# Model constructors

The LinearMixedModel type represents a linear mixed-effects model. Typically it is constructed from a Formula and an appropriate Table type, usually a DataFrame.

## Examples of linear mixed-effects model fits

For illustration, several data sets from the lme4 package for R are made available in .arrow format in this package. Often, for convenience, we will convert these to DataFrames. These data sets include the dyestuff and dyestuff2 data sets.

using DataFrames, MixedModels, StatsModels
dyestuff = MixedModels.dataset(:dyestuff)
Arrow.Table with 30 rows, 2 columns, and schema:
:batch  String
:yield  Int16
describe(DataFrame(dyestuff))
2×7 DataFrame
Rowvariablemeanminmedianmaxnmissingeltype
SymbolUnion…AnyUnion…AnyInt64DataType
1batchAF0String
2yield1527.514401530.016350Int16

### The @formula language in Julia

MixedModels.jl builds on the the Julia formula language provided by StatsModels.jl, which is similar to the formula language in R and is also based on the notation from Wilkinson and Rogers (1973). There are two ways to construct a formula in Julia. The first way is to enclose the formula expression in the @formula macro:

StatsModels.@formulaMacro
@formula(ex)

Capture and parse a formula expression as a Formula struct.

A formula is an abstract specification of a dependence between left-hand and right-hand side variables as in, e.g., a regression model. Each side specifies at a high level how tabular data is to be converted to a numerical matrix suitable for modeling. This specification looks something like Julia code, is represented as a Julia Expr, but uses special syntax. The @formula macro takes an expression like y ~ 1 + a*b, transforms it according to the formula syntax rules into a lowered form (like y ~ 1 + a + b + a&b), and constructs a Formula struct which captures the original expression, the lowered expression, and the left- and right-hand-side.

Operators that have special interpretations in this syntax are

• ~ is the formula separator, where it is a binary operator (the first argument is the left-hand side, and the second is the right-hand side.
• + concatenates variables as columns when generating a model matrix.
• & representes an interaction between two or more variables, which corresponds to a row-wise kronecker product of the individual terms (or element-wise product if all terms involved are continuous/scalar).
• * expands to all main effects and interactions: a*b is equivalent to a+b+a&b, a*b*c to a+b+c+a&b+a&c+b&c+a&b&c, etc.
• 1, 0, and -1 indicate the presence (for 1) or absence (for 0 and -1) of an intercept column.

The rules that are applied are

• The associative rule (un-nests nested calls to +, &, and *).
• The distributive rule (interactions & distribute over concatenation +).
• The * rule expands a*b to a+b+a&b (recursively).
• Subtraction is converted to addition and negation, so x-1 becomes x + -1 (applies only to subtraction of literal 1).
• Single-argument & calls are stripped, so &(x) becomes the main effect x.

The second way is to combine Terms with operators like +, &, ~, and others at "run time". This is especially useful if you wish to create a formula from a list a variable names. For instance, the following are equivalent:

@formula(y ~ 1 + a + b + a & b) == (term(:y) ~ term(1) + term(:a) + term(:b) + term(:a) & term(:b))
true

MixedModels.jl provides additional formula syntax for representing random-effects terms. Most importantly, | separates random effects and their grouping factors (as in the formula extension used by the R package lme4. Much like with the base formula language, | can be used within the @formula macro and to construct a formula programmatically:

@formula(y ~ 1 + a + b + (1 + a + b | g))
FormulaTerm
Response:
y(unknown)
Predictors:
1
a(unknown)
b(unknown)
(a,b,g)->(1 + a + b) | g
terms = sum(term(t) for t in [1, :a, :b])
group = term(:g)
response = term(:y)
response ~ terms + (terms | group)
FormulaTerm
Response:
y(unknown)
Predictors:
1
a(unknown)
b(unknown)
(1 + a + b | g)

### Models with simple, scalar random effects

A basic model with simple, scalar random effects for the levels of batch (the batch of an intermediate product, in this case) is declared and fit as

fm = @formula(yield ~ 1 + (1|batch))
fm1 = fit(MixedModel, fm, dyestuff)
Linear mixed model fit by maximum likelihood
yield ~ 1 + (1 | batch)
logLik   -2 logLik     AIC       AICc        BIC
-163.6635   327.3271   333.3271   334.2501   337.5307

Variance components:
Column    Variance Std.Dev.
batch    (Intercept)  1388.3332 37.2603
Residual              2451.2501 49.5101
Number of obs: 30; levels of grouping factors: 6

Fixed-effects parameters:
────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
────────────────────────────────────────────────
(Intercept)  1527.5     17.6946  86.33    <1e-99
────────────────────────────────────────────────

(If you are new to Julia you may find that this first fit takes an unexpectedly long time, due to Just-In-Time (JIT) compilation of the code. The subsequent calls to such functions are much faster.)

using BenchmarkTools
dyestuff2 = MixedModels.dataset(:dyestuff2)
@benchmark fit(MixedModel, $fm,$dyestuff2)
BenchmarkTools.Trial: 10000 samples with 1 evaluation.
Range (min … max):  175.802 μs …  50.019 ms  ┊ GC (min … max): 0.00% … 93.85%
Time  (median):     184.902 μs               ┊ GC (median):    0.00%
Time  (mean ± σ):   208.130 μs ± 945.135 μs  ┊ GC (mean ± σ):  8.58% ±  1.88%

▁▃▇█▆▅▃
▁▁▂▃▆███████▇▇▅▅▄▄▃▄▄▃▄▄▄▄▄▄▄▄▄▄▄▃▃▂▂▂▂▂▂▂▂▂▁▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▃
176 μs           Histogram: frequency by time          223 μs <

Memory estimate: 54.30 KiB, allocs estimate: 1024.

By default, the model is fit by maximum likelihood. To use the REML criterion instead, add the optional named argument REML=true to the call to fit

fm1reml = fit(MixedModel, fm, dyestuff, REML=true)
Linear mixed model fit by REML
yield ~ 1 + (1 | batch)
REML criterion at convergence: 319.6542768422582

Variance components:
Column    Variance Std.Dev.
batch    (Intercept)  1764.0507 42.0006
Residual              2451.2498 49.5101
Number of obs: 30; levels of grouping factors: 6

Fixed-effects parameters:
────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
────────────────────────────────────────────────
(Intercept)  1527.5     19.3834  78.80    <1e-99
────────────────────────────────────────────────

### Floating-point type in the model

The type of fm1

typeof(fm1)
LinearMixedModel{Float64}

includes the floating point type used internally for the various matrices, vectors, and scalars that represent the model. At present, this will always be Float64 because the parameter estimates are optimized using the NLopt package which calls compiled C code that only allows for optimization with respect to a Float64 parameter vector.

