Add `categorical`, `string` and `table` support to `crosstab`

[SergeantQuickscoper]

Currently, crosstab.m requires categorical and string vector support for input arguments and table support as an output argument to be MATLAB compatible, as observed in the docs. This issue proposes adding support and necessary tests to crosstab.m as part of the ongoing datatypes integration effort.

[pr0m1th3as]

table support will be added through an overloaded method in table class. string and categorical support has been added, but we need a better testing suit with respect to returning values p, chisq, and labels.

[pr0m1th3as]

@Sonu0305 can you handle adding some tests and cross-validate them with MATLAB?

[Sonu0305]

@Sonu0305 can you handle adding some tests and cross-validate them with MATLAB?

Hi, yes @pr0m1th3as, I will do that.

[Sonu0305]

Hi @pr0m1th3as,

While testing the very first case, I came across this.
Octave produces a warning from chi2test, and also labels dont match.
Should I move ahead only after you push a fix for this?
Thanks.

MATLAB:

>> smoker = [1 1 0 0 1 0 1 1 0 0 1 0]';
gender = [1 0 1 0 1 1 0 0 1 0 0 1]';
[t, chisq, p, labels] = crosstab (smoker, gender)

t =

     2     4
     4     2


chisq =

    1.3333


p =

    0.2482


labels =

  2×2 cell array

    {'0'}    {'0'}
    {'1'}    {'1'}

Octave:

octave:46> smoker = [1 1 0 0 1 0 1 1 0 0 1 0]';
gender = [1 0 1 0 1 1 0 0 1 0 0 1]';
[t, chisq, p, labels] = crosstab (smoker, gender)
warning: chi2test: Expected values less than 5.
warning: called from
    chi2test at line 339 column 5
    crosstab at line 104 column 8

t =

   2   4
   4   2

chisq = 1.3333
p = 0.2482
labels =
  12x2 cell array

    {'1'}    {'1'}
    {'1'}    {'0'}
    {'0'}    {'1'}
    {'0'}    {'0'}
    {'1'}    {'1'}
    {'0'}    {'1'}
    {'1'}    {'0'}
    {'1'}    {'0'}
    {'0'}    {'1'}
    {'0'}    {'0'}
    {'1'}    {'0'}
    {'0'}    {'1'}

[pr0m1th3as]

Octave produces a warning from chi2test

The warning should be there. If group samples are 5 or less, then Fisher's exact test should be used instead.

also labels dont match.

Can you work on this?

[Sonu0305]

The warning should be there. If group samples are 5 or less, then Fisher's exact test should be used instead.

Okay, noted.

Can you work on this?

Okay, I'll open a pr for this fix, and then I'll move on with further testing.

[Sonu0305]

Instead, I decide to open a pr altogether itself.

[Sonu0305]

Hi @pr0m1th3as,

Another doubt that I came across is this, that's a bug to be fixed in chi2test right?

MATLAB:

>> a = ones (15,1);
b = ones (15,1);
[t, chisq, p] = crosstab (a, b)

t =

    15


chisq =

   NaN


p =

   NaN

Octave:

octave:88> a = ones (15,1);
b = ones (15,1);
[t, chisq, p] = crosstab (a, b)
error: chi2test: X must be a matrix.
error: called from
    chi2test at line 100 column 5
    crosstab at line 117 column 8

[pr0m1th3as]

I am not sure whether this is a chi2test bug or just crosstab returning NaN values for ill-formed contingency tables.
I think it's the latter. Leave chi2test out of this, and handle edge cases like in your example inside crosstab. i.e. if you have a scalar contingency table, and p + chisq values are requested, just return NaN to them.

[Sonu0305]

I think it's the latter.

Yes, it's the latter.

handle edge cases like in your example inside crosstab

Okay, noted.

[Sonu0305]

Another bug in crosstab, when all is NaN, chi2sq should be returned as NaN. Fixing this.

matlab:

>> na = NaN (6,1);
nb = (1:6)';
[t, chisq, p] = crosstab (na, nb)

t =

  0×6 empty double matrix


chisq =

   NaN


p =

   NaN

octave:

octave:98> na = NaN (6,1);
nb = (1:6)';
[t, chisq, p] = crosstab (na, nb)
t = [](0x6)
chisq = 0
p = NaN