Why the Longitudinal Model Changes the Causal Story
The subject here is not whether contact and outgroup attitudes are related. It is what a repeated-measures model treats as change, and therefore what it is entitled to call a contact effect. This chapter focuses on the conventional cross-lagged panel model and its two alternatives so that the reader can tell whether a result describes different kinds of people, a person moving away from their usual pattern, or a relationship that remains ambiguous because stable differences were not separated.
When a lag is mistaken for a cause
A longitudinal study measures contact and attitudes at several points in time. Imagine that contact is recorded at time 1 and attitudes at time 2. If people who report more contact at the first point also report warmer attitudes at the second, the result appears directional: contact came first, attitude came later. The conventional cross-lagged panel model, or CLPM, formalizes that intuition. It estimates the path from earlier contact to later attitudes while accounting for earlier attitudes. It also estimates how stable each variable is over time and allows contact and attitudes measured at the same wave to covary.
That structure is more informative than a single snapshot, but the cross-lag is not automatically a causal effect. Controlling for attitudes at time 1 asks whether contact predicts later attitudes beyond what earlier attitudes predict. It does not necessarily remove every stable characteristic that shaped both the earlier contact and the later attitude. A model can therefore have time order without having a clean counterfactual comparison.
Consider a hypothetical participant who is consistently high in right-wing authoritarianism. That characteristic may be related to less intergroup contact and less favorable outgroup attitudes at every wave. If the characteristic is not measured, the data still contain its influence. The model may observe that lower contact at time 1 is followed by less favorable attitudes at time 2 and attribute part of the relationship to the cross-lag from contact to attitudes. Yet the pattern could be generated partly by a person who remains different from others throughout the study, not by a change in contact that moves their attitude.
The same problem can appear in the opposite direction. People who are already open to an outgroup may seek more contact, find it easier to sustain friendships, and report warmer attitudes later. Their openness is not necessarily a one-time disturbance that prior-attitude adjustment can remove. It can be a relatively stable source of both exposure and outcome. The resulting association may be real and socially meaningful while still failing to identify what would happen if contact were increased for someone else.
This is the crucial difference between a between-person pattern and a within-person process. A between-person result says that people who tend to have more contact also tend to have more favorable attitudes. A within-person result asks whether a particular person’s contact is higher or lower than their own usual level and whether their subsequent attitude is correspondingly higher or lower than their own usual level. The contact hypothesis is naturally phrased in the second way: an encounter changes a person’s relationship to an outgroup. But a CLPM can blend the two sources of variation.
A simple thought experiment makes the mixture visible. Imagine two participants observed at three waves. One has many outgroup friends and consistently warm attitudes; the other has little contact and consistently cold attitudes. If neither participant changes, the panel still contains a strong relationship between contact and attitudes. That relationship is informative about their different social positions, but there is no observed contact increase from which to infer an attitude response. Now imagine that one participant’s contact rises at the second wave. To interpret a later attitude difference as a response to that rise, the analysis must distinguish the new exposure from the person’s enduring tendency to seek contact, the circumstances that made the exposure possible, and other changes occurring at the same time. The labels “time 1” and “time 2” do not perform that separation by themselves.
The conventional model tries to make the comparison sharper by accounting for earlier attitudes and the stability of both variables. That adjustment is valuable: it prevents the analysis from treating an already-existing attitude level as if it were entirely new. Yet an unmeasured stable factor can still shape the levels that the model is adjusting around. It can also shape the probability that contact changes at all. A person’s openness may affect both their starting point and their later opportunities, while a school, neighborhood, or political event may alter contact and attitudes together. The cross-lag is then a conditional prediction within a complicated history, not a direct replay of an experiment.
Blending does not make the CLPM useless. It makes its causal interpretation conditional. If stable unobserved confounders are absent, adequately measured, or otherwise handled by the data-generating process, the cross-lag may be informative. If they are present and influence contact and attitudes across waves, the estimate can be biased. The point is not that every CLPM result is false; it is that the model’s familiar path does not, by itself, certify the causal story attached to it.
The distinction becomes even more important when the model is used to compare positive and negative contact. A significant path from positive contact to warmer attitudes, or from negative contact to colder attitudes, may fit contact theory. But the sign of a path does not reveal whether it was produced by a person’s changing experience or by enduring differences in who has which experiences. The causal question is carried not only by the direction of the coefficient, but also by the variance and confounding assumptions behind it.
