orthogonal_polynomial: Orthogonal-Polynomial (Trend) Contrast Coefficients

Description

Builds the set of orthogonal-polynomial contrasts (linear, quadratic, cubic, ...) for a factor whose \(a\) levels are quantitative, so that between-group variation can be decomposed into independent trend components. The columns are mutually orthogonal and each sums to zero (so each is a contrast on the group means). By default the coefficients are returned in orthonormal form, meaning each column also has unit length, \(\sum_i c_i^2 = 1\); the alternative type = "integer" rescales every column to the small whole numbers used in the published orthogonal-polynomial tables, which are easier to read by eye and match hand computation. The per-trend sum of squared coefficients \(\sum_i c_i^2\) is carried on the returned object and shown when it is printed.

Usage

orthogonal_polynomial(
  levels,
  scores = NULL,
  type = c("orthonormal", "integer"),
  degree = NULL
)

Value

A numeric \(a \times \text{degree}\) matrix with row names = the level labels and column names = the trend names ("linear", "quadratic", "cubic",

"quartic", ...). The matrix can be assigned directly to

contrasts(factor) or passed to

contrast_test, is_orthogonal_set, or

ci_c. The per-column sum of squared coefficients is stored as attr(*, "sum_sq") (a named numeric vector), and the level spacing and type are stored as

attr(*, "scores") and attr(*, "type"). The object carries class "orthogonal_polynomial" so that printing shows the coefficients alongside \(\sum_i c_i^2\). When printed, the object is shown in the textbook Table A.10 orientation (trends in rows, levels in columns) with the \(\sum_i c_i^2\) values as a final column; the stored matrix is the transpose of that display (levels in rows) so it can be assigned directly to contrasts().

Arguments

levels: One of: a single integer giving the number of levels \(a\) (labels default to "L1", "L2", ..., and the levels are treated as equally spaced); a character or factor vector of level labels (equally spaced unless scores is given); or a numeric vector of the quantitative level values themselves, which are used both as the labels and as the spacing.
scores: Optional numeric vector of length \(a\) giving the quantitative position of each level. Supply this when the labels are non-numeric but the spacing is unequal (e.g., doses 0, 1, 2, and 4). Defaults to equally spaced positions 1:a.
type: Either "orthonormal" (default) for unit-length columns (\(\sum_i c_i^2 = 1\), identical to contr.poly) or "integer" for the minimal whole-number coefficients of the classic trend-coefficient table. "integer" requires equally spaced scores.
degree: Highest-order trend to return, an integer between 1 and \(a - 1\). Defaults to \(a - 1\) (the full set the design can support).

Author

Ken Kelley [email protected]

Details

What a trend contrast is. When the levels of a factor are quantitative and ordered (minutes of study, dose, day), the omnibus between-group variation can be partitioned into a linear trend (does the mean rise or fall steadily?), a quadratic trend (is there curvature?), a cubic trend (an S-shape?), and so on, up to order \(a - 1\). Each trend is a single 1-df contrast on the group means, \(\hat\psi = \sum_i c_i \bar Y_i\), and because the contrasts are mutually orthogonal their sums of squares add up to the omnibus between-group sum of squares exactly.

Orthonormal versus integer scaling. The two type values return the same trends (same directions, same hypotheses, identical F, t, and p for a given trend); they differ only in how each column is scaled.

"orthonormal" normalizes each column to unit length, \(\sum_i c_i^2 = 1\). The trend portion of the design is then an orthonormal basis, the sum of squares for a trend is simply \(\hat\psi^2\), and in a fitted lm the .linear / .quadratic coefficients are the trend estimates on a common scale. This is the form lm and aov use internally and is defined for any spacing, equal or unequal.
"integer" rescales each column to the smallest whole numbers with the same ratios, reproducing the published orthogonal-polynomial coefficient table (e.g., for \(a = 4\) the linear contrast is \(-3, -1, 1, 3\)). These are easy to read and to compute with by hand, but the columns no longer share a common length: \(\sum_i c_i^2\) varies by trend (for \(a = 4\): 20, 4, and 20), so the sum of squares for a trend is \(n\, \hat\psi^2 / \sum_i c_i^2\). Integer coefficients exist only for equally spaced levels.