So in theory other floating point types, such as BigFloat or Float32, can be used to define a model but in practice only Float64 works at present.

In theory, theory and practice are the same. In practice, they aren't. – Anon

### Simple, scalar random effects

A simple, scalar random effects term in a mixed-effects model formula is of the form (1|G). All random effects terms end with |G where G is the grouping factor for the random effect. The name or, more generally the expression, G, should evaluate to a categorical array that has a distinct set of levels. The random effects are associated with the levels of the grouping factor.

A scalar random effect is, as the name implies, one scalar value for each level of the grouping factor. A simple, scalar random effects term is of the form, (1|G). It corresponds to a shift in the intercept for each level of the grouping factor.

### Models with vector-valued random effects

The sleepstudy data are observations of reaction time, reaction, on several subjects, subj, after 0 to 9 days of sleep deprivation, days. A model with random intercepts and random slopes for each subject, allowing for within-subject correlation of the slope and intercept, is fit as

sleepstudy = MixedModels.dataset(:sleepstudy)
fm2 = fit(MixedModel, @formula(reaction ~ 1 + days + (1 + days|subj)), sleepstudy)
Linear mixed model fit by maximum likelihood
reaction ~ 1 + days + (1 + days | subj)
logLik   -2 logLik     AIC       AICc        BIC
-875.9697  1751.9393  1763.9393  1764.4249  1783.0971

Variance components:
Column    Variance Std.Dev.   Corr.
subj     (Intercept)  565.51066 23.78047
days          32.68212  5.71683 +0.08
Residual              654.94145 25.59182
Number of obs: 180; levels of grouping factors: 18

Fixed-effects parameters:
──────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
──────────────────────────────────────────────────
(Intercept)  251.405      6.63226  37.91    <1e-99
days          10.4673     1.50224   6.97    <1e-11
──────────────────────────────────────────────────

### Models with multiple, scalar random-effects terms

A model for the Penicillin data incorporates random effects for the plate, and for the sample. As every sample is used on every plate these two factors are crossed.

penicillin = MixedModels.dataset(:penicillin)
fm3 = fit(MixedModel, @formula(diameter ~ 1 + (1|plate) + (1|sample)), penicillin)
Linear mixed model fit by maximum likelihood
diameter ~ 1 + (1 | plate) + (1 | sample)
logLik   -2 logLik     AIC       AICc        BIC
-166.0942   332.1883   340.1883   340.4761   352.0676

Variance components:
Column   Variance Std.Dev.
plate    (Intercept)  0.714980 0.845565
sample   (Intercept)  3.135193 1.770648
Residual              0.302426 0.549933
Number of obs: 144; levels of grouping factors: 24, 6

Fixed-effects parameters:
─────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
─────────────────────────────────────────────────
(Intercept)  22.9722    0.744596  30.85    <1e-99
─────────────────────────────────────────────────

In contrast, the cask grouping factor is nested within the batch grouping factor in the Pastes data.

pastes = DataFrame(MixedModels.dataset(:pastes))
describe(pastes)
3×7 DataFrame
Rowvariablemeanminmedianmaxnmissingeltype
SymbolUnion…AnyUnion…AnyInt64DataType
1batchAJ0String
3strength60.053354.259.366.00Float64

This can be expressed using the solidus (the "/" character) to separate grouping factors, read "cask nested within batch":

fm4a = fit(MixedModel, @formula(strength ~ 1 + (1|batch/cask)), pastes)
Linear mixed model fit by maximum likelihood
strength ~ 1 + (1 | batch) + (1 | batch & cask)
logLik   -2 logLik     AIC       AICc        BIC
-123.9972   247.9945   255.9945   256.7217   264.3718

Variance components:
Column   Variance Std.Dev.
batch & cask (Intercept)  8.433617 2.904069
batch        (Intercept)  1.199180 1.095071
Residual                  0.678002 0.823409
Number of obs: 60; levels of grouping factors: 30, 10

Fixed-effects parameters:
─────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
─────────────────────────────────────────────────
(Intercept)  60.0533    0.642136  93.52    <1e-99
─────────────────────────────────────────────────

If the levels of the inner grouping factor are unique across the levels of the outer grouping factor, then this nesting does not need to expressed explicitly in the model syntax. For example, defining sample to be the combination of batch and cask, yields a naming scheme where the nesting is apparent from the data even if not expressed in the formula. (That is, each level of sample occurs in conjunction with only one level of batch.) As such, this model is equivalent to the previous one.

pastes.sample = (string.(pastes.cask, "&",  pastes.batch))
fm4b = fit(MixedModel, @formula(strength ~ 1 + (1|sample) + (1|batch)), pastes)
Linear mixed model fit by maximum likelihood
strength ~ 1 + (1 | sample) + (1 | batch)
logLik   -2 logLik     AIC       AICc        BIC
-123.9972   247.9945   255.9945   256.7217   264.3718

Variance components:
Column   Variance Std.Dev.
sample   (Intercept)  8.433617 2.904069
batch    (Intercept)  1.199178 1.095070
Residual              0.678002 0.823409
Number of obs: 60; levels of grouping factors: 30, 10

Fixed-effects parameters:
─────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
─────────────────────────────────────────────────
(Intercept)  60.0533    0.642136  93.52    <1e-99
─────────────────────────────────────────────────

In observational studies it is common to encounter partially crossed grouping factors. For example, the InstEval data are course evaluations by students, s, of instructors, d. Additional covariates include the academic department, dept, in which the course was given and service, whether or not it was a service course.

insteval = MixedModels.dataset(:insteval)
fm5 = fit(MixedModel, @formula(y ~ 1 + service * dept + (1|s) + (1|d)), insteval)
Linear mixed model fit by maximum likelihood
y ~ 1 + service + dept + service & dept + (1 | s) + (1 | d)
logLik     -2 logLik       AIC         AICc          BIC
-118792.7767  237585.5534  237647.5534  237647.5804  237932.8763

Variance components:
Column   Variance Std.Dev.
s        (Intercept)  0.105418 0.324681
d        (Intercept)  0.258416 0.508347
Residual              1.384728 1.176745
Number of obs: 73421; levels of grouping factors: 2972, 1128

