Who do we ask?: answer key
Introduction to Social Research Methodology
Task 1: worked arithmetic
1a. Proportionate stratified sample
Sampling fraction: n / N = 90 / 450 = 1/5 (20%).
Apply the fraction to each stratum:
| Stratum | Population | Calculation | Stratum sample |
|---|---|---|---|
| Retail | 330 | 330 × 1/5 | 66 |
| E-commerce | 70 | 70 × 1/5 | 14 |
| HQ | 50 | 50 × 1/5 | 10 |
| Total | 450 | 90 ✓ |
The check matters: stratum samples must sum to the target n, and here they do without rounding problems. (If a group works with percentages — retail 73.3%, e-commerce 15.6%, HQ 11.1% of 90 — accept it, but the sampling-fraction route is cleaner and less error-prone.)
Q4 (what stratifying guarantees): the division composition of the sample exactly matches the workforce — sampling error on the stratifying variable is reduced to zero. A simple random sample of 90 would only match it on average: by chance it could contain 8 or 20 e-commerce staff. Credit any answer that names the guarantee-by-construction versus correct-on-average distinction.
Extension point worth raising in plenary: 14 e-commerce respondents is thin ground for a separate claim about e-commerce. If the board wants division-level findings, a disproportionate design that over-samples the small strata (and reweights in analysis) is the professional move. Students who spot this unprompted are applying the “plan around the smallest reportable subgroup” rule — reward it.
1b. What counts as a response?
- Completes only: 96 / 300 = 32%.
- Completes + partials: (96 + 12) / 300 = 108 / 300 = 36%.
Should partials count? There is no single right answer; mark the reasoning. A defensible rule: count a partial as a response for the items it answered (so item-level analyses use them) but report the headline completion rate on completes only — or count partials only if they passed a defined point (e.g. completed the key outcome section). What is not defensible is deciding after seeing the data, because then the researcher can pick whichever rate flatters the study. The rule should have been fixed in advance, before fieldwork closed — this is the answer the question is fishing for, and students should state it explicitly.
Q4 (the missing 200): the response rate is dangerous through nonresponse bias, not through its absolute level. If the 96 responders are systematically different from the 204 silent invitees on what is being measured — e.g. the engaged answer and the disengaged do not — the estimate is biased regardless of whether 32% is “acceptable”. Good answers mention comparing responders with the frame on known characteristics (division, tenure) and chasing reminders; excellent ones note that at Meridian the disengaged non-responders are precisely the likely leavers, so the bias runs exactly against the study’s purpose.
1c. Systematic sampling
Sampling interval: k = N / n = 450 / 90 = 5.
Procedure: (1) obtain the ordered roster of 450; (2) choose a random starting point between 1 and 5 inclusive; (3) select that element and every 5th element thereafter (start 3 → persons 3, 8, 13, …, 448), yielding exactly 90. The random start is what makes this a probability design: it must be chosen by chance, not by the researcher, otherwise whoever ordered the list effectively chose the sample and the equal-chance property is lost. Answers must include the random start and the reason for it to earn full credit.
Ordering check — periodicity: the danger is a recurring structure in the list whose cycle aligns with the interval. Alphabetical order is harmless. Trouble examples (any one suffices): a roster organised in five-person teams with the team leader listed first (k = 5 samples all leaders or none); a rota where every 5th entry is a weekend shift; a list ordered store-by-store in fixed-size blocks. The soldiers-by-squad example from the session is fine if students adapt it.
Task 2: defensible designs, with acceptable alternatives
Mark against the three required elements — population, frame, and what the design licenses — rather than against the “model” choice. A well-argued alternative earns full credit; a correct label with no reasoning does not.
Brief A — why do leavers leave? Model answer: the population is everyone who left in the past year — and no complete frame exists (HR has contact details for only some, which is a frame–population gap, not a frame). The realistic design is purposive/judgmental sampling from the partial list, extended by snowball referrals (leavers know other leavers), with the honest caveat that findings are exploratory, not population estimates. Acceptable alternatives: treating HR’s partial list as the frame and sampling randomly within it is defensible only if the student flags that reachable leavers may differ systematically from unreachable ones (e.g. those who left on good terms are easier to find); a census of all reachable leavers is also sensible given the small numbers. Not acceptable: any answer sampling from the current-staff roster — that is the frame–population trap the session set up, and it excludes the entire population of interest.
Brief B — testing the new rota app. Model answer: the population is store staff across all 32 stores, and the app’s usability may vary by store size, format, or region. A cluster (multistage) design fits the operational reality: sample a handful of stores, then staff within them — piloting an app store-by-store is how rollouts actually work, and no fieldwork team can visit 32 sites. Stratified sampling of staff from the roster (by store type or division) is an equally strong answer if fieldwork is remote/CAWI. Acceptable alternative: purposive selection of deliberately contrasting stores (biggest/smallest, best/worst connectivity) for a qualitative usability test — a fine design if the student frames the purpose as diagnosis rather than estimation. Watch for: students should note the cluster design’s cost (extra sampling error per stage) or the purposive design’s limit (no generalisation); either observation marks a strong answer.
