sjp.lmer: Plot estimates, predictions or effects of linear mixed effects models

Description

By default, this function plots estimates (coefficients) with confidence intervalls of either fixed effects or random effects of linear mixed effects models (that have been fitted with the lmer-function of the lme4-package). Furhermore, this function also plot predicted values or diagnostic plots.

Usage

sjp.lmer(fit, type = "re", vars = NULL, ri.nr = NULL,
  group.estimates = NULL, remove.estimates = NULL, emph.grp = NULL,
  sample.n = NULL, poly.term = NULL, sort.est = NULL, title = NULL,
  legend.title = NULL, axis.labels = NULL, axis.title = NULL,
  geom.size = NULL, geom.colors = "Set1", show.values = TRUE,
  show.p = TRUE, show.ci = FALSE, show.legend = FALSE,
  show.loess = FALSE, show.loess.ci = FALSE, show.intercept = FALSE,
  string.interc = "(Intercept)", p.kr = TRUE, show.scatter = TRUE,
  point.alpha = 0.2, point.color = NULL, jitter.ci = FALSE,
  fade.ns = FALSE, axis.lim = NULL, digits = 2, vline.type = 2,
  vline.color = "grey70", facet.grid = TRUE, free.scale = FALSE,
  y.offset = 0.1, prnt.plot = TRUE, ...)

Arguments

fit

a fitted model as returned by the lmer-function.

type

type of plot. Use one of following:

"re": (default) for conditional modes of random effects as forest plot
"fe": for estimates of fixed effects as forest plot
"fe.std": for standardized estimates of fixed effects as forest plot
"fe.slope": to plot regression lines (slopes) with confidence intervals for each single fixed effect, i.e. all fixed terms are extracted and each is plotted against the response variable (linear relationship between each fixed term and response)
"fe.resid": to plot regression lines (slopes) with confidence intervals for each single fixed effect (against residuals), i.e. all fixed terms are extracted and each is plotted against the model residuals (linear relationship between each fixed term and residuals)
"fe.cor": for correlation matrix of fixed effects
"re.qq": for a QQ-plot of random effects (random effects quantiles against standard normal quantiles)
"ri.slope": for fixed effects slopes depending on the random intercept.
"rs.ri": for fitted regression lines indicating the random slope-intercept pairs. Use this to visualize the random parts of random slope-intercept (or repeated measure) models. When having too many groups, use sample.n argument.
"coef": for joint (sum of) random and fixed effects coefficients for each explanatory variable for each level of each grouping factor as forest plot.
"pred": to plot predicted values for the response, related to specific model predictors and conditioned on random effects. See 'Details'.
"pred.fe": to plot predicted values for the response, related to specific model predictors and conditioned on fixed effects only. See 'Details'.
"eff": to plot marginal effects of all fixed terms in fit. Note that interaction terms are excluded from this plot; use sjp.int to plot effects of interaction terms. See also 'Details' of sjp.lm.
"eff.ri": to plot marginal effects of all fixed terms in fit, varying by the random intercepts.
"poly": to plot predicted values (marginal effects) of polynomial terms in fit. Use poly.term to specify the polynomial term in the fitted model (see 'Examples' here and 'Details' of sjp.lm).
"ma": to check model assumptions. Note that no further arguments except fit are relevant for this option. All other arguments are ignored.

vars

Numeric vector with column indices of selected variables or a character vector with variable names of selected variables from the fitted model, which should be used to plot - depending on type - estimates, fixed effects slopes or predicted values (mean, probabilities, incidents rates, ...). See 'Examples'.

ri.nr

Numeric vector. If type = "re" or type = "ri.slope", and fitted model has more than one random intercept, ri.nr indicates which random effects of which random intercept (or: which list elements of ranef) will be plotted. Default is NULL, so all random effects will be plotted.

group.estimates

Numeric or character vector, indicating a group identifier for each estimate. Dots and confidence intervals of estimates are coloured according to their group association. See 'Examples'.

remove.estimates

Character vector with coefficient names that indicate which estimates should be removed from the plot. remove.estimates = "est_name" would remove the estimate est_name. Default is NULL, i.e. all estimates are printed.

