error for anova with too few samples for experiment

Changes

@@ -6,6 +6,7 @@
 - fix anova_design_matrix for between-subject >2 values (rest of RT#97925)
 - add demo
 - remove filt_exp and filt_ma, deprecated in 2011
+- now an exception to give anova n-observations <= product of categories in IVs
 
 0.853 2025-01-03
 - uses PDL 2.096+ lib/*.pd format for quicker builds

lib/PDL/Stats/GLM.pd

@@ -1056,6 +1056,7 @@ sub PDL::anova_rptd {
           # highest ord inter for purely within design, (p-1)*(q-1)*(n-1)
         my $factor = (ref $full[-1] ? $full[-1] : $err_ref->[$full[-1]])->dim(1);
         my $df = $ret{ "$i_pref$e df" } = $factor * $within_df;
+        die "${i_pref}residual df = 0" if $df <= 0;
         $ret{ "$i_pref$e ms" } = $ret{ "$i_pref$e ss" } / $df;
       } elsif (ref $full[$k]) {       # unique error term
         next EFFECT
@@ -1066,13 +1067,15 @@ sub PDL::anova_rptd {
           = $y->sse(sumover($b_G * $G->transpose)) - $ret{ss_residual};
         if ($k != $#full) {
           my $df = $ret{"$i_pref$e df"} = $full[$k]->dim(1);
+          die "residual df = 0" if $df <= 0;
           $ret{"$i_pref$e ms"} = $ss / $df;
         }
       } else {                        # repeating error term
         my $ss = $ret{$k == $#full ? 'ss_subject' : "$i_pref$e ss"}
           = $ret{"| $idv->[$full[$k]] |$e ss"};
         if ($k != $#full) {
           my $df = $ret{"$i_pref$e df"} = $ret{"| $idv->[$full[$k]] |$e df"};
+          die "residual df = 0" if $df <= 0;
           $ret{"$i_pref$e ms"} = $ss / $df;
         }
       }
@@ -1095,6 +1098,7 @@ sub PDL::anova_rptd {
         = $y->sse( sumover($b_G * $G->transpose) ) - $ret{ss_residual};
       my $df0 = $ret{ "| $idv->[$k] | df" } = $ivs_ref->[$k]->dim(1);
       $ret{ "| $idv->[$k] || err df" } = $df1;
+      die "residual df = 0" if $df1 <= 0;
       $ret{ "| $idv->[$k] | ms" } = $ret{ "| $idv->[$k] | ss" } / $df0;
     }
   }

t/glm.t

@@ -394,6 +394,15 @@ is_pdl $b_bad->dvrs(ones(6) * .5), pdl( 'BAD -1.17741002251547 -1.17741002251547
 ';
 }
 
+{ # anova with too few samples for experiment (3*2*2 categories, 12 samples)
+my $y = pdl '[1 1 2 2 3 3 3 3 4 5 5 5]'; # ratings for 12 apples
+my $a = sequence(12) % 3 + 1; # IV for types of apple
+my @b = qw( y y y y y y n n n n n n ); # IV for whether we baked the apple
+my @c = qw( r g r g r g r g r g r g ); # IV for apple colour (red/green)
+eval {$y->anova( $a, \@b, \@c, { IVNM=>[qw(apple bake colour)], PLOT=>0 } )};
+like $@, qr/residual df = 0/, 'error when too few sample';
+}
+
 { # anova 1 way
   my $d = pdl qw( 3 2 1 5 2 1 5 3 1 4 1 2 3 5 5 );
   my $a = qsort sequence(15) % 3;