Employee engagement survey

What it is good for

• Summarise ordinal Likert items with medians, modes and stacked bar charts rather than only means.
• Test association between department and agreement with a chi-square test.
• Spot low-quality responses (straight-lining, very fast completion).
• Explain how unequal response rates can bias an overall engagement estimate (nonresponse bias) and correct it by weighting.

Methods: frequency tables, stacked bar chart, chi-square test of independence, Mann–Whitney test, Kruskal–Wallis test, Spearman correlation, Cronbach's alpha, post-stratification weights.

Data quality notes

• Item nonresponse: each Likert item has some nulls (2–6%); q5_workload has the most.
• 18 respondents straight-line (identical answers on all six items, no skipped items) and have completion times under 2 minutes.
• Response rates range from 45% (Operations) to 80% (Human Resources); the less engaged departments responded less, so the unweighted mean overstates engagement.
• Likert items are stored as integer codes 1–5; treat them as ordinal.

Provenance and licence

Respondent counts per department = headcount × response rate (see engagement_population). Each respondent has a latent engagement score ~ Normal(department mean + small tenure and work-mode effects, 1). Each item = latent score × item loading + item intercept + noise, cut into 1–5 at fixed thresholds, so items are positively correlated ordinal variables. 18 respondents straight-line (same answer on all six items) and finish very quickly. Each item is then skipped (null) with probability 2–6% (the workload item most often).

Synthetic data · CC0-1.0 (synthetic data generated by Statistics Quest) · generated deterministically (seed 20241011). It does not describe real people or organisations; any resemblance is coincidental.