emph.grp

Numeric vector with index numbers of grouping levels (from random effect). If type = "ri.slope" and facet.grid = FALSE, an integrated plot of predicted probabilities of fixed effects resp. fixed effects slopes for each grouping level is plotted. To better find certain groups, use this argument to emphasize these groups in the plot. See 'Examples'.

sample.n

Numeric vector. only applies, if type = "rs.ri". If plot has many random intercepts (grouping levels), overplotting of regression lines may occur. In this case, consider random sampling of grouping levels. If sample.n is of length 1, a random sample of sample.n observation is selected to plot random intercepts. If sample.n is of length > 1, random effects indicated by the values in sample.n are selected to plot random effects. Use the latter option to always select a fixed, identical set of random effects for plotting (useful when ecomparing multiple models).

poly.term

name of a polynomial term in fit as string. Needs to be specified, if type = "poly", in order to plot marginal effects for polynomial terms. See 'Examples'.

sort.est

Determines in which way estimates are sorted in the plot:

If NULL (default), no sorting is done and estimates are sorted in order of model coefficients.
If sort.est = "sort.all", estimates are re-sorted for each coefficient (only applies if type = "re" and facet.grid = FALSE), i.e. the estimates of the random effects for each predictor are sorted and plotted to an own plot.
If type = "fe" or type = "fe.std", TRUE will sort estimates
If type = "re", specify a predictor's / coefficient's name to sort estimates according to this coefficient.

See 'Examples'.

title

Character vector with one or more labels that are used as plot title.

legend.title

Character vector, used as title for the plot legend. Note that only some plot types have legends (e.g. type = "pred" or when grouping estimates with group.estimates).

axis.labels

Character vector with labels for the model terms, used as axis labels. For mixed models, should either be vector of fixed effects variable labels (if type = "fe" or type = "fe.std") or a vector of group (value) labels from the random intercept's categories (if type = "re").

axis.title

Character vector of length one or two (depending on the plot function and type), used as title(s) for the x and y axis. If not specified, a default labelling is chosen. To set multiple axis titles (e.g. with type = "eff" for many predictors), axis.title must be a character vector of same length of plots that are printed. In this case, each plot gets an own axis title (applying, for instance, to the y-axis for type = "eff"). Note: Some plot types do not support this argument. In such cases, use the return value and add axis titles manually with labs, e.g.: $plot.list[[1]] + labs(x = ...)

geom.size

size resp. width of the geoms (bar width, line thickness or point size, depending on plot type and function). Note that bar and bin widths mostly need smaller values than dot sizes.

geom.colors

User defined color palette for geoms. If group.estimates is not specified, must either be vector with two color values or a specific color palette code (see 'Details' in sjp.grpfrq). Else, if group.estimates is specified, geom.colors must be a vector of same length as groups. See 'Examples'.

show.values

logical, whether values should be plotted or not.

show.p

Logical, adds significance levels to values, or value and variable labels.

show.ci

Logical, if TRUE, depending on type, a confidence interval or region is added to the plot. For frequency plots, the confidence interval for the relative frequencies are shown.

show.legend

logical, if TRUE, and depending on plot type and function, a legend is added to the plot.

show.loess

logical, if TRUE, and depending on type, an additional loess-smoothed line is plotted.

show.loess.ci

logical, if TRUE, a confidence region for the loess-smoothed line will be plotted. Default is FALSE. Only applies, if show.loess = TRUE (and for sjp.lmer, only applies if type = "fe.slope" or type = "fe.resid").

show.intercept

Logical, if TRUE, the intercept of the fitted model is also plotted. Default is FALSE. For glm's, please note that due to exponential transformation of estimates, the intercept in some cases can not be calculated, thus the function call is interrupted and no plot printed.

string.interc

String, axis label of intercept estimate. Only applies, if show.intercept = TRUE and axis.labels is not NULL.

p.kr

logical, if TRUE, p-value estimation is based on conditional F-tests with Kenward-Roger approximation for the df. Caution: Computation may take very long time for large samples!

