Set aside about 15 minutes to read and reflect. The text follows the presentation “From a clinical question to a research plan” (practical class 8, Social Medicine with Medical Ethics, Dental Medicine). Afterwards you should be able to turn a clinical problem into an answerable question, choose a design that can answer it, and name the elements of a research plan-programme. The rule that runs through the whole class: the design is chosen by the question — never the other way round.
What makes an inquiry research
Scientific research is a systematic, planned and reproducible inquiry that collects and analyses data in order to produce new knowledge that is generalisable (quantitative research) or transferable (qualitative research) beyond the people studied. It has an explicit question and, where appropriate, a testable hypothesis; a method decided in advance and written down in a protocol; measurement that others can repeat and check; and honest reporting, including results that disappoint the authors. The research cycle runs from observation to question, hypothesis, design, data collection, analysis, interpretation and publication — and on to new questions.
Research is not the same as clinical audit. An audit assesses local practice against an explicit existing standard. Research addresses a question — about burden, cause, effect, mechanism or experience — to produce knowledge that applies beyond the setting studied.
A study can be classified along four independent axes: purpose (basic research on mechanisms, such as the biofilm ecology of S. mutans, or applied research on diagnosis, treatment and prevention, such as a fluoride varnish trial); domain (clinical, public health or health services — implant outcomes, a school caries survey, access to dental care by income); approach (quantitative, qualitative, or mixed methods that deliberately integrate both); and control and time (observational or experimental, cross-sectional or longitudinal, prospective or retrospective). Laboratory and animal studies are research too, but they answer questions about mechanisms, not about effects in patients.
The research question: FINER and PICO/PECO
A good question is FINER: Feasible (enough participants, time, money and skills), Interesting, Novel (it confirms, refutes or extends — so check the literature first), Ethical (acceptable risk, approvable by an ethics committee) and Relevant (it changes practice, policy or further research). A vague interest has to be narrowed step by step: “I am interested in caries” becomes “caries in teenagers in Plovdiv”, and finally “What proportion of 15-year-old schoolchildren in Plovdiv have caries experience (DMFT ≥ 1), and is it associated with family socioeconomic position and the frequency of sugar intake?”
The question is then structured. PICO fits an intervention question: P — children and adolescents; I — fluoride varnish applied two to four times a year; C — placebo or no treatment; O — caries increment after three years. The Cochrane review that answers it pooled 22 trials with 12,455 children; for permanent tooth surfaces the prevented fraction was 43% (95% CI 30–57%) — a 43% smaller mean caries increment, not 43% fewer children with caries (Marinho et al., 2013). PECO fits an exposure question: P — adults receiving dental implants; E — smoking; C — non-smoking; O — implant failure, infection, marginal bone loss. A meta-analysis of 107 studies found failure of 6.35% of 19,836 implants in smokers against 3.18% of 60,464 in non-smokers, with uncontrolled confounding (Chrcanovic et al., 2015). Add T (time) or S (setting or study type) when they matter; qualitative questions use PICo — population, phenomenon of interest, context.
The aim is the general purpose in one sentence. The objectives are specific, measurable steps towards it, and each needs its own variables and analysis. A hypothesis is a testable statement about a relationship, written before the data are seen: H0 says there is no difference or association, H1 that there is. Three common errors: writing a hypothesis for a purely descriptive estimate (a prevalence needs none unless it is compared with a preset benchmark), writing objectives that no variable measures, and promising “to prove that…” — research tests, it does not prove.
Types of studies: a decision tree
The design follows from a few questions. If the investigator assigns the exposure at random, the study is a randomised controlled trial, which gives the effect of an intervention with the strongest protection against confounding; if the assignment is not random, it is a non-randomised (quasi-)experiment. If the investigator only observes and follows exposed and unexposed people from exposure to outcome, it is a cohort study (incidence, relative risk, hazard ratio, the temporal sequence). Selecting people by outcome and comparing their past exposure is a case–control study (odds ratio; efficient for rare diseases). Measuring exposure and outcome at one time point is a cross-sectional study (prevalence; no temporal order). Measuring both for groups such as towns or countries is an ecological study, which allows no inference to individuals. Without a comparison group there are only case reports and series, which generate hypotheses. Systematic reviews and meta-analyses synthesise primary studies; qualitative designs answer “how” and “why” questions about experience.
