| Course | GB892 Doctoral Project III: Data Analysis |
|---|---|
| Unit | Unit 4 |
| Paper type | Assumption check summary |
| Length | About 910 words, 6 pages double-spaced |
| Format | APA 7 student paper |
| School | Purdue University Global |
| Program | MBA |
| Updated | October 2026 |
Free sample paper for GB892 Unit 4
Two of Six Assumptions Fail: A Unit 4 Assumption Check Summary for a Dock Appointment Comparison
[Student Name]
Purdue University Global
GB892: Doctoral Project III: Data Analysis
Unit 4 Assumption Check Summary
[Instructor Name]
[Date]
Haverly Logistics, its centers and every statistic are composite and illustrative; the methods sources cited are real.
Why Check Before Modeling
A difference-in-differences estimate is credible only if certain conditions hold. Checking them after seeing the estimate invites a researcher to look harder at assumptions when the result is disappointing and less hard when it is pleasing. This summary records each check, its result and its consequence before the primary model is run. Angrist and Pischke (2009) describe the parallel trends assumption as the heart of the design, but the other five matter too.
Assumption 1: Parallel Pre-Period Trends
What it requires. In the absence of the portal, dwell at adopting and comparison centers would have followed parallel paths. This cannot be tested directly for the post-period, but pre-period trends can be compared.
Check. Monthly mean log dwell for each group was regressed on month for the nine months before adoption, and the difference in slopes was estimated.
Result. Pooled across the nine months, the difference in slopes was 0.003 log points a month, with a 95 percent interval of minus 0.005 to 0.011. The groups moved together. But one comparison center, DC-09, showed a drop of 0.112 log points relative to the other comparison centers from month five, when it added a second receiving shift.
Verdict. Holds for the groups as a whole; fails for one comparison center from month five.
Consequence. The primary model keeps DC-09, since the pooled trend holds, but a sensitivity model excludes it. If DC-09's earlier improvement made comparison centers look better in the post-period, excluding it should make the estimated effect of the portal larger.
Assumption 2: Stable Composition
What it requires. The kinds of visits at each center did not change in ways related to adoption, since a shift toward faster loads at adopting centers could look like an effect of appointments.
Check. The shares of floor-loaded visits and night arrivals were compared before and after month ten in each group.
Result. At adopting centers, the floor-loaded share moved by 0.1 percentage points and the night share by minus 0.2 points.
Verdict. Holds.
Assumption 3: The Outcome's Shape
What it requires. The model's errors should not be dominated by a few extreme values.
Check. The skewness of dwell in minutes and in log minutes was computed.
Result. Skewness was 1.77 in minutes, a long right tail of very long waits, and 0.26 after logging.
Verdict. Fails for raw minutes.
Consequence. The primary outcome becomes log dwell, so estimates are proportional changes. Results will be converted back to minutes at the adopters' pre-adoption median for interpretation. The originally planned model in raw minutes will also be reported, so that readers can see what the change of outcome did.
Assumption 4: A Negative Control That Should Not Move
What it requires. Appointments change when trucks arrive and how long they wait for a door; they should not change how long a trailer takes to empty once at the door. Unload time can therefore serve as a negative control outcome. Lipsitch et al. (2010) describe how an outcome that the treatment should not affect can reveal hidden confounding: if it moves, something other than the treatment changed.
Check. The difference-in-differences model was run with unload time as the outcome.
Result. Unload time moved by 0.5 percent, with an interval spanning zero.
Verdict. Holds. No sign that something other than appointments changed operations at adopting centers.
Assumption 5: No Anticipation
What it requires. Centers did not change practices before month ten in anticipation of the portal.
Check. Dwell in months eight and nine at adopting centers was compared with months one to seven.
Result. No change beyond the pooled trend.
Verdict. Holds.
Assumption 6: Enough Clusters for Inference
What it requires. Standard errors account for correlation within centers over time.
Check. The number of clusters was counted: twelve centers, six treated.
Result. Twelve clusters are too few for conventional clustered standard errors to be reliable. Bertrand et al. (2004) showed that such errors can greatly overstate precision in difference-in-differences settings with few groups, and Cameron et al. (2008) showed that a wild cluster bootstrap performs much better with as few as six to ten clusters.
Verdict. Conventional inference fails.
Consequence. All intervals will be computed with a wild cluster bootstrap.
Two assumptions failed, and both failures were written down before the estimate existed, which is what makes the estimate worth trusting.
What the Checks Change Together
The two failures change the analysis in complementary ways. Logging the outcome reduces the influence of the longest waits, which tends to make the estimate smaller than a model of raw minutes would give. Excluding DC-09 in a sensitivity model removes a comparison center whose own improvement would make the portal look less effective, which tends to make the estimate larger. Reporting both versions lets a reader see the range the checks produce.
A Figure for the Reader
A figure plots the monthly difference in mean log dwell between adopting and comparison centers across all eighteen months, with month ten marked. The pre-period differences hover near zero, the DC-09 step is visible as a small shift in month five, and the post-period differences fall below zero.
References
Angrist, J. D., & Pischke, J.-S. (2009). Mostly harmless econometrics: An empiricist's companion. Princeton University Press.
