Note
Go to the end to download the full example code
3. ANOVA tables and post-hoc comparisons¶
Note
ANOVAs and post-hoc tests are only available for Lmer
models estimated using the factors
argument of model.fit()
and rely on implementations in R
In the previous tutorial where we looked at categorical predictors, behind the scenes pymer4
was using the factor
functionality in R. This means the output of model.fit()
looks a lot like summary()
in R applied to a model with categorical predictors. But what if we want to compute an F-test across all levels of our categorical predictor?
pymer4
makes this easy to do, and makes it easy to ensure Type III sums of squares infereces are valid. It also makes it easy to follow up omnibus tests with post-hoc pairwise comparisons.
ANOVA tables and orthogonal contrasts¶
Because ANOVA is just regression, pymer4
can estimate ANOVA tables with F-results using the .anova()
method on a fitted model. This will compute a Type-III SS table given the coding scheme provided when the model was initially fit. Based on the distribution of data across factor levels and the specific coding-scheme used, this may produce invalid Type-III SS computations. For this reason the .anova()
method has a force-orthogonal=True
argument that will reparameterize and refit the model using orthogonal polynomial contrasts prior to computing an ANOVA table.
Here we first estimate a mode with dummy-coded categories and suppress the summary output of .fit()
. Then we use .anova()
to examine the F-test results.
# import basic libraries and sample data
import os
import pandas as pd
from pymer4.utils import get_resource_path
from pymer4.models import Lmer
# IV3 is a categorical predictors with 3 levels in the sample data
df = pd.read_csv(os.path.join(get_resource_path(), "sample_data.csv"))
# # We're going to fit a multi-level regression using the
# categorical predictor (IV3) which has 3 levels
model = Lmer("DV ~ IV3 + (1|Group)", data=df)
# Using dummy-coding; suppress summary output
model.fit(factors={"IV3": ["1.0", "0.5", "1.5"]}, summarize=False)
# Get ANOVA table
print(model.anova())
SS Type III Analysis of Variance Table with Satterthwaite approximated degrees of freedom:
(NOTE: Using original model contrasts, orthogonality not guaranteed)
SS MS NumDF DenomDF F-stat P-val Sig
IV3 2359.778135 1179.889067 2 515.000001 5.296284 0.005287 **
Type III SS inferences will only be valid if data are fully balanced across levels or if contrasts between levels are orthogonally coded and sum to 0. Below we tell pymer4
to respecify our contrasts to ensure this before estimating the ANOVA. pymer4
also saves the last set of contrasts used priory to forcing orthogonality.
Because the sample data is balanced across factor levels and there are not interaction terms, in this case orthogonal contrast coding doesn’t change the results.
# Get ANOVA table, but this time force orthogonality
# for valid SS III inferences
# In this case the data are balanced so nothing changes
print(model.anova(force_orthogonal=True))
SS Type III Analysis of Variance Table with Satterthwaite approximated degrees of freedom:
(NOTE: Model refit with orthogonal polynomial contrasts)
SS MS NumDF DenomDF F-stat P-val Sig
IV3 2359.778135 1179.889067 2 515.000001 5.296284 0.005287 **
# Checkout current contrast scheme (for first contrast)
# Notice how it's simply a linear contrast across levels
print(model.factors)
{'IV3': ['0.5', '1.0', '1.5']}
# Checkout previous contrast scheme
# which was a treatment contrast with 1.0
# as the reference level
print(model.factors_prev_)
{'IV3': ['1.0', '0.5', '1.5']}
Marginal estimates and post-hoc comparisons¶
pymer4
leverages the emmeans
package in order to compute marginal estimates (“cell means” in ANOVA lingo) and pair-wise comparisons of models that contain categorical terms and/or interactions. This can be performed by using the .post_hoc()
method on fitted models. Let’s see an example:
First we’ll quickly create a second categorical IV to demo with and estimate a 3x3 ANOVA to get main effects and the interaction.
# Fix the random number generator
# for reproducibility
import numpy as np
np.random.seed(10)
# Create a new categorical variable with 3 levels
df = df.assign(IV4=np.random.choice(["1", "2", "3"], size=df.shape[0]))
# Estimate model with orthogonal polynomial contrasts
model = Lmer("DV ~ IV4*IV3 + (1|Group)", data=df)
model.fit(
factors={"IV4": ["1", "2", "3"], "IV3": ["1.0", "0.5", "1.5"]},
ordered=True,
summarize=False,
)
# Get ANOVA table
# We can ignore the note in the output because
# we manually specified polynomial contrasts
print(model.anova())
SS Type III Analysis of Variance Table with Satterthwaite approximated degrees of freedom:
(NOTE: Using original model contrasts, orthogonality not guaranteed)
SS MS NumDF DenomDF F-stat P-val Sig
IV4 449.771051 224.885525 2 510.897775 1.006943 0.366058
IV3 2486.124318 1243.062159 2 508.993080 5.565910 0.004063 **
IV4:IV3 553.852530 138.463132 4 511.073624 0.619980 0.648444
Example 1¶
Compare each level of IV3 to each other level of IV3, within each level of IV4. Use default Tukey HSD p-values.
