Proxies and Measurement Error

Unit ID: ML-M02-U04 Estimated active time: 25-30 minutes

Some columns stand in for something else

A proxy is a value that may indirectly represent another condition.

For example, access_bandwidth_band may reflect internet quality. It may also reflect location, income, device access, or work conditions. We do not know.

Proxy columns need caution because they can carry unfair patterns.

Measurement error

Measurement error means the recorded value is not exactly the true value.

Examples:

• day3_minutes_active may miss time if the timer fails.
• preferred_device may record the last device, not the usual device.
• day3_support_request may miss learners who needed help but did not ask.

The table may look clean while the measurement is weak.

Why this matters for ML

Models do not know why a value is wrong. They learn patterns from what you give them.

If a column measures access problems, a model may learn to treat access as ability. That would be harmful.

The solution is not always to delete the column immediately. The first step is to flag the risk.

Practice

Review these columns:

• preferred_device
• access_bandwidth_band
• day3_support_request
• day3_minutes_active

For each, write:

1. What does it appear to measure?
2. What might it actually reflect?
3. Should it be used, excluded, or reviewed later?

Takeaway

A clean-looking column can still be a weak or risky measurement. Flag proxy and measurement risks before modelling.