Bertrand, M., Duflo, E., & Mullainathan, S. (2004). How much should we trust differences-in-differences estimates? The Quarterly Journal of Economics, 119(1), 249-275. https://doi.org/10.1162/003355304772839588
Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2008). Bootstrap-based improvements for inference with clustered errors. The Review of Economics and Statistics, 90(3), 414-427. https://doi.org/10.1162/rest.90.3.414
Lipsitch, M., Tchetgen Tchetgen, E., & Cohen, T. (2010). Negative controls: A tool for detecting confounding and bias in observational studies. Epidemiology, 21(3), 383-388. https://doi.org/10.1097/EDE.0b013e3181d61eeb
What the GB892 Unit 4 assignment asks for
Unit 4 in GB892 asks you to check the assumptions your analysis depends on before running the main model. Explain why checking first matters. For each assumption, state what it requires, the check you ran, the result in figures, a verdict and the consequence for the analysis. Cover trends before treatment, including whether any single unit departs from its group, composition, the distribution of the outcome, any negative control outcome, anticipation and the number of clusters for inference. Where an assumption fails, change the analysis or add a sensitivity model, and say which way the change should move the estimate. Explain how the checks work together. Provide a figure that lets a reader see the pre-period pattern. Cite methods sources for each check. About four pages built around one table, with a row per assumption, is a common format.
How this GB892 Unit 4 example is built
The summary explains that checks run after estimation invite selective scrutiny, citing Angrist and Pischke on parallel trends. The pooled pre-period slope difference is tiny, but DC-09's step leads to a sensitivity model. Composition shares barely move, and a margin note explains why composition is checked separately. Skewness of 1.77 fails, so log dwell becomes primary. Unload time, a negative control in Lipsitch and colleagues' sense, moves 0.5 percent. Anticipation is absent. Twelve clusters fail conventional inference, so Bertrand and colleagues and Cameron and colleagues justify the wild bootstrap. The highlighted sentence explains why recording failures first matters, a section explains how the changes pull in opposite directions, and a figure shows the monthly gap.
GB892 Unit 4 rubric: what earns full points
The GB892 assumption check earns its points by changing the analysis where needed. Instructors give credit for checks run and recorded before the primary estimate. Each assumption should have a stated requirement, a check, a numerical result and a verdict. Detecting a single unit's departure within a passing group test is especially valued. Failures earn credit when they lead to a concrete change or sensitivity model, with the expected direction of its effect. A negative control outcome shows sophistication. Recognizing that few clusters require special inference is expected at doctoral level. A figure helps the reader judge pre-period trends. Methods sources should be cited for each check, and the table should make every verdict easy to find.
GB892 Unit 4 help from the desk
Many GB892 assumption checks report that parallel trends hold and move on, missing a single unit that departs from its group. Others report a failed check without changing anything. Students often check assumptions only after seeing the estimate. Skewed outcomes are modeled in raw units without comment. Negative controls are rarely used, though they are a cheap test for hidden changes. Inference with few clusters is treated as routine. Check before estimating, give each assumption a requirement, check, result and verdict, look inside pooled tests, change the analysis when a check fails, say which way each change should move the estimate, add a negative control and use inference suited to few clusters. If your model is ready, our team can help you run and write the checks.
Get a GB892 Unit 4 example written to your instructions
Share what your analysis plan assumed, the checks you have run or intend to run, and any results that worried you, along with the Unit 4 prompt and rubric. A summary comes back table-first, a verdict and consequence per assumption, with failures named rather than buried. Your first custom sample carries no cost; 24-48h. The paper above is an original model document written by our desk, not a submitted student paper and not an official Purdue University Global document.
More GB892 papers and MBA samples
- GB892 Unit 1: Discussion Board Post
- GB892 Unit 2: Data Preparation Record
- GB892 Unit 3: Codebook
- GB892 Unit 5: Descriptive Results Table
- GB892 Unit 6: Seminar Reflection
- GB892 Unit 7: Thematic Analysis
- GB892 Unit 8: Inferential Results Write-up
- GB892 Unit 9: Departure Log
- GB892 Unit 10: Results Chapter Draft
- GB891 Unit 4: Recruitment Plan
- GB500 Unit 4: Financial Statement Reading
- GB890 Unit 4: Closest Study Appraisal
- GB512 Unit 4: Bad-News Business Letter
GB892 Unit 4 questions, answered
What should happen when a key assumption fails?
State the failure, then change what the analysis claims or how it estimates, in proportion to the problem. A skewed outcome may call for a transformation; a comparison unit on a different trend may call for a sensitivity model without it. Abandoning the design is rarely required. What committees resist is a failure noticed and then quietly left out of the write-up.
Is a pooled pre-trend test enough to support parallel trends?
Usually not on its own. A pooled test averages across units and can pass while one unit diverges sharply, which is what DC-09 does in this sample. Plotting each unit's series against the comparison average, and naming any unit that departs, gives a reader far more than a single p value. Your section may specify which evidence it expects.
Why use a negative control outcome?
Because it tests the design rather than the hypothesis. An outcome the intervention should not affect, unload duration here, ought to show no change. If it moves, something else changed at the treated sites at the same time, and the main estimate would absorb that change. A near-zero result does not prove the design sound, but it removes one competing explanation.
Where can I find a free GB892 Unit 4 sample paper?
A complete GB892 Unit 4 assumption check summary is on this page, testing six assumptions behind a difference-in-differences comparison and recording two failures before estimation.
What is a negative control outcome?
It is an outcome the intervention should not affect. If it changes anyway, something other than the intervention probably changed, which warns of hidden confounding.