Three models, three answers
The full-forward CLPM, or FF-CLPM, addresses part of the confounding problem by adding higher-order lags. In a study with three measurement waves, it can include paths from contact and attitudes at time 1 directly to the corresponding variables at time 3, in addition to the usual paths between adjacent waves. With more waves, it adds the possible longer-lag stability and cross-lagged paths. These paths represent prior exposure and prior attitude history that a simpler model leaves less fully represented.
The causal-inference logic is subtle. Suppose a stable unobserved factor has influenced a person’s contact and attitude levels from the beginning. By controlling not only the immediately preceding values but also earlier values, the FF-CLPM tries to account for the accumulated history through which that factor could have operated. The model’s strongest test is usually the cross-lag in the final interval, because it has the greatest amount of prior exposure available for adjustment. That is a stronger control strategy than simply reading every adjacent-wave path as a separate causal effect.
But more paths do not make the model a transparent window onto the truth. The FF-CLPM is just-identified when all possible structural paths are freely estimated. In practical terms, it can reproduce the observed covariance structure perfectly, leaving no degrees of freedom with which to ask whether its fit is plausible relative to the data. Perfect fit in this setting is a property of the model’s flexibility, not independent evidence that the causal assumptions are correct. The model can be useful for a particular confounding-control purpose while offering no ordinary global fit test.
The random-intercept CLPM, or RI-CLPM, takes a different route. It represents each variable as two components. The random intercept captures a person’s stable level across the observed period: some participants generally report more contact than others, and some generally report warmer or colder attitudes. The within-person component captures departures from that personal baseline at each wave. A cross-lagged path in this part of the model asks whether being above one’s own average contact is followed by being above one’s own average attitude level, after accounting for the person’s prior deviation.
That is much closer to the process implied by the contact hypothesis. It also changes the interpretation of a familiar positive relationship. Suppose participants with more contact generally have warmer attitudes. In an RI-CLPM, that stable pattern can appear as a correlation between the contact random intercept and the attitude random intercept. It may be large and statistically significant. But it is a between-person association: it says that people’s enduring contact levels and enduring attitude levels move together. It does not say that increasing contact for an individual will produce a later attitude change.
The within-person path is the quantity more relevant to that causal question, but it is not an assumption-free answer. “Above one’s own average” is a relative statement, not a claim that contact reached some universal level of quality or intensity. A small, temporary increase may be meaningful for one participant and negligible for another. Separating stable differences helps with time-invariant confounding that is represented in the stable components; it does not guarantee that every time-varying confounder has disappeared, that the measurement interval matches the psychological process, or that the model’s decomposition is exactly how the data were generated. The authors of the re-analysis therefore treat the RI-CLPM as more strictly controlled for stable differences, not as a universal replacement for every other model.
The timing issue matters as much as the decomposition. A conversation can affect anxiety within hours, while a friendship may alter an attitude over months. If measurements are six months apart, a short-lived response may be missed or absorbed into later stability. If they are too close together, the survey may capture mood or a contemporaneous circumstance rather than durable change. The model can only estimate the process that the measurement schedule makes visible. A null lag therefore needs to be read alongside the exposure being measured and the interval over which the response was expected.
This is also why model fit and causal interpretation must be held apart. Fit asks how well a model represents the observed variance and covariance. Causal identification asks whether the assumptions make the estimated quantity correspond to the intervention or change we care about. A model can fit poorly and still produce a coefficient that resembles a theory prediction. A model can fit better because it represents stable person-level structure more flexibly, while its within-person estimate remains uncertain. Neither result can be understood from fit or significance alone.
The three models thus answer related but distinct questions. The CLPM asks whether earlier contact predicts later attitudes after conventional stability adjustment, while leaving stable between-person and within-person variation intertwined. The FF-CLPM adds prior-exposure paths to strengthen control of indirect confounding, at the cost of a model whose perfect fit is not informative. The RI-CLPM separates stable levels from situational deviations and makes the within-person cross-lag explicit, while retaining strong assumptions about how those components behave.
That difference is the reason the model choice can change the causal story without changing the underlying data. The observations are the same; what changes is which variation is treated as evidence for contact-driven attitude change and which variation is treated as a stable association. The next empirical question is therefore not which acronym sounds most advanced. It is what these models do when applied across many published longitudinal contact datasets, with their varied samples, measures, intervals, and imperfections.