The review examined research published between 2000 and December 2023, searching PubMed, Web of Science and IEEE Xplore for methods used to make individual-level data from independently designed studies comparable. From 1,585 records initially identified, duplicate removal left 1,234 papers for title and abstract screening. After full-text assessment, 35 papers met the eligibility criteria. The included studies involved longitudinal epidemiological data and described statistical procedures for pooling or harmonizing information collected from at least two separate datasets. The researchers excluded studies concerned only with straightforward recoding or calibration, as well as methods designed specifically for technical data types such as magnetic-resonance images or genotyping. Their focus was on the more difficult problem faced by large health cohorts: aligning tabular information such as questionnaire responses, clinical measurements and repeated assessments collected over years.

The need for such methods is accelerating. Large cohort studies can follow participants from childhood into old age, recording health, behavior, social conditions and biological measurements at multiple time points. When data from several cohorts are combined, researchers gain larger sample sizes, longer observation periods and a wider range of participants. These advantages can improve statistical power and allow scientists to investigate questions that no single study could answer. Pooling can also make expensive existing datasets useful for new research rather than requiring investigators to begin another lengthy and costly study. But the instruments used to collect the data are usually designed for the objectives of individual projects. One cohort may ask participants whether they feel “downhearted,” another may use a standardized depression scale, and a third may record only a binary yes-or-no response. Treating those variables as identical can blur important distinctions in what was actually measured.

The simplest solutions are often appropriate when variables are directly observable and their relationship is clear. A rule-based transformation can recode categories into a shared system, while calibration can convert a measurement using a known relationship, such as changing kilograms into pounds. For continuous variables whose distributions are reasonably comparable, researchers can use distribution-based methods. Linear equating transforms one variable so that it has the same mean and standard deviation as another. Standardized z-scores and t-scores are familiar examples of this approach. The method is computationally easy, but it implicitly assumes that the variables are sufficiently similar and that their distributions can be represented effectively by their first two statistical moments. If the underlying distributions are strongly skewed or have different shapes, matching only their means and standard deviations may conceal important differences.

Equipercentile equating offers a more flexible alternative by matching percentile ranks rather than simply aligning averages and variability. A participant at the 70th percentile in one study, for example, is mapped to the corresponding percentile in another study. This permits nonlinear transformations and does not require the variables to follow a normal distribution. It may therefore be useful when two questionnaires or scales have substantially different shapes but are believed to measure the same target. The trade-off is that percentile estimates become unstable in small samples, particularly when there is limited variation in the measured variable. Researchers must also be confident that the two measures genuinely represent the same underlying quantity; forcing two unrelated variables into matching distributions can manufacture comparability where none exists.

A more serious challenge arises when researchers want to harmonize a latent construct—something that cannot be observed directly but is inferred from several indicators. Depression, cognitive ability and well-being are not measured by a single physical unit. Instead, they are represented through collections of questions or tasks, each capturing part of the underlying trait. The proportion score method provides a simple solution: convert items to a common binary format, count the number positively endorsed and divide by the number of available items. The resulting score is easy to understand and can be calculated even when some items are unavailable. But it gives every item equal weight. A mild symptom and a severe symptom may contribute identically, and the method assumes that an item functions in the same way for different ages, populations and studies—an assumption that may be unrealistic.

Latent variable models attempt to preserve more of the information contained in multi-item assessments. Linear factor analysis represents observed responses as functions of one or more unobserved factors and can test whether the same construct is being measured across studies and over time. Item response theory, by contrast, models the probability of a particular response as a function of a person’s position on a latent trait and the characteristics of the item, including its difficulty and ability to distinguish between participants. A two-parameter logistic model can allow items to differ in discrimination, while a simpler one-parameter model assumes that they discriminate equally. These approaches can provide more precise harmonized scores, but they typically require large samples and advanced statistical expertise. They also need at least some common items linking the datasets.

The review highlights moderated nonlinear factor analysis, or MNLFA, as a particularly adaptable option when measurement conditions vary. Unlike conventional factor analysis, it can model nonlinear relationships between observed responses and latent traits. It can also account for differential item functioning—situations in which people with the same underlying level of a trait respond differently because of characteristics such as age, sex or study membership. Whereas many item response theory applications focus on discrete subgroups, MNLFA can incorporate both categorical and continuous moderators, including age as a continuous variable. It can further handle mixtures of continuous, binary and ordinal indicators, making it useful when different cohorts use incompatible response formats. The flexibility comes at a cost: MNLFA is computationally demanding, requires substantial sample sizes for stable estimates and is difficult to implement without specialized knowledge. An R package called aMNLFA was developed to automate parts of the model-fitting and scoring process.