Brief C — the courier experience. Model answer: population is the gig couriers; no frame exists and none can be built from HQ records — this is the textbook trigger for snowball sampling, starting from the few couriers who can be contacted (e.g. via the dispatch app) and following referrals. The brief asks for depth of understanding, so the data collection is qualitative (interviews), which suits a small snowball sample; conclusions are exploratory and describe the network reached, not “all couriers”. Acceptable alternative: recruiting through the delivery app itself (a channel-based convenience/volunteer approach) — credit it if the self-selection bias is named. Not acceptable: any probability design — there is nothing to draw it from; students proposing one have missed the frame requirement.
Brief D — the pulse check. Model answer: the honest tension in this brief is speed versus warrant, and the best answers surface it. Within the constraint, the defensible options are: a short CAWI questionnaire to the full roster (a census attempt — with 450 staff and e-mail, inviting everyone is faster than sampling, and the risk shifts to nonresponse bias, which should be named); or a small stratified random sample with reminders, which trades breadth for a cleaner claim. A quota design mirroring divisions is acceptable for speed if the non-random fill is acknowledged. The trap: an open “click here” link in the newsletter — pure self-selection; students who choose it must explain why it would over-hear the happiest and angriest and miss the drifting middle, at which point it stops being a defensible choice and becomes a diagnosed one. Reward especially: answers that say what they would tell the CEO about the confidence the pulse check does and does not support — that is the session’s cardinal-sin point applied under deadline pressure.
Task 3: annotated model recruitment plan
Channel
Strong plans combine two or three of: exit-interview consent lists / personal e-mail retained with consent (most direct; bias — only covers those who consented, likely the amicable leavers; note GDPR requires that consent); LinkedIn / professional networks (reaches those who cut ties; bias — over-represents white-collar and HQ leavers, under-represents shop-floor staff who may not maintain profiles); referrals from current staff or other leavers, i.e. snowballing (reaches the otherwise unreachable; bias — travels along friendship networks); social media (broadest; severe self-selection). Full credit requires naming a bias per channel, not just listing channels. Plans relying on intranet or company e-mail fail the task — leavers cannot see either.
Model invitation (82 words), with annotations
Dear [name],
I’m [researcher], from the independent research team working with Meridian. You’re one of a small number of former employees we’re inviting because you left in the past year — and we want to understand why people leave, directly from those who did. The online survey takes about 10 minutes. Your answers are confidential: results reach Meridian only in combined form, never linked to you, with no effect on references. Findings will shape how Meridian treats staff; we’ll send you a summary in March.
Annotations — the four jobs, mapped:
- Who is asking and why this person: named researcher, independent team (not Meridian management — frankness and voluntariness both improve), and the selection rationale (“because you left in the past year”) — being told why you specifically were chosen raises response.
- Topic and honest time estimate: the subject is stated plainly, not euphemised, and “about 10 minutes” is a concrete, keepable promise.
- What happens to answers: “confidential… combined form, never linked to you” — note this correctly promises confidentiality, not anonymity (the researcher knows who responded); penalise invitations that promise “anonymous” while addressing the respondent by name. The reference reassurance speaks to this population’s specific fear.
- What it is for: a concrete purpose plus a results promise with a date — the most-forgotten element; its presence distinguishes excellent invitations.
Marking the word count: 60–90 words is a hard constraint of the task; invitations far outside it lose the brevity point the constraint exists to teach. Tone should be neutral-warm, not corporate (“we value your feedback”) and not grovelling.
Incentive
Either decision is markable; the justification carries the credit. Model answer: a guaranteed small incentive (e.g. a 50 zł voucher or a charity donation per complete): leavers owe Meridian nothing, are being asked for a favour by an organisation they chose to leave, and a guaranteed-for-all incentive produces the most reliable lift without the lottery’s prize-chasers. Students should name a trade-off: e.g. the incentive may slightly over-recruit the money-motivated, or a Meridian-branded voucher may feel unwelcome to bitter leavers (a cash-equivalent or donation option handles this). “No incentive” is defensible if argued (topic salience is high for recent leavers; some genuinely want to be heard) — but for interviews rather than a 10-minute CAWI, paid time should be offered, and answers proposing interviews without payment should lose the point.
Safeguard against hierarchy pressure
Any one concrete, operational safeguard suffices; vague gestures (“make it voluntary”) do not. Strong options: invitations sent and data held by the independent/external team, with Meridian never seeing who was invited or who responded; an explicit written statement that participation and answers cannot affect references or rehiring, with references handled by a process that has no access to the study; no manager involvement in choosing whom to contact (former store managers must not nominate “suitable” leavers — this is both a coercion and a selection-bias safeguard); reporting only in aggregates with a minimum cell size so no small team’s leavers are identifiable. The link back to session 4 should name a principle: this is voluntary informed consent (freedom from coercion) and/or confidentiality in operation — one sentence naming the principle earns the point.
Timing and plenary guidance
Task 1 is deliberately quick arithmetic; if pairs finish early, push them to the 1a extension (disproportionate stratification). Task 2 generates the best discussion on Briefs A and D — take one group that fell into the roster trap on A (or invent one) and one that chose the newsletter link on D, and let the room diagnose them. For Task 3, hearing two or three invitations read aloud is more instructive than collecting them: the room hears immediately which ones they would delete. Close by pointing forward: the invitations promise “about 10 minutes” — next session is about making the questionnaire worth those minutes.