ExcelStan

[Andrew Gelman]

Hi all. Is there an ExcelStan? Could be worth doing, no?
A

[Daniel Lee]

Given our difficulties with Windows installs and RStan, I'd say this would be much too much of a maintenance burden to even start. And I'm not sure if Stan would actually compile under MSVC now.

If you convince me that Windows installs are not a problem, then I'd say someone should look into it.

Daniel

On Tue, Sep 13, 2016 at 1:02 AM, Andrew Gelman <gel...@stat.columbia.edu> wrote:

Hi all.  Is there an ExcelStan?  Could be worth doing, no?

A


--
You received this message because you are subscribed to the Google Groups "stan development mailing list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Andrew Gelman]

I have no idea!  It was just a thought.  Perhaps better just to do CloudStan and then the Excel users (i.e., the people who are scared of Stata and R) could just go that route.

To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.

[Daniel Lee]

That would be a lot easier to handle.

By the way, if we have funds for computation on a cloud, I think CloudStan is doable now. Ask Eric. I think he's almost there.

Daniel

[Avraham Adler]

On Tuesday, September 13, 2016 at 1:02:41 AM UTC-4, Andrew Gelman wrote:
> Hi all. Is there an ExcelStan? Could be worth doing, no?
> A

Why in the name of all that is holy and good would you want this? We need to wean people off Excel for statistical work (for goodness sakes, even gammaln.precise really isn't, although it's loads better than ye old gammaln). We should be weaning them off Matlab too, but that's a different story.

Need a table? Use Rmarkdown or Python or go whole hog and knit a LaTeX file, but speaking as someone who spends at between 3 & 6 hours a day in Excel make this as low a priority as you can 8-).

Avi

[Bob Carpenter]

I think Andrew means something different by CloudStan---something
with a GUI he can use to demo Stan and then let users play with who
are afraid of installs and afraid or R, Stata, etc. So it has to
be a very basic Stan IDE if I have the right end of the stick here.

Andrew's given us several different back-of-the-napkin
designs, all of which seem very complicated if we want users to
save models for the future, modify data, and we still have no idea
how to pay for hosting. It's not clear what the minimally useful first
step would be that would make Andrew happy and not devolve into a full
time fundraising (for server time) and maintenance/build-out job.

There are at least three perfectly functional "cloud" Stan solutions
I've seen:

* Alan Riddell's: Jupyter in Docker for RStan, Stan.jl,
and PyStan

* Jon Zelner's: complete end-to-end analyses via GitLab(s?),
and knitr using RStan)

* Eric Novick's: more Docker containers, but I don't know what
interface he went with.

All allow people to use Stan from a browser without a local
install, but do require users be knowledgeable in one of the interface
languages.

- Bob

> On Sep 13, 2016, at 1:40 AM, Daniel Lee <bea...@alum.mit.edu> wrote:
>
> That would be a lot easier to handle.
>
> By the way, if we have funds for computation on a cloud, I think CloudStan is doable now. Ask Eric. I think he's almost there.
>
>
> Daniel
>
>
> On Tue, Sep 13, 2016 at 1:35 AM, Andrew Gelman <gel...@stat.columbia.edu> wrote:
> I have no idea! It was just a thought. Perhaps better just to do CloudStan and then the Excel users (i.e., the people who are scared of Stata and R) could just go that route.
>
>
>> On Sep 13, 2016, at 1:14 AM, Daniel Lee <bea...@alum.mit.edu> wrote:
>>
>> Given our difficulties with Windows installs and RStan, I'd say this would be much too much of a maintenance burden to even start. And I'm not sure if Stan would actually compile under MSVC now.
>>
>> If you convince me that Windows installs are not a problem, then I'd say someone should look into it.
>>
>>
>> Daniel
>>
>>
>> On Tue, Sep 13, 2016 at 1:02 AM, Andrew Gelman <gel...@stat.columbia.edu> wrote:
>> Hi all. Is there an ExcelStan? Could be worth doing, no?
>> A
>>
>>
>> --
>> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.

>> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.

>> For more options, visit https://groups.google.com/d/optout.
>>
>>
>> --
>> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.

>> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.

>> For more options, visit https://groups.google.com/d/optout.
>
>
> --
> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.

> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.

> For more options, visit https://groups.google.com/d/optout.
>
>
> --
> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.

> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.

[Bob Carpenter]

I'd say no. We can't support the users we have at
the moment---opening this floodgate would be a disaster.

- Bob

> On Sep 13, 2016, at 1:02 AM, Andrew Gelman <Gel...@stat.columbia.edu> wrote:
>
> Hi all. Is there an ExcelStan? Could be worth doing, no?
> A
>
>

[Charles Margossian]

We need to wean people off Excel for statistical work

We should be weaning them off Matlab too, but that's a different story.

@Avi: I'd actually be curious to understand why. I don't use Excel, but I have friends in finance who seem pretty content with it. The MatLab case intrigues me even more.

On Tue, Sep 13, 2016 at 10:50 AM, Bob Carpenter <ca...@alias-i.com> wrote:

I'd say no.  We can't support the users we have at
the moment---opening this floodgate would be a disaster.

