# Group transform ​

TIP

The group transform is for aggregating ordinal or nominal data. For quantitative or temporal data, use the bin transform.

The group transform groups ordinal or nominal data — discrete values such as name, type, or category. You can then compute summary statistics for each group, such as a count, sum, or proportion. The group transform is most often used to make bar charts with the bar mark.

For example, the bar chart below shows a distribution of Olympic athletes by sport.

Fork
js
``````Plot.plot({
marginBottom: 100,
x: {label: null, tickRotate: 90},
y: {grid: true},
marks: [
Plot.barY(olympians, Plot.groupX({y: "count"}, {x: "sport"})),
Plot.ruleY([0])
]
})``````

TIP

Ordinal domains are sorted naturally (alphabetically) by default. Either set the scale domain explicitly to change the order, or use the mark sort option to derive the scale domain from a channel.

The groupX transform groups on x. The outputs argument (here `{y: "count"}`) declares desired output channels (y) and the associated reducer (count). Hence the height of each bar above represents the number of Olympic athletes by sport.

While the groupX transform is often used to generate y, it can output to any channel. For example, by declaring r in outputs, we can generate dots of size proportional to the number of athletes in each sport.

Fork
js
``````Plot.plot({
marginBottom: 100,
x: {label: null, tickRotate: 90},
r: {range: [0, 14]},
marks: [
Plot.dot(olympians, Plot.groupX({r: "count"}, {x: "sport"}))
]
})``````

The fill channel meanwhile will produce a one-dimensional heatmap. Since there is no y channel below, we use a cell instead of a bar.

Fork
js
``````Plot.plot({
marginBottom: 80,
x: {tickRotate: 90},
color: {scheme: "YlGnBu"},
marks: [
Plot.cell(olympians, Plot.groupX({fill: "count"}, {x: "sport"}))
]
})``````

We aren’t limited to the count reducer. We can use the mode reducer, for example, to show which sex is more prevalent in each sport: men are represented more often than women in every sport except gymnastics and fencing.

Fork
js
``````Plot.plot({
marginBottom: 80,
x: {tickRotate: 90},
marks: [
Plot.cell(olympians, Plot.groupX({fill: "mode"}, {fill: "sex", x: "sport"}))
]
})``````

You can partition groups using z. If z is undefined, it defaults to fill or stroke, if any. In conjunction with the barY mark’s implicit stackY transform, this will produce stacked bars.

Fork
js
``````Plot.plot({
marginBottom: 100,
x: {label: null, tickRotate: 90},
y: {grid: true},
color: {legend: true},
marks: [
Plot.barY(olympians, Plot.groupX({y: "count"}, {x: "sport", fill: "sex"})),
Plot.ruleY([0])
]
})``````

TIP

You can invoke the stack transform explicitly as `Plot.stackY(Plot.groupX({y: "count"}, {x: "sport", fill: "sex"}))`, producing an identical chart.

You can opt-out of the implicit stackY transform by having groupX generate y1 or y2 instead of y (and similarly x1 or x2 for stackX and groupY). When overlapping marks, use either opacity or blending to make the overlap visible.

Fork
js
``````Plot.plot({
marginBottom: 100,
x: {label: null, tickRotate: 90},
y: {grid: true},
color: {legend: true},
marks: [
Plot.barY(olympians, Plot.groupX({y2: "count"}, {x: "sport", fill: "sex", mixBlendMode: "multiply"})),
Plot.ruleY([0])
]
})``````

CAUTION

While the mixBlendMode option is useful for mitigating occlusion, it can be slow to render if there are many elements. More than two overlapping histograms may also be hard to read.

Perhaps better would be to make a grouped bar chart using faceting. This is accomplished by setting the fx channel to facet horizontally on sport, while the x channel is used within each facet to draw side-by-side bars for each sex. The group transform automatically partitions groups by facet.

Fork
js
``````Plot.plot({
marginBottom: 100,
fx: {padding: 0, label: null, tickRotate: 90, tickSize: 6},
y: {grid: true},
color: {legend: true},
marks: [
Plot.barY(olympians, Plot.groupX({y2: "count"}, {x: "sex", fx: "sport", fill: "sex"})),
Plot.ruleY([0])
]
})``````

Alternatively, below we use directional arrows (a link mark with markers) to indicate the difference in counts of male and female athletes by sport. The color of the arrow indicates which sex is more prevalent, while its length is proportional to the difference.

Fork
js
``````Plot.plot({
marginBottom: 100,
x: {label: null, tickRotate: 90},
y: {grid: true, label: "Frequency"},
color: {type: "categorical", domain: [-1, 1], unknown: "#aaa", transform: Math.sign},
marks: [
Plot.ruleY([0]),
olympians,
Plot.groupX(
{
y1: (D) => d3.sum(D, (d) => d === "female"),
y2: (D) => d3.sum(D, (d) => d === "male"),
stroke: (D) => d3.sum(D, (d) => d === "male") - d3.sum(D, (d) => d === "female")
},
{
x: "sport",
y1: "sex",
y2: "sex",
markerStart: "dot",
markerEnd: "arrow",
stroke: "sex",
strokeWidth: 2
}
)
)
]
})``````

The group transform comes in four orientations:

• groupX groups on x, and often outputs y as in a vertical↑ bar chart;
• groupY groups on y, and often outputs x as in a horizontal→ bar chart;
• groupZ groups on neither x nor y, combining everything into one group; and
• group groups on both x and y, and often outputs to fill or r as in a heatmap.

