lattice (version 0.22-6)

B_00_xyplot: Common Bivariate Trellis Plots


This help page documents several commonly used high-level Lattice functions. xyplot produces bivariate scatterplots or time-series plots, bwplot produces box-and-whisker plots, dotplot produces Cleveland dot plots, barchart produces bar plots, and stripplot produces one-dimensional scatterplots. All these functions, along with other high-level Lattice functions, respond to a common set of arguments that control conditioning, layout, aspect ratio, legends, axis annotation, and many other details in a consistent manner. These arguments are described extensively in this help page, and should be used as the reference for other high-level functions as well.

For control and customization of the actual display in each panel, the help page of the respective default panel function will often be more informative. In particular, these help pages describe many arguments commonly used when calling the corresponding high-level function but are specific to them.


xyplot(x, data, ...)
dotplot(x, data, ...)
barchart(x, data, ...)
stripplot(x, data, ...)
bwplot(x, data, ...)

# S3 method for formula xyplot(x, data, allow.multiple = is.null(groups) || outer, outer = !is.null(groups), auto.key = lattice.getOption("default.args")$auto.key, aspect = "fill", panel = lattice.getOption("panel.xyplot"), prepanel = NULL, scales = list(), strip = TRUE, groups = NULL, xlab, xlim, ylab, ylim, drop.unused.levels = lattice.getOption("drop.unused.levels"), ..., lattice.options = NULL, default.scales, default.prepanel = lattice.getOption("prepanel.default.xyplot"), subscripts = !is.null(groups), subset = TRUE)

# S3 method for data.frame xyplot(x, data = NULL, formula = data, ...)

# S3 method for formula dotplot(x, data, panel = lattice.getOption("panel.dotplot"), default.prepanel = lattice.getOption("prepanel.default.dotplot"), ...)

# S3 method for data.frame dotplot(x, data = NULL, formula = data, ...)

# S3 method for formula barchart(x, data, panel = lattice.getOption("panel.barchart"), default.prepanel = lattice.getOption("prepanel.default.barchart"), box.ratio = 2, ...)

# S3 method for data.frame barchart(x, data = NULL, formula = data, ...)

# S3 method for formula stripplot(x, data, panel = lattice.getOption("panel.stripplot"), default.prepanel = lattice.getOption("prepanel.default.stripplot"), ...)

# S3 method for data.frame stripplot(x, data = NULL, formula = data, ...)

# S3 method for formula bwplot(x, data, allow.multiple = is.null(groups) || outer, outer = FALSE, auto.key = lattice.getOption("default.args")$auto.key, aspect = "fill", panel = lattice.getOption("panel.bwplot"), prepanel = NULL, scales = list(), strip = TRUE, groups = NULL, xlab, xlim, ylab, ylim, box.ratio = 1, horizontal = NULL, drop.unused.levels = lattice.getOption("drop.unused.levels"), ..., lattice.options = NULL, default.scales, default.prepanel = lattice.getOption("prepanel.default.bwplot"), subscripts = !is.null(groups), subset = TRUE)

# S3 method for data.frame bwplot(x, data = NULL, formula = data, ...)


The high-level functions documented here, as well as other high-level Lattice functions, return an object of class "trellis". The

update method can be used to subsequently update components of the object, and the

print method (usually called by default) will plot it on an appropriate plotting device.



All high-level function in lattice are generic. x is the object on which method dispatch is carried out.

For the "formula" methods, x must be a formula describing the primary variables (used for the per-panel display) and the optional conditioning variables (which define the subsets plotted in different panels) to be used in the plot. Conditioning is described in the “Details” section below.

For the functions documented here, the formula is generally of the form y ~ x | g1 * g2 * ... (or equivalently, y ~ x | g1 + g2 + ...), indicating that plots of y (on the y-axis) versus x (on the x-axis) should be produced conditional on the variables g1, g2, .... Here x and y are the primary variables, and g1, g2, ... are the conditioning variables. The conditioning variables may be omitted to give a formula of the form y ~ x, in which case the plot will consist of a single panel with the full dataset. The formula can also involve expressions, e.g., sqrt(), log(), etc. See the data argument below for rules regarding evaluation of the terms in the formula.

With the exception of xyplot, the functions documented here may also be supplied a formula of the form ~ x | g1 * g2 * .... In that case, y defaults to names(x) if x is named, and a factor with a single level otherwise.

Cases where x is not a formula is handled by appropriate methods. The numeric methods are equivalent to a call with no left hand side and no conditioning variables in the formula. For barchart and dotplot, non-trivial methods exist for tables and arrays, documented at barchart.table.

The conditioning variables g1, g2, ... must be either factors or shingles. Shingles provide a way of using numeric variables for conditioning; see the help page of shingle for details. Like factors, they have a "levels" attribute, which is used in producing the conditional plots. If necessary, numeric conditioning variables are converted to shingles using the shingle function; however, using equal.count may be more appropriate in many cases. Character variables are coerced to factors.

