RES 720 Week 8 Quantitative Analysis Plan Example

A Quantitative Analysis Plan for a Dissertation on Early Warehouse Turnover

[Student Name]

University of Phoenix

RES/720: Statistical Research Methods and Design II

Week 8 Assignment

[Instructor Name]

[Date]

The learner, the company, its associates and all data are composites written for a model paper.

Over seven weeks, the composite HR director at a nine-warehouse Louisville logistics company ran preliminary analyses of early turnover. Her dissertation will extend those analyses to a third year of hires, a full 12 months of follow-up and the completed onboarding rollout. This plan specifies, before those data are delivered, how every question will be answered.

Refined Research Questions and Hypotheses

Course results led to two refinements: the outcome becomes time to quit over 12 months rather than a 90-day yes or no, and the shift-by-program interaction becomes a stated hypothesis.

H1: Rotating-shift associates have a higher hazard of quitting within 12 months than fixed-day associates.

H2: Associates hired under the new onboarding program have a lower hazard of quitting than those hired under the old program.

H3: The onboarding program's effect on 90-day productivity is larger for night and rotating associates than for day associates.

H4: Longer commutes, lower pay and less experience are associated with a higher hazard of quitting.

Exploratory questions: whether incident risk in the first month falls after safety coaching is added and whether night supervisor staffing relates to the program's effect.

Data and Variables

Data cover about 5,400 associates hired over three years. Outcomes: days from hire to quitting, censored at 365 days or at the data cutoff for later hires; third-month pick rate. Predictors: shift, three categories; onboarding, two; pay, dollars per hour; commute, miles; experience, months; age, years; assessment score at hiring. Context: warehouse, nine sites; hiring quarter; rollout period.

Analysis for H1, H2 and H4: Time to Quit

Singer and Willett (2003) explain that survival methods model the timing of an event while using information from people who have not yet experienced it, whose times are censored, and that discarding or misclassifying censored cases biases estimates. The primary model is a Cox proportional hazards regression with shift, onboarding, pay, commute, experience, age and assessment score, stratified by warehouse so each site has its own baseline hazard. Hazard ratios with 95 percent confidence intervals will be reported. Proportional hazards will be checked with Schoenfeld residuals; if shift violates the assumption, as is plausible if rotating shifts drive quits mainly in the first weeks, the model will allow its effect to differ before and after 90 days.

Analysis for H3: Productivity

Raudenbush and Bryk (2002) describe how hierarchical linear models handle individuals nested in groups, separating within-group from between-group variation and producing correct standard errors when observations within groups are correlated. The H3 model is a linear mixed model with productivity as the outcome; shift, program and their interaction as fixed effects; rollout period as a fixed effect; and warehouse as a random intercept. Because nine warehouses is a small number of groups, the plan includes a fixed-effects version as a check, and the interaction is tested with a small-sample correction to the degrees of freedom.

Nine warehouses are enough to hold sites constant but not enough to generalize about sites, so the plan does both and says which is which.

Data Screening

Before analysis, the learner will check ranges and impossible values, such as hire dates after quit dates; examine distributions of continuous predictors; identify duplicate records for rehired associates and keep the first hire; and document every change in a log.

Missing Data

Preliminary files show missing commute data for about 6 percent of associates and missing assessment scores for 9 percent, mostly at two warehouses that changed systems. Schafer and Graham (2002) reviewed methods for missing data and concluded that deleting incomplete cases can bias results unless data are missing completely at random, while maximum likelihood and multiple imputation perform well under the weaker assumption that missingness depends only on observed variables. The plan uses multiple imputation with 20 data sets, including outcomes and warehouse in the imputation model, and reports complete-case results as a sensitivity check.

Assumptions and Fallbacks

Cox model: proportional hazards, checked by residuals; fallback, time-varying effects.

Mixed model: normal residuals and equal variance, checked by plots; fallback, robust standard errors or a log transformation of productivity.

Linearity of continuous predictors: checked with spline terms; fallback, keep splines if they improve fit substantially.

Collinearity: VIFs; pay and warehouse are expected to overlap, and the plan reports pay effects as within-warehouse associations.

Multiple Testing

The four hypotheses yield seven confirmatory tests. The Benjamini-Hochberg procedure will control the false discovery rate at .05 across them. Exploratory analyses will be reported with uncorrected p values, clearly labeled, and treated as hypotheses for future work.

Power

With about 1,800 quits expected among 5,400 associates, the chance of detecting a hazard ratio as small as 1.3 for rotating versus day shifts is above .90. For H3, simulation based on the course data, which preserves the clustering by warehouse, gives power of about .80 for an interaction of the size observed in Week 3, so the dissertation can test it but could miss a smaller one.

Robustness Checks

The plan lists checks in advance: results without the two warehouses that changed systems; results excluding rehires; a logistic model for 90-day quitting to connect with the course results; and a model adding assessment score to test whether candidate quality explains shift differences.

Reporting

Each confirmatory result will be reported with its estimate, 95 percent interval, p value before and after correction and effect in practical terms, such as the predicted 12-month retention for typical associates. Kaplan-Meier curves by shift and program will display survival over time.

Why Not Simpler Alternatives

A committee may ask why the plan does not simply repeat the course's logistic regression with 12-month quitting as the outcome. That approach would treat an associate who quits on day 20 the same as one who quits on day 350 and would drop or misclassify later hires whose 12 months are incomplete. Survival analysis uses the timing and keeps every associate. Similarly, a single regression ignoring warehouses would understate standard errors, which the mixed model corrects.

Ethics and Data Protection

The company will supply de-identified records under a data use agreement, with codes replacing names and home addresses converted to commute distances before transfer. Results for any subgroup smaller than 20 associates will not be reported. The university's ethics board will review the plan before data delivery, and the company has agreed that findings may be published without naming it.

Timeline and Software

Data delivery in month 1, screening and imputation in months 2 and 3, confirmatory analyses in months 4 and 5, robustness and exploratory work in month 6. Analyses will be run in R, with code shared with the committee.

Conclusion

The plan turns preliminary course findings into four confirmatory hypotheses, matches them to survival and mixed models suited to time-to-event outcomes and nested data and specifies screening, missing data, assumptions, multiple testing, power and reporting in advance. Exploratory questions are named as such. A committee member could carry out every step without further instructions.

References

Raudenbush, S. W., & Bryk, A. S. (2002). Hierarchical linear models: Applications and data analysis methods (2nd ed.). Sage.

Schafer, J. L., & Graham, J. W. (2002). Missing data: Our view of the state of the art. Psychological Methods, 7(2), 147-177. https://doi.org/10.1037/1082-989X.7.2.147

Singer, J. D., & Willett, J. B. (2003). Applied longitudinal data analysis: Modeling change and event occurrence. Oxford University Press.