Fixed-effects parameters:
────────────────────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
────────────────────────────────────────────────────────────────
(Intercept)              3.27628      0.0793647  41.28    <1e-99
service: Y               0.0116044    0.0699321   0.17    0.8682
dept: D02               -0.0411091    0.120331   -0.34    0.7326
dept: D03                0.00967412   0.108411    0.09    0.9289
dept: D04                0.105017     0.0944964   1.11    0.2664
dept: D05                0.0828643    0.11148     0.74    0.4573
dept: D06               -0.01194      0.0978342  -0.12    0.9029
dept: D07                0.0992679    0.110598    0.90    0.3694
dept: D08                0.0575337    0.127935    0.45    0.6529
dept: D09               -0.00263182   0.107085   -0.02    0.9804
dept: D10               -0.223423     0.099838   -2.24    0.0252
dept: D11                0.0129816    0.110639    0.12    0.9066
dept: D12                0.00503824   0.0944243   0.05    0.9574
dept: D14                0.00508269   0.109041    0.05    0.9628
dept: D15               -0.0466719    0.101942   -0.46    0.6471
service: Y & dept: D02  -0.144352     0.0929527  -1.55    0.1204
service: Y & dept: D03   0.0174078    0.0927237   0.19    0.8511
service: Y & dept: D04  -0.0381262    0.0810901  -0.47    0.6382
service: Y & dept: D05   0.0596632    0.123952    0.48    0.6303
service: Y & dept: D06  -0.254044     0.080781   -3.14    0.0017
service: Y & dept: D07  -0.151634     0.11157    -1.36    0.1741
service: Y & dept: D08   0.0508942    0.112189    0.45    0.6501
service: Y & dept: D09  -0.259448     0.0899448  -2.88    0.0039
service: Y & dept: D10   0.25907      0.111369    2.33    0.0200
service: Y & dept: D11  -0.276577     0.0819621  -3.37    0.0007
service: Y & dept: D12  -0.0418489    0.0792928  -0.53    0.5977
service: Y & dept: D14  -0.256742     0.0931016  -2.76    0.0058
service: Y & dept: D15   0.24042      0.0982071   2.45    0.0144
────────────────────────────────────────────────────────────────

### Simplifying the random effect correlation structure

MixedModels.jl estimates not only the variance of the effects for each random effect level, but also the correlation between the random effects for different predictors. So, for the model of the sleepstudy data above, one of the parameters that is estimated is the correlation between each subject's random intercept (i.e., their baseline reaction time) and slope (i.e., their particular change in reaction time per day of sleep deprivation). In some cases, you may wish to simplify the random effects structure by removing these correlation parameters. This often arises when there are many random effects you want to estimate (as is common in psychological experiments with many conditions and covariates), since the number of random effects parameters increases as the square of the number of predictors, making these models difficult to estimate from limited data.

The special syntax zerocorr can be applied to individual random effects terms inside the @formula:

fm2zerocorr_fm = fit(MixedModel, @formula(reaction ~ 1 + days + zerocorr(1 + days|subj)), sleepstudy)
Linear mixed model fit by maximum likelihood
reaction ~ 1 + days + MixedModels.ZeroCorr((1 + days | subj))
logLik   -2 logLik     AIC       AICc        BIC
-876.0016  1752.0033  1762.0033  1762.3481  1777.9680

Variance components:
Column    Variance Std.Dev.   Corr.
subj     (Intercept)  584.25897 24.17145
days          33.63281  5.79938   .
Residual              653.11578 25.55613
Number of obs: 180; levels of grouping factors: 18

Fixed-effects parameters:
──────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
──────────────────────────────────────────────────
(Intercept)  251.405      6.70771  37.48    <1e-99
days          10.4673     1.51931   6.89    <1e-11
──────────────────────────────────────────────────

Alternatively, correlations between parameters can be removed by including them as separate random effects terms:

fit(MixedModel, @formula(reaction ~ 1 + days + (1|subj) + (days|subj)), sleepstudy)
Linear mixed model fit by maximum likelihood
reaction ~ 1 + days + (1 | subj) + (days | subj)
logLik   -2 logLik     AIC       AICc        BIC
-876.0016  1752.0033  1762.0033  1762.3481  1777.9680

Variance components:
Column    Variance Std.Dev.   Corr.
subj     (Intercept)  584.25897 24.17145
days          33.63281  5.79938   .
Residual              653.11578 25.55613
Number of obs: 180; levels of grouping factors: 18

Fixed-effects parameters:
──────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
──────────────────────────────────────────────────
(Intercept)  251.405      6.70771  37.48    <1e-99
days          10.4673     1.51931   6.89    <1e-11
──────────────────────────────────────────────────

Finally, for predictors that are categorical, MixedModels.jl will estimate correlations between each level. Notice the large number of correlation parameters if we treat days as a categorical variable by giving it contrasts:

fit(MixedModel, @formula(reaction ~ 1 + days + (1 + days|subj)), sleepstudy,
contrasts = Dict(:days => DummyCoding()))
Linear mixed model fit by maximum likelihood
reaction ~ 1 + days + (1 + days | subj)
logLik   -2 logLik     AIC       AICc        BIC
-805.3997  1610.7994  1742.7994  1821.0649  1953.5346

Variance components:
Column    Variance  Std.Dev.   Corr.
subj     (Intercept)   955.84564 30.91675
days: 1       497.30180 22.30026 -0.30
days: 2       916.07104 30.26667 -0.57 +0.75
days: 3      1268.58086 35.61714 -0.37 +0.72 +0.87
days: 4      1485.82531 38.54640 -0.31 +0.59 +0.67 +0.91
days: 5      2298.45189 47.94217 -0.25 +0.46 +0.45 +0.70 +0.85
days: 6      3850.28623 62.05067 -0.27 +0.30 +0.48 +0.70 +0.77 +0.75
days: 7      1805.73363 42.49392 -0.15 +0.23 +0.47 +0.50 +0.63 +0.64 +0.71
days: 8      3151.96665 56.14238 -0.20 +0.29 +0.36 +0.56 +0.73 +0.90 +0.73 +0.74
days: 9      3083.58992 55.53008 +0.06 +0.26 +0.16 +0.38 +0.59 +0.78 +0.38 +0.53 +0.85
Residual                19.07837  4.36788
Number of obs: 180; levels of grouping factors: 18

Fixed-effects parameters:
───────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
───────────────────────────────────────────────────
(Intercept)  256.652       7.35951  34.87    <1e-99
days: 1        7.84395     5.45414   1.44    0.1504
days: 2        8.71009     7.28098   1.20    0.2316
days: 3       26.3402      8.52036   3.09    0.0020
days: 4       31.9976      9.20139   3.48    0.0005
days: 5       51.8667     11.3935    4.55    <1e-05
days: 6       55.5265     14.6978    3.78    0.0002
days: 7       62.0988     10.1212    6.14    <1e-09
days: 8       79.9777     13.3127    6.01    <1e-08
days: 9       94.1994     13.1693    7.15    <1e-12
───────────────────────────────────────────────────

