Tell me more ×
Cross Validated is a question and answer site for statisticians, data analysts, data miners and data visualization experts. It's 100% free, no registration required.
up vote 1 down vote favorite
1

Suppose I have the conditional probabilities of $B$ given $A$ (both categorical): $$\Pr(B|A).$$ The probabilities are given as a contingency table.

Using these conditional probabilities, I can fudge a $B$ column into a dataset that has only an $A$ column. For each row in the dataset, I choose a value for $B$ at random based on $P(B|A)$ for the value of $A$ of the current row.

From my understanding, this can be seen as a model, although a quite simple one. Adding the $B$ column could be thought of as "simulating" the model.

My question: Is the term "model" applicable in this case? If so, how would such a model be called?

Somebody mentioned the term "Poisson distribution" in this context, but I was unable to align these two concepts with my limited knowledge of statistics.

See also this question on how to implement this in R.

share|improve this question

1 Answer

up vote 1 down vote accepted

I would call it a multinomial model. For each possible value of $A$, $B$ follows a different multinomial distribution. But that doesn't really add anything except jargon. You might say more about the nature of $A$, $B$ and $\Pr(B | A)$.

share|improve this answer
$A$ and $B$ are both categorical (discrete) variables. $\Pr(B|A)$ is given as map from $A \times B$ to $[0, 1]$ and is otherwise unspecified. – krlmlr Sep 25 '11 at 6:37
@user946850 - Yes, but what is the scientific (or otherwise) context for A and B? I guess I still don't really understand what you're after in your question. – Karl Sep 26 '11 at 13:20
Suppose I have a large sample with an $A$ attribute and a smaller, but still representative, sample with $A$ and $B$ attributes. I want to add a $B$ attribute to the large sample using the conditional probabilities suggested by the small sample and the values of the $A$ column in the large sample. – krlmlr Sep 26 '11 at 22:20
@user946850 - There's no real name for this model that is more precise than multinomial (which covers any discrete distribution), unless $\Pr(B|A)$ had a particular form. Your particular would be called imputation. The Poisson distribution is a particular discrete distribution that could match $\Pr(B|A)$ but probably is not at all related to this. – Karl Sep 27 '11 at 3:12
1  
@user946850 - yes, though actually, if I were writing it myself, I would omit the phrase "using a simple multinomial model". – Karl Sep 27 '11 at 7:59
show 1 more comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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