thanks for your note. pls see below...
On Wed, 27 Jul 2005, Parry Clarke wrote:
> Dear Professor King,
>
>
> My name is Parry Clarke and I am currently writing up my PhD on baboon
> intersexual conflict. Im sorry to bother you but I was wondering if you
> could answer some questions regarding Logistic Regression of Rare events.
>
> I am currently attempting to model the occurrence of male aggression
> directed at oestrous females. The behaviour itself is fairly rare and so
> I have binary coded its occurrence. In total I have 1745 sampling units,
> or rows of data, with only 21 of these coded 1 and 1724 coded 0.
> Initially I carried out standard logistic regression, but found that
> when it came to diagnosis all my influential points were my 1s. So
> that, if deletion diagnostics were performed I was left with no variance
> In addition, my intercepts did not seem very convincing.
>
> However, having now discovered your work on the subject and the package you
> have written for R things are looking up!, but I just have a couple of
> questions:
>
> 1)Are there diagnostics unique to the rare event analysis?
> 2)Is influential point examination redundant in rare event logistic
> regression?
> 3)How do I get deviance estimates of the final model?
relogit estimates the same coefficients as logit from the same model.
they do differ tho in order to get better properties. so just as using
weighted least squares will give different answers -- and will fit the
data less well than least squares -- we generally prefer wls to ls when
there are weights available. so in both relogit and wls, you have the
same issue of how to deal with diagnostics. there are no special
diagnostics, but in both (and lots of other methods) the issue is that you
can't really treat all the observations equally and an outlier for one
observation isn't the same as another. so in relogit for example, an
extra 1 inadvertently included in the dataset will be much more
consequential than an extra 0.
> 4)Part of my analysis has to been trying to relate other forms of male
> aggression to oestrous female-directed aggression. However, these other forms
> are also rare and so when I enter them into the analysis as a dichotomous
> explanatory variable I get answers that are not really supported by the
> observed data: For example: overlap between the 1s in the response and
> explanatory variables may only be 2 or 3 data points but the final model
> suggests that a large amount of the deviance is explained and the
> coefficients are highly significant: Is this simply an artifact of the rarity
> of both the dependent and independent variable and am I, therefore, better
> off excluding them from the multivariate analysis and doing separate
> contingency table analysis with them?
rare events in explanatory variables is a different issue involving the
sensitivity of the estimates to the coding of X. since almost all
relevant models are conditional on X, you have little choice but to run
the analysis as is unless you reconceptualize the project (such as having
both variables being dependent variables).
> 5)In a number of your articles (e.g. Explaining Rare Events In
> International Relations) your talk at length about sampling strategies
> and database trimming to create favourable ratio of 0s to 1s and to cut
> down on costs. In a situation such as mine where the data is collected
> and the database fixed is there any need to trim and subset. If so how
> do you suggest I go about that and is there a command in R for randomly
> selecting subsets of data?
if you have the data, you should use it. no reason to subsample at that
point. we do it to demonstrate what would happen if you couldn't afford
to collect all the data, as is the case in many fields. but more data are
better generally and here too.
>
> Once again I am sorry to bother you with such a lengthy email and I hope it
> is not too much of an imposition.
best of luck with your research,
Gary King
>
>
> Yours sincerely,
>
> Parry Clarke.
>
> _________________________________________________________________
> Winks & nudges are here - download MSN Messenger 7.0 today!
> http://messenger.msn.co.uk
>
>
On Wed, 13 Jul 2005, Barbaresco Gabriele wrote:
>> Dear Prof. King,
>>
>> I have read with great interest the following your papers on logit and
>> rare events:
>>
>> Logistic regression in rare events Explaining rare events in
>> international relations
>>
>> I really appreciated them, also due to the lack of academic literature
>> on this important topics. I am currently working on a model to forecast
>> companies failure and so I am dealing with logit and rare events.
>>
>> Referring to the second paper, the less technical one, I'd like to be
>> advised on the following points:
>>
>> 1) Pag. 697, I see that the standard errors of betas depend on the
>> estimated probability, that in rare events is usually very far from
>> 0.5 , but greater (if the model works well) for Y=1 than for Y=0.
>> So, you argue, the ones have more informative power. To reduce these
>> errors, can I balance the sample (50% of ones and 50% of zeros) or
>> better use the original sample balancing observations that they are
>> equally weighted? It seems from my simulation that rebalancing the
>> sample lead to better Wald test and also to better pseudo Rsquared
>> measures of fitting. In doing so the estimated probability range
>> from 0 to 1, and 0.5 could be used as cutoff.
if you can 'balance' the sample by collecting additional data (the 1s),
then that's great. but if you're talking about discarding 0s, that's not
generally a good idea (although there are possible exceptions unrelated to
your point). in statistics, more data are better.
>> 2) Page 704, here again it seems to me that the bias coefficient on the
>> costant is smaller the bigger is the sample (n stand at denominator)
>> and the closer the probability to 0.5. I work with a sample of
>> around 180.000 items, if I rebalance the sample as above, can I
>> reduce this source of bias too?
probably not.
>>
>> In a nutshell, it seems to me that rebalancing the sample can reduce
>> some of the problems you highlight, but it seem also that you never
>> suggest such a solution, so I wonder if I have misunderstood the
>> meaning of your paper.
we thought about it and tried it when writing the paper, and there is
probably still something going on that might be useful from a theoreetical
perspective, but from you perspective, I wouldn't discard any data if you
don't have to.
Gary King
---
Gary King
Institute for Quantitative Social Science
Harvard University, 34 Kirkland St, Cambridge, MA 02138
http://GKing.Harvard.Edu, King(a)Harvard.Edu
Direct 617-495-2027, Assistant 495-9271, eFax 812-8581
>>
>> Sorry for my boring questions, and thank you very much for your kind
>> assistance.
>>
>> Best regards,
>>
>> Mr. Gabriele Barbaresco
>> Research Dept. - Mediobanca
>> Milan - Italy
>>
>> www.mediobanca.it
>> www.mbres.it
>
-
relogit mailing list served by Harvard-MIT Data Center
List Address: relogit(a)latte.harvard.edu
Subscribe/Unsubscribe: http://lists.hmdc.harvard.edu/?info=relogit