# Compute post-hoc tests
marginal_estimates, comparisons = model.post_hoc(
marginal_vars="IV3", grouping_vars="IV4"
)
# "Cell" means of the ANOVA
print(marginal_estimates)
P-values adjusted by tukey method for family of 3 estimates
IV3 IV4 Estimate 2.5_ci 97.5_ci SE DF
1 1.0 1 42.554 33.778 51.330 4.398 68.140
2 0.5 1 45.455 36.644 54.266 4.417 69.299
3 1.5 1 40.904 32.196 49.612 4.361 65.943
4 1.0 2 42.092 33.301 50.882 4.406 68.609
5 0.5 2 41.495 32.829 50.161 4.339 64.626
6 1.5 2 38.786 29.961 47.612 4.425 69.746
7 1.0 3 43.424 34.741 52.107 4.348 65.149
8 0.5 3 46.008 37.261 54.755 4.383 67.208
9 1.5 3 38.119 29.384 46.854 4.376 66.801
# Pairwise comparisons
print(comparisons)
Contrast IV4 Estimate 2.5_ci ... DF T-stat P-val Sig
1 IV31.0 - IV30.5 1 -2.901 -9.523 ... 510.016 -1.030 0.558
2 IV31.0 - IV31.5 1 1.650 -4.750 ... 510.137 0.606 0.817
3 IV30.5 - IV31.5 1 4.552 -1.951 ... 510.267 1.645 0.228
4 IV31.0 - IV30.5 2 0.596 -5.749 ... 510.249 0.221 0.973
5 IV31.0 - IV31.5 2 3.305 -3.387 ... 510.883 1.161 0.477
6 IV30.5 - IV31.5 2 2.709 -3.749 ... 510.732 0.986 0.586
7 IV31.0 - IV30.5 3 -2.584 -8.893 ... 510.213 -0.963 0.601
8 IV31.0 - IV31.5 3 5.305 -1.006 ... 510.710 1.976 0.119
9 IV30.5 - IV31.5 3 7.889 1.437 ... 510.663 2.874 0.012 *
[9 rows x 10 columns]
Example 2¶
Compare each unique IV3,IV4 “cell mean” to every other IV3,IV4 “cell mean” and used FDR correction for multiple comparisons:
# Compute post-hoc tests
marginal_estimates, comparisons = model.post_hoc(
marginal_vars=["IV3", "IV4"], p_adjust="fdr"
)
# Pairwise comparisons
print(comparisons)
P-values adjusted by fdr method for 36 comparisons
Contrast Estimate 2.5_ci ... T-stat P-val Sig
1 IV31.0 IV41 - IV30.5 IV41 -2.901 -11.957 ... -1.030 0.535
2 IV31.0 IV41 - IV31.5 IV41 1.650 -7.102 ... 0.606 0.726
3 IV31.0 IV41 - IV31.0 IV42 0.463 -8.657 ... 0.163 0.871
4 IV31.0 IV41 - IV30.5 IV42 1.059 -7.649 ... 0.391 0.835
5 IV31.0 IV41 - IV31.5 IV42 3.768 -5.364 ... 1.326 0.473
6 IV31.0 IV41 - IV31.0 IV43 -0.870 -9.659 ... -0.318 0.869
7 IV31.0 IV41 - IV30.5 IV43 -3.454 -12.306 ... -1.254 0.473
8 IV31.0 IV41 - IV31.5 IV43 4.435 -4.426 ... 1.609 0.390
9 IV30.5 IV41 - IV31.5 IV41 4.552 -4.341 ... 1.645 0.390
10 IV30.5 IV41 - IV31.0 IV42 3.364 -5.732 ... 1.189 0.493
11 IV30.5 IV41 - IV30.5 IV42 3.960 -4.883 ... 1.440 0.446
12 IV30.5 IV41 - IV31.5 IV42 6.669 -2.568 ... 2.321 0.186
13 IV30.5 IV41 - IV31.0 IV43 2.031 -6.796 ... 0.740 0.637
14 IV30.5 IV41 - IV30.5 IV43 -0.552 -9.603 ... -0.196 0.869
15 IV30.5 IV41 - IV31.5 IV43 7.336 -1.568 ... 2.648 0.118
16 IV31.5 IV41 - IV31.0 IV42 -1.188 -10.044 ... -0.431 0.827
17 IV31.5 IV41 - IV30.5 IV42 -0.591 -9.041 ... -0.225 0.869
18 IV31.5 IV41 - IV31.5 IV42 2.117 -6.937 ... 0.752 0.637
19 IV31.5 IV41 - IV31.0 IV43 -2.520 -11.037 ... -0.951 0.535
20 IV31.5 IV41 - IV30.5 IV43 -5.104 -13.818 ... -1.883 0.362
21 IV31.5 IV41 - IV31.5 IV43 2.785 -5.986 ... 1.021 0.535
22 IV31.0 IV42 - IV30.5 IV42 0.596 -8.082 ... 0.221 0.869
23 IV31.0 IV42 - IV31.5 IV42 3.305 -5.848 ... 1.161 0.493