Extended formula interface: As a useful extension of the interface described above, the primary variable terms (both the LHS y and RHS x) may consist of multiple terms separated by a ‘+’ sign, e.g., y1 + y2 ~ x | a * b. This formula would be taken to mean that the user wants to plot both y1 ~ x | a * b and y2 ~ x | a * b, but with the y1 ~ x and y2 ~ x superposed in each panel. The two groups will be distinguished by different graphical parameters. This is essentially what the groups argument (see below) would produce, if y1 and y2 were concatenated to produce a longer vector, with the groups argument being an indicator of which rows come from which variable. In fact, this is exactly what is done internally using the reshape function. This feature cannot be used in conjunction with the groups argument.

To interpret y1 + y2 as a sum, one can either set allow.multiple=FALSE or use I(y1+y2).

A variation on this feature is when the outer argument is set to TRUE. In that case, the plots are not superposed in each panel, but instead separated into different panels (as if a new conditioning variable had been added).

Primary variables: The x and y variables should both be numeric in xyplot, and an attempt is made to coerce them if not. However, if either is a factor, the levels of that factor are used as axis labels. In the other four functions documented here, exactly one of x and y should be numeric, and the other a factor or shingle. Which of these will happen is determined by the horizontal argument --- if horizontal=TRUE, then y will be coerced to be a factor or shingle, otherwise x. The default value of horizontal is FALSE if x is a factor or shingle, TRUE otherwise. (The functionality provided by horizontal=FALSE is not S-compatible.)

Note that the x argument used to be called formula in earlier versions (when the high-level functions were not generic and the formula method was essentially the only method). This is no longer allowed. It is recommended that this argument not be named in any case, but instead be the first (unnamed) argument.


For the formula methods, a data frame (or more precisely, anything that is a valid envir argument in eval, e.g., a list or an environment) containing values for any variables in the formula, as well as groups and subset if applicable. If not found in data, or if data is unspecified, the variables are looked for in the environment of the formula. For other methods (where x is not a formula), data is usually ignored, often with a warning if it is explicitly specified.


The formula to be used for the "data.frame" methods. See documentation for argument x for details.


Logical flag specifying whether the extended formula interface described above should be in effect. Defaults to TRUE whenever sensible.


Logical flag controlling what happens with formulas using the extended interface described above (see the entry for x for details). Defaults to FALSE, except when groups is explicitly specified or grouping does not make sense for the default panel function.


Applicable to barchart and bwplot. Specifies the ratio of the width of the rectangles to the inter-rectangle space. See also the box.width argument in the respective default panel functions.


Logical flag applicable to bwplot, dotplot, barchart, and stripplot. Determines which of x and y is to be a factor or shingle (y if TRUE, x otherwise). Defaults to FALSE if x is a factor or shingle, TRUE otherwise. This argument is used to process the arguments to these high-level functions, but more importantly, it is passed as an argument to the panel function, which is expected to use it as appropriate.

A potentially useful component of scales in this case may be abbreviate = TRUE, in which case long labels which would usually overlap will be abbreviated. scales could also contain a minlength argument in this case, which would be passed to the abbreviate function.

Common arguments: The following arguments are common to all the functions documented here, as well as most other high-level Trellis functions. These are not documented elsewhere, except to override the usage given here.


Once the subset of rows defined by each unique combination of the levels of the grouping variables are obtained (see “Details”), the corresponding x and y variables (or other variables, as appropriate, in the case of other high-level functions) are passed on to be plotted in each panel. The actual plotting is done by the function specified by the panel argument. The argument may be a function object or a character string giving the name of a predefined function. Each high-level function has its own default panel function, named as “panel.” followed by the name of the corresponding high-level function (e.g., panel.xyplot, panel.barchart, etc).

Much of the power of Trellis Graphics comes from the ability to define customized panel functions. A panel function appropriate for the functions described here would usually expect arguments named x and y, which would be provided by the conditioning process. It can also have other arguments. It is useful to know in this context that all arguments passed to a high-level Lattice function (such as xyplot) that are not recognized by it are passed through to the panel function. It is thus generally good practice when defining panel functions to allow a ... argument. Such extra arguments typically control graphical parameters, but other uses are also common. See documentation for individual panel functions for specifics.

Note that unlike in S-PLUS, it is not guaranteed that panel functions will be supplied only numeric vectors for the x and y arguments; they can be factors as well (but not shingles). Panel functions need to handle this case, which in most cases can be done by simply coercing them to numeric.

Technically speaking, panel functions must be written using Grid graphics functions. However, knowledge of Grid is usually not necessary to construct new custom panel functions, as there are several predefined panel functions which can help; for example, panel.grid, panel.loess, etc. There are also some grid-compatible replacements of commonly used traditional graphics functions useful for this purpose. For example, lines can be replaced by llines (or equivalently, panel.lines). Note that traditional graphics functions like lines will not work in a lattice panel function.

One case where a bit more is required of the panel function is when the groups argument is not NULL. In that case, the panel function should also accept arguments named groups and subscripts (see below for details). A useful panel function predefined for use in such cases is panel.superpose, which can be combined with different panel.groups functions to determine what is plotted for each group. See the “Examples” section for an interaction plot constructed in this way. Several other panel functions can also handle the groups argument, including the default ones for xyplot, barchart, dotplot, and stripplot.