Separating the 1 and days random effects into separate terms removes the correlations between the intercept and the levels of days, but not between the levels themselves:

fit(MixedModel, @formula(reaction ~ 1 + days + (1|subj) + (days|subj)), sleepstudy,
contrasts = Dict(:days => DummyCoding()))
Linear mixed model fit by maximum likelihood
reaction ~ 1 + days + (1 | subj) + (days | subj)
logLik   -2 logLik     AIC       AICc        BIC
-827.7828  1655.5655  1769.5655  1823.7623  1951.5641

Variance components:
Column    Variance Std.Dev.  Corr.
subj     (Intercept)   790.1204 28.1091
days: 1         0.0000  0.0000   .
days: 2       357.5656 18.9094   .     NaN
days: 3       684.8436 26.1695   .     NaN +0.81
days: 4       949.8596 30.8198   .     NaN +0.57 +0.91
days: 5      1752.2705 41.8601   .     NaN +0.26 +0.66 +0.87
days: 6      3355.9085 57.9302   .     NaN +0.45 +0.72 +0.80 +0.76
days: 7      1538.4506 39.2231   .     NaN +0.39 +0.42 +0.59 +0.62 +0.71
days: 8      2737.2502 52.3187   .     NaN +0.22 +0.51 +0.75 +0.93 +0.72 +0.75
days: 9      2768.1150 52.6129   .     NaN -0.05 +0.28 +0.57 +0.80 +0.34 +0.52 +0.87
Residual               135.1577 11.6257
Number of obs: 180; levels of grouping factors: 18

Fixed-effects parameters:
───────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
───────────────────────────────────────────────────
(Intercept)  256.652       7.16968  35.80    <1e-99
days: 1        7.84395     3.87524   2.02    0.0430
days: 2        8.71009     5.90612   1.47    0.1403
days: 3       26.3402      7.28453   3.62    0.0003
days: 4       31.9976      8.23332   3.89    0.0001
days: 5       51.8666     10.6003    4.89    <1e-06
days: 6       55.5264     14.1936    3.91    <1e-04
days: 7       62.0988     10.0243    6.19    <1e-09
days: 8       79.9777     12.9262    6.19    <1e-09
days: 9       94.1994     12.9924    7.25    <1e-12
───────────────────────────────────────────────────

(Notice that the variance component for days: 1 is estimated as zero, so the correlations for this component are undefined and expressed as NaN, not a number.)

An alternative is to force all the levels of days as indicators using fulldummy encoding.

MixedModels.fulldummyFunction
fulldummy(term::CategoricalTerm)

Assign "contrasts" that include all indicator columns (dummy variables) and an intercept column.

This will result in an under-determined set of contrasts, which is not a problem in the random effects because of the regularization, or "shrinkage", of the conditional modes.

The interaction of fulldummy with complex random effects is subtle and complex with numerous potential edge cases. As we discover these edge cases, we will document and determine their behavior. Until such time, please check the model summary to verify that the expansion is working as you expected. If it is not, please report a use case on GitHub.

source
fit(MixedModel, @formula(reaction ~ 1 + days + (1 + fulldummy(days)|subj)), sleepstudy,
contrasts = Dict(:days => DummyCoding()))
Linear mixed model fit by maximum likelihood
reaction ~ 1 + days + (1 + days | subj)
logLik   -2 logLik     AIC       AICc        BIC
-805.3992  1610.7984  1764.7984  1882.5631  2010.6561

Variance components:
Column    Variance  Std.Dev.   Corr.
subj     (Intercept)   800.57914 28.29451
days: 0       497.56202 22.30610 -0.27
days: 1       410.15355 20.25225 -0.15 +0.45
days: 2       434.76249 20.85096 -0.37 +0.02 +0.52
days: 3       310.18945 17.61220 +0.30 -0.59 +0.13 +0.58
days: 4       290.91418 17.05621 +0.62 -0.92 -0.40 -0.17 +0.59
days: 5      1032.57421 32.13369 +0.37 -0.54 -0.29 -0.32 +0.15 +0.61
days: 6      2242.26772 47.35259 +0.26 -0.53 -0.46 -0.14 +0.31 +0.56 +0.53
days: 7      1152.26742 33.94506 +0.21 -0.10 -0.23 +0.06 -0.08 +0.17 +0.32 +0.47
days: 8      1875.70614 43.30942 +0.29 -0.41 -0.37 -0.29 +0.01 +0.46 +0.82 +0.54 +0.55
days: 9      2236.70020 47.29376 +0.44 -0.16 -0.15 -0.39 -0.18 +0.29 +0.67 +0.04 +0.31 +0.78
Residual                18.76427  4.33177
Number of obs: 180; levels of grouping factors: 18

Fixed-effects parameters:
───────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
───────────────────────────────────────────────────
(Intercept)  256.652       7.35989  34.87    <1e-99
days: 1        7.84395     5.45718   1.44    0.1506
days: 2        8.71009     7.28496   1.20    0.2318
days: 3       26.3402      8.52416   3.09    0.0020
days: 4       31.9976      9.20461   3.48    0.0005
days: 5       51.8666     11.3943    4.55    <1e-05
days: 6       55.5264     14.7049    3.78    0.0002
days: 7       62.0988     10.1251    6.13    <1e-09
days: 8       79.9777     13.3201    6.00    <1e-08
days: 9       94.1994     13.1501    7.16    <1e-12
───────────────────────────────────────────────────

This fit produces a better fit as measured by the objective (negative twice the log-likelihood is 1610.8) but at the expense of adding many more parameters to the model. As a result, model comparison criteria such, as AIC and BIC, are inflated.

But using zerocorr on the individual terms does remove the correlations between the levels:

fit(MixedModel, @formula(reaction ~ 1 + days + zerocorr(1 + days|subj)), sleepstudy,
contrasts = Dict(:days => DummyCoding()))
Linear mixed model fit by maximum likelihood
reaction ~ 1 + days + MixedModels.ZeroCorr((1 + days | subj))
logLik   -2 logLik     AIC       AICc        BIC
-882.9138  1765.8276  1807.8276  1813.6757  1874.8797

Variance components:
Column    Variance Std.Dev.  Corr.
subj     (Intercept)   958.5560 30.9606
days: 1         0.0000  0.0000   .
days: 2         0.0000  0.0000   .     .
days: 3         0.0000  0.0000   .     .     .
days: 4         0.0000  0.0000   .     .     .     .
days: 5       519.6013 22.7948   .     .     .     .     .
days: 6      1703.8595 41.2778   .     .     .     .     .     .
days: 7       608.7620 24.6731   .     .     .     .     .     .     .
days: 8      1273.1231 35.6809   .     .     .     .     .     .     .     .
days: 9      1754.0280 41.8811   .     .     .     .     .     .     .     .     .
Residual               434.8635 20.8534
Number of obs: 180; levels of grouping factors: 18