As you might guess, the groupY transform with the barX mark produces a horizontal→ bar chart. (We must increase the marginLeft to avoid the y axis labels from being cut off.)

Fork
js
``````Plot.plot({
marginLeft: 100,
x: {grid: true},
y: {label: null},
marks: [
Plot.barX(olympians, Plot.groupY({x: "count"}, {y: "sport", sort: {y: "x"}})),
Plot.ruleX([0])
]
})``````

You can produce a two-dimensional heatmap with group transform and a cell mark by generating a fill output channel. For example, we could show the median weight of athletes by sport (x) and sex (y).

Fork
js
``````Plot.plot({
marginBottom: 80,
x: {label: null, tickRotate: 90},
y: {label: null},
color: {label: "Median weight (kg)", legend: true, scheme: "YlGnBu"},
marks: [
Plot.cell(olympians, Plot.group({fill: "median"}, {fill: "weight", x: "sport", y: "sex"}))
]
})``````

Or, we could group athletes by sport and the number of gold medals 🥇 won. (Michael Phelps, the most decorated Olympian of all time, won five gold medals in the 2016 Summer Olympics. Simone Biles and Katie Ledecky each won four.)

Fork
js
``````Plot.plot({
marginBottom: 100,
x: {label: null, tickRotate: 90},
y: {label: "gold", labelAnchor: "top", labelArrow: true, reverse: true},
color: {type: "sqrt", scheme: "YlGnBu"},
marks: [
Plot.cell(olympians, Plot.group({fill: "count"}, {x: "sport", y: "gold"}))
]
})``````

We could instead output r and use a dot mark whose size again represents the number of athletes in each group.

Fork
js
``````Plot.plot({
marginBottom: 100,
x: {label: null, tickRotate: 90},
y: {type: "point", label: "gold", labelAnchor: "top", labelArrow: true, reverse: true},
r: {range: [0, 12]},
marks: [
Plot.dot(olympians, Plot.group({r: "count"}, {x: "sport", y: "gold"}))
]
})``````

We can add the stroke channel to show overlapping distributions by sex.

Fork
js
``````Plot.plot({
marginBottom: 100,
x: {label: null, tickRotate: 90},
y: {type: "point", label: "gold", labelAnchor: "top", labelArrow: true, reverse: true},
r: {range: [0, 12]},
marks: [
Plot.dot(olympians, Plot.group({r: "count"}, {x: "sport", y: "gold", stroke: "sex"}))
]
})``````

To group solely on z (or fill or stroke), use groupZ. The single stacked bar chart below (an alternative to a pie chart) shows the proportion of athletes by sport. The proportion reducer converts counts into normalized proportions adding up to 1, while the first reducer pulls out the name of the sport for labeling.

Fork
js
``````Plot.plot({
height: 100,
x: {percent: true},
marks: [
Plot.barX(
olympians,
Plot.stackX(
{order: "x", reverse: true},
Plot.groupZ(
{x: "proportion"},
{z: "sport", fillOpacity: 0.2, inset: 0.5}
)
)
),
Plot.text(
olympians,
Plot.filter(
(D) => D.length > 200,
Plot.stackX(
{order: "x", reverse: true},
Plot.groupZ(
{x: "proportion", text: "first"},
{z: "sport", text: "sport", rotate: 90}
)
)
)
),
Plot.ruleX([0, 1])
]
})``````

INFO

Although barX applies an implicit stackX transform, textX does not; this example uses an explicit stackX transform in both cases for clarity, and to pass the additional order and reverse options to place the largest sport on the left. The filter transform is applied after the stack transform to hide the labels on the smallest sports where the bars are too thin.

## Group options ​

Given input data = [d₀, d₁, d₂, …], by default the resulting grouped data is an array of arrays where each inner array is a subset of the input data such as [[d₁, d₂, …], [d₀, …], …]. Each inner array is in input order. The outer array is in input order according to the first element of each group.

By specifying a different reducer for the data output, as described below, you can change how the grouped data is computed. The outputs may also include filter and sort options specified as reducers, and a reverse option to reverse the order of generated groups. By default, empty groups are omitted, and non-empty groups are generated in ascending (natural) order.