Even when groups is not present, the panel function can have subscripts as a formal argument. In either case, the subscripts argument passed to the panel function are the indices of the x and y data for that panel in the original data, BEFORE taking into account the effect of the subset argument. Note that groups remains unaffected by any subsetting operations, so groups[subscripts] gives the values of groups that correspond to the data in that panel.

This interpretation of subscripts does not hold when the extended formula interface is in use (i.e., when allow.multiple is in effect). A comprehensive description would be too complicated (details can be found in the source code of the function latticeParseFormula), but in short, the extended interface works by creating an artificial grouping variable that is longer than the original data frame, and consequently, subscripts needs to refer to rows beyond those in the original data. To further complicate matters, the artificial grouping variable is created after any effect of subset, in which case subscripts may have no relationship with corresponding rows in the original data frame.

One can also use functions called panel.number and packet.number, representing panel order and packet order respectively, inside the panel function (as well as the strip function or while interacting with a lattice display using trellis.focus etc). Both provide a simple integer index indicating which panel is currently being drawn, but differ in how the count is calculated. The panel number is a simple incremental counter that starts with 1 and is incremented each time a panel is drawn. The packet number on the other hand indexes the combination of levels of the conditioning variables that is represented by that panel. The two indices coincide unless the order of conditioning variables is permuted and/or the plotting order of levels within one or more conditioning variables is altered (using perm.cond and index.cond respectively), in which case packet.number gives the index corresponding to the ‘natural’ ordering of that combination of levels of the conditioning variables.

panel.xyplot has an argument called type which is worth mentioning here because it is quite frequently used (and as mentioned above, can be passed to xyplot directly). In the event that a groups variable is used, panel.xyplot calls panel.superpose, arguments of which can also be passed directly to xyplot. Panel functions for bwplot and friends should have an argument called horizontal to account for the cases when x is the factor or shingle.


This argument controls the physical aspect ratio of the panels, which is usually the same for all the panels. It can be specified as a ratio (vertical size/horizontal size) or as a character string. In the latter case, legitimate values are "fill" (the default) which tries to make the panels as big as possible to fill the available space; "xy", which computes the aspect ratio based on the 45 degree banking rule (see banking); and "iso" for isometric scales, where the relation between physical distance on the device and distance in the data scale are forced to be the same for both axes.

If a prepanel function is specified and it returns components dx and dy, these are used for banking calculations. Otherwise, values from the default prepanel function are used. Not all default prepanel functions produce sensible banking calculations.


A variable or expression to be evaluated in data, expected to act as a grouping variable within each panel, typically used to distinguish different groups by varying graphical parameters like color and line type. Formally, if groups is specified, then groups along with subscripts is passed to the panel function, which is expected to handle these arguments. For high level functions where grouping is appropriate, the default panel functions can handle grouping.

It is very common to use a key (legend) when a grouping variable is specified. See entries for key, auto.key and simpleKey for how to draw a key.


A logical, or a list containing components to be used as arguments to simpleKey. The default can be set using lattice.options.

auto.key = TRUE is equivalent to auto.key = list(), in which case simpleKey is called with a set of default arguments (which may depend on the relevant high-level function). Most valid components to the key argument can be specified in this manner, as simpleKey will simply add unrecognized arguments to the list it produces.

auto.key is typically used to automatically produce a suitable legend in conjunction with a grouping variable. If auto.key = TRUE, a suitable legend will be drawn if a groups argument is also provided, and not otherwise. In list form, auto.key will modify the default legend thus produced. For example, auto.key=list(columns = 2) will create a legend split into two columns (columns is documented in the entry for key).

More precisely, if auto.key is not FALSE, groups is non-null, and there is no key or legend argument specified in the call, a key is created with simpleKey with levels(groups) as the first (text) argument. (Note: this may not work in all high-level functions, but it does work for the ones where grouping makes sense with the default panel function). If auto.key is provided as a list and includes a text component, then that is used instead as the text labels in the key, and the key is drawn even if groups is not specified.

Note that simpleKey uses the default settings (see trellis.par.get) to determine the graphical parameters in the key, so the resulting legend will be meaningful only if the same settings are used in the plot as well. The par.settings argument, possibly in conjunction with simpleTheme, may be useful to temporarily modify the default settings for this purpose.

One disadvantage to using key (or even simpleKey) directly is that the graphical parameters used in the key are absolutely determined at the time when the "trellis" object is created. Consequently, if a plot once created is re-plot-ted with different settings, the original parameter settings will be used for the key even though the new settings are used for the actual display. However, with auto.key, the key is actually created at plotting time, so the settings will match.


A function that takes the same arguments as the panel function and returns a list, possibly containing components named xlim, ylim, dx, and dy (and less frequently, xat and yat). The return value of a user-supplied prepanel function need not contain all these components; in case some are missing, they are replaced by the component-wise defaults.

The xlim and ylim components are similar to the high level xlim and ylim arguments (i.e., they are usually a numeric vector of length 2 defining a range, or a character vector representing levels of a factor). If the xlim and ylim arguments are not explicitly specified (possibly as components in scales) in the high-level call, then the actual limits of the panels are guaranteed to include the limits returned by the prepanel function. This happens globally if the relation component of scales is "same", and on a per-panel basis otherwise.

