When Trend Lines Lie: Time-Series in Fragmented Reporting

Reporting artefacts look like epidemiology

When a health facility reports monthly, misses two months, then submits all three months at once, the dashboard shows a flat line followed by a spike. To someone reading the chart quickly, that spike looks like a surge in cases. In reality, nothing changed on the ground. The reporting caught up.

This happens constantly in systems where facilities report on paper, where internet connectivity is unreliable, or where data clerks are stretched across multiple responsibilities. The data is not fabricated. It is just delayed, and the delay creates a pattern that looks like something it is not.

The pattern is particularly dangerous during disease outbreak investigations. A district surveillance officer sees a threefold increase in suspected cholera cases in one week and triggers an alert. Outbreak response teams are mobilised. Supplies are redirected. Two days later, the district data manager explains that a rural health centre submitted its backlog after the DHIS2 server came back online. The cases were spread across the previous quarter, not concentrated in a single week. The response was based on a reporting artefact, not an outbreak.

Seasonal reporting patterns add another layer of confusion. In agricultural regions, health workers who double as community volunteers report less during planting and harvest seasons. The drop in reported cases during those months is not a decline in disease; it is a decline in data entry. Six weeks later, when the backlog is cleared, the apparent spike mirrors the reporting gap, not the disease trajectory.

Completeness is the missing chart

Most dashboards show the trend line but not the reporting completeness behind it. If you see that cases dropped by 40% last month, you need to know whether 40% fewer facilities reported. Without that context, the drop could mean the disease is declining or it could mean the reporting system had a bad month.

We treat completeness as a first-class indicator. It appears alongside the case counts, not buried in a footnote. When completeness drops below a threshold, the trend line is flagged so nobody interprets it as a real epidemiological shift.

Building this requires knowing the expected number of reports for each period. That sounds simple, but it requires a registry of reporting facilities, their expected reporting frequency, and their operational status. If a facility closed last month but nobody updated the registry, the completeness calculation is wrong. We maintain facility registries as part of the analytics infrastructure and flag facilities that have not reported in three consecutive periods for status verification.

• Show reporting completeness alongside every trend line
• Flag periods where fewer than 80% of expected reports were received
• Distinguish between zero reports (no cases) and missing reports (no data submitted)
• Use rolling averages to smooth reporting artefacts when appropriate

Techniques for separating signal from noise

One practical approach is to assign reported cases to the period when they occurred, not the period when the report was submitted. This requires the reporting form to capture the date of onset or date of diagnosis separately from the date of submission. Many HMIS forms already have this field, but dashboards often ignore it and plot cases by submission date because it is easier to implement.

Another approach is to calculate reporting-adjusted rates. If only 60% of facilities reported in a given month, the raw case count can be adjusted upward based on historical reporting patterns from the missing facilities. This introduces assumptions, and those assumptions should be stated, but it produces a more stable trend line than plotting raw counts that swing with reporting completeness.

We also use statistical process control methods adapted for health surveillance. Instead of drawing a simple trend line, the dashboard shows upper and lower control limits based on historical variation. A data point that falls outside the control limits after adjusting for reporting completeness is more likely to represent a genuine epidemiological event than one that falls within expected bounds.

Honest uncertainty is better than false confidence

The goal is not to produce charts that look clean. The goal is to produce charts that tell the truth about what the data can and cannot support. A trend line with a confidence band and a completeness flag is more useful than a crisp line that hides the mess underneath.

Programme managers who work with this data every day already know it is messy. They trust tools that acknowledge that reality more than tools that pretend it does not exist.

There is a practical payoff to this honesty. When a dashboard consistently labels data quality, programme managers stop second-guessing the numbers and start using them. They learn that a flagged period means 'wait for more data before acting' and an unflagged spike means 'investigate this now.' That distinction saves time, resources, and credibility. It also makes it possible to run post-hoc analyses that separate genuine outbreaks from reporting artefacts, which improves the evidence base for future planning.