If you’ve looked at the example metric trees, you’ve probably noticed a few patterns. Here are two that come up most frequently and a third that is the most powerful.
Conversion rates
Conversions rates are the best way to model the quantity vs quality tradeoff. Each input metric represents performance against one of the levers and the output metric represents both.
A common question about modeling conversion rates this way is that the conversion rate metric actually requires the output metric to calculate the input. Why do it this way instead of inverting it? The answer: each input metric represents its own collection of levers that interact with the metric.
Funnels
You may recognize that funnels are an extension of the conversion rate model as a larger sub graph (tree). They represent the quantity/quality tradeoff of a series of conversions that can represent complex processes or critical handoffs.
The benefit of representing funnels this way is that they surface the impact of every part of a process explicitly while making it easy to quantify each compared to the others. (More on this soon)
Growth loops
Loops are the most powerful mechanism of growth. They are the only way to increase output disproportionally to input. This works because the output metric is a input to itself (usually mediated by other processes). As the output increases, that impact loops back on itself. Examples of this are invitations, referrals, viral social media, and even reinvesting profits.
Can you find opportunities to create (viable) loops in your growth model?
Influence Metrics
Influence metrics act as the "fulcrum" between performance metrics and our efforts to improve them. Also known as influences, these metrics exert some impact on the Component metrics that make up a tree's scaffolding.
Unlike Component metrics, the relationship between an influence metric and a corresponding output metric is empirical rather than mathematical.An empirical relationship is based on observations, experiments, or experiences. The strength of that relationship may vary over time and will likely vary between companies.
The power of influence metrics is that they provide a framework for effective experimentation. In other words, you can run experiments to move influence metrics, and that effect ripples up the tree to the metrics that encapsulate performance.It'd be nice to be able to pull the "Customers" lever and gain more customers, but in reality, as Abhi Sivasailam says, the job of a data team is to identify real levers (influence relationships) and help the business pull them (experiments).
Components vs Slices
Designing metric trees will inevitably lead to a common modeling decision: should a metric be represented as a sum of several component metrics or as a single metric that can be sliced along a dimension?
First, some terminology, and then we’ll get to the principles behind it.
Components are the constituent metrics that, in this case, are summed together to produce a compound metric. (e.g. New Customers = Self Serve Signups + Closed Won Accounts)
Slices are metrics segmented along a dimension. Why not just call them segments? Segments are a subset of data, while slices are a subscope of a metric.
The example above shows the two different ways to represent Web Signups. One is a summation of signups from each channel, and the other is a single metric subdivided into slices for each channel.
How do you decide which approach to use?
The best guidance comes from Abhi Sivasailam: choose the more actionable approach. If the quantity has an owner who is responsible for it, it's better as a metric. If the subdivision is driven by analysis, it's better as a sliceable metric. The Web Signups example is a good case for component metrics—the metrics are strategically separate, and tactical levers are almost entirely distinct. Web User Agent (IOS, Windows, etc.), on the other hand, would certainly be a slice because the dimension doesn’t affect ownership or leverage.
A simple rule of thumb is illustrated in the B2B New MRR tree (miro link below). If the constituent metrics have different subtrees, they should be treated as metrics. If the subdivision terminates the tree, they are likely better as slices.
The decision ultimately affects things like causal analysis and forecasting. Still, the best way to get started is to represent trees in a way that stakeholders can understand and align with. 🍕
Have questions about these patterns? Feel free to reach out.