The dx and dy components are used for banking computations in case aspect is specified as "xy". See documentation of banking for details.

If xlim or ylim is a character vector (which is appropriate when the corresponding variable is a factor), this implicitly indicates that the scale should include the first n integers, where n is the length of xlim or ylim, as the case may be. The elements of the character vector are used as the default labels for these n integers. Thus, to make this information consistent between panels, the xlim or ylim values should represent all the levels of the corresponding factor, even if some are not used within that particular panel.

In such cases, an additional component xat or yat may be returned by the prepanel function, which should be a subset of 1:n, indicating which of the n values (levels) are actually represented in the panel. This is useful when calculating the limits with relation="free" or relation="sliced" in scales.

The prepanel function is responsible for providing a meaningful return value when the x, y (etc.) variables are zero-length vectors. When nothing else is appropriate, values of NA should be returned for the xlim and ylim components.


A logical flag or function. If FALSE, strips are not drawn. Otherwise, strips are drawn using the strip function, which defaults to strip.default. See documentation of strip.default to see the arguments that are available to the strip function. This description also applies to the strip.left argument (see ... below), which can be used to draw strips on the left of each panel (useful for wide short panels, e.g., in time-series plots).


Character or expression (or a "grob") giving label(s) for the x-axis. Generally defaults to the expression for x in the formula defining the plot. Can be specified as NULL to omit the label altogether. Finer control is possible, as described in the entry for main, with the modification that if the label component is omitted from the list, it is replaced by the default xlab.


Character or expression (or "grob") giving label for the y-axis. Generally defaults to the expression for y in the formula defining the plot. Finer control is possible, see entries for main and xlab.


Generally a list determining how the x- and y-axes (tick marks and labels) are drawn. The list contains parameters in name=value form, and may also contain two other lists called x and y of the same form (described below). Components of x and y affect the respective axes only, while those in scales affect both. When parameters are specified in both lists, the values in x or y are used. Note that certain high-level functions have defaults that are specific to a particular axis (e.g., bwplot has alternating=FALSE for the categorical axis only); these can only be overridden by an entry in the corresponding component of scales.

As a special exception, scales (or its x and y components) can also be a character string, in which case it is interpreted as the relation component.

The possible components are :


A character string that determines how axis limits are calculated for each panel. Possible values are "same" (default), "free" and "sliced". For relation="same", the same limits, usually large enough to encompass all the data, are used for all the panels. For relation="free", limits for each panel is determined by just the points in that panel. Behavior for relation="sliced" is similar, except that the length (max - min) of the scales are constrained to remain the same across panels.

The determination of what axis limits are suitable for each panel can be controlled by the prepanel function, which can be overridden by xlim, ylim or scales$limits (except when relation="sliced", in which case explicitly specified limits are ignored with a warning). When relation is "free", xlim or ylim can be a list, in which case it is treated as if its components were the limit values obtained from the prepanel calculations for each panel (after being replicated if necessary).


An integer, giving the suggested number of intervals between ticks. This is ignored for a factor, shingle, or character vector, for in these cases there is no natural rule for leaving out some of the labels. But see xlim.


A logical flag, defaulting to TRUE, that determines whether to draw the axis (i.e., tick marks and labels) at all.


Usually a logical flag specifying whether axis labels should alternate from one side of the group of panels to the other. For finer control, alternating can also be a vector (replicated to be as long as the number of rows or columns per page) consisting of the following numbers

  • 0: do not draw tick labels

  • 1: bottom/left

  • 2: top/right

  • 3: both.

alternating applies only when relation="same". The default is TRUE, or equivalently, c(1, 2)


Same as xlim and ylim.


The location of tick marks along the axis (in native coordinates), or a list as long as the number of panels describing tick locations for each panel.


Vector of labels (characters or expressions) to go along with at. Can also be a list like at.


A numeric multiplier to control character sizes for axis labels. Can be a vector of length 2, to control left/bottom and right/top labels separately.

font, fontface, fontfamily

Specifies the font to be used for axis labels.


Specifies the line height parameter (height of line as a multiple of the size of text); relevant for multi-line labels. (This is currently ignored for cloud.)


Usually a numeric scalar controlling the length of tick marks. Can also be a vector of length 2, to control the length of left/bottom and right/top tick marks separately.


Color of tick marks and labels.


Angle (in degrees) by which the axis labels are to be rotated. Can be a vector of length 2, to control left/bottom and right/top axes separately.


A logical flag, indicating whether to abbreviate the labels using the abbreviate function. Can be useful for long labels (e.g., in factors), especially on the x-axis.


Argument passed to abbreviate if abbreviate=TRUE.


Controls whether the corresponding variable (x or y) will be log transformed before being passed to the panel function. Defaults to FALSE, in which case the data are not transformed. Other possible values are any number that works as a base for taking logarithm, TRUE (which is equivalent to 10), and "e" (for the natural logarithm). As a side effect, the corresponding axis is labeled differently. Note that this is in reality a transformation of the data, not the axes. Other than the axis labeling, using this feature is no different than transforming the data in the formula; e.g., scales=list(x = list(log = 2)) is equivalent to y ~ log2(x).

