support pander

[jankatins]

pander returns markdown (and is used extensively with knitr to display rich tables), this should be displayed as markdown to be rendered in the notebook:

df <- data.frame(a=1:8, b=factor(c(0,1,0,1,0,1,0,1), levels=c(0,1), labels=c("a", "b")) )
pander(df %>% t.test(a~b, data = .) %>% tidy())

[jankatins]

[From the problems I run into, I get the feeling that I'm the first to actually use irkernel to do data analysis with the current hadleyverse stack :-/ ]

[jankatins]

Ok, this looks like something which is not so nice:

df <- data.frame(a=1:8, b=factor(c(0,1,0,1,0,1,0,1), levels=c(0,1), labels=c("a", "b")) )
t <- pander(df %>% t.test(a~b, data = .) %>% tidy()) # prints the markdown
str(t) # NULL

shows that t does not get anything assigned :-( Looks like this isn't easily changed without a change in pander itself :-( E.g. it would be easier if pander would return a object and have a default print method for that object and we could then add a repr_markdown.pander_object method.

CC: @Rapporter

function (x = NULL, ...) 
{
    kaao <- panderOptions("knitr.auto.asis")
    if (isTRUE(panderOptions("knitr.auto.asis")) && isTRUE(getOption("knitr.in.progress")) && 
        requireNamespace("knitr", quietly = TRUE)) {
        panderOptions("knitr.auto.asis", FALSE)
        stdout <- vector("character")
        con <- textConnection("stdout", "wr", local = TRUE)
        sink(con)
        sink(con, type = "message")
        on.exit({
            if (panderOptions("knitr.auto.asis") != kaao) {
                panderOptions("knitr.auto.asis", kaao)
            }
            sink()
            sink(type = "message")
            close(con)
            if (tail(stdout, 1) == "") {
                stdout <- c(stdout, "")
            }
            return(knitr::asis_output(paste(stdout, collapse = "\n")))
        })
    }
    UseMethod("pander", x)
}

[jankatins]

[Edit: stupidity on my side with cat removed and added the correct workaround...]

Workaround:

t <- pander_return(df %>% t.test(a~b, data = .) %>% tidy())
x <- paste(t, collapse = '\n')
display_markdown(x)

Unfortunately, this concrete exemple still renders badly:

Looks like an incompatibility of pandoc and the lib which is used in the jupyter notebook :-(

----
text
----

is interpreted as a headline, not a single cell table...

[flying-sheep]

i don’t think there’s a possibility to do single table cells in markdown.

[jankatins]

I think the problem already starts above:

-------------------------------------------------------------------------------
estimate   estimate1   estimate2   statistic   p.value   parameter   conf.low 
---------- ----------- ----------- ----------- --------- ----------- ----------
-1          4           5        -0.5477    0.6036        6        -5.467  
-------------------------------------------------------------------------------

Table: Table continues below

-----------
conf.high 
-----------
3.467   
-----------

In this case, even the first header is not displayed as header, but works in pandoc (multiple rows in that format also do not work in the notebook) :-(

[jankatins]

So, to summarize, supporting pander out of the box doesn't work because:

• pander prints markdown to stdout (=displayed in the notebook as "text/plain") so we would have to replace the pander function with our own function to use the display system instead or upstream a patch which would reverse the pander logic (pander = cat the MD to stdout-> sink stdout to a string -> return string vs pander_new = return object -> print object) or at least do something similar to the knitr workaround (which returns a asis object when an option is set and it detects a knitr run) to hook in our own repr_markdown.pander_object.
• the resulting markdown is incompatible with the markdown display library in the jupyter notebook because it doesn't understand the table format

Bummer :-(

[jankatins]

Damn:

stargazer uses cat() to output LaTeX/HTML code or ASCII text for the table.

[flying-sheep]

why? capture.output is our friend!

[jankatins]

... but this needs special treatment in the form of a wrapper function and therefore the nice examples (e.g. pander(models...) do not work :-( I think it would be really nice if we could go and advertise a better patter for this (e.g. return an object and include a print function for that object; maybe talk to the knitr people to unify the repr/knitr asis functionality).

[I also find using capture.output slightly revolting: :-(]

Anyway: the more depressing aspect here is that the notebook doesn't understand the markdown table syntax and therefore it seems it can't produce nicely formatted tables :-( I haven't yet found a way to convert the markdown to html via pandoc (e.g. knitr includes a pandocfunction, but it requires tempfiles :-()

@takluyver Do you think there is a chance to get a better markdown library in the notebook?

[flying-sheep]

I also find using capture.output slightly revolting: :-(

sure, but we gotta use what we have.

[takluyver]

The current CommonMark spec does not describe any Markdown table syntax, and we'd broadly like to stick to that standard.

[jankatins]

I can understand the "sticking to standards" argument, but I think this will become a problem for the adoption of the R kernel/notebook in the R world: pandoc is the defacto standard of markdown in knitr and IMO the notebook is competing with knitr. Not having nice summary tables (in both html and the PDF generated by nbconvert) is IMO a bad signal if one want's to use the notebook together with statistical research. :-( Lets hope that this is fixed in the commonmark spec in the near future...

On a sidenote: having nicer tables in general would be nice... Bootstrap has nice table styles, maybe adding that to the notebook would already by a nice thing...

[jankatins]

Here is my current version of a pander based table generation:

df <- data.frame(a=1:8, b=factor(c(0,1,0,1,0,1,0,1), levels=c(0,1), labels=c("a", "b")) )
nice_pander <- function(...){
    t = capture.output(pander(...))
    mdFile <- tempfile(fileext=".md")
    cat(paste(t, collapse="\n"), file=mdFile)
    htmlFile <- knitr::pandoc(input = mdFile, format = "html")
    t = readr::read_file(htmlFile)
    #print(t)
    display_html(paste(t, collapse="\n"))
    display_html("<link rel='stylesheet' href='http://maxcdn.bootstrapcdn.com/bootstrap/3.3.6/css/bootstrap.min.css'>
<script src='http://maxcdn.bootstrapcdn.com/bootstrap/3.3.6/js/bootstrap.min.js'></script>
<script>$('tr.header').parent('thead').parent('table').addClass('table table-condensed');</script>
")
}
nice_pander(df %>% t.test(a~b, data = .) %>% tidy())

and one for stargazer:

display_html(paste(capture.output(stargazer(df %>% t.test(a~b, data = .) %>% tidy(), type="html")), collapse="\n"))

Both currently don't work:

• In the pander case, it seems the classes get assigned to the new appended table, but there is no styling. Not sure why that is, could be a different .table css rule in the default notebook css...
• the stargazer case produces html which has inline css but it seems that this is overwritten somewhere, resulting in a horrible styled table :-(

[jankatins]

The last is now jupyter/notebook#1182

[daroczig]

I've just found this thread thanks to Google Alerts :) I am the maintainer of the pander package, and happy to help you fine-tine the output for your needs -- just please let me know what's the optimal output.

Eg if you need another markdown table format instead of the default multiline, you could set a global option as panderOptions('table.style', 'grid') for the Emacs-like grid format, which will be used in all future pander calls, or eg rmarkdown for the PHP Extra markdown version. To disable splitting the tables, you could do panderOptions('table.split.table', Inf)

Please let me know if I can help/contribute with something here.

Best,
Gergely

PS: thanks for the @Rapporter above, but this did not trigger any mail to me or the other collaborators :( Please feel free to ping my handle directly.