Fixed-effects parameters:
───────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
───────────────────────────────────────────────────
(Intercept)  256.652       8.79842  29.17    <1e-99
days: 1        7.84395     6.95113   1.13    0.2591
days: 2        8.71009     6.95113   1.25    0.2102
days: 3       26.3402      6.95113   3.79    0.0002
days: 4       31.9976      6.95113   4.60    <1e-05
days: 5       51.8667      8.78549   5.90    <1e-08
days: 6       55.5265     11.9573    4.64    <1e-05
days: 7       62.0988      9.06302   6.85    <1e-11
days: 8       79.9777     10.9109    7.33    <1e-12
days: 9       94.1994     12.0733    7.80    <1e-14
───────────────────────────────────────────────────
fit(MixedModel, @formula(reaction ~ 1 + days + (1|subj) + zerocorr(days|subj)), sleepstudy,
contrasts = Dict(:days => DummyCoding()))
Linear mixed model fit by maximum likelihood
reaction ~ 1 + days + (1 | subj) + MixedModels.ZeroCorr((days | subj))
logLik   -2 logLik     AIC       AICc        BIC
-882.9138  1765.8276  1807.8276  1813.6757  1874.8797

Variance components:
Column    Variance Std.Dev.  Corr.
subj     (Intercept)   958.5560 30.9606
days: 1         0.0000  0.0000   .
days: 2         0.0000  0.0000   .     .
days: 3         0.0000  0.0000   .     .     .
days: 4         0.0000  0.0000   .     .     .     .
days: 5       519.6013 22.7948   .     .     .     .     .
days: 6      1703.8595 41.2778   .     .     .     .     .     .
days: 7       608.7620 24.6731   .     .     .     .     .     .     .
days: 8      1273.1231 35.6809   .     .     .     .     .     .     .     .
days: 9      1754.0280 41.8811   .     .     .     .     .     .     .     .     .
Residual               434.8635 20.8534
Number of obs: 180; levels of grouping factors: 18

Fixed-effects parameters:
───────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
───────────────────────────────────────────────────
(Intercept)  256.652       8.79842  29.17    <1e-99
days: 1        7.84395     6.95113   1.13    0.2591
days: 2        8.71009     6.95113   1.25    0.2102
days: 3       26.3402      6.95113   3.79    0.0002
days: 4       31.9976      6.95113   4.60    <1e-05
days: 5       51.8667      8.78549   5.90    <1e-08
days: 6       55.5265     11.9573    4.64    <1e-05
days: 7       62.0988      9.06302   6.85    <1e-11
days: 8       79.9777     10.9109    7.33    <1e-12
days: 9       94.1994     12.0733    7.80    <1e-14
───────────────────────────────────────────────────
fit(MixedModel, @formula(reaction ~ 1 + days + zerocorr(1 + fulldummy(days)|subj)), sleepstudy,
contrasts = Dict(:days => DummyCoding()))
Linear mixed model fit by maximum likelihood
reaction ~ 1 + days + MixedModels.ZeroCorr((1 + days | subj))
logLik   -2 logLik     AIC       AICc        BIC
-878.9843  1757.9686  1801.9686  1808.4145  1872.2137

Variance components:
Column     Variance   Std.Dev.   Corr.
subj     (Intercept)  1134.919182 33.688562
days: 0       775.847069 27.854031   .
days: 1       357.517914 18.908144   .     .
days: 2       221.013083 14.866509   .     .     .
days: 3         0.017031  0.130503   .     .     .     .
days: 4        44.205157  6.648696   .     .     .     .     .
days: 5       670.050361 25.885331   .     .     .     .     .     .
days: 6      1739.539393 41.707786   .     .     .     .     .     .     .
days: 7       908.310283 30.138186   .     .     .     .     .     .     .     .
days: 8      1457.484304 38.177013   .     .     .     .     .     .     .     .     .
days: 9      2028.306513 45.036724   .     .     .     .     .     .     .     .     .     .
Residual               181.102557 13.457435
Number of obs: 180; levels of grouping factors: 18

Fixed-effects parameters:
───────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
───────────────────────────────────────────────────
(Intercept)  256.652      10.7803   23.81    <1e-99
days: 1        7.84395     9.11522   0.86    0.3895
days: 2        8.71009     8.68928   1.00    0.3162
days: 3       26.3402      7.95148   3.31    0.0009
days: 4       31.9976      8.10438   3.95    <1e-04
days: 5       51.8666     10.0225    5.18    <1e-06
days: 6       55.5264     12.6438    4.39    <1e-04
days: 7       62.0988     10.6624    5.82    <1e-08
days: 8       79.9777     12.0082    6.66    <1e-10
days: 9       94.1994     13.2631    7.10    <1e-11
───────────────────────────────────────────────────

## Fitting generalized linear mixed models

To create a GLMM representation, the distribution family for the response, and possibly the link function, must be specified.

verbagg = MixedModels.dataset(:verbagg)
verbaggform = @formula(r2 ~ 1 + anger + gender + btype + situ + mode + (1|subj) + (1|item));
gm1 = fit(MixedModel, verbaggform, verbagg, Bernoulli())
Generalized Linear Mixed Model fit by maximum likelihood (nAGQ = 1)
r2 ~ 1 + anger + gender + btype + situ + mode + (1 | subj) + (1 | item)
Distribution: Bernoulli{Float64}

logLik    deviance     AIC       AICc        BIC
-4067.9165  8135.8331  8153.8331  8153.8568  8216.2372

Variance components:
Column   Variance Std.Dev.
subj (Intercept)  1.793965 1.339390
item (Intercept)  0.117103 0.342203

Number of obs: 7584; levels of grouping factors: 316, 24

Fixed-effects parameters:
──────────────────────────────────────────────────────
Coef.  Std. Error       z  Pr(>|z|)
──────────────────────────────────────────────────────
(Intercept)   -0.149071    0.385246    -0.39    0.6988
anger          0.0572458   0.0167547    3.42    0.0006
gender: M      0.319978    0.19123      1.67    0.0943
btype: scold  -1.06091     0.184128    -5.76    <1e-08
btype: shout  -2.10545     0.186487   -11.29    <1e-28
situ: self    -1.0546      0.151169    -6.98    <1e-11
mode: want     0.707419    0.150982     4.69    <1e-05
──────────────────────────────────────────────────────

The canonical link, which is LogitLink for the Bernoulli distribution, is used if no explicit link is specified.

