What Social Science Gets Wrong When It Treats Reality as Linear

Apr 25, 2026

10 min read

7 views

88%

The hidden mistake: assuming the world speaks in straight lines

What if the biggest source of error in social science is not bad data, bad models, or even bad theory, but a deeper habit of thought: treating nonlinear realities as if they were linear?

That mistake is easy to miss because linear methods are comforting. They are fast, interpretable, and often surprisingly effective. But the comfort comes at a cost. When outcomes are binary, fractional, counted, capped, or lumpy, the world does not move in neat additive steps. A policy that changes the chance of receiving bail, the probability of hiring, or the number of sales calls does not behave like a simple increase in income or temperature. The effect depends on the scale you are measuring, and sometimes the scale itself is the message.

This is where a deeper tension appears. On one side is the classic econometric instinct: identify causal effects with fixed effects, trends, and difference in differences. On the other side is a newer computational instinct: simulate social processes with language models, structural causal models, and in silico experiments. At first glance these look like separate worlds. One is about estimation from data, the other about synthetic experimentation. But both are wrestling with the same problem: causality is not just about what happens, but about the form in which it happens.

That insight changes the game. The real question is not whether a model is linear or nonlinear in some abstract sense. The real question is: what relationship between treatment and outcome is stable enough to survive the measurement scale, the data-generating process, and the inferential strategy?

Parallel trends is not one assumption, it is a family of assumptions

Difference in differences is often taught as if it were a single, clean idea: treated and untreated groups would have followed parallel paths in the absence of treatment. But that simplicity hides a critical ambiguity. Parallel in what sense? On the raw mean scale? On a log scale? In a ratio of means? The answer matters more than many practitioners realize.

If the outcome is binary, the familiar additive version of parallel trends may be an awkward fit. The same is true for fractional responses, count data, or corner solutions, where many observations sit at zero and a few jump upward dramatically. In these cases, a treatment may not add a constant amount to the mean, but it may multiply the baseline risk, rate, or intensity. A hiring intervention might double the chance of selection without adding a fixed number of hires. A policy might reduce expected hospitalizations proportionally rather than subtracting a constant number of admissions.

This is why the distinction between linear parallel trends and conditional or multiplicative parallel trends is so consequential. The first says that untreated trends differ by an additive constant on the mean scale. The second says that trends align after a transformation, such as a log or other canonical link. Those are not cosmetic differences. They imply different estimands, different estimators, and different interpretations of treatment effect.

The choice of scale is not a technical afterthought. It is a claim about how the world actually changes.

Think of rainfall. Measuring in millimeters invites one kind of analysis. Measuring in log millimeters invites another. If a drought policy changes the relative growth of vegetation, the ratio scale may be more natural than the raw scale. Likewise, if a public health intervention affects infection rates multiplicatively, then treating the outcome as if it should shift additively can produce misleading estimates even when the regression looks perfectly polished.

The deeper point is that identification depends on the right invariance. Some causal patterns remain visible only after transformation. Others disappear when transformed. The analyst’s job is therefore not to force reality into a convenient linear mold, but to find the scale on which the untreated counterfactual evolves plausibly and coherently.

The best model is often the one that knows how to disappear

There is a striking computational lesson here. In nonlinear settings, fixed effects can become problematic because the model tries to estimate too many nuisance parameters from too little time variation, creating the infamous incidental parameters problem. Yet the practical response is not to abandon panel data or treatment effects. It is to choose a formulation in which pooled estimation, conditional means, and quasi likelihood methods can recover the causal quantities cleanly.

That is more than a statistical trick. It reveals a principle: good causal methods often work by making structure do the heavy lifting.

Suppose you are studying the effect of a job training program on the probability of employment. If the true untreated process is multiplicative, then a logit or Poisson style model may preserve the relevant parallel trends better than a linear probability model. In that setting, pooled estimation can be attractive because it respects the nonlinear data generating process while still allowing staggered interventions and cohort specific effects. The result is not just computational convenience. It is a better alignment between assumption and measurement.

This is a useful mental model: imagine trying to compare the growth of two startups. If both are compounding at different rates, subtracting their revenues each quarter may tell you less than comparing their percentage growth. The right unit reveals the structure. The wrong unit distorts it.

What makes this especially important is that a model can look elegant while quietly failing on the outcome scale that matters. A researcher may estimate a beautiful fixed effects regression only to learn later that the treated group’s trend was parallel to the control group only after a transformation. In that case the model was not merely inefficient. It was aiming at the wrong target.

So the useful question is not, “Can I put fixed effects in this regression?” The better question is, “What scale makes the untreated process stable enough to identify a counterfactual?”