24 IV31.0 IV42 - IV31.0 IV43 -1.333 -10.235 ... -0.481 0.811
25 IV31.0 IV42 - IV30.5 IV43 -3.916 -12.888 ... -1.403 0.446
26 IV31.0 IV42 - IV31.5 IV43 3.972 -4.883 ... 1.442 0.446
27 IV30.5 IV42 - IV31.5 IV42 2.709 -6.123 ... 0.986 0.535
28 IV30.5 IV42 - IV31.0 IV43 -1.929 -10.318 ... -0.739 0.637
29 IV30.5 IV42 - IV30.5 IV43 -4.513 -13.207 ... -1.669 0.390
30 IV30.5 IV42 - IV31.5 IV43 3.376 -5.172 ... 1.270 0.473
31 IV31.5 IV42 - IV31.0 IV43 -4.638 -13.457 ... -1.691 0.390
32 IV31.5 IV42 - IV30.5 IV43 -7.222 -16.183 ... -2.590 0.118
33 IV31.5 IV42 - IV31.5 IV43 0.667 -8.475 ... 0.235 0.869
34 IV31.0 IV43 - IV30.5 IV43 -2.584 -11.212 ... -0.963 0.535
35 IV31.0 IV43 - IV31.5 IV43 5.305 -3.325 ... 1.976 0.351
36 IV30.5 IV43 - IV31.5 IV43 7.889 -0.935 ... 2.874 0.118
[36 rows x 9 columns]
Example 3¶
For this example we’ll estimate a more complicated ANOVA with 1 continuous IV and 2 categorical IVs with 3 levels each. This is the same model as before but with IV2 thrown into the mix. Now, pairwise comparisons reflect changes in the slope of the continuous IV (IV2) between levels of the categorical IVs (IV3 and IV4).
First let’s get the ANOVA table
model = Lmer("DV ~ IV2*IV3*IV4 + (1|Group)", data=df)
# Only need to polynomial contrasts for IV3 and IV4
# because IV2 is continuous
model.fit(
factors={"IV4": ["1", "2", "3"], "IV3": ["1.0", "0.5", "1.5"]},
ordered=True,
summarize=False,
)
# Get ANOVA table
print(model.anova())
SS Type III Analysis of Variance Table with Satterthwaite approximated degrees of freedom:
(NOTE: Using original model contrasts, orthogonality not guaranteed)
SS MS NumDF ... F-stat P-val Sig
IV2 46010.245470 46010.245470 1 ... 306.765451 1.220547e-54 ***
IV3 726.318000 363.159000 2 ... 2.421301 8.984551e-02 .
IV4 143.379932 71.689966 2 ... 0.477981 6.203159e-01
IV2:IV3 613.455876 306.727938 2 ... 2.045056 1.304528e-01
IV2:IV4 4.914900 2.457450 2 ... 0.016385 9.837494e-01
IV3:IV4 92.225327 23.056332 4 ... 0.153724 9.612985e-01
IV2:IV3:IV4 368.085569 92.021392 4 ... 0.613537 6.530638e-01
[7 rows x 7 columns]
Now we can compute the pairwise difference in slopes
# Compute post-hoc tests with bonferroni correction
marginal_estimates, comparisons = model.post_hoc(
marginal_vars="IV2", grouping_vars=["IV3", "IV4"], p_adjust="bonf"
)
# Pairwise comparisons
print(comparisons)
P-values adjusted by bonf method for 3 comparisons
Contrast IV4 Estimate 2.5_ci ... DF T-stat P-val Sig
1 IV31.0 - IV30.5 1 -0.053 -0.254 ... 502.345 -0.638 1.000
2 IV31.0 - IV31.5 1 -0.131 -0.313 ... 502.494 -1.734 0.250
3 IV30.5 - IV31.5 1 -0.078 -0.278 ... 502.821 -0.933 1.000
4 IV31.0 - IV30.5 2 -0.038 -0.210 ... 501.096 -0.526 1.000
5 IV31.0 - IV31.5 2 0.002 -0.184 ... 502.745 0.031 1.000
6 IV30.5 - IV31.5 2 0.040 -0.142 ... 502.836 0.530 1.000
7 IV31.0 - IV30.5 3 -0.134 -0.329 ... 502.956 -1.646 0.301
8 IV31.0 - IV31.5 3 -0.110 -0.302 ... 502.109 -1.368 0.516
9 IV30.5 - IV31.5 3 0.024 -0.166 ... 502.538 0.304 1.000
[9 rows x 10 columns]