ECON 371 Class Recording 11/12
Overview
The class distinguishes classical from nonclassical measurement error in regression: random error in an independent variable attenuates its coefficient toward zero, while random error in a dependent variable leaves the coefficient unbiased but increases its standard error. Stata simulations using Fitbit step counts and country BMI/GDP illustrate these results, and the 1936 Literary Digest poll, which mailed 10 million ballots and received 2.5 million, demonstrates how selection and nonresponse bias can overwhelm a large sample.
Key takeaways
- With classical error in an independent variable, the expected regression coefficient is the true coefficient multiplied by Var(X) / [Var(X) + Var(error)], so greater error variance means greater attenuation toward zero.
- Classical measurement error in a dependent variable preserves an unbiased coefficient when its conditional mean is zero, but increases standard errors and makes statistical significance less likely.
- Nonclassical measurement error is correlated with true X or included predictors, so its bias can be positive or negative and may even reverse the apparent relationship.
- The 1936 Literary Digest poll's 10 million mailings and 2.5 million responses did not make it representative: sampling from car registrations and telephone directories disproportionately captured wealthier voters.
- Nonresponse can create bias when response probability depends on the outcome or predictors; for example, dissatisfied customers may be more likely than satisfied ones to submit product reviews.
- The reported-data coefficient, the true-data coefficient, and Var(X) can be used to solve for unknown measurement-error variance by rearranging the classical attenuation equation.
Chapters
- The class reviews the distinction between observed values, written as X with a tilde, and the underlying true X values.
- With classical, random error in X, the estimated coefficient is multiplied by Var(X) / [Var(X) + Var(measurement error)], pulling it toward zero.
- As measurement-error variance grows, the attenuation becomes stronger; with no measurement error, the fraction equals one.
- A student dataset pairs self-reported daily steps with Fitbit-recorded steps, allowing the reporting error to be measured directly.
- Regressing GPA on reported steps produces a coefficient of about 0.016, compared with about 0.022 using Fitbit steps.
- The observed difference is broadly consistent with attenuation, though omitted-variable bias can keep the estimates from matching the theoretical formula exactly.
- Misreporting may depend on actual phone use: someone who unlocks a phone frequently may report using it less.
- For phone unlocks as X, the self-reported regression coefficient is about 0.006, while the coefficient using recorded values is about 0.003.
- The self-reported estimate is farther from zero, showing that the toward-zero result applies to classical error, not error correlated with true X.
- If observed Y equals true Y plus measurement error, the measurement error becomes part of the regression disturbance.
- When that error has conditional mean zero given the included X variables, it does not violate the zero-conditional-mean assumption, so the coefficient remains unbiased.
- The added noise increases the standard error, lowers the test statistic, and raises the p-value, making statistical significance harder to detect.
- A drug-use survey may underreport use among law-enforcement officers because disclosure could threaten their jobs, making reporting error related to an included X variable.
- For random error in Y, reported and actual step counts yield similar GPA coefficients, around 2.03 and a nearby value.
- For reported phone unlocks as Y, the GPA relationship differs substantially from the relationship using actual unlocks, illustrating that nonclassical error can shift estimates in either direction.
- A Stata exercise uses country-level BMI to predict GDP and compares simulated estimates with the coefficient from the original data.
- Adding random measurement error with variance 1 to BMI produces an average estimate near the theoretical attenuated value of about 49,700.
- Repeating the simulation across more samples brings the average closer to theory, while using only 100 simulations leaves more sampling variation.
- Increasing the BMI measurement-error variance from 1 to 100 pushes the simulated GDP coefficient closer to zero, to roughly 2,700.
- Adding random error to GDP instead leaves the coefficient close to its original value of about 60,889.
- In the GDP example, very large random Y error raises the standard error to roughly 11,000, demonstrating its effect on precision rather than the coefficient.
- The simulation makes GDP measurement error depend on BMI, directly violating the condition that the error has mean zero conditional on X.
- One version yields an estimate near 60,631, while changing the error function can move the estimate to the opposite side of the true effect.
- The direction of bias from nonclassical measurement error cannot be inferred from the classical attenuation formula.
- A quiz problem gives a reported-steps coefficient of 0.016, a true-data coefficient of 0.023, and true step-count variance of 5.7.
- Rearranging the attenuation equation gives measurement-error variance = (true coefficient / reported coefficient) × Var(X) − Var(X).
- Substituting the rounded values yields an estimated measurement-error variance of about 2.5.
- The value from the regression using reported data is the coefficient affected by measurement error; the actual-data coefficient is the unattenuated reference.
- Measurement error can arise from rounding company reports, not only from survey self-reports.
- Height measurements illustrate how recording protocols matter: NBA measurements may differ from earlier university measurements because the procedures are more standardized.
- The Literary Digest mailed 10 million survey cards and received about 2.5 million responses, yet incorrectly predicted a landslide for Republican Alf Landon over Franklin D. Roosevelt.
- Its sampling lists used telephone directories and car registrations, overrepresenting wealthier households and excluding people without cars or telephone service.
- The poll's huge sample could not correct the selection bias: voting patterns differed across income and urban-rural groups, and the excluded groups included many Roosevelt voters.
- A follow-up to the Literary Digest example found that Landon voters were more likely than Roosevelt voters to return surveys, creating nonresponse bias even within sampled groups.
- Amazon-style product reviews can skew negative because dissatisfied customers may be more motivated to post than satisfied customers.
- Credit-card purchase datasets also have selection limits because cash purchases, including some illicit or sensitive transactions, are not recorded.
- The next topics are missing data and panel data, followed by instrumental variables; the posted quiz is due Tuesday and the problem set is due the following day.
Transcript excerpt
0:01 So, we'll kind of pick up where we left off. Um, we were talking about measurement error. So, I'm going to kind of review that a little bit and then talk about different types of measurement error we could have. So, the first thing that we'll do is I'm looking at that homepage. Going to come down here to the current week we're in. I'm actually going to use this measurement error D file and this Excel file. So, open that up here once again. A little annoying since I had to download all of these. Now, it's not going to let me access the downloads. Okay. Uh let's see if this will There we go. Okay. So, go here and then
0:55 looks like it's catching up. I'm going to open up this file that was on Canvas. So, this is the do file I put up there. Just going to run this kind of scaling the variables a little bit so we can kind of look at some interesting stuff. So we had talked about our measurement error when it was in our independent variable. Right? So we said if this is completely like random measurement error or what we would call non-classical, right? Then we said the estimate that we're going to end up seeing maybe [clears throat]
2:04 See if it matches up. All right. Switch gears a little bit here. I'll just kind of do this from the the do file and and I think maybe my pencil. Um, so we got this data set. It was kind of interesting. And I forget where exactly it was from, but they asked students to report different things, and then they tracked them with a Fitbit. So, I think this was like the number of steps they did every day, the number of hours they
This is the start of the transcript. Save this video free for the whole transcript, searchable and time-coded.
In this course
Related notes
Summary, takeaways, and chapters were generated by AI from the video's transcript and may contain errors. The video belongs to its creator, Professor Lantis.