how can we actually enable CUPED for our experiments?

The knowledge sources explain what CUPED is and when to use it, but they do not contain explicit step-by-step instructions on how to enable CUPED within the Mixpanel UI for your experiments. [Advanced Statistical Methods]

What the documentation does confirm is that CUPED is one of the advanced statistical methods available in the Experiments report, and that it uses a pre-exposure period of your choosing to adjust for pre-experiment user behavior. [CUPED]

For precise UI instructions on enabling CUPED, I'd recommend:

What the docs do tell us about CUPED:

kapa.ai do I need to use CUPED if I already monitors new visitors and existing visitors separately?

Great question. The knowledge sources don't directly address whether segmenting new vs. existing visitors replaces the need for CUPED, but here's what the documentation does tell us that's relevant:

What CUPED does:

CUPED reduces variance by accounting for how users were already behaving before the experiment. It helps isolate the true treatment effect by controlling for pre-existing behavioral differences between groups — not by segmenting users, but by adjusting the outcome measurement. [CUPED]

Where your segmentation approach helps:

Monitoring new and existing visitors separately is a sound practice — it avoids survivorship bias and surfaces different behavioral patterns. The documentation explicitly recommends using segmentation (e.g., new vs. returning users) to understand nuances in experiment results. [Reading Results]

Why CUPED may still be useful:

Bottom line: Segmenting new vs. existing visitors and using CUPED serve different purposes — they're complementary, not interchangeable. Segmentation helps you analyze subgroups correctly; CUPED helps you detect real effects faster within a group by reducing noise. The knowledge sources don't state that one eliminates the need for the other.

kapa.ai why do I get an Imbalance alert?

An "imbalance alert" in the context of Mixpanel Experiments refers to a Sample Ratio Mismatch (SRM) — a health check that Mixpanel runs automatically on all experiments.

What triggers an SRM (Imbalance) alert?

You get this alert when the actual distribution of users across your variants doesn't match your configured split in a statistically significant way. For example, if you set up a 50/50 split but end up with 60% in control and 40% in treatment, Mixpanel flags this as a potential problem. [Health Checks]

Mixpanel uses a Chi-squared test to determine whether the observed imbalance is likely due to random chance or signals a real issue. [Health Checks]

Common causes

What should you do?

The recommended action is to:

💡 As noted in the previous conversation, enabling CUPED can help mitigate the impact of pre-existing group bias, but it doesn't fix a broken randomization setup — that needs to be resolved at the source.

but my ratio is actually 50/50 kapa.ai

Even if your observed ratio appears to be 50/50, an SRM alert can still fire if the imbalance is detected at a statistically significant level — meaning even a small but consistent skew can trigger it.

The knowledge sources don't go into further detail about why an SRM alert might appear despite an apparently balanced split, but based on what's documented, here are the most likely explanations to investigate:

The recommended course of action remains the same regardless: pause the experiment, investigate the root cause, and restart after fixing it. [Advanced Statistical Methods]

For a deeper diagnosis, I'd recommend checking whether your exposure event ($experiment_started) is firing at the right moment and consistently for both variants — not at assignment time, but at actual exposure. [Experiments]