show.scatter

Logical, if TRUE (default), adds a scatter plot of data points to the plot. Only applies for slope-type or predictions plots. For most plot types, dots are jittered to avoid overplotting, hence the points don't reflect exact values in the data.

point.alpha

Alpha value of point-geoms in the scatter plots. Only applies, if show.scatter = TRUE.

point.color

Color of of point-geoms in the scatter plots. Only applies, if show.scatter = TRUE.

jitter.ci

Logical, if TRUE and show.ci = TRUE and confidence bands are displayed as error bars, adds jittering to lines and error bars to avoid overlapping.

fade.ns

Logical, if TRUE, non significant estimates will be printed in slightly faded colors.

axis.lim

Numeric vector of length 2, defining the range of the plot axis. Depending on plot type, may effect either x- or y-axis, or both. For multiple plot outputs (e.g., from type = "eff" or type = "slope" in sjp.glm), axis.lim may also be a list of vectors of length 2, defining axis limits for each plot (only if non-faceted).

digits

Numeric, amount of digits after decimal point when rounding estimates and values.

vline.type

Linetype of the vertical "zero point" line. Default is 2 (dashed line).

vline.color

Color of the vertical "zero point" line. Default value is "grey70".

facet.grid

TRUE to arrange the lay out of of multiple plots in a grid of an integrated single plot. This argument calls facet_wrap or facet_grid to arrange plots. Use plot_grid to plot multiple plot-objects as an arranged grid with grid.arrange.

free.scale

Logical, if TRUE and facet.grid = TRUE, each facet grid gets its own fitted scale. If free.scale = FALSE, each facet in the grid has the same scale range.

y.offset

numeric, offset for text labels when their alignment is adjusted to the top/bottom of the geom (see hjust and vjust).

prnt.plot

logical, if TRUE (default), plots the results as graph. Use FALSE if you don't want to plot any graphs. In either case, the ggplot-object will be returned as value.

...

Other arguments passed down to further functions. Currently, following arguments are supported:

?effects::effect: Any arguments accepted by the effect resp. allEffects function, for type = "eff".
width: The width-argument for error bars.
alpha: The alpha-argument for confidence bands.
level: The level-argument confidence bands.

Value

(Insisibily) returns

the ggplot-object (plot), if type = "fe" or if type = "re" and facet.grid = TRUE). Multiple plots (type = "re" and if facet.grid = FALSE) are returned in the object plot.list.
a list of ggplot-objects (plot.list). see plot for details.
a data frame data with the data used to build the ggplot-object(s).

Details

type = "re": plots the conditional modes of the random effects, inclduing predicion intervals. It basically does the same as dotplot(ranef(fit, condVar = TRUE)[[i]]), where i denotes the random effect index.
type = "fe.slope": plots the linear relationship between each fixed effect and the response. The regression lines are not based on the fitted model's fixed effects estimates (though they may be similar). This plot type just computes a simple linear model for each fixed effect and response. Hence, it's intended for checking model assumptions, i.e. if predictor and respone are in a linear relationship. You may use the show.loess argument to see whether the linear line differs from the best fitting line.
type = "fe.resid": Similar to type = "fe.slope", this this type is intended for checking model assumptions. However, fitted values are plotted against the residuals instead of response.
type = "eff": plots the adjusted (marginal) effects for each fixed effect, with all co-variates set to the mean, as returned by the allEffects function. You can pass further arguments down to allEffects for flexible function call via the ...-argument.
type = "eff.ri": plots the adjusted (marginal) effects for each fixed effect, with all co-variates set to the mean, varying by the random intercepts. This plot type basically does the same as type = "ri.slope", except that the co-variates are not set to zero, but adjusted for. This plot type differs from type = "ri.slope" only in the adjusted y-axis-scale
type = "rs.ri": plots regression lines for the random parts of the model, i.e. all random slopes for each random intercept. As the random intercepts describe the deviation from the global intercept, the regression lines are computed as global intercept + random intercept + random slope. In case of overplotting, use the sample.n argument to randomly sample a limited amount of groups.
type = "ri.slope": plots regression lines for each fixed effect (slopes) within each random intercept. Lines are based on the fixed effects intercept, plus each random intercept and each specific fixed term's estimate. All other fixed effects are set to zero (i.e. ignored), which corresponds to b0 + b0[r1-rn] + bi * xi) (where xi is the estimate of fixed effects, b0 is the intercept of the fixed effects and b0[r1-rn] are all random intercepts).
type = "coef": forest plot of joint fixed and random effect coefficients, as retrieved by coef.merMod, it's simply ranef + fixef.
type = "pred" or type = "pred.fe": predicted values for response, conditional on fixed effects only or on random intercept. It's calling predict(fit, type = "response", re.form = NA) resp. predict(fit, type = "response", re.form = NULL) to compute the values. This plot type requires the vars argument to select specific terms that should be used for the x-axis and - optional - as grouping factor. Hence, vars must be a character vector with the names of one or two model predictors. See 'Examples'.