The reported \(\sum_i c_i^2\) (shown on printing and stored in attr(*, "sum_sq")) is exactly the divisor in that sum-of-squares formula; for the orthonormal form it is 1 for every trend.

Unequal spacing. With unequally spaced scores the orthonormal polynomials are still uniquely defined and are returned by the "orthonormal" type. Whole-number coefficients generally do not exist in that case, so type = "integer" is an error.

Relation to base R. For equally spaced levels the orthonormal output is identical to contr.poly(a); orthogonal_polynomial adds the integer table form, an explicit scores argument for unequal spacing, meaningful trend names, and the \(\sum_i c_i^2\) report.

References

Fisher, R. A., & Yates, F. (1953). Statistical tables for biological, agricultural and medical research (4th ed.). Oliver and Boyd. (Origin of the tabulated integer coefficients.)

Maxwell, S. E., Delaney, H. D., & Kelley, K. (2027). Designing experiments and analyzing data: A model comparison perspective (4th ed.). Routledge. (See Chapter 6 on trend analysis; Appendix Table A.10 reports these coefficients.)

Examples

Run this code

# Trend (orthogonal-polynomial) analysis decomposes the omnibus effect of a
# quantitative factor (dose, time, trial block, stimulus intensity) into
# independent linear, quadratic, cubic, ... components, as in the trend
# analysis of Maxwell, Delaney, and Kelley (2027, Chapter 6).

# 1. The a = 4 coefficient table in the integer form of the textbook
#    appendix (Table A.10). Printing puts the trends in rows and appends a
#    final sum-of-squared-coefficients (sum c^2) column: linear -3 -1 1 3,
#    and per-trend sum c^2 of 20, 4, 20.
orthogonal_polynomial(4, type = "integer")

# 2. The default orthonormal form (identical to stats::contr.poly(4)); every
#    trend has sum c^2 = 1, the scale lm() and aov() use internally.
orthogonal_polynomial(4)

# 3. A worked trend analysis in the style of MDK Chapter 6. An outcome is
#    measured at a = 5 equally spaced levels of a quantitative factor (say
#    stimulus intensity 1..5), n = 8 per level, from a population with a
#    strong linear and a mild quadratic trend. Assigning the trend contrasts
#    to the factor makes lm() report each trend test directly.
set.seed(113)
intensity <- factor(rep(1:5, each = 8))
contrasts(intensity) <- orthogonal_polynomial(levels(intensity))
y <- c(10, 9, 7, 6, 6)[as.integer(intensity)] +
       rnorm(length(intensity), sd = 1.2)
round(coef(summary(lm(y ~ intensity))), 3)

# 4. The same trends through DMAR's contrast_test(), which reports each
#    trend's estimate, standard error, t, p, and confidence interval from a
#    fitted one-way model. The integer columns are the contrast weights.
op <- orthogonal_polynomial(5, type = "integer")
contrast_test(aov(y ~ intensity),
              contrasts = list(linear    = op[, "linear"],
                               quadratic = op[, "quadratic"],
                               cubic     = op[, "cubic"]))

# 5. Unequally spaced doses (0, 1, 2, 4 mg): integer coefficients no longer
#    exist, but the orthonormal trends remain uniquely defined.
orthogonal_polynomial(c("0 mg", "1 mg", "2 mg", "4 mg"),
                      scores = c(0, 1, 2, 4))

# 6. Confirm a returned set is mutually orthogonal.
is_orthogonal_set(orthogonal_polynomial(5, type = "integer"))

Run the code above in your browser using DataLab