Simulation does not replace theory. It reveals what theory can and cannot say

This same issue appears in a very different setting: synthetic social science with language models. At first, it might seem that simulation has the opposite problem from econometrics. Econometrics worries about too little flexibility and too much structure. Simulation worries about too much flexibility and too little discipline. Yet both domains converge on a common lesson: causal understanding requires a model of the mechanism, not just a prediction machine.

A large language model can be used not only as a text generator, but as an agent in a structural causal model. That makes it possible to encode hypotheses, run experiments in silico, and estimate effects under controlled scenarios such as negotiations, bail hearings, job interviews, or auctions. The point is not that the model magically knows social reality. The point is that once the causal structure is specified, the model can help explore the implications of that structure.

And yet a surprising asymmetry appears. The model may be able to predict the sign of an effect, but not its magnitude. In other words, it can often tell you whether a factor helps or hurts, but not by how much. Once conditioned on a fitted structural causal model, however, the predictions improve dramatically. This suggests something profound: the model contains fragments of causal knowledge that are not directly accessible without the right scaffold.

That is a deep parallel with nonlinear difference in differences. In both cases, the issue is not whether the system is intelligent or whether the data are rich. The issue is whether the underlying causal structure has been expressed in a form that the method can actually exploit. A language model, like a fixed effects estimator, can fail simply because the problem is posed in the wrong coordinates.

Intelligence is often latent, but not directly usable. Structure is what converts latent capacity into reliable inference.

Consider a bail hearing. A language model may intuit that prior record increases detention likelihood, but without a causal blueprint it may miss how that effect varies across contexts or how interventions alter the distribution of outcomes. Now imagine the same logic in policy evaluation. A model may understand that a program helps, but unless the effect is linked to a credible scale and a transparent intervention path, the estimate can be misleading. In both cases, the crucial step is not merely prediction, but causal indexing: locating the effect in the right mechanism and the right metric.

A new framework: causal literacy means thinking in transformations

The deepest connection between these two bodies of work is not methodological. It is epistemic. Both suggest that mature causal reasoning requires comfort with transformations, because transformations reveal which parts of the world are invariant and which are not.

Here is a practical framework.

1. Ask what quantity is naturally additive, and what quantity is naturally multiplicative

Some processes accumulate by addition, such as dollars spent in a budget. Others compound, such as infections, adoption, or wealth. If a policy acts on the second type of process, additive comparisons may disguise the real causal structure.

2. Identify the scale on which parallel trends is plausible

Parallel trends is not a universal law. It is a claim about the untreated evolution of outcomes on a chosen scale. Before estimating, ask whether that scale should be the level, the log, the odds, the rate, or something else.

3. Separate the model of the outcome from the model of the mechanism

A binary variable can be modeled linearly, but the mechanism behind it may be threshold based or multiplicative. The same applies to synthetic agents: an LLM may generate surface level behavior, but the structural causal model tells you what the behavior means.

4. Use estimation to test structure, not just to produce coefficients

A good estimator is not merely a way to get a number. It is a way to test whether your assumed invariances fit the data. If your model only works on the transformed scale, that is information, not inconvenience.

5. Treat prediction and explanation as allies, not substitutes

The language model can improve when given the causal graph. The panel data estimator improves when the right functional form is chosen. In both cases, prediction becomes more trustworthy when explanation is built in.

This is why “nonlinear” should not be treated as a special case. Nonlinearity is often the normal case. The linear model is the approximation. Sometimes it is excellent. Sometimes it is the wrong language entirely.

Key Takeaways

Do not assume additive effects just because they are easy to estimate. Ask whether the outcome evolves more naturally in levels, ratios, or logs.
Parallel trends is scale dependent. If it fails in levels, it may hold after a transformation, and that can change the right estimator completely.
Good causal structure improves both human inference and machine inference. Language models perform better when given a structural causal model, just as panel estimators perform better when the functional form matches the data generating process.
Prediction is not enough. Whether using regressions or simulated agents, the goal is to identify mechanisms that remain stable under intervention.
Think in invariances. The most important modeling question is often not what is changing, but what remains comparable when the world changes.

Conclusion: the real skill is choosing the right coordinate system

We tend to talk about causality as if the challenge were discovering hidden truth. But often the harder task is finding the coordinate system in which truth becomes visible. Sometimes that means moving from levels to logs. Sometimes it means moving from raw predictions to structural causal models. Sometimes it means recognizing that the model already “knows” something, but cannot use it until the structure is made explicit.

That is the unifying lesson here: many causal failures are not failures of intelligence, but failures of representation. The data may be informative, the model may be sophisticated, and the intuition may be sound. Yet if the scale is wrong, the effect hides in plain sight.

Once you see that, a lot of social science looks different. The question is no longer whether the world is linear or nonlinear in some abstract sense. The question is whether you have found the transformation that lets the world tell the truth.