- Bob

> On Sep 13, 2016, at 1:02 AM, Andrew Gelman <Gel...@stat.columbia.edu> wrote:
>
> Hi all.  Is there an ExcelStan?  Could be worth doing, no?
> A
>
>
> --
> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.

> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+unsubscribe@googlegroups.com.

> For more options, visit https://groups.google.com/d/optout.
>

--
You received this message because you are subscribed to the Google Groups "stan development mailing list" group.

To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+unsubscribe@googlegroups.com.

[Allen Riddell]

On 09/13/2016 10:49 AM, Bob Carpenter wrote:
> I think Andrew means something different by CloudStan---something
> with a GUI he can use to demo Stan and then let users play with who
> are afraid of installs and afraid or R, Stata, etc. So it has to
> be a very basic Stan IDE if I have the right end of the stick here.
>
> Andrew's given us several different back-of-the-napkin
> designs, all of which seem very complicated if we want users to
> save models for the future, modify data, and we still have no idea
> how to pay for hosting. It's not clear what the minimally useful first
> step would be that would make Andrew happy and not devolve into a full
> time fundraising (for server time) and maintenance/build-out job.
>
> There are at least three perfectly functional "cloud" Stan solutions
> I've seen:
>
> * Alan Riddell's: Jupyter in Docker for RStan, Stan.jl,
> and PyStan
>

This does work well. The only problem was that with the public setup,
all your data got deleted after 10m of inactivity. There was also the
question of limited resources. I think both could be solved with some
scheme where Andrew (or someone) could distribute invite codes and
individuals with invite codes could get a long-lasting account.

> * Jon Zelner's: complete end-to-end analyses via GitLab(s?),
> and knitr using RStan)
>
> * Eric Novick's: more Docker containers, but I don't know what
> interface he went with.
>
> All allow people to use Stan from a browser without a local
> install, but do require users be knowledgeable in one of the interface
> languages.
>

Who are these people who don't know R or Python but are very interested
in Bayesian inference?

[Bob Carpenter]

Many people only work in SAS, only work in Stata, only work
in Matlab, etc. But Excel? Not sure how many would want Bayes per se.
But there are lots and lots of packages people write with GUIs for
non-computer-savvy people to do statistics, such as MLwiN
(even I am pained by their capitalization scheme) and the one for
ecologists we were discussion a few weeks ago.

- Bob

[Seth Flaxman]

While we're listing them, my colleague Louis Aslett has another
"cloud" Stan solution, an Amazon Machine Image with a bunch of useful
things for doing statistics with python and R pre-installed, including
RStan:

http://www.louisaslett.com/RStudio_AMI/

[Avraham Adler]

On Tuesday, September 13, 2016 at 11:02:40 AM UTC-4, Charles Margossian wrote:
> We need to wean people off Excel for statistical work
> We should be weaning them off Matlab too, but that's a different story.
>
>
> @Avi: I'd actually be curious to understand why. I don't use Excel, but I have friends in finance who seem pretty content with it. The MatLab case intrigues me even more.
>

Excel's statistical formulæ are sketchy; they are not precise and often not accurate to machine precision. Yes, this was worse pre 2007, but it is not ironclad now. Same with their random number generator. If I HAVE to use RNGs in excel, I use a VBA implementation of the Mersenne-Twister. Similarly with gammaln, NormalCDF, inversenormalCDF, T, and inverseT to name a few. We often deal with numbers very close to 0 in left and right tails, so precision is important. Let's not get started on how weak Solver is once you leave two variables. Yes, you can do stuff in Excel (I once coded the Marsaglia implementation of Anderson-Darling using only offsets and array functions because I was told I couldn't use VBA. Probably couldn't do that anymore :( ). but it is more meant as a large calculator. Algebra, no problem. Calculus, maybe. Probability, weak. If we can train people to use proper tools, we are better off.

<rant>
As for Matlab; perhaps this is my bias as a non-academic but someone who works in industry. COnsider that over the past decade I have saved my company probably near $75K by dropping my MatLab license after 2 years (>6K per annum) and never getting a SAS license (>2K per year, IIRC) by using R/Python instead. Perhaps the sweet deals they give students continues in academia, but for those who will need to wrestle with statistics outside of academia, being able to perform the same function in the same timeframe for nearly five figures cheaper per annum is a GOOD thing, IMO. Even for students, free >> cheap, especially as students' budgets tend to be severely limited.
</rant>

Avi

[Ben Goodrich]

On Tuesday, September 13, 2016 at 10:50:30 AM UTC-4, Bob Carpenter wrote:

There are at least three perfectly functional "cloud" Stan solutions
I've seen:

Four including the one I wrote on the wiki

https://github.com/stan-dev/stan/wiki/CloudStan-functional-specification#mostly-complete-free-implementation

[Bob Carpenter]

Thanks for the reminder. We should get feedback from Andrew
on what he wants, too.

Could everyone list ones they've created on the Wiki if they're
available either as a service or as code.

We may want to break the wiki apart into GUI/IDE CloudStan vs.
something more direct like Jupyter.

Thanks.

- Bob

[Michael Betancourt]

+1 for Louis’s AMI. Definitely the solution if you’re using Amazon’s resources.