Ranking vectors of data

From Free Pascal wiki
Jump to navigationJump to search

Ranking is an important method in statistics and data science. It is a preparative step that replaces each value by a rank in the order of magnitude of the original data. This operation is necessary for many non-parametric methods, e.g. the U test by Wilcoxon-Mann-Whitney or Spearman's rank correlation.

Simple algorithm not caring for ties

In the simplest case, where every value is distinct without identical data, the following frugal algorithm may be sufficient:

function rank(x: array of longint): array of real;
var
  i, j, n: integer;
  lessCount: integer; { number of observations less than a given observation }
begin
  n := length(x);
  SetLength(result, n);
  for i := 0 to n - 1 do
  begin
    lessCount := 0;
    for j := 0 to n - 1 do
      if x[j] < x[i] then
        inc(lessCount);
      result[i] := lessCount + 1;
  end;
end;

This function delivers a list of ranks that replace the original data. It doesn't deliver satisfactory results if the data contain identical values, e.g. in the vector 3, 4, 4, 9, 0, 1. In statistical terminology, repeated observation-values are referred to as ties. The potential presence of ties requires a more complex algorithm:

Advanced algorithm addressing potential ties

type
  TTiesMethod = (ascend, descend, average, min, max);
  
function rank(x: array of longint; 
  TiesMethod: TTiesMethod = average): array of real;
var
  i, j, n: integer;
  equalCount, lessCount: array of integer;
begin
  n := length(x);
  SetLength(lessCount, n);
  SetLength(equalCount, n);
  SetLength(Result, n);
  for i := 0 to n - 1 do
  begin
    equalCount[i] := 0;
    lessCount[i] := 0;
    for j := 0 to n - 1 do
      case TiesMethod of
        average, min, max:
          if x[j] = x[i] then
            Inc(equalCount[i])
          else if x[j] < x[i] then
            Inc(lessCount[i]);
        ascend:
          if (j <= i) and (x[j] = x[i]) then
            Inc(equalCount[i])
          else if x[j] < x[i] then
            Inc(lessCount[i]);
        descend:
          if (j >= i) and (x[j] = x[i]) then
            Inc(equalCount[i])
          else if x[j] < x[i] then
            Inc(lessCount[i]);
      end;
  end;
  for i := 0 to n - 1 do
    case TiesMethod of
      average:
        Result[i] := lessCount[i] + (equalCount[i] + 1) / 2;
      min:
        Result[i] := lessCount[i] + 1;
      max, ascend, descend:
        Result[i] := lessCount[i] + equalCount[i];
    end;
end;

This function delivers a vector of ranks that is corrected for ties. Similar to the function rank of the statistical programming language S (e.g. in the environment R) it supports different methods to address potential ties. They include average (default method, the rank for each occurrence of a repeated number is given as rank[i] = lessCount + (equalCount + 1) / 2), ascend (permutation with increasing values at each index set of ties), descend (decreasing values), min (replaces values with ties with their minimum) and max (replacement by maximum).

See also

References

  1. Cooke D, Craven AH, Clarke GM. Statistical Computing in Pascal. Edward Arnold (Publishers), London 1985
  2. Becker, R. A., Chambers, J. M. and Wilks, A. R. (1988) The New S Language. Wadsworth & Brooks/Cole.
  3. R Documentation: Sample Ranks.