Omitted Variable Bias: How to Predict the Direction Before You Run the Regression

Here’s a question that separates people who ran a regression from people who understand one.

You regress wages on years of schooling. You get a coefficient of 0.11 — about an 11% return per year. Your supervisor looks at it and says:

“You’ve left out ability. Which way is that biasing your estimate?”

Now. You don’t have data on ability. Nobody does. You cannot run the regression with ability included, because ability isn’t in your dataset, isn’t in anybody’s dataset, and arguably isn’t measurable at all.

And yet you can answer the question. Confidently. Right now, before touching a keyboard.

The answer is “upward,” and by the end of this post you’ll be able to say why in one sentence, and do the same thing for any omitted variable in any regression you ever run.

This is the single most useful skill in applied econometrics, and it’s usually taught as a formula to memorise rather than a thing you can actually do. Let’s fix that.

Start with the model you wish you could run — the true one:

You can’t run it. Ability is unobserved. So you run this instead:

And here’s the problem: ability didn’t disappear. It got shoved into the error term. Your u now contains ability plus everything else.

That would be fine — genuinely fine — if ability had nothing to do with schooling. Error terms are allowed to contain things. That’s what they’re for.

But ability is not unrelated to schooling. More able people tend to get more schooling. So your regressor is correlated with your error term.

And that breaks the one assumption that actually matters.

Here it is. It’s short, and every part of it is something you can reason about.

Bias = β₂ × δ₁. Two terms. Two signs. Multiply them. That’s it. That’s the whole thing.

And here’s the part that makes it powerful: you don’t need data to sign either term. You need economics.

Multiply the two signs. That’s your bias direction.

Read it as ordinary multiplication. Positive × positive = positive bias. Positive × negative = negative bias. Negative × negative = positive bias.

Now let’s use it.

Two questions, both answerable from theory alone.

Does ability raise wages, holding schooling constant?
Almost certainly yes. More able workers are more productive; more productive workers earn more. β₂ > 0.

Are more able people more likely to get more schooling?
Almost certainly yes. School is easier and more rewarding if you’re good at it; able students get encouraged, admitted, and funded. δ₁ > 0.

So your 11% estimate is too big. Part of what you’re attributing to schooling is really ability wearing schooling’s clothes. The true causal return is somewhere below 11%.

You just answered your supervisor’s question. No data. No regression. Two sentences of economics and a multiplication.

Get quick at this. It’s a two-line habit.

You regress test scores on class size and find smaller classes → higher scores. You’ve omitted family income.

• Does family income raise scores? Yes. β₂ > 0.
• Is family income related to class size? Richer areas fund smaller classes, so income ↓ class size. δ₁ < 0.
• Bias = (+) × (−) = negative. Your estimate is too negative — you’re overstating the benefit of small classes. Part of it is just wealthy families.

You regress crime rates on police numbers and find more police → more crime. Absurd on its face. You’ve omitted underlying crime propensity.

• Does crime propensity raise crime? By definition. β₂ > 0.
• Do high-crime places hire more police? Obviously. δ₁ > 0.
• Bias = (+) × (+) = upward. Your positive coefficient is bias, possibly entirely. This is simultaneity — police respond to crime — and it’s why this literature lives on natural experiments like terror alerts and election cycles.

You regress employment on minimum wage across US states and find no effect. You’ve omitted local economic conditions.

• Do strong local economies raise employment? Yes. β₂ > 0.
• Do states raise minimum wages when their economies are strong? Generally yes — it’s politically easier. δ₁ > 0.
• Bias = (+) × (+) = upward. Your estimate is biased towards finding no harm. Which means a null result is weaker evidence than it looks — and it’s precisely why Card and Krueger needed a neighbouring-state design rather than a cross-state regression.

You regress sales on ad spend. You’ve omitted product quality.

• Does quality raise sales? Yes. β₂ > 0.
• Do firms with better products advertise more? Usually — they have more to shout about and more margin to spend. δ₁ > 0.
• Bias = (+) × (+) = upward. Your ROI on advertising is overstated. Every marketing department in the world has this problem and most of them don’t know it.

Notice the pattern: upward bias is the common case. Good things cluster with good things. The omitted variable usually helps y and correlates positively with the treatment, so naive regressions usually flatter the treatment.

That’s a genuinely useful prior. When you read a paper with no identification strategy claiming a large positive effect, your default suspicion should be that it’s smaller than reported.

No. And this is where a lot of otherwise-good dissertations die.

Adding controls is not free. Some controls remove bias. Some controls create it. Knowing which is which requires — again — theory, not software.

