Why the Direction of Data Collection Matters
A regression question can concern an association between a predictor and an outcome, but the way the data are assembled affects what can be learned from that association. In particular, it matters whether people join a study before the outcome of interest occurs or are selected for study after their outcome status is already known. It also matters whether predictor and outcome information comes from planned follow-up or from records of past events.
“Prospective” and “retrospective” describe a study’s direction relative to the events of interest—not simply whether a researcher typed information into a computer recently. A newly created database can contain retrospective information, while an older record system may have collected information prospectively. The key questions are: when was participation established, and had the relevant predictor and outcome histories already occurred by then?
A prospective study often starts by recording a predictor and then follows participants to see which outcomes occur. In a retrospective study, the same kinds of events may be reconstructed from earlier records. In both cases, a predictor and an outcome can be related in the data; the design determines how that relationship was assembled and what limitations apply to its interpretation.
Two Retrospective Designs
Retrospective studies do not all select participants in the same way. A retrospective cohort study identifies a group from the past, classifies its members by a predictor or exposure, and determines their later outcomes from records. The people or records are brought into the study after the events, but the historical sequence can still be predictor first, outcome later.
A case-control study instead selects participants according to outcome status: cases have the outcome, while controls do not. The study then examines earlier predictor or exposure information. Thus, a case-control study is retrospective with respect to the predictor history, but it is not simply a cohort with a shorter follow-up period. Its sample is assembled around the outcome.
| Design | How participants are assembled | What is typically recorded |
|---|---|---|
| Prospective cohort | Participation is established before the outcome occurs | Predictor information at or before follow-up, then outcomes as they occur |
| Retrospective cohort | A past group is identified after its outcome history has occurred | Historical predictor status and subsequent outcomes from records |
| Case-control | Participants are selected after outcome status is known | Past predictors or exposures for cases and controls |
A study can combine prospective and retrospective components—for example, using existing records for earlier history and then continuing follow-up. In such a design, describe which information was already available and which outcomes were observed after participation began, rather than forcing the entire project into an oversimplified label.
Worked Example: Following a Cohort Forward
Worked Example: Following a Cohort Forward
Suppose, hypothetically, that researchers enroll 100 adults who do not currently have a particular symptom. At enrollment they record whether each person usually gets at least 150 minutes of weekly activity. They then follow everyone for six months and record whether the symptom develops. There are 40 more-active participants, of whom 12 develop the symptom, and 60 less-active participants, of whom 9 develop it. Assume all 100 participants have a recorded six-month outcome and the same outcome definition is used for everyone.
Participation was established before the new symptom outcomes occurred. Predictor information was recorded at enrollment, and outcomes were recorded during follow-up, so this is a prospective cohort design.
The observed outcome proportion among the more-active participants is \(12/40=0.300\), or 30.0%. Among the less-active participants it is \(9/60=0.150\), or 15.0%. The difference between these observed proportions is
In this hypothetical cohort, the six-month symptom proportion is 15.0 percentage points higher in the more-active group. The ratio of the two proportions is \(0.300/0.150=2.00\): the observed proportion is twice as large in the more-active group. These are descriptive comparisons for this cohort. They do not show that activity caused the symptom; the design did not randomly assign activity, and the groups may differ in other ways.
Worked Example: Reconstructing a Historical Cohort
Worked Example: Reconstructing a Historical Cohort
Imagine, hypothetically, that a research team begins in 2025 by using archived records for 200 employees who were part of a workplace program in 2018. The records classify 80 employees as having used a particular workstation setup in 2018 and 120 as not having used it. Records through 2022 show that 16 of the 80 setup users and 12 of the 120 nonusers reported a specified symptom. Assume the archive identifies the relevant group and provides the same outcome information for every member.
The team starts the study after the 2018 predictor information and the outcome histories through 2022 already exist. This is a retrospective cohort study: the historical group is classified by predictor status, and later outcomes are reconstructed from records.