See entry for equispaced.log below for details on how to control axis labeling.


A logical flag indicating whether tick mark locations should be equispaced when ‘log scales’ are in use. Defaults to TRUE.

Tick marks are always labeled in the original (untransformed) scale, but this makes the choice of tick mark locations nontrivial. If equispaced.log is FALSE, the choice made is similar to how log scales are annotated in traditional graphics. If TRUE, tick mark locations are chosen as ‘pretty’ equispaced values in the transformed scale, and labeled in the form "base^loc", where base is the base of the logarithm transformation, and loc are the locations in the transformed scale.

See also xscale.components.logpower in the latticeExtra package.


The format to use for POSIXct variables. See strptime for description of valid values.


A character string, "r" (default) or "i". In the latter case, the axis limits are calculated as the exact data range, instead of being padded on either side. (May not always work as expected.)

Note that much of the function of scales is accomplished by pscales in splom.


A logical flag specifying whether or not a vector named subscripts should be passed to the panel function. Defaults to FALSE, unless groups is specified, or if the panel function accepts an argument named subscripts. This argument is useful if one wants the subscripts to be passed on even if these conditions do not hold; a typical example is when one wishes to augment a Lattice plot after it has been drawn, e.g., using panel.identify.


An expression that evaluates to a logical or integer indexing vector. Like groups, it is evaluated in data. Only the resulting rows of data are used for the plot. If subscripts is TRUE, the subscripts provided to the panel function will be indices referring to the rows of data prior to the subsetting. Whether levels of factors in the data frame that are unused after the subsetting will be dropped depends on the drop.unused.levels argument.


Normally a numeric vector (or a DateTime object) of length 2 giving left and right limits for the x-axis, or a character vector, expected to denote the levels of x. The latter form is interpreted as a range containing c(1, length(xlim)), with the character vector determining labels at tick positions 1:length(xlim).

xlim could also be a list, with as many components as the number of panels (recycled if necessary), with each component as described above. This is meaningful only when scales$x$relation is "free", in which case these are treated as if they were the corresponding limit components returned by prepanel calculations.


Similar to xlim, applied to the y-axis.


A logical flag indicating whether the unused levels of factors will be dropped, usually relevant when a subsetting operation is performed or an interaction is created. Unused levels are usually dropped, but it is sometimes appropriate to suppress dropping to preserve a useful layout. For finer control, this argument could also be list containing components cond and data, both logical, indicating desired behavior for conditioning variables and primary variables respectively. The default is given by lattice.getOption("drop.unused.levels"), which is initially set to TRUE for both components. Note that this argument does not control dropping of levels of the groups argument.


A list giving the default values of scales for a particular high-level function. This is rarely of interest to the end-user, but may be helpful when defining other functions that act as a wrapper to one of the high-level Lattice functions.


A function or character string giving the name of a function that serves as the (component-wise) fallback prepanel function when the prepanel argument is not specified, or does not return all necessary components. The main purpose of this argument is to enable the defaults to be overridden through the use of lattice.options.


A list that could be supplied to lattice.options. These options are applied temporarily for the duration of the call, after which the settings revert back to what they were before. The options are retained along with the object and reused during plotting. This enables the user to attach options settings to the trellis object itself rather than change the settings globally. See also the par.settings argument described below for a similar treatment of graphical settings.


Further arguments, usually not directly processed by the high-level functions documented here, but instead passed on to other functions. Such arguments can be broadly categorized into two types: those that affect all high-level Lattice functions in a similar manner, and those that are meant for the specific panel function being used.

The first group of arguments are processed by a common, unexported function called trellis.skeleton. These arguments affect all high-level functions, but are only documented here (except to override the behaviour described here). All other arguments specified in a high-level call, specifically those neither described here nor in the help page of the relevant high-level function, are passed unchanged to the panel function used. By convention, the default panel function used for any high-level function is named as “panel.” followed by the name of the high-level function; for example, the default panel function for bwplot is panel.bwplot. In practical terms, this means that in addition to the help page of the high-level function being used, the user should also consult the help page of the corresponding panel function for arguments that may be specified in the high-level call.

The effect of the first group of common arguments are as follows:


A logical flag that controls the order in which panels should be displayed: if FALSE (the default), panels are drawn left to right, bottom to top (as in a graph); if TRUE, left to right, top to bottom (as in a table).


A list with components x and y (both usually 0 by default), numeric vectors specifying the space between the panels (units are character heights). x and y are repeated to account for all panels in a page and any extra components are ignored. The result is used for all pages in a multi page display. In other words, it is not possible to use different between values for different pages.


A list that defines a legend to be drawn on the plot. This list is used as an argument to the draw.key function, which produces a "grob" (grid object) eventually plotted by the print method for "trellis" objects. The structure of the legend is constrained in the ways described below.

Although such a list can be and often is created explicitly, it is also possible to generate such a list using the simpleKey function; the latter is more convenient but less flexible. The auto.key argument can be even more convenient for the most common situation where legends are used, namely, in conjunction with a grouping variable. To use more than one legend, or to have arbitrary legends not constrained by the structure imposed by key, use the legend argument.

