mediamix is deliberately narrower than the full marketing mix modelling frameworks. This page sets out what each tool is for, where they overlap, and when to reach for which. Tools change quickly; this reflects them as of 2026, and each project’s own documentation is the authority on its current features.
Other tools
Robyn (Meta, R, with a Python version in beta). An
end-to-end MMM: geometric and Weibull adstock, Hill saturation, ridge
regression, multi-objective hyperparameter search with the Python
library nevergrad (reached from R through
reticulate), calibration to experiments, and a budget
allocator. Use it when you want Meta’s full, opinionated pipeline.
Meridian (Google, Python). A Bayesian hierarchical MMM, built for geo-level data, with priors on ROI, reach and frequency support and experiment calibration. Its documentation lists Jin et al. (2017) among the papers it builds on.
PyMC-Marketing (PyMC Labs, Python). Bayesian MMM with configurable adstock and saturation, budget optimisation, and customer-lifetime-value models, built on PyMC.
ChannelAttribution (R and Python). Markov-chain attribution of any order, and heuristic models, in C++, starting from aggregated path strings.
recipes / tune (R, tidymodels). General-purpose preprocessing and hyperparameter tuning, with no media-specific steps.
Where mediamix fits
mediamix is the transform, attribution and reporting layer on its own, in R, with no compiled code and no Python:
| Need | mediamix | Full framework |
|---|---|---|
Adstock and saturation inside your own lm(),
brms, glmnet or Stan model |
Yes: plain functions on vectors | Usually inside the framework’s model |
| Carryover tuned jointly with a model penalty in tidymodels |
step_adstock() keeps state across resampling
boundaries |
Framework-specific search |
| Per-channel carryover by cross-validation in your own model | tune_carryover_joint() |
Framework-specific search |
| Marginal ROI by counterfactual simulation | marginal_roi() |
Meridian reports one, from a 1% spend increase by default; see each framework’s documentation for its definition |
| Journeys built from a raw event log | build_paths() |
Not covered by the MMM frameworks; ChannelAttribution starts from paths |
| Heuristic and Markov attribution side by side |
attribute(), attribution_spread()
|
ChannelAttribution (Markov and heuristics) |
| Uncertainty intervals |
block_bootstrap() around any model |
Posterior intervals in Meridian and PyMC-Marketing; in Robyn, bootstrapped intervals across each cluster of candidate models |
| Bayesian estimation with priors | No; use brms or Stan on mediamix transforms |
Meridian, PyMC-Marketing |
| Budget optimisation |
allocate_budget(): the split of a given total that
maximises predicted media response, keeping each channel’s flighting and
within bounds on each channel’s total; it optimises point estimates.
simulate_plan() and reallocate() test single
moves |
Robyn, Meridian, PyMC-Marketing |
| Reach and frequency, geo hierarchies | No | Meridian |
Choosing
- You want a complete, supported MMM product and are happy with its modelling choices: Robyn in R, or Meridian or PyMC-Marketing in Python.
- You already have a modelling approach, such as a regression, a
brmsmodel or a tidymodels workflow, and need correct media transforms and reporting around it: mediamix. - You have an event log and need journeys, rule-based credit and
Markov removal effects without building the paths by hand: mediamix,
with
as_channel_paths()to hand off to ChannelAttribution for higher-order chains. - You need incrementality: none of these on their own. Run an experiment, and use it to check or calibrate whichever model you choose.
Moving parameters between tools
- Geometric decay means the same thing everywhere, but mediamix
normalises the kernel by default and Robyn does not. Use
normalise = FALSEto reproduce Robyn’s adstocked series; the decay itself is unchanged. - Robyn’s Weibull
scaleis a quantile of the series length, not a number of periods.from_robyn()converts it for Robyn’s CDF form, whichadstock_weibull(type = "cdf", normalise = FALSE)then reproduces. Robyn rescales its PDF form’s weights differently, so that form has no exact counterpart inadstock_weibull(). - Hill
half_maxis in the units of the adstocked media. Robyn’sgammais a fraction of the adstocked series’ maximum (Robyn 3.12.0 and later) or a point within its range (3.7.0 to 3.11.1), andfrom_robyn()converts either given the series.step_saturation()’shalf_maxis a fraction of each column’s reference level in the training data, its maximum by default. - For
order = 1,markov_removal()andChannelAttribution::markov_model()implement the same removal-effect definition, and the package’s tests compare the two whenever ChannelAttribution is installed; small differences come from ChannelAttribution’s simulation. Export withas_channel_paths(sep = ">")and pass the sameseptomarkov_model().