document linearRegression

src/marks/linearRegression.d.ts

@@ -2,23 +2,110 @@ import type {ChannelValue, ChannelValueDenseBinSpec, ChannelValueSpec} from "../
 import type {Data, MarkOptions, RenderableMark} from "../mark.js";
 import type {BinOptions, BinReducer} from "../transforms/bin.js";
 
+/** Options for the linearRegressionX and linearRegressionY marks. */
 interface LinearRegressionOptions extends MarkOptions, BinOptions {
-  x?: ChannelValueSpec;
-  y?: ChannelValueSpec;
-  z?: ChannelValue;
+  /** The confidence interval in (0, 1), or 0 to hide bands; defaults to 0.95. */
   ci?: number;
+  /** The distance in pixels between samples of the confidence band; defaults to 4. */
   precision?: number;
+
+  /**
+   * An optional ordinal channel for grouping data into (possibly stacked)
+   * series, producing an independent regression for each group. If not
+   * specified, it defaults to **fill** if a channel, or **stroke** if a
+   * channel.
+   */
+  z?: ChannelValue;
+
+  /**
+   * How to reduce the dependent variable (**y** for linearRegressionY, or **x**
+   * for linearRegressionX) when the independent variable (**x** for
+   * linearRegressionY, or **y** for linearRegressionX) is binned with an
+   * **interval**; defaults to *first*. For example, for a trend line on the
+   * maximum value of the Apple’s stack price by month:
+   *
+   * ```js
+   * Plot.linearRegressionY(aapl, {x: "Date", y: "Close", interval: "month", reduce: "max"})
+   * ```
+   *
+   * To default to zero instead of showing gaps in data, as when the observed
+   * value represents a quantity, use the *sum* reducer.
+   */
   reduce?: BinReducer;
 }
 
+/** Options for the linearRegressionX mark. */
 export interface LinearRegressionXOptions extends LinearRegressionOptions {
+  /**
+   * The independent variable vertical position channel, typically bound to the
+   * *y* scale; defaults to the zero-based index of the data [0, 1, 2, …].
+   *
+   * If an **interval** is specified, **y** values are binned accordingly,
+   * allowing zeroes for empty bins instead of interpolating across gaps. This
+   * is recommended to “regularize” sampled data; for example, if your data
+   * represents timestamped observations and you expect one observation per day,
+   * use *day* as the **interval**.
+   */
   y?: ChannelValueDenseBinSpec;
+
+  /**
+   * The dependent variable horizontal position channel, typically bound to the
+   * *x* scale; defaults to identity, assuming that *data* = [*x₀*, *x₁*, *x₂*,
+   * …].
+   */
+  x?: ChannelValueSpec;
 }
 
+/** Options for the linearRegressionY mark. */
 export interface LinearRegressionYOptions extends LinearRegressionOptions {
+  /**
+   * The independent variable horizontal position channel, typically bound to
+   * the *x* scale; defaults to the zero-based index of the data [0, 1, 2, …].
+   *
+   * If an **interval** is specified, **x** values are binned accordingly,
+   * allowing zeroes for empty bins instead of interpolating across gaps. This
+   * is recommended to “regularize” sampled data; for example, if your data
+   * represents timestamped observations and you expect one observation per day,
+   * use *day* as the **interval**.
+   */
   x?: ChannelValueDenseBinSpec;
+
+  /**
+   * The dependent variable vertical position channel, typically bound to the
+   * *y* scale; defaults to identity, assuming that *data* = [*y₀*, *y₁*, *y₂*,
+   * …].
+   */
+  y?: ChannelValueSpec;
 }
 
+/**
+ * Like linearRegressionY, but where *x* is the dependent variable and *y* is
+ * the independent variable. This orientation is infrequently used, but suitable
+ * for example when visualizing a time-series where time goes up↑; use
+ * linearRegressionY instead if time goes right→.
+ */
 export function linearRegressionX(data?: Data, options?: LinearRegressionXOptions): RenderableMark;
 
+/**
+ * Returns a mark that draws [linear regression][1] lines with confidence bands,
+ * representing the estimated relation of a dependent variable (*y*) on an
+ * independent variable (*x*). For example to estimate the linear dependence of
+ * horsepower (*hp*) on weight (*wt*):
+ *
+ * ```js
+ * Plot.linearRegressionY(mtcars, {x: "wt", y: "hp"})
+ * ```
+ *
+ * The linear regression line is fit using the [least squares][2] approach. See
+ * Torben Jansen’s [“Linear regression with confidence bands”][3] and [this
+ * StatExchange question][4] for details on the confidence interval calculation.
+ *
+ * Multiple regressions can be produced by specifying a **z**, **fill**, or
+ * **stroke** channel.
+ *
+ * [1]: https://en.wikipedia.org/wiki/Linear_regression
+ * [2]: https://en.wikipedia.org/wiki/Least_squares
+ * [3]: https://observablehq.com/@toja/linear-regression-with-confidence-bands
+ * [4]: https://stats.stackexchange.com/questions/101318/understanding-shape-and-calculation-of-confidence-bands-in-linear-regression
+ */
 export function linearRegressionY(data?: Data, options?: LinearRegressionYOptions): RenderableMark;