I think we’re still not really interested in promoting the use of NaN as a
surrogate for NULL, especially given that Nullable is going to be added to Base
in 0.4.
Your functions would perform substantially better if you iterated over the
values of A. For example,
function nanmean(A::Array)
s, n = 0.0, 0
for val in A
if !isnan(val)
s += val
n += 1
end
end
return s / n
end
— John
On Sep 7, 2014, at 10:36 AM, Alex <hollina...@gmail.com> wrote:
> I know it's a little late, but I was looking for the same thing and couldn't
> find it. I've made some slight adjustments to some code I found on github and
> made functions for nanmean and nanstd. I did not optimize for performance and
> wanted them to be able to handle arrays of various sizes.
>
> NANMEAN:
>
> function nanmean(x::Array)
>
> z=similar(x)
> fill!(z,1)
> z[isnan(x)]=0
> numb_not_NaN_in_x=sum(z)
>
> nansum_x=sum(x) do x isnan(x) ? 0 : x end #from
> https://gist.github.com/milktrader/5213361
> nansum_x/numb_not_NaN_in_x
>
> end
>
>
> NANSTD
> function nanstd(x::Array)
>
> z=similar(x)
> fill!(z,1)
> z[isnan(x)]=0
> numb_not_NaN_in_x=sum(z)
>
> nansum_x=sum(x) do x isnan(x) ? 0 : x end #from
> https://gist.github.com/milktrader/5213361
> nanmean_x=nansum_x/numb_not_NaN_in_x
>
> y=(x-nanmean_x).*(x-nanmean_x)
>
> ## NanMean for Sample
> function nanmean_sample(y::Array)
> w=similar(y)
> fill!(w,1)
> w[isnan(y)]=0
> numb_not_NaN_in_y=sum(w)
>
> nansum_y=sum(y) do y isnan(y) ? 0 : y end #from
> https://gist.github.com/milktrader/5213361
> nansum_y/(numb_not_NaN_in_y-1)
>
> end
>
> nanstd_x=sqrt(nanmean_sample(y))
>
> end
>
>
>
>
> On Wednesday, February 5, 2014 9:11:45 PM UTC-7, Roger Herikstad wrote:
> Hi,
> Are there equivalent functions to Matlab's nanmean and nanstd, i.e.
> functions for computing mean and standard deviation while ignoring NaN's?
> It's simple to put something together, of course, e.g.
>
> function nanmean(x)
> mean(~isnan(x))
> end
>
> but it would nice to have as part of Base, or perhaps StatsBase?
>
> ~ Roger
