Time-Series Data and Observation Timing

Why Observation Timing Matters

A time series is more than a column of values with dates attached. Its order, the spacing between observations, and the timing represented by each value determine what a comparison means. A value recorded at the end of a month may describe a condition on that date; another value may summarize activity throughout the month. Those observations have different temporal meanings even if they appear in adjacent rows.

For an empirical question, the first task is to establish what each observation refers to in time. Are the values monthly or quarterly? Do the dates mark the start or end of a period? Is a value a point-in-time stock, a flow accumulated over an interval, or an average over that interval? Clear answers help distinguish a genuine change in the measured variable from a change caused by comparing different time windows.

Frequency, Span, and the Time Index

Three basic features describe the calendar structure of a series. Frequency is how often observations are recorded, such as daily, monthly, or quarterly. Time span is the period covered, usually described by the first and last observation dates. The time index identifies the date or period attached to each value. Reporting all three is more informative than saying only that a data set contains, for example, 100 observations.

Frequency does not tell us the time span by itself. One hundred monthly observations cover a different calendar duration from 100 daily observations. Nor does frequency guarantee regular spacing. A quarterly series may be recorded for every quarter, while a series of survey estimates may be available only for selected dates. The actual dates are therefore important: the ordinal labels \(1,2,3\) do not tell us whether the gaps are days, months, or years.

A series should also identify what its values represent in time. A point-in-time measurement records a quantity at a specific date, such as an inventory level on the last day of a month. A flow records activity over an interval, such as total sales during a month. An interval average summarizes values across a period, such as the average number of employees during a quarter. A date label alone can conceal these distinctions.

Point Measurements and Interval Measurements

The difference between a point measurement and an interval measurement affects which comparisons are natural. A stock is often measured at a date: a bank balance on the last day of a month, for example. A flow is accumulated over an interval: total deposits during that month. An average may summarize an interval but is not a total. Treating these as interchangeable can lead to incorrect interpretations of changes or totals.

Consider quarterly production. If each observation is total production during a quarter, the value represents a three-month flow. Adding the four quarterly totals gives an annual total, provided they cover the four quarters without overlap or gaps and use the same definition. By contrast, adding four quarterly measurements of a stock, such as end-of-quarter inventory, does not yield annual production or an annual inventory total. Those values describe levels at separate dates.

Even interval averages require care. The average for a quarter describes the quarter, while an end-of-quarter measurement describes one date. A quarterly average cannot generally be treated as though it were the value on the final day. To compare observations, state whether each one is a date-specific value, an interval total, or an interval average.

Irregular Intervals and Meaningful Changes

When observations are equally spaced, the difference \(y_t-y_{t-1}\) compares adjacent periods of the same length. With irregular timing, adjacent rows may represent unequal amounts of elapsed time. A raw difference then describes the total change between the two dates, but does not describe a change per unit of time.

If \(y\) is measured on a scale for which an additive rate is meaningful, and observations occur at calendar times \(t_1\) and \(t_2\), the average change per unit of time is the endpoint difference divided by elapsed time:

Here the time difference must be expressed in the intended units. For example, if \(y\) is measured in thousands of dollars and time is measured in years, the result is thousands of dollars per year. This average rate summarizes the endpoints; it does not assert that the variable changed at a constant rate between them.

Changing Definitions Across Time

A sequence of dates does not guarantee that the values measure the same thing. The population included, the rules for classification, or the method used to summarize observations may change. When a definition changes, a difference across the break can combine a change in the underlying quantity with a change in what is counted.

For example, a hypothetical monthly series might report average weekly hours for full-time employees through June and for all employees beginning in July. A drop from 40 hours in June to 35 hours in July cannot be read straightforwardly as a fall in the hours of a stable group: the population covered has changed. The values are not directly comparable unless the series is made comparable using appropriate information, or the comparison is restricted to a definition that is consistent across the dates.

When a definition changes, record where the break occurs and what changed. Do not silently treat the observations as a continuous series on an unchanged basis. If there is no information to align the definitions, the size or direction of the underlying change may not be recoverable from the reported values alone.

Common Pitfalls and a Practical Description

A frequent mistake is to treat consecutive rows as equally spaced periods. If the dates are March, April, and August, the second and third values are adjacent in the table but four months apart. A raw difference is still a valid endpoint difference, but calling it a one-month change is wrong. Another common mistake is to interpret a flow total as a point measurement, or to overlook that the group included in the definition has changed.

Before comparing time-series values, describe the series using a small set of questions:

• What variable is recorded, and what are its units?
• What is the exact date or interval attached to each observation?
• What are the frequency and the first and last dates in the span?
• Does each value describe a point in time, an interval total, or an interval average?
• Are the time gaps equal, and are the variable’s definition and scale consistent across dates?

These questions establish the temporal meaning of a comparison. They do not, on their own, establish a cause for a change or determine whether a series supports a particular empirical design. They make clear what the recorded values can be compared to describe.