The observed outcome proportion for setup users is \(16/80=0.200\), or 20.0%. For nonusers it is \(12/120=0.100\), or 10.0%. The difference is
The recorded symptom proportion is 10.0 percentage points higher among setup users in this hypothetical historical cohort. Because both groups’ denominators and outcomes are available, these cohort proportions can be calculated from the records. Their interpretation still depends on the source records being sufficiently complete and comparable. A historical record does not become prospective merely because researchers analyze it today.
Worked Example: Selecting Participants by Outcome
Worked Example: Selecting Participants by Outcome
Suppose, hypothetically, that investigators select 50 people with a specified condition as cases and 50 people without it as controls. From historical records, 30 cases and 15 controls had a particular earlier exposure. The remaining 20 cases and 35 controls did not have that exposure. Since participants are selected after their outcome status is known and the study looks back at earlier exposure, this is a case-control design.
The exposure odds among the cases are \(30/20=1.50\). The exposure odds among the controls are \(15/35\approx0.4286\). Their ratio is
In this hypothetical sample, the odds of the earlier exposure are 3.50 times as large among cases as among controls. This is a comparison of exposure odds in the assembled sample; it is not a calculation of the condition’s risk among exposed and unexposed people.
The reason is that the investigators fixed the numbers of cases and controls at 50 each. Those sample counts do not reveal how common the condition was in the population. For example, the sample’s case proportion, \(50/100=0.500\), is a result of the selection plan, not an estimate of population risk. Estimating a population association from a case-control sample requires an appropriate way of selecting cases and controls; the sample counts alone do not establish that.
What the Design Changes—and What It Does Not
A prospective cohort makes it possible to record predictors before the outcome occurs and to monitor outcomes as follow-up proceeds. It can reduce dependence on participants’ later recall of past predictor values, although it may require substantial time and continued participation. A retrospective design can use events that have already occurred and may therefore be faster to conduct. Its conclusions depend on what historical records or memories can establish and on how the study group was assembled.
For cohort designs, when the relevant denominators and outcomes are available, observed outcome proportions can be compared across predictor groups. In a case-control design, sampling by outcome means the case and control totals are set by the design, so the resulting sample does not directly give population outcome proportions. This is a difference in the information supplied by the design, not a claim that one design is always better.
Neither timing label determines whether a regression association is causal. A prospective study records the predictor before the future outcome, which helps establish temporal order, but temporal order alone does not rule out other explanations for an association. Likewise, a retrospective study can preserve a clear historical order when records document it. In either design, an association should be interpreted in light of how participants entered the study, what was measured, and what population the observations represent.
Common Mistakes
- Calling a study prospective because the data were entered recently. The relevant question is whether participation was established before the outcome occurred or was observed, not when the database was created.
- Calling every retrospective study case-control. A retrospective cohort starts with a historical group and can compare outcome proportions across predictor groups. A case-control study selects people according to outcome status.
- Calculating risk from the case-control sample fraction. If the numbers of cases and controls were set by design, the sample case fraction reflects that choice, not population risk.
- Treating prospective follow-up as proof of causation. Recording the predictor first establishes temporal order in the study, but does not by itself show that changing the predictor would change the outcome.
- Assuming retrospective information is automatically unusable. Historical records may support informative comparisons when the relevant groups, predictors, outcomes, and denominators are adequately documented.
Check Your Understanding
For each question, focus on the timing of participation and on how the study group is assembled.
- A team enrolls people before a health outcome occurs, records a predictor, and follows them for two years. What design is this, and what makes it prospective?
- A team uses archived records to identify a past workforce, groups employees by an earlier predictor, and compares later outcomes. Why is this a retrospective cohort rather than a case-control study?
- Investigators choose equal numbers of people with and without a condition, then compare an earlier exposure. Why can the sample fraction with the condition not be treated as the population risk?
- In a prospective observational cohort, the predictor is recorded before the outcome. Does that fact alone establish that the predictor caused the outcome? Explain.
- Can information collected in a newly created database be retrospective? Explain what determines the design label.