In addition to data, the following channels are automatically output:

• x - the horizontal position of the group
• y - the vertical position of the group
• z - the first value of the z channel, if any
• fill - the first value of the fill channel, if any
• stroke - the first value of the stroke channel, if any

The x output channel is only computed by the groupX and group transform; similarly the y output channel is only computed by the groupY and group transform.

You can declare additional output channels by specifying the channel name and desired reducer in the outputs object which is the first argument to the transform. For example, to use groupX to generate a y channel of group counts as in a frequency histogram:

js
``Plot.groupX({y: "count"}, {x: "species"})``

The following named reducers are supported:

• first - the first value, in input order
• last - the last value, in input order
• count - the number of elements (frequency)
• sum - the sum of values
• proportion - the sum proportional to the overall total (weighted frequency)
• proportion-facet - the sum proportional to the facet total
• min - the minimum value
• min-index - the zero-based index of the minimum value
• max - the maximum value
• max-index - the zero-based index of the maximum value
• mean - the mean value (average)
• median - the median value
• mode - the value with the most occurrences
• pXX - the percentile value, where XX is a number in [00,99]
• deviation - the standard deviation
• variance - the variance per Welford’s algorithm
• identity - the array of values
• x ^0.6.12 - the group’s x value (when grouping on x)
• y ^0.6.12 - the group’s y value (when grouping on y)
• z ^0.6.14 - the group’s z value (z, fill, or stroke)

In addition, a reducer may be specified as:

• a function to be passed the array of values for each group and the extent of the group
• an object with a reduceIndex method, an optionally a scope

In the last case, the reduceIndex method is repeatedly passed three arguments: the index for each group (an array of integers), the input channel’s array of values, and the extent of the group (an object {data, x, y}); it must then return the corresponding aggregate value for the group.

If the reducer object’s scope is data, then the reduceIndex method is first invoked for the full data; the return value of the reduceIndex method is then made available as a third argument (making the extent the fourth argument). Similarly if the scope is facet, then the reduceIndex method is invoked for each facet, and the resulting reduce value is made available while reducing the facet’s groups. (This optional scope is used by the proportion and proportion-facet reducers.)

Most reducers require binding the output channel to an input channel; for example, if you want the y output channel to be a sum (not merely a count), there should be a corresponding y input channel specifying which values to sum. If there is not, sum will be equivalent to count.

js
``Plot.groupX({y: "sum"}, {x: "species", y: "body_mass_g"})``

You can control whether a channel is computed before or after grouping. If a channel is declared only in options (and it is not a special group-eligible channel such as x, y, z, fill, or stroke), it will be computed after grouping and be passed the grouped data: each datum is the array of input data corresponding to the current group.

js
``Plot.groupX({y: "count"}, {x: "species", title: (group) => group.map((d) => d.body_mass_g).join("\n")})``

This is equivalent to declaring the channel only in outputs.

js
``Plot.groupX({y: "count", title: (group) => group.map((d) => d.body_mass_g).join("\n")}, {x: "species"})``

However, if a channel is declared in both outputs and options, then the channel in options is computed before grouping and can be aggregated using any built-in reducer (or a custom reducer function) during the group transform.

js
``Plot.groupX({y: "count", title: (masses) => masses.join("\n")}, {x: "species", title: "body_mass_g"})``

If any of z, fill, or stroke is a channel, the first of these channels is considered the z dimension and will be used to subdivide groups.

The default reducer for the title channel returns a summary list of the top 5 values with the corresponding number of occurrences.

## group(outputs, options) ​

js
``Plot.group({fill: "count"}, {x: "island", y: "species"})``

Groups on x, y, and the first channel of z, fill, or stroke, if any.

## groupX(outputs, options) ​

js
``Plot.groupX({y: "sum"}, {x: "species", y: "body_mass_g"})``

Groups on x and the first channel of z, fill, or stroke, if any.

## groupY(outputs, options) ​

js
``Plot.groupY({x: "sum"}, {y: "species", x: "body_mass_g"})``

Groups on y and the first channel of z, fill, or stroke, if any.

## groupZ(outputs, options) ​

js
``Plot.groupZ({x: "proportion"}, {fill: "species"})``

Groups on the first channel of z, fill, or stroke, if any. If none of z, fill, or stroke are channels, then all data (within each facet) is placed into a single group.

## find(test) ^0.6.12​

js
``````Plot.groupX(
{y1: Plot.find((d) => d.sex === "F"), y2: Plot.find((d) => d.sex === "M")},
{x: "date", y: "value"}
)``````

Returns a reducer that finds the first datum for which the given test function returns a truthy value, and returns the corresponding channel value. This may be used with the group or bin transform to implement a “pivot wider” transform; for example, a “tall” dataset with separate rows for male and female observations may be transformed into a “wide” dataset with separate columns for male and female values.

Resources
Observable