The position of the key can be controlled in either of two possible ways. If a component called space is present, the key is positioned outside the plot region, in one of the four sides, determined by the value of space, which can be one of "top", "bottom", "left" and "right". Alternatively, the key can be positioned inside the plot region by specifying components x, y and corner. x and y determine the location of the corner of the key given by corner, which is usually one of c(0,0), c(1,0), c(1,1) and c(0,1), which denote the corners of the unit square. Fractional values are also allowed, in which case x and y determine the position of an arbitrary point inside (or outside for values outside the unit interval) the key.

x and y should be numbers between 0 and 1, giving coordinates with respect to the “display area”. Depending on the value of the "legend.bbox" option (see lattice.getOption), this can be either the full figure region ("full"), or just the region that bounds the panels and strips ("panel").

The key essentially consists of a number of columns, possibly divided into blocks, each containing some rows. The contents of the key are determined by (possibly repeated) components named "rectangles", "lines", "points" or "text". Each of these must be lists with relevant graphical parameters (see later) controlling their appearance. The key list itself can contain graphical parameters, these would be used if relevant graphical components are omitted from the other components.

The length (number of rows) of each such column (except "text"s) is taken to be the largest of the lengths of the graphical components, including the ones specified outside (see the entry for rep below for details on this). The "text" component must have a character or expression vector as its first component, to be used as labels. The length of this vector determines the number of rows.

The graphical components that can be included in key and also in the components named "text", "lines", "points" and "rectangles" (as appropriate) are:

  • cex=1 (text, lines, points)

  • col="black" (text, rectangles, lines, points)

  • alpha=1 (text, rectangles, lines, points)

  • fill="transparent" (lines, points)

  • lty=1 (lines)

  • lwd=1 (lines, points)

  • font=1 (text, points)

  • fontface (text, points)

  • fontfamily (text, points)

  • pch=8 (lines, points)

  • adj=0 (text)

  • type="l" (lines)

  • size=5 (rectangles, lines)

  • height=1 (rectangles)

  • lineheight=1 (text)

  • angle=0 (rectangles, but ignored)

  • density=-1 (rectangles, but ignored)

In addition, the component border can be included inside the "rect" component to control the border color of the rectangles; when specified at the top level, border controls the border of the entire key (see below).

angle and density are unimplemented. size determines the width of columns of rectangles and lines in character widths. type is relevant for lines; "l" denotes a line, "p" denotes a point, and "b" and "o" both denote both together. height gives heights of rectangles as a fraction of the default.

Other possible components of key are:


Logical flag, defaulting to FALSE. If TRUE, all components are reversed after being replicated (the details of which may depend on the value of rep). This is useful in certain situations, e.g., with a grouped barchart with stack = TRUE with the categorical variable on the vertical axis, where the bars in the plot will usually be ordered from bottom to top, but the corresponding legend will have the levels from top to bottom unless reverse.rows = TRUE. Note that in this case, unless all columns have the same number or rows, they will no longer be aligned.


Numeric vector giving the amount of space (character widths) surrounding each column (split equally on both sides).


String or expression giving a title for the key.


Logical flag, defaults to TRUE. By default, it is assumed that all columns in the key (except the "text"s) will have the same number of rows, and all components are replicated to be as long as the longest. This can be suppressed by specifying rep=FALSE, in which case the length of each column will be determined by components of that column alone.


Zoom factor for the title.


The amount of vertical space to be occupied by the title in lines (in multiples of itself). Defaults to 2.


The amount of space (padding) to be used above and below each row containing text, in multiples of the default, which is currently 0.2 * "lines". This padding is in addition to the normal height of any row that contains text, which is the minimum amount necessary to contain all the text entries.


Background color for the legend. Defaults to the global background color.


An alpha transparency value between 0 and 1 for the background.


Either a color for the border, or a logical flag. In the latter case, the border color is black if border is TRUE, and no border is drawn if it is FALSE (the default).


Logical flag, whether legend should have a transparent background.


A character or numeric vector of length one or two giving horizontal and vertical justification for the placement of the legend. See grid.layout for more precise details.


The number of column-blocks (drawn side by side) the legend is to be divided into.


Space between column blocks, in addition to between.


Number of point symbols to divide each line when type is "b" or "o" in lines.


The legend argument can be useful if one wants to place more than one key. It also allows the use of arbitrary "grob"s (grid objects) as legends.

If used, legend must be a list, with an arbitrary number of components. Each component must be named one of "left", "right", "top", "bottom", or "inside". The name "inside" can be repeated, but not the others. This name will be used to determine the location for that component, and is similar to the space component of key. If key (or colorkey for levelplot and wireframe) is specified, their space component must not conflict with the name of any component of legend.

Each component of legend must have a component called fun. This can be a "grob", or a function (or the name of a function) that produces a "grob" when called. If this function expects any arguments, they must be supplied as a list in another component called args. For components named "inside", there can be additional components called x, y and corner, which work in the same way as for key.


A function of one argument (page number) to be called after drawing each page. The function must be ‘grid-compliant’, and is called with the whole display area as the default viewport., ylab.right:

Labels for the x-axis on top, and y-axis on the right. Similar to xlab and ylab, but less commonly used.


