re: st: Question on a quasi-time series

["Ariel Linden, DrPH" <ariel.linden@gmail.com>]

Hi Rachel,

I think the regression approach is suitable here, as opposed to time series
analysis. However, I am not clear on how exactly you are using nbreg? 

If I read your posting correctly, you are taking what data you have and
converting that into a daily rate? My concern would be that you have a
tremendous amount of missing data for all the days in which you don't (or
can't) collect data. As a result, your "rate" is miss-specified because of
the missing data (unless you can make an argument that the data are missing
at random, and all observations across all years are equally subject to this
rule).

Because I don't know the content of your data (or the limitations of your
data collection), I would hate to steer you in the wrong direction. It seems
to me that if you have certain months (or days) every year in which you are
certain that your data are accurate (and representative), you could limit
your analysis to only those months. Then you could compare your counts (or
rates) year over year using panel type models (assuming these are the same
animals in your data set year over year), or straight count models such as
nbreg or poission (clustered to account for serial dependency within units).

As for setting the p value at 0.01, I am not sure why that would be
considered a fix for the problem of serial dependency? Clustering, or using
longitudinal models are certainly more appropriate (and defensible)
approaches.

I hope this helps.

Ariel 

 
Date: Wed, 23 Mar 2011 16:17:06 +0000
From: rachel grant <rachelannegrant@gmail.com>
Subject: Fwd: st: Question on a quasi-time series

Hi Ariel
I have been using regression approaches and only including the data
that I have for each year. Ideally I would like to keep the regression
based approach as I have now gone so far along that path...I am not
able to start a new statitistical technique in the time I have
available so is there an easy way fo correcting for the serial
dependency within the negative binomial regression. I have seen in
some papers they set the confidence limit at p=0.01 to cover this. Is
that a valid approach?

many thans
Rachel Grant

On 23 March 2011 15:43, Ariel Linden, DrPH <ariel.linden@gmail.com> wrote:
>
> It seems to me that time series analysis is not the appropriate approach
> here. If there are only a couple of monthly observations available for
each
> year of data, I cannot imagine how the model will fit the missing 10
months
> in any accurate way.
>
> I would consider comparing only the observations for "like" months in each
> year so that you are limiting the evaluation to data you actually have.
> Similarly, you'd have a good rationale for limiting the analysis to only
> those months of observation if indeed no other observations occur outside
> the timeframe.
>
> As to which statistical approaches to take, I think that perhaps a
> difference-in-differences approach could be considered, or perhaps some
type
> of correlation approach (i.e., intra-class, controlling for within year
> serial dependency)
>
> Just some food for thought...
>
>  Ariel



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/