You want the effect of education on earnings. You control for occupation.

But education works through occupation. Getting a degree gets you a graduate job, and the graduate job pays more. By controlling for occupation, you’ve removed the main channel by which education operates. Your coefficient now answers: “among people in the same job, does education pay?” — which is a real question, but a completely different one from the one you asked.

This one is genuinely counterintuitive and it’s worth slowing down for.

Suppose a firm hires people based on two independent things: ability and interview charm. In the general population these are unrelated.

Now look only at people the firm hired. Among the hired, ability and charm are negatively correlated — because if you got hired despite being unremarkable at interview, you must have been very able, and vice versa.

You didn’t discover a relationship. You manufactured one, by conditioning on a variable that both causes affect.

This is also why studies of hospitalised patients, hired employees, admitted students, or surviving firms are treacherous: the sample is defined by a collider.

When you discuss a potential omitted variable — and you should, in every empirical write-up — use this structure:

“The estimated coefficient on [x] is likely to be biased [upward/downward] due to the omission of [variable]. [Variable] plausibly [raises/lowers] [y] (β₂ [>/<] 0), and is likely to be [positively/negatively] correlated with [x] (δ₁ [>/<] 0), implying a bias of [sign]. The estimate should therefore be interpreted as [an upper/a lower] bound on the true causal effect.”

That paragraph, written well, is worth more than a page of robustness checks. It demonstrates that you know what your regression can and cannot claim — which is the thing being examined.

Then say what you’d do about it. Even if you can’t do it. “With access to [data], I would [instrument / use fixed effects / exploit reform X].” Markers reward knowing the fix exists.

A1. β₂ > 0: better management raises profits. δ₁ > 0: better-managed firms are more likely to invest in R&D — they’re better at spotting opportunities and executing on them. Bias = (+)(+) = upward. The R&D effect is overstated; part of it is management quality showing up in both places. This is a well-known problem in the innovation literature, which is why it leans on R&D tax credit changes as natural experiments.

A2. β₂ > 0: more productive workers earn more. δ₁ > 0 by the assumption given: unions are more prevalent where productivity is high. Bias = (+)(+) = upward. The union wage premium is overstated. Note the assumption is doing real work — if you instead argued unions organise where workers are weakest, δ₁ flips and so does the bias. Which is the point: the sign of your conclusion depends on the economics you argue for, so argue for it explicitly rather than asserting it.

A3. Two good answers from many:

• Reverse causality. Rich, healthy countries can afford more healthcare, and healthier populations are more productive. Spending and life expectancy determine each other. No number of controls fixes simultaneity.
• Bad controls. GDP per capita is plausibly a mediator — healthcare spending improves health, healthier populations are more productive, productivity raises GDP. Controlling for GDP removes part of the effect being measured.
• (Also acceptable: unobservable confounders like diet, climate, or institutional quality; measurement heterogeneity in how countries record deaths; and the ecological fallacy of drawing individual-level conclusions from country-level data.)

The general lesson: the number of controls is not evidence of causal credibility. Twenty controls and no identification strategy is twenty controls and no identification strategy.

A4. You’ve controlled for a mediator. Education raises earnings substantially by getting people into better-paid occupations. By holding occupation fixed, you removed that channel. Your 0.04 is the direct effect — the return to education among people in the same job — not the total return to education.

Both numbers are meaningful. They answer different questions. The mistake isn’t the regression; it’s calling 0.04 “the return to education” when it’s “the within-occupation return to education.” Say which one you’ve estimated.

A5. Because bias and result are pointing in opposite directions.

If your estimate is biased upward and you still fail to find a significant positive effect, then the true effect is even smaller than the small thing you measured. The bias was working in favour of finding something, and it didn’t. That’s genuinely informative — much more so than a significant positive result, which is exactly what upward bias would produce anyway.

This is why the minimum wage example matters. If cross-state regressions are biased towards finding no employment harm, then a null result is weak evidence. But if the design is biased towards finding harm and you find none — that’s a real finding.

The general principle: always ask whether your bias runs with or against your conclusion. A result that survives a bias working against it is worth ten that agree with one.

• Bias = β₂ × δ₁. Sign both from theory. Multiply.
• β₂: does the omitted variable affect y?
• δ₁: does the omitted variable correlate with your x?
• Upward bias is the default case — good things cluster with good things
• You need no data to do this. You need economics.
• More controls ≠ less bias. Mediators remove real effects; colliders create fake ones
• Robust standard errors do nothing for this
• Bias has a direction for free. Magnitude costs a research design.