The resident's depression prevalence study and the ethics committee's question
A second-year family medicine resident proposes a dissertation: "Prevalence and predictors of undiagnosed depression among adults attending three urban primary health centres." She will enrol 400 consecutive adults in the waiting room over eight weeks, administer the PHQ-9 in the local language, and define depression as a score of 10 or more. She will record age, sex, income, chronic illness and family history, and analyse predictors with multivariable logistic regression. Her supervisor suggests switching to a case-control design because it would be "faster and cheaper". The institutional ethics committee has asked her to justify why written informed consent is needed when the PHQ-9 is "just a questionnaire" and is already used in routine care. She has no budget for a psychiatric interview and no validated local translation of the PHQ-9. Her statistician estimates that 400 participants will give 10 events per variable for her ten predictors.
1. Is her design fit for her stated aim? What can and cannot be concluded from it?
Read the worked answer
A cross-sectional design is the correct choice for prevalence, and the supervisor's case-control suggestion should be rejected: case-control sampling selects on outcome, so it cannot estimate prevalence at all, and choosing controls from clinic attenders would introduce selection bias that no analysis can remove.
But the resident must be precise about what she is measuring. PHQ-9 of 10 or more is a screening threshold, not a diagnosis. Without a confirmatory interview, her headline figure is the prevalence of a positive screen, which will overestimate depression because specificity is imperfect. If PHQ-9 sensitivity is 88% and specificity 82%, a true prevalence of 12% yields a positive predictive value of only about 40% - most screen-positives are not cases. She should either say "probable depression" throughout, or build in a validation substudy.
"Predictors" is also the wrong word for a cross-sectional analysis: she can show associations, not prediction of future onset, since exposure and outcome are measured simultaneously. Because depression will not be rare, odds ratios from logistic regression will exaggerate the prevalence ratios she actually wants; modified Poisson regression or log-binomial models would be more honest.
Finally, the statistician's arithmetic is suspect: 400 participants with 10 covariates needs about 100 events for 10 events per variable. If prevalence is 15%, she will have roughly 60 events - too few. She must reduce the number of covariates or increase the sample.
2. Redesign the study to strengthen measurement and inference within her budget.
Read the worked answer
Keep the cross-sectional frame but add a two-stage design. Screen all participants with the PHQ-9, then invite a random subsample - say 100 of the 400 - to a blinded structured interview such as the MINI conducted by a trained interviewer. That yields local estimates of sensitivity and specificity, allows a corrected prevalence using the Rogan-Gladen formula, and gives the committee a defensible case definition. If the interview is impossible, she should report the screen-positive prevalence with its confidence interval and explicitly state the limitation rather than overclaim.
Fix the instrument before data collection. A validated local-language PHQ-9 requires forward and back translation, cognitive interviews with a dozen patients, and a pilot; using an unvalidated translation guarantees uninterpretable data.
Fix the sampling. Consecutive waiting-room recruitment over-samples frequent attenders and people who can afford to wait. Better: random selection from clinic registers with telephone or household follow-up, or systematic sampling with a random start and documented non-response, so she can compare responders and non-responders.
Fix the sample size and analysis. For prevalence within ±5% at an expected 15% requires about 196 participants, so 400 is comfortable; for the regression she should cut to five or six pre-specified covariates or increase recruitment. Pre-specify the analysis plan, and report absolute numbers with confidence intervals, not just p-values.
3. How do you answer the ethics committee on consent and confidentiality?
Read the worked answer
The committee's premise is wrong: the fact that a questionnaire is already used in routine care does not make its use in research consent-free. Research uses the data for a different purpose, and the PHQ-9 identifies distress and suicidal ideation - that is not a trivial disclosure. Written consent, with an information sheet in the local language, is the default.
Where literacy is low, a witnessed verbal consent with a thumbprint and an independent witness is acceptable and should be described in the protocol. Consent must cover: the purpose, what participation involves, the voluntary nature and the right to withdraw without affecting care, how confidentiality will be protected, and what will happen if the score is high.
Confidentiality in a waiting room is a real problem. Administer the questionnaire in a private space, use study identifiers rather than names on data sheets, keep the linkage file separately under lock, and train recruiters not to call out scores.
Most importantly, the protocol needs a referral pathway. Screening for depression in a population with no access to treatment is ethically indefensible. A named clinician, a written pathway for scores suggesting suicidality, and a plan for follow-up should be in the submission. Risk of distress during completion, and time burden, should be disclosed.
The ethics committee should also note that this is a student project with no inducement to participate; any payment should cover time and travel only, not act as an incentive.