Skip to main content

2  Daily 14 — Oct 8

2.1 Class Performance

Students: 191 | Content mean: 1.62 / 2 | Median: 1.5 | SD: 0.45 | Mean daily score: 9.62 / 10

Content scores ranged from 0 to 2 out of 2. (Your daily score adds the 8-point attendance credit for submitting; each question is worth 1 of the remaining 2 points.)

NoteTwo questions, and no Daily 13 chapter

Only the first two questions on the slide were asked on Oct 8, so this daily is graded on two. Daily 13 (Oct 6) has no chapter: no questions were asked that day, and it was scored as full credit.

Q1 was close to universal: 97 percent of the class named both ways the bias can vanish. Q2 carried nearly all the lost credit, and its most common error is worth a careful look. Students knew both ingredients of the bias and still got its direction wrong.

2.2 Score Distribution

2.3 Performance by Question

ImportantThe pattern worth taking from this daily

The direction of omitted-variable bias is the product of two signs.

\[\text{bias} = \underbrace{\beta_{exper}}_{\text{effect on wages}} \times \underbrace{\delta_{exper,\,educ}}_{\text{relation to schooling}}\]

On Q2, about one student in seven wrote both signs correctly. They said experience is negatively correlated with schooling and positively with wages, and then called the bias upward. They had both ingredients and skipped the last step: a positive times a negative is negative, so the bias is downward.

Table 5 shows the same thing without any algebra. Adding exper raises the schooling coefficient from 0.052 to 0.093, so leaving it out had pushed the estimate down. When the table and your sign product disagree, one of them has been misread.

2.4 Questions

2.4.1 Q1: In \(y_i = \beta_0 + \beta_1 x_{1i} + \beta_2 x_{2i} + u_i\), if \(x_2\) is omitted, \(\hat\beta_1\) is biased unless \(\beta_2 =\) ___ or \(x_1\) and \(x_2\) are ___.

\(\beta_2 =\) 0, or \(x_1\) and \(x_2\) uncorrelated.

The omitted-variable bias formula says

\[E(\hat\beta_1) = \beta_1 + \beta_2\, \delta_{21},\]

where \(\delta_{21}\) is the slope from regressing \(x_2\) on \(x_1\). The bias is the product \(\beta_2\, \delta_{21}\), so it vanishes when either factor is zero: the omitted variable does not affect \(y\) (\(\beta_2 = 0\)), or it is unrelated to the included one (\(\delta_{21} = 0\)).

There were very few errors here, and they were scattered rather than systematic:

  • A number from the regression output in place of the condition. The blanks ask when the bias disappears, not what a coefficient is.
  • “Unbiased” in the second blank. That names the result, not the condition on \(x_1\) and \(x_2\).
  • A look-alike word. “Uncorrected” is a different word from “uncorrelated”, not a misspelling of it.

“Zero” written out, “independent”, a slashed zero, and creative spellings of uncorrelated all cost nothing.

2.4.2 Q2: Table 5 shows that omitting exper biased the educ coefficient ___, because exper is ___ correlated with educ and ___ correlated with log wages.

Biased downward: exper is negatively correlated with educ and positively correlated with log wages.

\[E(\hat\beta_{educ}) = \beta_{educ} + \underbrace{\beta_{exper}}_{+}\cdot\underbrace{\delta_{exper,\,educ}}_{-}\]

At a given age, more years in school leave fewer years of work experience, so the correlation is negative. Experience raises wages, so its coefficient is positive. A positive times a negative is negative, so leaving exper out pulls the schooling estimate down. That is the drop Table 5 shows between its columns.

  • “Upward” with both signs right. The most common error; see the callout above. Several students wrote “downward” first and then changed it.
  • Experience positively correlated with schooling. The intuitive guess, but backwards: at a given age, more years in school means fewer years at work. Paired with “downward”, it contradicts itself, because positive times positive is upward. Paired with “upward”, it is consistent but wrong about the data.
  • The two signs swapped. Some students wrote negative with wages and positive with schooling next to a correct “downward”. The product comes out right, but each sign is attached to the wrong relationship, so this earned half credit.
  • A number in the direction blank, usually the size of the change in the coefficient. “Biased ___” asks which way, not how much.
  • A variable name or a yes/no in the sign blanks, such as “exper”, “is” or “not”. “Exper is ___ correlated with educ” asks for the sign of the correlation.
  • Non-directions in the first blank, such as “biased”, “right” or “error”.

“Down”, “toward zero”, “negatively” for the direction, and “inversely” for the negative sign all earned credit.

2.5 What the Class Did Well

The conditions for omitted-variable bias are secure. Nearly everyone named both ways it can vanish: the omitted variable has no effect on \(y\), or it is unrelated to the included \(x\).

Many of the Q2 misses were one step from right. The students who wrote “upward” with both signs correct had done the hard part, reading two relationships off the data. What remained was a multiplication.

2.6 What to Review

  1. Bias = effect × correlation. Write down both signs, then multiply them. Positive times negative is negative.
  2. Check the direction against the table. If adding a control raises a coefficient, leaving that control out had pushed the estimate down.
  3. Schooling and experience at a given age. Every year in school is a year not working, so the two move in opposite directions.
  4. “Biased ___” asks for a direction, not a magnitude.
NoteHow this was graded

Every paper was scored twice, independently, against the same published rubric, by graders who could not see each other’s scores or your name. The two passes agreed on more than 99% of all scores; the two disagreements were each reviewed a third time against the rubric and settled with a written reason. Submitting the daily earns 8 of 10 regardless of content.