P L E A S E  W A I T...

Interaction Effects in Marketing: Understanding How Organic, Paid Channels Work Towards Customer Acquisition


The Client

The client, an electronics and electrical equipment company in the US, is one of the largest printer manufacturing companies worldwide. With more than a billion dollars in revenue from the US, the client manufactures consumer/industrial machinery and accessories.


The Objective

The client was running paid and organic digital marketing campaigns on various media channels, including Google Ads, Facebook, Instagram, LinkedIn, Email, Pinterest, and YouTube. Nabler was already working with the client to measure its online advertising effectiveness.

In addition to this, the client wanted to understand interaction effect: How the paid channels, owned channels, and website engagement, interacted with each other in driving the consumer purchase.


The Challenges in Identifying the Interaction Effect

  1. The media spending on some of the digital channels was very sporadic. This posed a challenge to model interactions of these channels with other features.

  2. The client had two years of data (January 2019 to May 2021) to be aggregated at a week level (~126 weeks) as the point of sales (POS) numbers were available at weekly granularity. With only 126 data points, it was difficult for normal Media Mix Models to capture the interaction effects.

  3. The Covid-19 pandemic had significantly impacted the client’s sales, inventory, and media spend.
Data Unification Challenges:

  1. Sales data was scattered across multiple sources at various levels. This required complex joins and aggregations to stitch the data together.

  2. Mapping SKUs (POS data) to the relevant category required rules/values that were provided by the client.

  3. Website engagement and activity data, captured using Google Analytics, was extracted from Big Query. This involved writing several, complex queries to get the data to the required granularity.
Data Summary:


Nabler’s Approach to Identify the Interaction Effect

Model Dev
  1. Bayesian Network Model:
    • The data science consulting team at Nabler used a Bayesian network (BN) model to identify the interaction effect between the variables. BN is a probabilistic graphical model comprised of nodes and directed edges.

    • There are two main components involved in learning a BN – structure learning, which involves discovering the Directed Acyclic Graph(DAG) and parameter learning, which involves learning about the conditional probability distributions. E.g., Below is a sample BN with 4 nodes and 4 edges.


Why Did Nabler Choose Bayesian Network Models to Identify Interaction Effect?
  • It helps capture causal relationships. Both conditionally dependent and conditionally independent relationships between random variables can be captured using the network.

  • It inherently solved the issue of multicollinearity, which is a problem for other class of models. Multicollinearity exists whenever an independent variable is highly correlated with one or more of the other independent variables in a multiple regression model.

  • It is useful in understanding the interaction effect of variables in a dynamic system. BN models can be built using expert inputs or can be learned from data or a mix of both.

  • Constraints can be added to the network so that unrealistic connections would not be learned, and some important connections are enforced, e.g., website visits via Paid search Ads cannot happen before the media budget is spent on paid search campaigns.
Model Development:
  • The Hill Climbing algorithm along with required constraints was used to learn the structure of the network. Once the DAG structure was determined, robust regression was used to learn the parameters of the network.

Interaction Effect Observation: Graphical Network For A Product Category X

The above graph represents the interaction effect between the different variables for one of the six product categories. The above structure was learned for the category. Some of the interactions which are evident from the graph are:

  • Spending on display affected paid search spending, which in turn affected the visitors coming to the website via paid search campaigns. This affected the (Product Detail Page) PDP views which finally influenced the revenue.
  • The display media spend influenced the YouTube video views, which in turn drove organic traffic to the website. This affected the PDP views which affected the revenue.
  • The display spending directly affected the organic visitors coming to the website, this, in turn, affected the PDP views which influenced the revenue.

Although display spending didn’t have a direct impact on the revenue, it affected other nodes which in turn contributed to driving the overall revenue. It was also evident that external factors like Covid-19 cases affected the organic visitors, and pricing and promotions on products also directly affected the overall revenue.

Interpretation of interactions between the different variables in the data and the strength of these interactions
  • Nabler built a mutilated version of the BN to understand how the overall revenue changed in response to the changes in spending for that channel.

  • This process helped determine how the influence of a channel(node) propagated downstream to revenue.

E.g., The below graph shows how the effect of reducing spend on Social Media Ads propagates through the entire network, affecting all the connected nodes.


How It Helped the Client
  • The client was able to understand how marketing influence propagates through the different nodes in the network through the interaction effect. (The interaction between the various channels and the website that leads to consumer purchase)

  • The client was able to understand how media channels interact with each other to drive revenue.

  • The client was also able to understand how website activity and engagement, eventually impacts the revenue – both online & offline.


Want to learn more? Let's Talk.