Original practice scenario inspired by Abbott Laboratories's public business context; this is not a reported interview question or a description of its internal systems. A fictional healthcare-product analytics model uses measurements from several device channels. One channel is often missing in a particular site. Design an evaluation of missing-input handling without making clinical efficacy claims.
Read model answer +
Missingness can encode site, workflow, or equipment differences. Evaluate the complete pipeline under realistic missing patterns, rather than judging an imputation formula only on randomly hidden values.
Work through it
Describe which inputs are optional and record why values are absent when known. Keep sites or devices separated if deployment targets new settings.
Fit imputation and scaling within training data only. Compare an explicit missingness indicator, a simpler reduced-feature baseline, and an abstention path for unsupported patterns.
Report performance and coverage by missingness pattern and site. Stress-test channel outages and ensure the interface identifies unavailable or unsupported predictions.
Tradeoffs to discuss
Abstention reduces coverage but may avoid unjustified outputs. A model that exploits site-specific missingness may fail when collection workflows change.
Common pitfalls
Replacing missing with zero can create a plausible but false measurement.
Offline predictive accuracy does not establish clinical safety or authorize automated care decisions.
Original practice scenario inspired by Abbott Laboratories's public business context; this is not a reported interview question or a description of its internal systems. A synthetic operations report shows an increasing retest rate. The data includes samples, test attempts, invalid attempts, and repeat orders for new samples. How would you define and investigate the metric?
Read model answer +
Define the unit and reason for a retest before counting. A repeated attempt on the same sample differs from a new sample ordered later, and invalid first attempts change the interpretation of quality.
Work through it
Create a sample-level timeline with stable identifiers, attempt ordering, and outcome categories. Separate technical reruns, confirmation workflows, and newly ordered samples.
Choose a denominator such as eligible first attempts, then report the proportion needing a technical rerun within a fixed window. Keep unresolved samples visible.
Stratify by instrument, site, specimen type, and time while checking changes in mix and coding. Investigate abrupt shifts against maintenance and process changes.
Tradeoffs to discuss
A simple aggregate supports monitoring, but reason-specific metrics are more actionable. Small strata require uncertainty and careful interpretation.
Common pitfalls
Counting every repeated row as a retest confuses duplicates with work.
Removing invalid attempts can conceal the operational problem the metric should reveal.
