Thanks to everyone who replied to my query: here's a summary of the
responses I received
cheers
Chris
***********************************************************
> Sorry I forgot to sign my previous email. Consequently I will
> resubmit.
>
> My name is Nelse Grundvig and I work as a research analyst.
>
> I hesitate to respond because I do not know how your data is
> collected nor the structure of your database. If your data is
> monthly with classification of workers one of your variables. You
> could use trend analysis to forecast furture absenteeism. The
> technique would depend on the number of observations (generally you
> need 60 observations). I do similar work when projecting short-term
> employment patterns by industry. This is done in SPSS with the
> trends model.
***********************************************************
> Date: Mon, 26 Apr 1999 10:41:15 -0400
> From: Alan Hutson <[log in to unmask]>
> The regression models for exponentially distributed data are
> typically found in survival texts even though the outcome
> doesn't necessarily need to be censored/survival data.
>
> I don't know about SPSS, but in SAS you can fit the type of
> model you describe straightforward in proc lifereg.....
> otherwise just take a log transform of the data
>
> Best of luck
> Alan
***********************************************************
> Date: Mon, 26 Apr 1999 15:42:53 +0100 (GMT)
> From: Jane Hutton <[log in to unmask]>
> Dear Chris
>
> It sounds like a poisson regression problem to me - try SAS or
> GLIM?
>
> At a guess, you might be interested in or need to model zero
> absences and non-zero separately.
>
> There are standard models for exponential data in GLIM, or SAS - you
> can use survival models wihtout censoring, if you wish to treat this
> as continuous exponential rather than Poisson.
>
> best wishes
> Jane
***********************************************************
> From: Robert Shelly <[log in to unmask]>
> You might try redefining the data so you have the ln of the observed
> values as the dependent variable in your analysis...I have had some
> success with this in some data I have worked with. RS
***********************************************************
> From: ozcan <[log in to unmask]>
> Organization: METU-ODTU
>
> Did you try tobit analysis?
> Regards,
***********************************************************
> Try using a Gamma error and a Link function on which you are happy
> to interpret differences in any group comparisons you wish to make.
> It effectively does what a log-transformation would do without
> leaving you with estimated effects in terms of geometric means.
> Yours Tony Swan
>
> PHLS Statistics Unit, e-mail: [log in to unmask]
***********************************************************
> Is this not likely to be a fairly classical situation for using
> Poisson Regression?
>
> Kind Regards,
>
> John
>
>Dr John Whittington,
***********************************************************
> From: "J.M.Russell" <[log in to unmask]>
> Have you looked up the literature on transforming variables. You
> might care to look at Practical Statistics for Medical Research by
> Douglas G. Altman pg 143 following.
>
> If you really are desperate you can fiddle with the loss function in
> no linear regression or maybe use the weighted least squares.
> Jean.
***********************************************************
> From: "Ridout, Martin" <[log in to unmask]>
> I think the exponential distribution fits into the generalised
> linear model framework (special case of gamma distribution,
> with dispersion parameter=1).
>
> Martin
***********************************************************
> From: "PAUL MCGALE" <[log in to unmask]>
>
> Hi,
>
> Not entirely sure I know what you mean, but if
> your response variable is a set of failure times
> and these are exponetially distributed then you can
> fit an exponential regression model. This is essentially
> the same as a doing a poisson regression using the
> log of the fail times as the offset. Any standard stats
> package should let you do either of these.
>
> Hope this is of some help.
>
> Cheers
> Paul
***********************************************************
> From: David McGeoghegan <[log in to unmask]>
> At Westlakes we (seem to) spend a lot of time on the analysis of
> sickness absence data. The duration of absence is conveniently
> analysed using Tweedie compound poisson models. The Glim 4 macro
> library (release 2) contains a macro for this type of modelling,
> called Tweedie, Deviation, written by Bent Jorgensen. There is also
> a book 'The theory of dispersion models' by Bent Jorgensen, Chapman
> & Hall, Monographs in statistics and applied probability, Vol 76,
> 1997, which may be of interest.
>
> Hope this helps,
>
> Dave
***********************************************************
> From: "Max Martin" <[log in to unmask]>
> Chris,
>
> If you get any useful off-list replies to your query, would you
> please forward them? I too work with absence/attendance data in
> predictive and explanatory models and the severe skew always
> presents problems, even after transforming the data by taking the
> inverse (recommended in Tabachnick & Fidell)
>
> Messy, messy!
>
> Thanks
>
> Max
***********************************************************
> You might want to try a poisson regression.
>
> Ashley Blackman
> The Faneuil Group
***********************************************************
> Chris,
>
> Had you tried stochastic modelling, with Negative ou Beta Binomiale
> Distribution? I haven't developped such approach using SPSS but you
> only need a MLE syntax..
>
> Good uck
> Naji
***********************************************************
> Date: Mon, 26 Apr 1999 20:01:47 -0700
> From: Florabelle Gagalac <[log in to unmask]>
> try tobit of logit models. These are used for highly skewed data.
>
> SAS/STAT Manual has a good explanation and example of this kind.
>
> Hope it helps.
>
> Cheers,
> Olla
***********************************************************
> From: Jay Warner <[log in to unmask]>
> No refs., per se. But what I did was to do a log transform.
> Bingo!
> Then examine for serious outliers. True case: The manager of the
> dept. sat down with the one outlier, after transform, showed him the
> oriignal histogram, explained that 1 out of ordinary point suggested
> something amiss, indicated some possible reasons, some +, some -
> from the view of the individual. No further absenses noted.
> Conclusion, individual had been gold bricking, didn't know that the
> boss was noticing.
***********************************************************
> From: "Beaney, Steve J" <[log in to unmask]>
> Sounds like you want to look at
> Generalized Linear Models, MCCullagh and Nelder
***********************************************************
> My immediate thought is to use a generalized liear model -
> exponential is just a special case of the Gamma. The software is no
> longer a problem as R is now free Jan.
>
> G.Janacek e-mail [log in to unmask]
> School of Mathematics tel 44-(0)1603-592577
> UEA, Norwich N
***********************************************************
> Dear Chris
>
> Usually count data analysed by using log-linear models, easily
> fitted in a number of well-known statistical packages: SAS, SPSS
> e.t.c. An excellent reference for this kind of analysis is :
>
> Title: Modelling frequency and count data
> Author: J K Lindsey
> Publisher: Oxford university Press
>
> Have a look at Chapter 1, Page 17, where an analysis for a dataset
> similar to the one you described is given. From there on you can
> easily find your way through the literature to analyse your data.
> Another good reference on log-linear models is the one by Agresti,
> "Categorica Data Analysis", John Wiley & Sons, where you can also
> find similar type of examples.
>
> Hope this helps
>
> Regards
>
> Dimitrios Lambrou
> Statistician
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|