Observation Frequency: Planning Your Daily Rounds
Distribute observations across days and rounds without confusing total records with independent observations.
Updated:
There is no universal minimum of four daily rounds or maximum of ten that makes a study valid. Frequency must cover the target period and population with feasible effort. Sample size, scheduling and dependence between readings are related decisions.
Turn precision into a feasible schedule
For a proportion under approximate independence, n=ceil(z²p₀(1−p₀)/e²). With z=1.96, p₀=0.5 and e=0.035, ceil(784)=784. Planning 800 may be a budget choice. It does not establish that 800 clustered readings contain the information of 800 independent ones.
If a round obtains m valid readings over D study days, the required average rounds per day are R=N/(D×m). This is a logistical identity, not a statistical test.
In a simulated plan, N=800, D=5 and m=20 give R=8 rounds/day. For an assumed 480-minute shift, T/R=60 minutes is an average budget spacing. It does not require starting exactly on the hour or imply simultaneous readings.
If all 20 people share one stoppage, their states may be correlated. A simple equal-cluster model gives an approximate design effect of 1+(m−1)ρ, where ρ is within-cluster correlation. With m=20 and ρ=0.10, that is 2.9; treating N=800 as fully independent exaggerates information. This illustrates dependence, not a known property of your plant.
Randomize times deliberately
You may select uniform moments within chosen windows or stratify by shift and time band. Retain inclusion probabilities and account for access time. Systematic sampling with a random start may be valid in some designs but can align with process cycles; it does not ensure every state is covered.
In a homogeneous Poisson process, event counts in an interval follow a Poisson distribution, while interarrival times are exponential. They are not “Poisson-distributed duration intervals.” No distribution automatically removes coverage or behavioral problems.
Minimum gaps imposed for logistics change the design. Document the rule and its selection consequences. No universal gap equal to the shortest activity duration ensures independence: long states and shared conditions matter too.
Missing readings and upward rounding
If the expected missing fraction is r, obtaining N valid readings on average requires ceil(N/(1−r)) attempts. For N=800 and r=0.15: ceil(800/0.85)=942 attempts. Multiplying by1.15 gives only 920 and does not exactly compensate for that expected loss.
At20 attempts per round over five days, 942/100=9.42 rounds/day on average. If the same integer count is required daily, use ten:1000 attempts with 850 expected valid readings under that assumption. Actual loss varies. Oversampling does not repair selective missingness or justify treating an included absence state as invalid.
Measure constraints in a pilot
| Constraint | Information for the decision |
|---|---|
| Access and travel | Actual route duration and inaccessible targets |
| Shared states | Readings affected by the same incident |
| Temporal coverage | Excluded shifts, tasks and products |
| Observer capacity | Alert-to-reading delays and classification errors |
| Safety | Access restrictions and tasks that must not be interrupted |
More small rounds may broaden temporal coverage compared with a few large rounds, but are not always optimal: travel cost and dependence matter. Unsourced sector rates do not replace a plant-specific design.
For an analyst covering two areas, interleaving may be convenient. Alternate days can confound area with day; a morning/afternoon split can confound area with time. Compare schedules with the target population and balance days and hours as appropriate rather than declaring one option universally superior.
Assess coverage and dependence
Review planned/actual times, losses and inclusion probabilities. No statistically significant morning/afternoon difference does not prove schedule adequacy; power may be low. A genuine difference does not prove undercoverage either: the populations may differ.
Durbin–Watson concerns regression residuals under specified conditions. Applying it to arbitrary category codes with1.5/2.5 cutoffs does not validate rounds or determine excessive spacing. Analyze binary indicators and cluster structure with a suitable model. No detected pattern is not proof of independence.
WorkSamp supports visit planning and recording. Notifications depend on device and permissions; check pending records and synchronization. The application does not replace sampling design. Cronometras complements duration analysis and ASETEMYT lists methods resources.