What Does It Mean for Data to Be Time-Indexed?
A time-indexed data set records a value together with the time to which that value belongs. This pairing is essential: the number alone does not say whether it is a weekly count, a month-end balance, or a measurement taken at a particular instant. Before analyzing such data, identify both what was observed and when it was observed.
For example, a hypothetical insurer might record the number of new claims reported in each calendar month. The observation is the monthly count; its time index is the month to which that count refers. The row number in a spreadsheet is not necessarily the time index. Rows can be sorted by another field, entered out of sequence, or copied from separate files.
A time index also needs a clear interpretation. A value dated June 30 might be a balance measured at the end of June, while a value labelled June might be the total accumulated over the entire month. Those are different kinds of observations. A point-in-time measurement refers to a moment; a total or average may refer to an interval. State which interpretation applies before comparing values.
Regular and Irregular Observation Times
Observations are regularly sampled when successive time indices are separated by the same fixed time interval, such as every day, every calendar month, or every 15 minutes. The unit and schedule matter. Monthly observations do not occur at equal numbers of days apart, but they are regularly sampled by calendar month if each value refers to the same point or interval in each successive month.
Observations are irregularly sampled when the elapsed time between successive observations is not constant. This can happen because measurements are taken only when an event occurs, because collection times vary, or because some scheduled observations are unavailable. A gap in a data table should not automatically be filled with a zero: no recorded value is not the same thing as a measured value of zero.
To check a schedule, first put the time indices in chronological order, then compare the intervals between adjacent times. For calendar data, use the stated calendar unit rather than assuming all months or years contain the same number of days. For elapsed-time comparisons, express each gap in a consistent unit, such as days or hours.
Worked Example: Reading Quarterly Observations
A hypothetical pension administrator lists the number of benefit applications received in each quarter: 12 in the first quarter, 15 in the second, 11 in the third, and 18 in the fourth. Each count is a total for its named quarter. Identify the observations and time indices, determine whether sampling is regular, and calculate the change from one quarter to the next.
Method. Treat each quarter as one time interval and attach each count to the quarter it summarizes. The time indices, in order, are Q1, Q2, Q3, and Q4. Because the quarters are consecutive and each covers one calendar quarter, the sampling is regular by quarter.
The observations are \(y_1=12\), \(y_2=15\), \(y_3=11\), and \(y_4=18\). The consecutive changes are:
Thus the count increased by 3 applications from Q1 to Q2, decreased by 4 from Q2 to Q3, and increased by 7 from Q3 to Q4. The method compares neighboring quarters in calendar order, not neighboring values after sorting the counts. No probability distribution is assumed; these are hypothetical observed counts.
Check. The changes sum to \(3+(-4)+7=6\), which equals the overall change \(18-12=6\). This verifies the arithmetic. A common mistake would be to sort the values as 11, 12, 15, 18 and treat their differences as quarterly changes; that would discard the quarters to which the counts belong.
Chronological Order and Unequal Time Gaps
When observations are irregularly sampled, the raw difference between two values does not by itself describe how quickly the measured quantity changed. A difference of 9 over nine days is not the same pace as a difference of 9 over one day. If a per-unit rate of change is appropriate, divide the value difference by the elapsed time in that unit:
This is an average over the interval between two measurements. It does not establish that the value changed at a constant rate at every moment within that interval. Also, a rate calculation is meaningful only when the measured quantity and time unit are correctly identified.
Worked Example: Comparing Changes at Irregular Times
A hypothetical monitoring device records a water-level index of 120 on March 2, 132 on March 5, 141 on March 14, and 145 on March 15. Determine whether the observations are regularly sampled and compare the average change per day over each interval. Treat the dates as measurements taken at the start of each named day.
Method. The time indices are the four dates, already in chronological order. The elapsed gaps are 3 days from March 2 to March 5, 9 days from March 5 to March 14, and 1 day from March 14 to March 15. Since these gaps are unequal, the sampling is irregular.
For March 2 to March 5, the value change is \(132-120=12\). The average change per day is \(12/3=4\) index units per day.
For March 5 to March 14, the value change is \(141-132=9\). The average change per day is \(9/9=1\) index unit per day.
For March 14 to March 15, the value change is \(145-141=4\). The average change per day is \(4/1=4\) index units per day.
The raw changes are 12, 9, and 4, but the corresponding average daily changes are 4, 1, and 4. The middle interval has the largest duration and the smallest daily rate. A common mistake is to compare raw changes as if all three intervals had the same length. No distribution or stochastic model is specified or needed for these calculations.
Check. The total change is \(145-120=25\) over \(13\) days. The interval changes add to \(12+9+4=25\), and the interval lengths add to \(3+9+1=13\), confirming that the endpoints and gaps have been accounted for consistently.
When Rows Are Not in Time Order
A data file can be arranged by customer, location, data-entry time, or an arbitrary row identifier rather than by the time of the observation. To describe what happened over time, sort by the actual observation time, not by the order in which records happen to appear. If two observations share a date, check whether their timestamps provide a further order. If they are genuinely simultaneous or only recorded to the same date, do not invent an order between them.
