Which sampling procedure to use for a patient-satisfaction study — and why

Recommended method: Stratified random sampling (probability sampling).

Why:

1. Representative coverage of key subgroups. In an organization (hospital/clinic) there are natural subgroups that can differ substantially in satisfaction (e.g., inpatient vs. outpatient; departments such as maternity, emergency, pediatrics; age groups; language/ethnicity). Stratifying by those meaningful categories and then randomly sampling within each stratum ensures you get adequate representation from each subgroup and reduces sampling error versus simple random sampling when subgroup responses vary.

1. Better precision & comparability. Stratified designs give more precise estimates for each subgroup and allow direct comparison between units (e.g., emergency vs. primary care). If you expect some strata to be small (e.g., specialty clinics), you can oversample those strata and then apply sampling weights during analysis so overall estimates remain unbiased.

2. Operational practicality in organizations. Healthcare organizations maintain rosters or visit logs (electronic medical records, appointment lists) that serve as a sampling frame. These lists make it straightforward to define strata (clinic/service line) and draw random samples within each stratum. If no roster exists, cluster sampling (selecting whole clinic-days) is a backup option but is usually less efficient.

How to implement (practical steps):

1. Define population and strata. E.g., all patients seen in last 3 months. Strata: (a) service type — inpatient, outpatient, ED; (b) specialty clinics (maternity, pediatrics, chronic care); or combine by clinic × visit type if sample logistics allow.

2. Construct sampling frame. Use appointment logs or discharge lists to create a roster for each stratum.

3. Decide sample size. Use standard sample-size calculations for proportions (desired margin of error, confidence level, expected satisfaction proportion). Adjust for design effect if clustering or weighting is used, and inflate for expected nonresponse (e.g., +10–30%).

4. Random selection within strata. Use random-number generation to select patients from each stratum roster. If electronic lists are large, systematic sampling (every kᵗʰ) after random start works too.

5. Nonresponse management. Follow up via multiple contact attempts, offer translated surveys, and compare responders vs. nonresponders on basic demographics to check bias. If some strata have low response, use weighting in analysis. APA

When you might not use stratified sampling:

• If you lack a reliable sampling frame, or resources are extremely constrained, you may need cluster sampling (sample by clinic-day) or, less ideally, convenience sampling — but these introduce bias and reduce generalizability. Aim for probability sampling whenever feasible.