1
$\begingroup$

Hi,

I have a set of N objects randomly distributed in a 2D physical space. Each object (i) generates a bernoulli random number (0 or 1) based on a marginal probability Pr(xi = 1) = p. These objects a correlated by physical distance. The closer the objects are, the larger their correlation is.

E.g. If objects i and j are co-located, they are expected to generate correlated results. For Example, if P(Xi=1)= 0.6 and P(Xj=1)=0.3 they would produce something like:

Xi= 0 1 0 1 1 1 0 1 0 1

Xj= 0 1 0 0 0 1 0 1 0 0

Such that Pr(Xi|Xj)=1

On the other hand if i and j are distant they would produce uncorrelated results such that Pr(Xi|Xj)=Pr(Xi)

I have tried to use some of the packages in Matlab (Sampling from multivariate correlated binary and poisson random variables) and R (bindata) but I could not produce an acceptable correlation matrix.

Any ideas how I can produce an acceptable correlation matrix?

BTW, I have checked the following earlier posts discrete stochastic process: exponentially correlated Bernoulli?

and

Constructing Bernoulli random variables with prescribed correlation

But I am not sure how I can relate to them.

Thanks

Improve this question
$\endgroup$

1 Answer 1

Reset to default
2
$\begingroup$

Here's a suggestion:

Define a non-negative decreasing function $w(r)$ measuring interaction strength. Given each object its own independent $N(0,1)$ random variable $N_i$. Now set $$ Y_i=\frac{\sum_{j}w(\|x_i-x_j\|)N_j}{\sqrt{\sum_j w(\|x_i-x_j\|)^2}}, $$ where $x_i$ denotes the location of the $i$th object.

Then the $Y_i$ are correlated $N(0,1)$ random variables. If two objects are co-located the normal random variables agree.

Finally set $t_i=\Phi^{-1}(p_i)$ (i.e. $\mathbb P(N < t_i)=p_i$) and set $X_i=1$ if $Y_i < p_i$ and 0 otherwise.

With this setup you can write down the covariance of $Y_i$ and $Y_k$ explicitly: it's just $$ \text{Cov}(Y_i,Y_k)=\frac{\sum_j w(\|x_i-x_j\|)w(\|x_k-x_j\|)} {\sqrt{\sum_j w(\|x_i-x_j\|)^2\sum_j w(\|x_k-x_j\|)^2}}. $$

If you write this as $\cos\theta_{ik}$ then you can write the covariance of $X_i$ and $X_k$ as an integral: $$ 1/(2\pi)\int_{ x < t_1\;,\; cos\theta_{ik}x+\sin\theta_{ik}y < t_2} e^{-(x^2+y^2)/2}\,dxdy-p_ip_k. $$

Improve this answer
$\endgroup$
6
  • $\begingroup$ If i understood you correctly, what you are saying is to simply use the Cov(Yi,Yk) above as the covariance matrix and plug it in the bernoulli generator. If so, then Unfortunately, I tried that but I keep getting an Unacceptable correlation matrix, which seams to be not positive definite. $\endgroup$
    – alandalusi
    Commented Apr 26, 2011 at 21:44
  • $\begingroup$ No you didn't understand me correctly. Here is my concrete suggestion. (1) Compute $t_i=\Phi^{-1}(p_i)$; (2) Compute the matrix $Cov(Y_i,Y_k)$ as above (3) Use a multivariate normal generator to build some $Y_i$'s. (4) Set $X_i=1$ if $Y_i<t_i$ and 0 otherwise. You could compute the covariance of the Bernoulli's (using the nasty integral formula I wrote down), but if your purpose is to just generate the random #s, then the procedure I described will work just as well. $\endgroup$
    – Anthony Quas
    Commented Apr 26, 2011 at 22:59
  • $\begingroup$ Acually, It is positive definite, but does not have a dichotomized Gaussian distribution for the correlation matrix. $\endgroup$
    – alandalusi
    Commented Apr 26, 2011 at 23:03
  • $\begingroup$ OK, I think I understand your suggestion now. I will test it right away. $\endgroup$
    – alandalusi
    Commented Apr 26, 2011 at 23:32
  • $\begingroup$ It works!, but I noticed that the Correlation (or Conditional Probabilities) $Pr(X{_i}=1|X{_k}=1)$ between any fixed pair (i and k) slightly change as we add more objects into the physical space. So, I changed $Cov(Y{_i},Y{_k})$ to be $w(||x{_i}-x{_k}||)$ and it worked with fixed conditional probabilities, even if I add more objects. I am still testing it, but does this make sense to you? For your information, my distance functions is $e^{(-\frac{eucaliandistance}{dcorr})}$ $\endgroup$
    – alandalusi
    Commented Apr 27, 2011 at 16:33

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .