For those who want to use LMM to construct a full fixed model and refit random model in a full-to-null order.
Note
ezLMM currently supports auto reports on simple effect analysis of 2-way interaction.
If you ever find significance in 3-way or more interactions, the report should come up with a blank space:
[Not yet ready for simple simple effect analysis (interaction item). Construct individual models by subsetting your data.]
For example, if you find a significant 3-way interaction with a model of 3 fixed factors (like
2 * 2 * 2), you then have 6 models to run:
a1 -> b * c, a2 -> b * c, b1 -> a * c, b2 -> a * c, c1 -> a * b, c2 -> a * b.
Subset your data, construct new models on each of them, and replace the "[XXX]" with individual results.
Make sure you have downloaded R interpreter in your system and add it to PATH.
pip install ezlmmTip
Example: LMM on Reaction Time
from ezlmm import LinearMixedModel, GeneralizedLinearMixedModel
lmm = LinearMixedModel()
# glmm = GeneralizedLinearMixedModel()
lmm.read_data("data_path.csv") # Change with your own data path
# Dependant variable
lmm.dep_var = "rt"
# Independent variable(s)
lmm.indep_var = ["syllable_number", "priming_effect", "speech_isochrony"]
# Random factor(s)
lmm.random_var = ["subject", "word"]
# [Optional]
# --------------------
lmm.exclude_trial_SD(target="rt", subject="subject", SD=2.5)
lmm.data = lmm.data[lmm.data['acc'] == 1]
# Descriptive statistics
descriptive_stats = lmm.descriptive_stats()
print(descriptive_stats)
# Code your variables right before fitting
lmm.code_variables({
"syllable_number": {"disyllabic": -0.5, "trisyllabic": 0.5},
"speech_isochrony": {"averaged": -0.5, "original": 0.5},
"priming_effect": {"inconsistent": -0.5, "consistent": 0.5}
})
# --------------------
# Fitting the model until it is converged
lmm.fit()
# Print output
print(lmm.report) # plain text for your paper
print(lmm.summary) # model fitting
print(lmm.anova) # anova on fixed model- ezlmm executes from full model to null model for random model.
- ezlmm operates on full fixed model and won't change your fixed model. (i.e., It stays the same.)
- ezlmm currently considers random intercept as default for each random effect.
In other words,
- ezlmm goes from
dv ~ iv1 * iv2 + (1+iv1:iv2+iv1+iv2|rf)todv ~ iv1 * iv2 + (1|rf). - It only stops and generates output until (1) it meets a model that fits, or (2) not a single model fits.
# In my research area:
# continuous variables (e.g., reaction time) use lmm
# categorical variables use glmm(family="possion")
# binomial variables use glmm(family="binomial") * default
from ezlmm import LinearMixedModel, GeneralizedLinearMixedModelFor linear mixed model,
lmm = LinearMixedModel()For generalized linear mixed model,
glmm = GeneralizedLinearMixedModel()lmm.read_data("data.csv") # Change with your own data pathI recommend you use the exact var name you want to use in your paper.
More specifically, use "_" (underscore) to replace " " (space) in column name and ezlmm will help you replace it back in the final report.
# Dependant variable
lmm.dep_var = "rt"
# Independent variable(s)
lmm.indep_var = ["syllable_number", "priming_effect", "speech_isochrony"]
# Random factor(s)
lmm.random_var = ["subject", "word"]lmm.exclude_trial_SD(target="rt", subject="subject", SD=2.5)I use pandas here, so you could edit it in the pandas way
lmm.data = lmm.data[lmm.data['acc'] == 1]Subsetting your data would be the same here. For example,
lmm.data = lmm.data[lmm.data['syllable'] == "disyllabic"] # df = df[df[colname] == value]descriptive_stats = lmm.descriptive_stats()
print(descriptive_stats)The structure should be: var_name: {lvl_name: lvl_name_coded}
lmm.code_variables({
"syllable_number": {"disyllabic": -0.5, "trisyllabic": 0.5},
"speech_isochrony": {"averaged": -0.5, "original": 0.5},
"priming_effect": {"inconsistent": -0.5, "consistent": 0.5}
})In case you do not want to code these variables, use this instead:
lmm.code_variables() # or lmm.code_variables({})To start from the full model, use this
# Fitting the model until it is converged
lmm.fit()In case you want to use optimizer, add optimizer parameter (I have only tested "bobyqa" yet XD)
# give it a list, i.e., [optimizer name, iteration times]
lmm.fit(optimizer=["bobyqa", 20000])Running the first few formulas can take quite long time. To start from the midpoint,
# This is just an example, type in any formula you want it to start from
lmm.fit(prev_formula="rt ~ syllable_number * priming_effect * speech_isochrony + (1 + speech_isochrony | subject) + (1 | word)")ezlmm currently does not support you add an extra variable that does not appear in lmm.indep_var to random model.
I add a function that can generate report, which you can simply copy-paste to your paper.
print(lmm.report)To see the model fitting results, and F-test like anova results on fixed model,
# model fitting
print(lmm.summary)
# anova on fixed model
print(lmm.anova)You see, my procedure was to start from full random model to null random model, and discard random slope with the least variance. It is not a big deal if you have only 1 or 2 independent variables, but not for 3 and above. You can easily misjudge which item has the least variance. It is annoying and tiring. That is the main reason.
One more thing is that, R language was simply not my thing. But still, I think the best way you can do LMM analysis is to
use R packages like lme4, lmerTest, etc. Therefore, I created ezlmm, so that you can enjoy the convenience of Python, and
at the same time cite these R packages. ezlmm is just a Python interface, it's not a big deal; you should credit them for their great work.
There was one day I was having a talk with my gf about independent T test and paired T test. I came to understand that, although paired T test can take between-subject difference into consideration, it needs to average a lot of trials, which might not be a good thing; for independent T test, it preserves the integrity of your data, but fails to take between-subject difference into account.
And, same thing for repeated-measures ANOVA and regular ANOVA: rm ANOVA needs averaging on each condition; regular ANOVA fails to consider the effect of repeated measures.
Therefore, for my understanding, I believe LMM is the combination of these two advantages:
- not averaging your data
- taking the consideration of between-subject/item/XXXX differences (depending on what's your random factors).
ezlmm uses numpy, pandas and rpy2 in Python, and lme4, lmerTest, stats, car, and emmeans from R. It's better that you cite these packages. If you wish to cite ezlmm, feel free to link this repo to your paper.
Here is description about R packages used in ezlmm:
For LinearMixedModel(), ezlmm first used lmerTest package (Kuznetsova et al., 2017) to construct a linear mixed-effects model (LMM, estimated using REML). Subsequently, an F test was performed on the optimal model using anova function from stats package (Team, 2024).
For GeneralizedLinearMixedModel(), ezlmm first used lme4 package (Bates et al., 2015) to construct a generalized linear mixed model (GLMM), following a Chi-squared test using Anova function from the car package (Fox & Weisberg, 2019).
If a significant main effect or interaction was found, post-hoc and simple effects analyses would be conducted leveraging emmeans package (Lenth, 2024).
I am a Psychology/NeuroScience student currently doing my Master program. My area is Psycholinguistics. That is to say, I am not a professional in terms of neither statistics nor machine learning. I wrote a lot of scripts for my own research as well as my colleagues'. So this is kind of my first Python package (lol I don't even quite understand how to use GitHub's pull request)
Feel free to reach out to me at paradeluxe3726@gmail.com or zhengyuan.liu@connect.um.edu.mo!