Note that, in keeping with convention in the GLM package, the distribution family for a binary (i.e. 0/1) response is the Bernoulli distribution. The Binomial distribution is only used when the response is the fraction of trials returning a positive, in which case the number of trials must be specified as the case weights.

### Optional arguments to fit

An alternative approach is to create the GeneralizedLinearMixedModel object then call fit! on it. The optional arguments fast and/or nAGQ can be passed to the optimization process via both fit and fit! (i.e these optimization settings are not used nor recognized when constructing the model).

As the name implies, fast=true, provides a faster but somewhat less accurate fit. These fits may suffice for model comparisons.

gm1a = fit(MixedModel, verbaggform, verbagg, Bernoulli(), fast = true)
deviance(gm1a) - deviance(gm1)
0.33779754808256257
@benchmark fit(MixedModel, $verbaggform,$verbagg, Bernoulli())
BenchmarkTools.Trial: 2 samples with 1 evaluation.
Range (min … max):  3.507 s …   3.523 s  ┊ GC (min … max): 0.00% … 0.00%
Time  (median):     3.515 s              ┊ GC (median):    0.00%
Time  (mean ± σ):   3.515 s ± 10.893 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

█                                                       █
█▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁█ ▁
3.51 s         Histogram: frequency by time        3.52 s <

Memory estimate: 23.91 MiB, allocs estimate: 432624.
@benchmark fit(MixedModel, $verbaggform,$verbagg, Bernoulli(), fast = true)
BenchmarkTools.Trial: 19 samples with 1 evaluation.
Range (min … max):  270.744 ms … 290.521 ms  ┊ GC (min … max): 0.00% … 6.09%
Time  (median):     275.392 ms               ┊ GC (median):    0.00%
Time  (mean ± σ):   276.271 ms ±   4.682 ms  ┊ GC (mean ± σ):  0.34% ± 1.40%

▁█  ▁ ▁ ▁ ▁ ▁▁█  ▁ ▁   ▁  ▁ ▁▁ ▁                            ▁
██▁▁█▁█▁█▁█▁███▁▁█▁█▁▁▁█▁▁█▁██▁█▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁█ ▁
271 ms           Histogram: frequency by time          291 ms <

Memory estimate: 11.62 MiB, allocs estimate: 84969.

The optional argument nAGQ=k causes evaluation of the deviance function to use a k point adaptive Gauss-Hermite quadrature rule. This method only applies to models with a single, simple, scalar random-effects term, such as

contraception = MixedModels.dataset(:contra)
contraform = @formula(use ~ 1 + age + abs2(age) + livch + urban + (1|dist));
bernoulli = Bernoulli()
deviances = Dict{Symbol,Float64}()
b = @benchmarkable deviances[:default] = deviance(fit(MixedModel, $contraform,$contraception, $bernoulli)); run(b) b = @benchmarkable deviances[:fast] = deviance(fit(MixedModel,$contraform, $contraception,$bernoulli, fast = true));
run(b)
b = @benchmarkable deviances[:nAGQ] = deviance(fit(MixedModel, $contraform,$contraception, $bernoulli, nAGQ=9)); run(b) b = @benchmarkable deviances[:nAGQ_fast] = deviance(fit(MixedModel,$contraform, $contraception,$bernoulli, nAGQ=9, fast=true));
run(b)
sort(deviances)
OrderedCollections.OrderedDict{Symbol, Float64} with 4 entries:
:default   => 2372.73
:fast      => 2372.78
:nAGQ      => 2372.46
:nAGQ_fast => 2372.51

# Extractor functions

LinearMixedModel and GeneralizedLinearMixedModel are subtypes of StatsAPI.RegressionModel which, in turn, is a subtype of StatsBase.StatisticalModel. Many of the generic extractors defined in the StatsBase package have methods for these models.

## Model-fit statistics

The statistics describing the quality of the model fit include

StatsAPI.loglikelihoodFunction
loglikelihood(model::StatisticalModel)
loglikelihood(model::StatisticalModel, observation)

Return the log-likelihood of the model.

With an observation argument, return the contribution of observation to the log-likelihood of model.

If observation is a Colon, return a vector of each observation's contribution to the log-likelihood of the model. In other words, this is the vector of the pointwise log-likelihood contributions.

In general, sum(loglikehood(model, :)) == loglikelihood(model).

StatsAPI.aicFunction
aic(model::StatisticalModel)

Akaike's Information Criterion, defined as $-2 \log L + 2k$, with $L$ the likelihood of the model, and k its number of consumed degrees of freedom (as returned by dof).

StatsAPI.bicFunction
bic(model::StatisticalModel)

Bayesian Information Criterion, defined as $-2 \log L + k \log n$, with $L$ the likelihood of the model, $k$ its number of consumed degrees of freedom (as returned by dof), and $n$ the number of observations (as returned by nobs).

StatsAPI.dofFunction
dof(model::StatisticalModel)

Return the number of degrees of freedom consumed in the model, including when applicable the intercept and the distribution's dispersion parameter.

StatsAPI.nobsFunction
nobs(model::StatisticalModel)

Return the number of independent observations on which the model was fitted. Be careful when using this information, as the definition of an independent observation may vary depending on the model, on the format used to pass the data, on the sampling plan (if specified), etc.

loglikelihood(fm1)
-163.6635299405715
aic(fm1)
333.327059881143
bic(fm1)
337.5306520261295
dof(fm1)   # 1 fixed effect, 2 variances
3
nobs(fm1)  # 30 observations
30
loglikelihood(gm1)
-4067.916537078975

In general the deviance of a statistical model fit is negative twice the log-likelihood adjusting for the saturated model.

StatsAPI.devianceMethod
deviance(m::GeneralizedLinearMixedModel{T}, nAGQ=1)::T where {T}

Return the deviance of m evaluated by the Laplace approximation (nAGQ=1) or nAGQ-point adaptive Gauss-Hermite quadrature.

If the distribution D does not have a scale parameter the Laplace approximation is the squared length of the conditional modes, $u$, plus the determinant of $Λ'Z'WZΛ + I$, plus the sum of the squared deviance residuals.

source

Because it is not clear what the saturated model corresponding to a particular LinearMixedModel should be, negative twice the log-likelihood is called the objective.

This value is also accessible as the deviance but the user should bear in mind that this doesn't have all the properties of a deviance which is corrected for the saturated model. For example, it is not necessarily non-negative.

objective(fm1)
327.327059881143
deviance(fm1)
327.327059881143

The value optimized when fitting a GeneralizedLinearMixedModel is the Laplace approximation to the deviance or an adaptive Gauss-Hermite evaluation.

