSTOR paper he explicity states that ' this edge correction will not concern us.' FWIW-I am currently using a correction weight of 2, but intuitively that seems excessive in my case.

  • $\begingroup$ I am confused by the apparent contradiction between "1D" (one-dimensional?) and "bivariate" (two-dimensional, of course). Would you mind explaining what you mean by these? $\endgroup$ – whuber Aug 22 '12 at 17:05
  • 2
    $\begingroup$ @whuber: there seems to be various ways to refer this situation (bivariate, cross, multivariate, etc.) and, with apologies, I tried to hit them all. Hopefully the edit will clear up the confusion. I believe you have referred to it as 'cross-K' in a slightly related posting. $\endgroup$ – Aengus Aug 22 '12 at 19:35
  • $\begingroup$ You may be interested in some more recent work that looks at Ripley's K for point processes on 1-D road networks. See Ang et al. 2011 for one example. They have implemented their corrections in the spatstat package for R. $\endgroup$ – Andy W Aug 23 '12 at 12:34
  • $\begingroup$ @Andy: Thanks much for the reference. I had not seen this; there's some good ideas in here and, while not directly applicable to the immediate problem, will apply to some related work. $\endgroup$ – Aengus Aug 23 '12 at 13:39
  • $\begingroup$ The best edge correction for 1D Ripley's functions seems to be to weight by dividing the counted points by the proportion of the interval that was actually sampled. For a 2t interval around i that gets out of the transect start (zero) it will be w=(i+t)/2t; on the other side, w=(t-i+[top limit])/2t. So this "inflates" the actual counted points within 2t to compensate for the "uncounted" (though expected under stationarity) that were out of bounds [Sorry, I can't find the proper reference now, but will add it if it eventually finds me again] $\endgroup$ – FairMiles Sep 2 '15 at 16:06

Have you seen Gavin's K1D code?

In his vignette, synchronicity is $K(t) > 2t$ and asynchrony between $2t$ and $0$. I believe the $2t$ arises from the evaluation of the pair in both ways which is necessary as Gavin explains that sometimes the edge correction $w(t_i,t_j)\neq w(t_j,t_i)$.

| cite | improve this answer | |
  • $\begingroup$ Yes, the algorithm I used was more general than Gavin’s $\endgroup$ – Aengus Jun 18 '19 at 1:42

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.