is it necessary to test autocorrelation for panel data in binary logistic regression?

[derya]

Dear members,

I ran a binary logistic model using panel data. I collected quarterly
mutual fund data. But it`s unequal, what I mean by that is for some funds
there is data for only couple of months because the fund is terminated but
for others there is data for the full sample period. I am trying to
estimate the probability of fund termination by logistic model.
Significant independent variables are lagged returns (5., 6. and 7. lags)
and net asset value. The other two variables (age and net fund flow) are
not significant. I checked some text books and googled and found out that
there is no need to test for heteroskerasticity in logistic regression
analysis, there is a need to test for multicollinearity but could not find
out if there is a need to test for autocorrelation and how to do it for
binary logistic model. I used PcGive for my model but if I have to I may
try using SPSS. Please let me know if you have an answer to my question.

There is one more thing i would like to ask: if a variable is already
transformed (log net asset value), how can I get log(net asset value)^2 to
test for nonlinearity? Do I need to multiply it by itself?

Thanks to you all.

Derya.

[Christian Ganser]

Dear Derya,

I'm sorry this isn't a direct answer to your question, but to me it
seems you should ask yourself if a logistic regression model is
appropriate. As far as I understand, your dependent variable in fact is
the time it takes for a fund to be terminated? So for funds which
weren't terminated during your sample period, you have censored data. In
this case, you should think about using event history analsysis.

Regards
Christian

derya wrote:

> Dear members,
>
> I ran a binary logistic model using panel data. I collected quarterly
> mutual fund data. But it`s unequal, what I mean by that is for some funds
> there is data for only couple of months because the fund is terminated but
> for others there is data for the full sample period. I am trying to
> estimate the probability of fund termination by logistic model.
> Significant independent variables are lagged returns (5., 6. and 7. lags)
> and net asset value. The other two variables (age and net fund flow) are
> not significant. I checked some text books and googled and found out that
> there is no need to test for heteroskerasticity in logistic regression
> analysis, there is a need to test for multicollinearity but could not find
> out if there is a need to test for autocorrelation and how to do it for
> binary logistic model. I used PcGive for my model but if I have to I may
> try using SPSS. Please let me know if you have an answer to my question.
>
> There is one more thing i would like to ask: if a variable is already
> transformed (log net asset value), how can I get log(net asset value)^2 to
> test for nonlinearity? Do I need to multiply it by itself?
>
> Thanks to you all.
>
> Derya.
>
>

[derya]

Thanks for your answer, Christian.

I should have written in my message that my focus was on survivorship bias
literature and that my aim was to examine the determinants of fund
dissappearance for a developing country mutual funds. I know that survival
analysis can also be used and it probably is more appropriate. But I may
study it as another paper.

Derya.