Typically a character string or expression describing the main title to be placed on top of each page. Defaults to NULL.

main (as well as xlab, ylab and sub) is usually a character string or an expression that gets used as the label, but can also be a list that controls further details. Expressions are treated as specification of LaTeX-like markup as described in plotmath. The label can be a vector, in which case the components will be spaced out horizontally (or vertically for ylab). This feature can be used to provide column or row labels rather than a single axis label.

When main (etc.) is a list, the actual label should be specified as the label component (which may be unnamed if it is the first component). The label can be missing, in which case the default will be used (xlab and ylab usually have defaults, but main and sub do not). Further named arguments are passed on to textGrob; this can include arguments controlling positioning like just and rot as well as graphical parameters such as col and font (see gpar for a full list).

main, sub, xlab, ylab,, and ylab.right can also be arbitrary "grob"s (grid graphical objects).


Character string or expression (or a list or "grob") for a subtitle to be placed at the bottom of each page. See entry for main for finer control options.


A list of parameters to control the appearance of strip text. Notable components are col, cex, font, and lines. The first three control graphical parameters while the last is a means of altering the height of the strips. This can be useful, for example, if the strip labels (derived from factor levels, say) are double height (i.e., contains "\n"-s) or if the default height seems too small or too large.

Additionally, the lineheight component can control the space between multiple lines. The labels can be abbreviated when shown by specifying abbreviate = TRUE, in which case the components minlength and dot (passed along to the abbreviate function) can be specified to control the details of how this is done.


In general, a conditioning plot in Lattice consists of several panels arranged in a rectangular array, possibly spanning multiple pages. layout determines this arrangement.

layout is a numeric vector of length 2 or 3 giving the number of columns, rows, and pages (optional) in a multipanel display. By default, the number of columns is the number of levels of the first conditioning variable and the number of rows is the number of levels of the second conditioning variable. If there is only one conditioning variable, the default layout vector is c(0,n), where n is the number of levels of the given vector. Any time the first value in the layout vector is 0, the second value is used as the desired number of panels per page and the actual layout is computed from this, taking into account the aspect ratio of the panels and the device dimensions (via par("din")). If NA is specified for the number of rows or columns (but not both), that dimension will be filled out according to the number of panels.

The number of pages is by default set to as many as is required to plot all the panels, and so rarely needs to be specified. However, in certain situations the default calculation may be incorrect, and in that case the number of pages needs to be specified explicitly.


A logical vector (default FALSE), replicated to be as long as the number of panels (spanning all pages). For elements that are TRUE, the corresponding panel position is skipped; i.e., nothing is plotted in that position. The panel that was supposed to be drawn there is now drawn in the next available panel position, and the positions of all the subsequent panels are bumped up accordingly. This may be useful for arranging plots in an informative manner.


strip.left can be used to draw strips on the left of each panel, which can be useful for wide short panels, as in time-series (or similar) plots. See the entry for strip for detailed usage.

xlab.default, ylab.default:

Fallback default for xlab and ylab when they are not specified. If NULL, the defaults are parsed from the Trellis formula. This is rarely useful for the end-user, but can be helpful when developing new Lattice functions.

xscale.components, yscale.components:

Functions that determine axis annotation for the x and y axes respectively. See documentation for xscale.components.default, the default values of these arguments, to learn more.


Function responsible for drawing axis annotation. See documentation for axis.default, the default value of this argument, to learn more.


An integer vector, a permutation of 1:n, where n is the number of conditioning variables. By default, the order in which panels are drawn depends on the order of the conditioning variables specified in the formula. perm.cond can modify this order. If the trellis display is thought of as an n-dimensional array, then during printing, its dimensions are permuted using perm.cond as the perm argument does in aperm.


Whereas perm.cond permutes the dimensions of the multidimensional array of panels, index.cond can be used to subset (or reorder) margins of that array. index.cond can be a list or a function, with behavior in each case described below.

The panel display order within each conditioning variable depends on the order of their levels. index.cond can be used to choose a ‘subset’ (in the R sense) of these levels, which is then used as the display order for that variable. If index.cond is a list, it has to be as long as the number of conditioning variables, and the i-th component has to be a valid indexing vector for levels(g_i), where g_i is the i-th conditioning variable in the plot (note that these levels may not contain all levels of the original variable, depending on the effects of the subset and drop.unused.levels arguments). In particular, this indexing may repeat levels, or drop some altogether. The result of this indexing determines the order of panels within that conditioning variable. To keep the order of a particular variable unchanged, the corresponding component must be set to TRUE.

Note that the components of index.cond are interpreted in the order of the conditioning variables in the original call, and is not affected by perm.cond.

Another possibility is to specify index.cond as a function. In this case, this function is called once for each panel, potentially with all arguments that are passed to the panel function for that panel. (More specifically, if this function has a ... argument, then all panel arguments are passed, otherwise, only named arguments that match are passed.) If there is only one conditioning variable, the levels of that variable are then sorted so that these values are in ascending order. For multiple conditioning variables, the order for each variable is determined by first taking the average over all other conditioning variables.