Missing data create a second layer of difficulty. When one study collected a variable that other cohorts never measured, the pooled dataset contains systematic rather than random missingness. Multiple imputation can estimate absent values using overlapping variables and related information, but the review found that the proportion of systematically missing data that can be validly imputed has not been established through sufficient simulation research. For latent constructs, however, some missing items can be handled during the measurement process. Studies do not necessarily need to share every question if they can be linked through overlapping sets of items. One study may share one group of questions with a second study, while the second shares another group with a third. Those overlaps can form a statistical chain through which comparable factor scores are derived. This approach can retain study-specific items, but it depends on strong assumptions about the links between measurements.

The authors’ roadmap therefore begins with the target of the eventual analysis: is it an observed variable, such as height or weight, or a latent construct, such as mental health? For single-item measures, the scale type and distribution determine whether simple transformation, calibration, linear equating or equipercentile equating is most suitable. Binary, nominal and ordinal variables generally require rule-based recoding, while continuous variables demand closer examination of their calibration and distribution. For multi-item latent constructs, researchers must consider the number and type of indicators, sample size, the availability of overlapping items and whether responses behave consistently across cohorts. The review also emphasizes that harmonization should never be treated as a neutral technical step. Converting diverse raw measurements into a common score can discard information or introduce bias, especially when the original indicators capture subtly different concepts.

Quality control is consequently essential, yet it is inconsistently reported. Of the 25 pooling studies and empirical methodological papers included in the review, only five either performed sensitivity analyses comparing different harmonization decisions or reported descriptive comparisons between harmonized variables and the original cohort-specific data. The authors recommend documenting every transformation, including the original variables, their scales, the reasoning behind the chosen method and the variables lost during processing. Researchers could report correlations between harmonized and original measurements when both are available. For latent constructs, they might compare harmonized scores with related measures, examine measurement invariance and provide model-fit statistics. Such checks would make it possible to determine whether the resulting variable still represents the original data or has become a statistical artifact.

The review found no included study that used a fully automated machine-learning procedure to derive a harmonized epidemiological dataset. Tools such as Opal and Mica can support data management, vocabulary standardization and metadata dissemination, while Athena and Usagi assist with terminology mapping. Natural-language-processing systems such as Harmony can help identify potentially equivalent questionnaire items based on their semantic content. Yet matching words is not the same as determining whether two measurements are scientifically interchangeable. A question about sleep, for instance, may refer to duration in one cohort and perceived sleep quality in another. Important details—including whether blood glucose was measured while participants were fasting, how follow-up visits were scheduled or which population a questionnaire was validated in—are often buried in narrative cohort descriptions and codebooks. Automation can accelerate discovery, but researchers still need to judge context, measurement validity and acceptable information loss.

The field also faces a fundamental statistical limitation: longitudinal observations from the same participant are not independent. Many existing harmonization procedures address this by selecting one observation per person to create a calibration sample, estimating measurement properties from that reduced dataset and then applying the resulting parameters to all available observations. This avoids violating independence assumptions but throws away some of the repeated-measures information that makes longitudinal studies valuable. The authors call for new models that can account directly for within-person dependence while estimating item parameters from the full record. They also note that when cohorts share no common items at all, methods such as Linear Linking for Related Traits may provide a possible bridge through related constructs, although these approaches rely on strong assumptions and remain insufficiently tested in longitudinal settings.

For epidemiologists, the message is both urgent and practical. Combining datasets can reveal patterns in disease risk, aging, mental health and development that remain invisible within isolated cohorts, but larger numbers do not automatically produce more reliable evidence. Statistical harmonization determines what the merged data mean, which participants can be compared and which distinctions disappear. The review’s roadmap offers a way to make those decisions more explicit, while its call for rigorous documentation and quality metrics addresses a weakness that could undermine reproducibility across the field. As research communities adopt FAIR principles and share increasingly complex longitudinal resources, harmonization will become a central scientific task rather than a behind-the-scenes cleaning exercise. The next breakthrough may not be a new statistical model alone, but an automated system capable of combining computational speed with the contextual judgment required to understand what health measurements truly represent.