A tied timestamp may also indicate that records should be combined, for example when several transactions are meant to produce a daily total. Whether to combine them depends on what each row represents and what the analysis requires; it is not a purely mechanical sorting decision. Similarly, if a scheduled time has no record, distinguish that missing observation from a recorded zero.
Worked Example: Sorting Records Before Calculating Changes
A hypothetical sensor file contains four readings, shown in the order they were entered: May 1, value 80; May 4, value 92; May 2, value 86; May 9, value 110. Identify the chronological sequence and calculate the change per day between each successive observation. Assume all dates are in the same year and each reading is taken at the start of the date shown.
Method. Sort by observation date. The chronological records are May 1: 80, May 2: 86, May 4: 92, and May 9: 110. Relative to May 1, their elapsed times are 0, 1, 3, and 8 days. The adjacent gaps are therefore 1, 2, and 5 days.
From May 1 to May 2, the change is \(86-80=6\) over 1 day, or \(6/1=6\) units per day.
From May 2 to May 4, the change is \(92-86=6\) over 2 days, or \(6/2=3\) units per day.
From May 4 to May 9, the change is \(110-92=18\) over 5 days, or \(18/5=3.6\) units per day.
The observations are irregularly sampled because the gaps are 1, 2, and 5 days. The original row order would incorrectly make the sequence appear to jump from May 4 back to May 2. For instance, subtracting 92 from 86 in that order produces \(-6\), but it is not a chronological change from May 4 to May 2; the earlier date must come first. The correct method avoids treating file order as time order.
Check. The three chronological changes sum to \(6+6+18=30\), equal to \(110-80=30\). The total elapsed time is \(1+2+5=8\) days, equal to the difference between May 1 and May 9.
Why Order Changes the Interpretation
The values in a time-indexed data set are not just a collection of numbers. Their order determines which observations are adjacent and which changes occurred between successive times. Reordering the same values can therefore change the apparent sequence, even though the smallest value, largest value, and overall collection remain unchanged.
For a direct check, consider the hypothetical values 2, 5, and 9. In the order 2, 5, 9, the successive changes are \(5-2=3\) and \(9-5=4\). If the same values are instead placed in the order 2, 9, 5, the successive changes are \(9-2=7\) and \(5-9=-4\). The total change from the first to last value differs too: it is \(9-2=7\) in the first ordering and \(5-2=3\) in the second.
This demonstrates why order is part of the information being analyzed. A summary that ignores timestamps cannot, by itself, establish what changed first, what followed, or how long a change took. The timestamps are not labels to discard after reading the values; they define the sequence and the intervals over which comparisons are made.
The same principle applies when a report presents weekly or monthly aggregates. The time label must match the period represented by the value, and comparisons must respect the order and spacing of those periods. A missing week cannot be silently treated as a week with zero activity; nor should a monthly total be interpreted as if it were a measurement at one particular instant.
Common Mistakes to Avoid
- Using row order as time order. Sort by the observation timestamp when the file is not already chronological.
- Ignoring the time unit. A change over several days should not be compared as a one-day change without accounting for the elapsed time.
- Calling unequal gaps regular. Check the time indices, not the pattern of the measured values.
- Replacing missing data with zero. Zero is a measured value; a missing record means the value is unavailable unless the data definition says otherwise.
- Confusing an interval total with a point measurement. A monthly count summarizes a period, whereas a month-end balance refers to a point in time.
- Inventing an order for tied times. Use finer timestamps if available, or retain the tie and explain how it is handled.
Check Your Understanding
For each problem, identify the relevant time structure before interpreting the values.
- A hypothetical call center reports 44, 51, and 47 calls for three consecutive weeks, with each value a weekly total. State whether the observations are regularly sampled and give the two consecutive changes. (Answer: Regularly sampled weekly; changes are \(+7\) calls and \(-4\) calls.)
- A device records a reading of 30 on August 3, 42 on August 7, and 48 on August 10. Find the average change per day in each interval and state whether sampling is regular. (Answer: Irregular; \(12/4=3\) units per day, then \(6/3=2\) units per day.)
- A file lists September 6: 70, September 2: 58, and September 4: 64. Put the records in chronological order and calculate the overall change per day from the first to the last date. (Answer: September 2: 58, September 4: 64, September 6: 70; \((70-58)/4=3\) units per day.)
- A scheduled daily report has values for October 10 and October 12, with no October 11 record. The two observed values are 18 and 0, respectively. Is the October 11 value known to be zero, and what is the average change per day across the two-day interval? (Answer: No; October 11 is missing, not known to be zero. The average change is \((0-18)/2=-9\) units per day.)
- A hypothetical account balance is recorded at 09:00 with value 200 and at 15:00 with value 212 on the same day. What is the elapsed time, the average change per hour, and whether these two observations alone establish a constant rate throughout the interval? (Answer: 6 hours; \(12/6=2\) units per hour; no, the endpoint readings establish only the average rate.)