Although they can be supplied in high-level function calls directly, it is more typical to use perm.cond and index.cond to update an existing "trellis" object, thus allowing it to be displayed in a different arrangement without re-calculating the data subsets that go into each panel. In the update.trellis method, both can be set to NULL, which reverts these back to their defaults.


A list that could be supplied to trellis.par.set. When the resulting object is plotted, these options are applied temporarily for the duration of the plotting, after which the settings revert back to what they were before. This enables the user to attach some display settings to the trellis object itself rather than change the settings globally. See also the lattice.options argument described above for a similar treatment of non-graphical options.


A list containing possible arguments to plot.trellis, which will be used by the plot or print methods when drawing the object, unless overridden explicitly. This enables the user to attach such arguments to the trellis object itself. Partial matching is not performed.


Deepayan Sarkar


The high-level functions documented here, as well as other high-level Lattice functions, are generic, with the formula method usually doing the most substantial work. The structure of the plot that is produced is mostly controlled by the formula (implicitly in the case of the non-formula methods). For each unique combination of the levels of the conditioning variables g1, g2, ..., a separate “packet” is produced, consisting of the points (x,y) for the subset of the data defined by that combination. The display can be thought of as a three-dimensional array of panels, consisting of one two-dimensional matrix per page. The dimensions of this array are determined by the layout argument. If there are no conditioning variables, the plot produced consists of a single packet. Each packet usually corresponds to one panel, but this is not strictly necessary (see the entry for index.cond above).

The coordinate system used by lattice by default is like a graph, with the origin at the bottom left, with axes increasing to the right and top. In particular, panels are by default drawn starting from the bottom left corner, going right and then up, unless as.table = TRUE, in which case panels are drawn from the top left corner, going right and then down. It is possible to set a global preference for the table-like arrangement by changing the default to as.table=TRUE; this can be done by setting lattice.options(default.args = list(as.table = TRUE)). Default values can be set in this manner for the following arguments: as.table, aspect, between, page, main, sub, par.strip.text, layout, skip and strip. Note that these global defaults are sometimes overridden by individual functions.

The order of the panels depends on the order in which the conditioning variables are specified, with g1 varying fastest, followed by g2, and so on. Within a conditioning variable, the order depends on the order of the levels (which for factors is usually in alphabetical order). Both of these orders can be modified using the index.cond and perm.cond arguments, possibly using the update (and other related) method(s).


Sarkar, Deepayan (2008) Lattice: Multivariate Data Visualization with R, Springer.

See Also

Lattice for an overview of the package, as well as barchart.table, print.trellis, shingle, banking, reshape, panel.xyplot, panel.bwplot, panel.barchart, panel.dotplot, panel.stripplot, panel.superpose, panel.loess, panel.average, strip.default, simpleKey trellis.par.set


Run this code

## Tonga Trench Earthquakes

Depth <- equal.count(quakes$depth, number=8, overlap=.1)
xyplot(lat ~ long | Depth, data = quakes)
       strip = strip.custom(strip.names = TRUE, strip.levels = TRUE),
       par.strip.text = list(cex = 0.75),
       aspect = "iso")

## Extended formula interface 

xyplot(Sepal.Length + Sepal.Width ~ Petal.Length + Petal.Width | Species,
       data = iris, scales = "free", layout = c(2, 2),
       auto.key = list(x = .75, y = .75, corner = c(0.5, 0.5)))

## user defined panel functions

states <- data.frame(state.x77,
            = dimnames(state.x77)[[1]],
                     state.region = state.region)
xyplot(Murder ~ Population | state.region, data = states,
       snames = states$,
       panel = function(x, y, subscripts, snames) {
           panel.text(x = x, y = y, labels = snames[subscripts], cex = 1,
                      fontfamily = "HersheySans")

## Stacked bar chart

barchart(yield ~ variety | site, data = barley,
         groups = year, layout = c(1,6), stack = TRUE,
         auto.key = list(space = "right"),
         ylab = "Barley Yield (bushels/acre)",
         scales = list(x = list(rot = 45)))

bwplot(voice.part ~ height, data = singer, xlab = "Height (inches)")

dotplot(variety ~ yield | year * site, data=barley)

## Grouped dot plot showing anomaly at Morris

dotplot(variety ~ yield | site, data = barley, groups = year,
        key = simpleKey(levels(barley$year), space = "right"),
        xlab = "Barley Yield (bushels/acre) ",
        aspect=0.5, layout = c(1,6), ylab=NULL)

stripplot(voice.part ~ jitter(height), data = singer, aspect = 1,
 = TRUE, xlab = "Height (inches)")

## Interaction Plot

xyplot(decrease ~ treatment, OrchardSprays, groups = rowpos,
       type = "a",
       auto.key =
       list(space = "right", points = FALSE, lines = TRUE))

## longer version with no x-ticks

if (FALSE) {
bwplot(decrease ~ treatment, OrchardSprays, groups = rowpos,
       panel = "panel.superpose",
       panel.groups = "panel.linejoin",
       xlab = "treatment",
       key = list(lines = Rows(trellis.par.get("superpose.line"),
                  c(1:7, 1)),
                  text = list(lab = as.character(unique(OrchardSprays$rowpos))),
                  columns = 4, title = "Row position"))

Run the code above in your browser using DataLab