lightweight_mmm icon indicating copy to clipboard operation
lightweight_mmm copied to clipboard

Can I add other kpi in the extra_features?

Open jannnnabi opened this issue 2 years ago • 1 comments
trafficstars

image

extra_features – Other variables to add to the model.

Your explanation is too simple, if my target is visit, can I put other variables like order, revenue, cart add, etc?

jannnnabi avatar Aug 16 '23 23:08 jannnnabi

All extra_features are modelled as linear predictors, with 0 centred Normal distribution Beta priors. So in theory you can model any numerical variable, allowing for both +/- linear impact on the target. However, whether you want to do that, depends on whether the 'other variables' are causational or intermediary stages between the paid advertisement and the target you ultimately want to predict. For instance, typically Paid media advertisement spend, drives people to a company's services, which leads to people to visit their site, then adding items to cart then leading to sales. All three of these arguably 'target' actions are causally downstream from the original paid advertisement. It's a little unclear from your example, but if you're trying to predict daily total website? visits from advertisement, I wouldn't include other target actions as there is confounding, they are both caused by the advertisement.

There will likely be a strong correlation between revenue and number of people visiting a company's site, but it doesn't make causal sense for the revenue to cause website visits, quite the opposite, so revenue shouldn't be a feature to predict the visits target. So typically extra features should be outside the causal path from paid advertisement to target, or it will cannibalise credit that should be going to the paid advertisement. Where we use regression in MMM, we are also trying to establish causation from association, so it's important to only use features that make sense to be causing changes in the target.

Common extra features I've seen in MMM's are typically things like price, promotions, exchange rates, competitor prices, and GDP (market performance). These impact the target, but are mutually exclusive to and not impacted by the paid advertisement itself.

becksimpson avatar Aug 17 '23 20:08 becksimpson