Examples

Run this code

# NOT RUN {
# fit model
library(lme4)
fit <- lmer(Reaction ~ Days + (Days | Subject), sleepstudy)

# simple plot
sjp.lmer(fit)

# plot fixed effects
sjp.lmer(fit, type = "fe")

# sort by predictor Days
sjp.lmer(fit, sort.est = "Days")

# plot each predictor as own plot
# sort each plot
sjp.lmer(fit, facet.grid = FALSE, sort.est = "sort.all")

library(sjmisc)
library(sjlabelled)
data(efc)
# prepare group variable
efc$grp = as.factor(efc$e15relat)
levels(x = efc$grp) <- get_labels(efc$e15relat)
# data frame for fitted model
mydf <- data.frame(neg_c_7 = as.numeric(efc$neg_c_7),
                   sex = as.factor(efc$c161sex),
                   c12hour = as.numeric(efc$c12hour),
                   barthel = as.numeric(efc$barthtot),
                   grp = efc$grp)
# fit lmer
fit <- lmer(neg_c_7 ~ sex + c12hour + barthel + (1|grp), data = mydf)

sjp.lmer(fit, type = "fe.std", sort.est = TRUE)

# highlight specific grouping levels, in this case we compare
# spouses, children and children-in-law
sjp.lmer(fit, type = "ri.slope", emph.grp = c(1, 2, 4), vars = "c12hour")

# }
# NOT RUN {
# plotting polynomial terms
# check linear relation between predictors and response
sjp.lmer(fit, type = "fe.slope", show.loess = TRUE)

# "barthel" does not seem to be linear correlated to response
# try to find appropiate polynomial. Grey line (loess smoothed)
# indicates best fit. Looks like x^4 has the best fit,
# however, x^2 seems to be suitable according to p-values.
sjp.poly(fit, "barthel", 2:4, show.scatter = FALSE)

# fit new model
fit <- lmer(neg_c_7 ~ sex + c12hour + barthel +
            I(barthel^2) + (1|grp), data = mydf)

# plot marginal effects of polynomial term
sjp.lmer(fit, type = "poly", poly.term = "barthel")

# lme4 complaints about scale of polynomial term, so
# try centering this predictor
mydf$barthel_s <- sjmisc::std(mydf$barthel)

# re-fit model
fit_s <- lmer(neg_c_7 ~ sex + c12hour + barthel_s +
              I(barthel_s^2) + (1|grp), data = mydf)

# plot marginal effects of centered, scaled polynomial term
sjp.lmer(fit_s, type = "poly", poly.term = "barthel_s")

# scaling also improved p-values
sjt.lmer(fit, fit_s)

# plotting predicted values for response
# conditioned on random effects
sjp.lmer(fit, type = "pred", vars = "c12hour")

# grouped, for fixed effects only
sjp.lmer(fit, type = "pred.fe", vars = c("c12hour", "sex"))

# grouped, for fixed effects only, non-facted
sjp.lmer(fit, type = "pred.fe", vars = c("c12hour", "sex"),
         facet.grid = FALSE, show.ci = FALSE)
# }
# NOT RUN {
# }