st: Re: simple way to create missing data that is "missing at random" from a small datset

[Suzy <scott_788@wowway.com>]

Thanks Maarten for providing me more detail on your command. I worked 
with the constant and now have the correct proportion of missingness, 
although I'm not sure what the implications are of the std dev and the 
max values of p (.549). Now that I better understand what the command is 
doing, I will continue to work with the values and look at the outcomes. 
I really appreciate your help!

. gen p = invlogit( -8 +.1*age )

. sum p

   Variable |       Obs        Mean    Std. Dev.       Min        Max
-------------+--------------------------------------------------------
          p |       332    .0999268     .113432   .0054863    .549834

. replace bmi = . if uniform() < p
(27 real changes made, 27 to missing)




Maarten buis wrote:


> Suzy:
>
> No problem, but if you find my reply puzzeling than chances are that someone else on statalist
> might find it puzzeling too, so I also sent my reply (and your full question underneath) to the
> statalist. 
> The variable p is the probability of missingness, so the mean of p should be .1 if you want
> apporximately 10% missingness. Your mean is .99, so most people will be made missing.  -invlogit-
> transforms a linear function of "explanatory variables" (in yourcase .1*age) to lie between zero
> and one according to 1/(1+exp{-xb}), so the values you plug in (in your case .1 for age and 0 for
> the constant) are "logistic regression coefficients". I would play around with values of the
> constant so that you get a mean p of about .1 (the more negative the constant the lower the
> probability), For instance look at the mean of p if you do -gen p =invlogit(-10 + .1*age)-
>
> Afterwards I would look if there is enough variation in the values of p. If the value of p is
> approximately constant than the influence of age on the probability of missingness is probably not
> strong enough to show up in your simulations. If p is approximately constant you should increase
> the parameter of age. This might than mess up the mean probability of missingness a bit, so than
> it would be good to check if the mean probability of missingness is still close to .1
>
> HTH,
> Maarten
>
> --- Suzy <scott_788@wowway.com> wrote:
>
>  
>
> > Dear Maarten:
> >
> > Hope you don't mind the direct e-mail. I tried your code based on my 
> > dataset and what I thought I should do and all of my BMI observations 
> > went missing rather than say 5-10%. I have obviously done something 
> > wrong with it. I'm hoping you can help. I would like about 10% of the 
> > BMI variable to be missing. I want the missingness to be associated with 
> > older age, but not dependent on the value of BMI - thus hopefully 
> > satisfying the MAR assumption.
> >
> > I've included the summary stats of the variables, the code you provided 
> > (I modified it somewhat) and the result...
> > can you see what I did wrong??
> >
> > summarize
> >
> >    Variable |       Obs        Mean    Std. Dev.       Min        Max
> > -------------+--------------------------------------------------------
> >         sex |       332    .4849398    .5005275          0          1
> >        race |       332    .3253012    .4691944          0          1
> >         age |       332    52.06024     12.6857         28         82
> >        fhdm |       332    .3373494    .4735189          0          1
> >         bmi |       332    30.98795     6.18837         18         48
> > -------------+--------------------------------------------------------
> >       dmcat |       332    .2771084    .4482461          0          1
> >
> > . gen p = invlogit(.1*age)
> >
> > . sum p
> >
> >    Variable |       Obs        Mean    Std. Dev.       Min        Max
> > -------------+--------------------------------------------------------
> >           p |       332    .9894261    .0121324   .9426758   .9997254
> >
> >
> > . replace bmi = . if uniform() < p
> > (332 real changes made, 332 to missing)
> >
> > . summarize
> >
> >    Variable |       Obs        Mean    Std. Dev.       Min        Max
> > -------------+--------------------------------------------------------
> >         sex |       332    .4849398    .5005275          0          1
> >        race |       332    .3253012    .4691944          0          1
> >         age |       332    52.06024     12.6857         28         82
> >        fhdm |       332    .3373494    .4735189          0          1
> >         bmi |         0
> > -------------+--------------------------------------------------------
> >       dmcat |       332    .2771084    .4482461          0          1
> >           p |       332    .9894261    .0121324   .9426758   .9997254
> >
> >
> >
> >
> >    
>
> -----------------------------------------
> between 1/2/2006 and 31/3/2006 I will be
> visiting the UCLA, during this time the
> best way to reach me is by email
>
> Maarten L. Buis
> Department of Social Research Methodology
> Vrije Universiteit Amsterdam
> Boelelaan 1081
> 1081 HV Amsterdam
> The Netherlands
>
> visiting adress:
> Buitenveldertselaan 3 (Metropolitan), room Z214
>
> +31 20 5986715
>
> http://home.fsw.vu.nl/m.buis/
> -----------------------------------------
>
>
> 		
> ___________________________________________________________ 
> Win a BlackBerry device from O2 with Yahoo!. Enter now. http://www.yahoo.co.uk/blackberry
> *
> *   For searches and help try:
> *   http://www.stata.com/support/faqs/res/findit.html
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
>
>
>
>  

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/