MixedModels.deviance!Function
deviance!(m::GeneralizedLinearMixedModel, nAGQ=1)

Update m.η, m.μ, etc., install the working response and working weights in m.LMM, update m.LMM.A and m.LMM.R, then evaluate the deviance.

source
MixedModels.deviance!(gm1)
8135.833074157906

## Fixed-effects parameter estimates

The coef and fixef extractors both return the maximum likelihood estimates of the fixed-effects coefficients. They differ in their behavior in the rank-deficient case. The associated coefnames and fixefnames return the corresponding coefficient names.

MixedModels.fixefFunction
fixef(m::MixedModel)

Return the fixed-effects parameter vector estimate of m.

In the rank-deficient case the truncated parameter vector, of length rank(m) is returned. This is unlike coef which always returns a vector whose length matches the number of columns in X.

source
coef(fm1)
coefnames(fm1)
1-element Vector{String}:
"(Intercept)"
fixef(fm1)
fixefnames(fm1)
1-element Vector{String}:
"(Intercept)"

An alternative extractor for the fixed-effects coefficient is the β property. Properties whose names are Greek letters usually have an alternative spelling, which is the name of the Greek letter.

fm1.β
1-element Vector{Float64}:
1527.4999999999989
fm1.beta
1-element Vector{Float64}:
1527.4999999999989
gm1.β
7-element Vector{Float64}:
-0.14907072096016058
0.05724583641105297
0.3199781484791075
-1.060913337389384
-2.105448522512181
-1.0545990047936376
0.7074189332976233

A full list of property names is returned by propertynames

propertynames(fm1)
(:formula, :reterms, :Xymat, :feterm, :sqrtwts, :parmap, :dims, :A, :L, :optsum, :θ, :theta, :β, :beta, :βs, :betas, :λ, :lambda, :stderror, :σ, :sigma, :σs, :sigmas, :σρs, :sigmarhos, :b, :u, :lowerbd, :X, :y, :corr, :vcov, :PCA, :rePCA, :objective, :pvalues)
propertynames(gm1)
(:A, :L, :theta, :beta, :coef, :λ, :σ, :sigma, :X, :y, :lowerbd, :objective, :σρs, :σs, :corr, :vcov, :PCA, :rePCA, :LMM, :β, :θ, :b, :u, :resp, :wt)

The variance-covariance matrix of the fixed-effects coefficients is returned by

StatsAPI.vcovFunction
vcov(model::StatisticalModel)

Return the variance-covariance matrix for the coefficients of the model.

vcov(fm2)
2×2 Matrix{Float64}:
43.9868   -1.37039
-1.37039   2.25671
vcov(gm1)
7×7 Matrix{Float64}:
0.148414    -0.00561618  -0.00982564   …  -0.0112035    -0.0113431
-0.00561618   0.00028072   7.19224e-5      -1.47936e-5    1.02396e-5
-0.00982564   7.19224e-5   0.0365688       -8.0425e-5     5.25782e-5
-0.0167932   -1.43694e-5  -9.25592e-5       0.000265874  -0.000172156
-0.0166173   -2.90515e-5  -0.000162368      0.00065912   -0.000520662
-0.0112035   -1.47936e-5  -8.0425e-5    …   0.0228521    -0.000247839
-0.0113431    1.02396e-5   5.25782e-5      -0.000247839   0.0227956

The standard errors are the square roots of the diagonal elements of the estimated variance-covariance matrix of the fixed-effects coefficient estimators.

StatsAPI.stderrorFunction
stderror(model::StatisticalModel)

Return the standard errors for the coefficients of the model.

stderror(fm2)
2-element Vector{Float64}:
6.632257709585527
1.5022355185633753
stderror(gm1)
7-element Vector{Float64}:
0.38524589846458285
0.016754701870612242
0.1912297567875182
0.18412753302858795
0.18648723294534778
0.15116915202907324
0.15098202380617845

Finally, the coeftable generic produces a table of coefficient estimates, their standard errors, and their ratio. The p-values quoted here should be regarded as approximations.

StatsAPI.coeftableFunction
coeftable(model::StatisticalModel; level::Real=0.95)

Return a table with coefficients and related statistics of the model. level determines the level for confidence intervals (by default, 95%).

The returned CoefTable object implements the Tables.jl interface, and can be converted e.g. to a DataFrame via using DataFrames; DataFrame(coeftable(model)).

coeftable(fm2)
──────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
──────────────────────────────────────────────────
(Intercept)  251.405      6.63226  37.91    <1e-99
days          10.4673     1.50224   6.97    <1e-11
──────────────────────────────────────────────────

## Covariance parameter estimates

The covariance parameters estimates, in the form shown in the model summary, are a VarCorr object

VarCorr(fm2)
Variance components:
Column    Variance Std.Dev.   Corr.
subj     (Intercept)  565.51066 23.78047
days          32.68212  5.71683 +0.08
Residual              654.94145 25.59182

VarCorr(gm1)
Variance components:
Column   Variance Std.Dev.
subj (Intercept)  1.793965 1.339390
item (Intercept)  0.117103 0.342203



Individual components are returned by other extractors

MixedModels.varestFunction
varest(m::LinearMixedModel)

Returns the estimate of σ², the variance of the conditional distribution of Y given B.

source
varest(m::GeneralizedLinearMixedModel)

Returns the estimate of ϕ², the variance of the conditional distribution of Y given B.

For models with a dispersion parameter ϕ, this is simply ϕ². For models without a dispersion parameter, this value is missing. This differs from disperion, which returns 1 for models without a dispersion parameter.

For Gaussian models, this parameter is often called σ².

source
MixedModels.sdestFunction
sdest(m::LinearMixedModel)

Return the estimate of σ, the standard deviation of the per-observation noise.

source
sdest(m::GeneralizedLinearMixedModel)

Return the estimate of the dispersion, i.e. the standard deviation of the per-observation noise.

For models with a dispersion parameter ϕ, this is simply ϕ. For models without a dispersion parameter, this value is missing. This differs from disperion, which returns 1 for models without a dispersion parameter.

For Gaussian models, this parameter is often called σ.

source
varest(fm2)
654.9414493336943
sdest(fm2)
25.591823876654324
fm2.σ
25.591823876654324

## Conditional modes of the random effects

The ranef extractor

MixedModels.ranefFunction
ranef(m::LinearMixedModel; uscale=false)

Return, as a Vector{Matrix{T}}, the conditional modes of the random effects in model m.

If uscale is true the random effects are on the spherical (i.e. u) scale, otherwise on the original scale.

