The main issue here is the nature of the omitted variable bias. Wikipedia states:
Two conditions must hold true for omitted-variable bias to exist in
linear regression:
- the omitted variable must be a determinant of the dependent variable (i.e., its true regression coefficient is not zero); and
- the omitted variable must be correlated with one or more of the included independent variables (i.e. cov(z,x) is not equal to zero).
It's important to carefully note the second criterion. Your betas will only be biased under certain circumstances. Specifically, if there are two variables that contribute to the response that are correlated with each other, but you only include one of them, then (in essence) the effects of both will be attributed to the included variable, causing bias in the estimation of that parameter. So perhaps only some of your betas are biased, not necessarily all of them.
Another disturbing possibility is that if your sample is not representative of the population (which it rarely really is), and you omit a relevant variable, even if it's uncorrelated with the other variables, this could cause a vertical shift which biases your estimate of the intercept. For example, imagine a variable, $Z$, increases the level of the response, and that your sample is drawn from the upper half of the $Z$ distribution, but $Z$ is not included in your model. Then, your estimate of the population mean response (and the intercept) will be biased high despite the fact that $Z$ is uncorrelated with the other variables. Additionally, there is the possibility that there is an interaction between $Z$ and variables in your model. This can also cause bias without $Z$ being correlated with your variables (I discuss this idea in my answer here.)
Now, given that in its equilibrium state, everything is ultimately correlated with everything in the world, we might find this all very troubling. Indeed, when doing observational research, it is best to always assume that every variable is endogenous.
There are, however, limits to this (c.f., Cornfield's Inequality). First, conducting true experiments breaks the correlation between a focal variable (the treatment) and any otherwise relevant, but unobserved, explanatory variables. There are some statistical techniques that can be used with observational data to account for such unobserved confounds (prototypically: instrumental variables regression, but also others).
Setting these possibilities aside (they probably do represent a minority of modeling approaches), what is the long-run prospect for science? This depends on the magnitude of the bias, and the volume of exploratory research that gets done. Even if the numbers are somewhat off, they may often be in the neighborhood, and sufficiently close that relationships can be discovered. Then, in the long run, researchers can become clearer on which variables are relevant. Indeed, modelers sometimes explicitly trade off increased bias for decreased variance in the sampling distributions of their parameters (c.f., my answer here). In the short run, it's worth always remembering the famous quote from Box:
All models are wrong, but some are useful.
There is also a potentially deeper philosophical question here: What does it mean that the estimate is being biased? What is supposed to be the 'correct' answer? If you gather some observational data about the association between two variables (call them $X$ & $Y$), what you are getting is ultimately the marginal correlation between those two variables. This is only the 'wrong' number if you think you are doing something else, and getting the direct association instead. Likewise, in a study to develop a predictive model, what you care about is whether, in the future, you will be able to accurately guess the value of an unknown $Y$ from a known $X$. If you can, it doesn't matter if that's (in part) because $X$ is correlated with $Z$ which is contributing to the resulting value of $Y$. You wanted to be able to predict $Y$, and you can.
To prove this, start from the probability limit of the OLS estimator. Let $X$ denote the full matrix of regressors to be used, $[1,X_1,X_2]$, and let $e \equiv u + b_3 X_3$. Also, let $b$ be the parameters we are trying to estimate, i.e. $b = (b_0,b_1,b_2)$.
\begin{align*}
p\lim \hat{\beta} &= p\lim \left[ (X'X)^{-1}X'Y \right]
\\ &= p\lim \left[ (X'X)^{-1}X'Y \right]
\\ &= p\lim \left[ (X'X)^{-1}X'(Xb + e) \right]
\\ &= p\lim \left[ (X'X)^{-1}X'Xb \right] + p\lim \left[ (X'X)^{-1}X'e \right]
\\ &= p\lim \left[ (X'X)^{-1}X'X \right] b + p\lim \left[ (X'X)^{-1}X'(b_3 X_3 + u) \right]
\\ &= b + b_3 p\lim \left[ (X'X)^{-1}X' X_3 \right] + p\lim \left[ (X'X)^{-1}X'u \right]
\\ &= b + b_3 p\lim \left[ (X'X)^{-1}X' X_3 \right]
\\ &= b + b_3 \mathbb{E}(X'X)]^{-1} \mathbb{E}(X' X_3)
\end{align*}
Above, a key step is of course that $p\lim \left[ (X'X)^{-1}X'u \right] =0$, which happens because
$$ p\lim \left[ (X'X)^{-1}X'u \right] = (p\lim X'X)^{-1} p\lim (X'u) = [\mathbb{E}(X'X)]^{-1} \mathbb{E}(X'u) $$, since $\mathbb{E}(X'u)=0$, which holds because the original assumption is that each of the regressors are uncorrelated with $u$ (but not necessarily $e$).
Now we see that $p\lim \hat{\beta} \ne b$ whenever $\mathbb{E}(X'X_3) \ne 0$, that is whenever there is correlation between $X_1$ and $X_3$ or between $X_2$ and $X_3$.
Best Answer
You are correct on both accounts, but omitting the variable is still a very bad idea. Increasing your sample size is not going to 'fix' omitted variable bias. Consider the semi-classic example of drowning deaths and temperature (because people go to swimming pools when it's warm but not when it's cold).
We estimate one model:
$$ drowning.deaths = \alpha + \beta_1 temperature + \epsilon $$
We estimate a second model:
$$ drowning.deaths = \alpha + \beta_1 temperature + \beta_2 pool.in.area + \epsilon $$
If we increase our sample size for the first model, yes, we will reduce our standard errors, but absolutely our model will still suffer from omitted variable bias. Theory first! Remember that models like this are for hypothesis testing, and the 'best fit' model isn't always the correct model. No matter how small your standard errors are, and no matter how big your sample size is, if you've modeled it wrong, your results are not going to give you the 'right story.'
In the case of swimming pools, say you have 50 communities with swimming pools, and 50 communities without swimming pools, our first model is going to underestimate the relationship between drowning deaths and temperature-given-swimming-pool and overestimate the relationship between drowning deaths and temperature-given-no-swimming-pool. Thus, missing a really important piece of the puzzle (interaction in this specific hypothetical a plus).
If you really wanted to demonstrate a strong relationship between $drowning.deaths$ and $temperature$ (depending on the number of communities that had a pool), you could drop $pool.in.area$, and $\beta_1$ might be larger. But, that would be very bad science if you also knew there was a relationship between $drowning.deaths$ and $pool.in.area$.
In general, if a variable is plausibly related to your outcome, include it in your model. With maybe a few exceptions: some instrumental variable considerations (which are beyond the scope of this question) or severe multicollinearity between the predictor and the outcome.
On the sign of the omitted variable bias - that depends on some other factors, including both the relationship between the omitted variable and the outcome, and relationship between the omitted variable and other covariates. It could go either way.
To conclude, in a summary of a summary: increasing the sample size will not solve this problem, and you should include the variable unless there are other very compelling reasons not to.