Overcoming Multicollinearity in Marketing Mix Models

Marketing mix models (MMMs for short) are a way for companies to figure out how their marketing affects their sales. They look at how much a company spends on different kinds of ads, like TV commercials, online ads, billboards and stuff like that. Then they use math to see which types of marketing make people buy more of the company’s products.

MMMs can tell a business if they should put more money into certain kinds of ads to boost their sales. It’s a handy tool to have. But sometimes the math gets messy and confusing. One big problem is called multicollinearity. It makes the models harder to understand.

The multicollinearity problem

Multicollinearity happens when the different things you’re looking at in your MMM are too similar to each other. Let’s say a company runs TV ads and YouTube ads at the same time. The number of TV ads and YouTube ads they run each week might be really close. When two things are that alike, the model has a tough time telling them apart.

Think of it like identical twins. If they wore the exact same clothes every day, it’d be hard to tell who was who. That’s kind of what multicollinearity does to marketing mix models. It makes it hard to figure out which type of ad is really driving sales.

Why multicollinearity messes up MMMs

When you have multicollinearity in a marketing mix model, a few things can go wrong:

The model might think one kind of ad is way more important than it really is. Or it could think an ad type doesn’t matter at all, even though it does. You could end up putting too much or too little money in the wrong places. Not good.

Sometimes the model just breaks. It can’t handle all the overlapping data, so it spits out crazy numbers that make no sense. Definitely not what you want when you’re trying to make big budget choices.

Even if the model’s predictions are okay, multicollinearity makes it hard to trust what it’s saying. You can’t be sure if a certain type of ad is actually working, or if it just looks that way because it’s so similar to another ad type.

Dealing with multicollinearity

Multicollinearity is tricky, but there are some ways to handle it. Here are a few of the big ones:

Combine the troublemakers

Let’s go back to that TV ad and YouTube ad example. If those are the variables messing up your model, just smoosh them together into one mega-variable called “video advertising”. Problem solved! Well, sort of. You won’t know if TV or YouTube is doing better, but at least your model won’t be freaking out anymore.

Take stuff out

Go through your marketing mix model and get rid of anything that looks too similar to something else. Just toss it out. Yeah, you’re losing some data, but it’s better than keeping variables that are going to turn your model into a confused mess.

Fancier math

If you want to get really fancy with it, there are some big brain math tricks that can help with multicollinearity. Stuff like regularization or partial least squares regression. They can balance out your model without making you delete data. But you’ll probably need to hire a math genius to figure it out.

Is it worth the trouble?

Dealing with multicollinearity can be a big pain. It takes time and effort to fix. Plus, if you have to combine or remove variables, you might not get as much detail from your model.

You could argue it’s better to just live with a little multicollinearity sometimes. As long as the model’s predictions are still solid, who cares if the variables are hugging? Might not be worth driving yourself nuts trying to untangle them.

But here’s the thing. If your MMM has a bad case of multicollinearity, it’s hard to trust anything it says – even the overall predictions. The whole model is built on shaky ground.

Say the model tells you to dump a ton of cash into TV and radio ads. You do it, but sales don’t jump like you expected. Turns out TV and radio were so intertwined that the model couldn’t really tell what was working. Now you’ve blown your budget on the wrong things. Ouch.

Keeping your models clean

The best way to handle multicollinearity is to stop it before it starts. When you’re building an MMM, really think about which marketing variables to include. Are any of them too similar? Could you use a broader category instead of two specific ones?

Also, take a close look at your data. Plot it out and see if any variables are moving in lockstep. Tons of overlap is a red flag that multicollinearity is lurking.

If you spot potential problems early, you can rejigger your model before it’s too late. Swap in different variables, combine a few, or just skip the ones that are more trouble than they’re worth. Slapping together an MMM and hoping for the best is a recipe for disaster.

Multicollinearity is manageable

The reality is, multicollinearity will probably always be a thorn in the side of marketing mix modeling. There’s no perfect solution.

But don’t let that scare you off! MMMs are still super valuable. As long as you’re smart about how you build your model and interpret the results, multicollinearity won’t steer you wrong. Stay vigilant, use the fancy math when you need to, and don’t be afraid to simplify.

With a sharp eye and some elbow grease, you can keep multicollinearity from turning your marketing mix model into a muddled mess. It’s worth the effort to get clear, trustworthy insights into your marketing magic. Your bottom line will thank you.