For a named variant, see raneftables.

source
ranef(fm1)
1-element Vector{Matrix{Float64}}:
[-16.628221011733622 0.36951602248394705 … 53.57982326003441 -42.49434258554293]
fm1.b
1-element Vector{Matrix{Float64}}:
[-16.628221011733622 0.36951602248394705 … 53.57982326003441 -42.49434258554293]

returns the conditional modes of the random effects given the observed data. That is, these are the values that maximize the conditional density of the random effects given the observed data. For a LinearMixedModel these are also the conditional means.

These are sometimes called the best linear unbiased predictors or BLUPs but that name is not particularly meaningful.

At a superficial level these can be considered as the "estimates" of the random effects, with a bit of hand waving, but pursuing this analogy too far usually results in confusion.

To obtain tables associating the values of the conditional modes with the levels of the grouping factor, use

MixedModels.raneftablesFunction
raneftables(m::MixedModel; uscale = false)

Return the conditional means of the random effects as a NamedTuple of columntables

source

as in

DataFrame(only(raneftables(fm1)))
6×2 DataFrame
Rowbatch(Intercept)
StringFloat64
1A-16.6282
2B0.369516
3C26.9747
4D-21.8014
5E53.5798
6F-42.4943

The corresponding conditional variances are returned by

MixedModels.condVarFunction
condVar(m::LinearMixedModel)

Return the conditional variances matrices of the random effects.

The random effects are returned by ranef as a vector of length k, where k is the number of random effects terms. The ith element is a matrix of size vᵢ × ℓᵢ where vᵢ is the size of the vector-valued random effects for each of the ℓᵢ levels of the grouping factor. Technically those values are the modes of the conditional distribution of the random effects given the observed data.

This function returns an array of k three dimensional arrays, where the ith array is of size vᵢ × vᵢ × ℓᵢ. These are the diagonal blocks from the conditional variance-covariance matrix,

s² Λ(Λ'Z'ZΛ + I)⁻¹Λ'
source
condVar(fm1)
1-element Vector{Array{Float64, 3}}:
[362.3104675622471]

[362.3104675622471]

[362.3104675622471]

[362.3104675622471]

[362.3104675622471]

[362.3104675622471]

## Case-wise diagnostics and residual degrees of freedom

The leverage values

StatsAPI.leverageFunction
leverage(model::RegressionModel)

Return the diagonal of the projection matrix of the model.

leverage(fm1)
30-element Vector{Float64}:
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
⋮
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315
0.15650534082766315

are used in diagnostics for linear regression models to determine cases that exert a strong influence on their own predicted response.

The documentation refers to a "projection". For a linear model without random effects the fitted values are obtained by orthogonal projection of the response onto the column span of the model matrix and the sum of the leverage values is the dimension of this column span. That is, the sum of the leverage values is the rank of the model matrix and n - sum(leverage(m)) is the degrees of freedom for residuals. The sum of the leverage values is also the trace of the so-called "hat" matrix, H. (The name "hat matrix" reflects the fact that $\hat{\mathbf{y}} = \mathbf{H} \mathbf{y}$. That is, H puts a hat on y.)

For a linear mixed model the sum of the leverage values will be between p, the rank of the fixed-effects model matrix, and p + q where q is the total number of random effects. This number does not represent a dimension (or "degrees of freedom") of a linear subspace of all possible fitted values because the projection is not an orthogonal projection. Nevertheless, it is a reasonable measure of the effective degrees of freedom of the model and n - sum(leverage(m)) can be considered the effective residual degrees of freedom.

For model fm1 the dimensions are

n, p, q, k = size(fm1)
(30, 1, 6, 1)

which implies that the sum of the leverage values should be in the range [1, 7]. The actual value is

sum(leverage(fm1))
4.695160224829894

For model fm2 the dimensions are

n, p, q, k = size(fm2)
(180, 2, 36, 1)

providing a range of [2, 38] for the effective degrees of freedom for the model. The observed value is

sum(leverage(fm2))
28.611526429924673

When a model converges to a singular covariance, such as

fm3 = fit(MixedModel, @formula(yield ~ 1+(1|batch)), MixedModels.dataset(:dyestuff2))
Linear mixed model fit by maximum likelihood
yield ~ 1 + (1 | batch)
logLik   -2 logLik     AIC       AICc        BIC
-81.4365   162.8730   168.8730   169.7961   173.0766

Variance components:
Column   Variance Std.Dev.
batch    (Intercept)   0.00000 0.00000
Residual              13.34610 3.65323
Number of obs: 30; levels of grouping factors: 6

Fixed-effects parameters:
───────────────────────────────────────────────
Coef.  Std. Error     z  Pr(>|z|)
───────────────────────────────────────────────
(Intercept)  5.6656    0.666986  8.49    <1e-16
───────────────────────────────────────────────

the effective degrees of freedom is the lower bound.

sum(leverage(fm3))
0.9999999999999998

Models for which the estimates of the variances of the random effects are large relative to the residual variance have effective degrees of freedom close to the upper bound.

fm4 = fit(MixedModel, @formula(diameter ~ 1+(1|plate)+(1|sample)),
MixedModels.dataset(:penicillin))
Linear mixed model fit by maximum likelihood
diameter ~ 1 + (1 | plate) + (1 | sample)
logLik   -2 logLik     AIC       AICc        BIC
-166.0942   332.1883   340.1883   340.4761   352.0676

Variance components:
Column   Variance Std.Dev.
plate    (Intercept)  0.714980 0.845565
sample   (Intercept)  3.135193 1.770648
Residual              0.302426 0.549933
Number of obs: 144; levels of grouping factors: 24, 6

Fixed-effects parameters:
─────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
─────────────────────────────────────────────────
(Intercept)  22.9722    0.744596  30.85    <1e-99
─────────────────────────────────────────────────
sum(leverage(fm4))
27.465317930118516

Also, a model fit by the REML criterion generally has larger estimates of the variance components and hence a larger effective degrees of freedom.

fm4r = fit(MixedModel, @formula(diameter ~ 1+(1|plate)+(1|sample)),
MixedModels.dataset(:penicillin), REML=true)
Linear mixed model fit by REML
diameter ~ 1 + (1 | plate) + (1 | sample)
REML criterion at convergence: 330.86058899093973

Variance components:
Column   Variance Std.Dev.
plate    (Intercept)  0.716908 0.846704
sample   (Intercept)  3.730909 1.931556
Residual              0.302415 0.549923
Number of obs: 144; levels of grouping factors: 24, 6

Fixed-effects parameters:
─────────────────────────────────────────────────
Coef.  Std. Error      z  Pr(>|z|)
─────────────────────────────────────────────────
(Intercept)  22.9722    0.808572  28.41    <1e-99
─────────────────────────────────────────────────
sum(leverage(fm4r))
27.472361780596334