Each design was illustrated with a real dental study:
- Cross-sectional. In NHANES 2009–2010, 47% of 3,742 dentate US adults aged 30 and over had periodontitis (Eke et al., 2012); oral HPV infection was found in 6.9% of 5,579 people aged 14–69, 10.1% of men and 3.6% of women (Gillison et al., 2012). Such surveys are quick and give prevalence for planning, but they cannot tell which came first.
- Case–control. One hundred patients with oropharyngeal cancer were compared with 200 controls; oral HPV-16 infection had an odds ratio of 14.6 (95% CI 6.3–36.6), with and without heavy tobacco and alcohol use (D’Souza et al., 2007). The typical threats are selection bias (controls must come from the population that produced the cases) and recall bias, and no incidence can be calculated.
- Cohort. The Dunedin birth cohort followed about 1,000 children born in 1972–73 to age 26. Those who grew up in low- rather than high-socioeconomic-status families had periodontal disease in 31.1% vs 11.9% and a high caries level in 32.2% vs 9.9% (Poulton et al., 2002). Cohorts measure exposure before the outcome but are long, costly and lose participants.
- Randomised trials. In the OPT trial, 823 pregnant women were randomised to periodontal treatment before 21 weeks or after delivery. Periodontal health improved, but preterm birth was 12.0% vs 12.8%, HR 0.93 (95% CI 0.63–1.37): a reduction was not demonstrated (Michalowicz et al., 2006). In another trial, intensive periodontal treatment improved flow-mediated dilatation by 2.0 percentage points at 180 days — a surrogate vascular outcome, not fewer heart attacks (Tonetti et al., 2007). The key features of a trial are randomisation, allocation concealment, blinding, intention-to-treat analysis and prospective registration.
- Synthesis and qualitative research. A Cochrane review of fluoride toothpastes found a lower caries increment with 1,450–1,500 ppm than with non-fluoride toothpaste (SMD −0.36) (Walsh et al., 2019). In-depth interviews with 20 adults at a sedation clinic showed that dental anxiety affects sleep, work, relationships, eating and oral hygiene (Cohen et al., 2000); qualitative sampling is purposive.
The lesson of the OPT trial is general: an association in observational studies does not guarantee that treating the exposure changes the outcome. And designs are not a rigid hierarchy — in the “new evidence pyramid” a well-conducted cohort can be more trustworthy than a poorly conducted trial (Murad et al., 2016).
The plan-programme
The research protocol, or plan-programme, has three layers. The conceptual layer says why: title, problem and rationale, aim and objectives, hypotheses, expected contribution. The methodological layer says how: design, object and unit of observation, population and sample, variables and their scales, methods and instruments, and the programme of data collection, processing and analysis. The organisational layer says who, where and when: place and time, team, pilot study and timetable, budget, ethics approval and data protection, dissemination.
Three terms need care. The object of observation is the set of persons studied (15-year-old pupils in Plovdiv); the unit of observation is the element on which the features are recorded (one pupil); the technical unit is where the data are gathered (the school). The unit of analysis must also be decided — several teeth or implants in one person are not independent.
Sampling. The chain runs from target population to accessible population, sampling frame, sample and respondents. Non-probability samples (convenience, purposive, snowball, quota) have unknown selection probabilities; probability samples (simple random, systematic, stratified, cluster) give every unit a known, non-zero chance. A convenience sample of university-clinic patients describes those patients, not the population of Plovdiv.
Sample size. To estimate a prevalence, n = Z² · p · (1 − p) / d². With p = 0.60, d = 0.05 and Z = 1.96, n = 369; doubled for the design effect of cluster sampling (738) and divided by 0.90 for 10% non-response, the survey invites 820 pupils. The planning value of p must come from a cited study or a pilot; p = 0.50 gives the largest n when nothing is known. Comparisons also need the expected difference and the power, calculated with software such as G*Power or OpenEpi.
Variables. In a caries study the outcome is caries experience measured as DMFT — the number of permanent teeth that are decayed, missing because of caries or filled, by WHO criteria. DT ≥ 1 means untreated caries; DMFT ≥ 1 means caries experience. The exposure might be the daily frequency of sugary snacks and drinks; age, sex, socioeconomic position and fluoride toothpaste use are confounders. The scale of measurement decides which statistics are valid.
Validity. Chance is reduced by an adequate sample and quantified by confidence intervals and p-values. Bias — selection bias (who gets into the study) or information bias (measurement error, recall, uncalibrated examiners) — is prevented by design and conduct; a larger sample does not remove it. Confounding is a common cause of exposure and outcome, such as socioeconomic position for smoking and implant failure, and is controlled by randomisation, restriction, matching, stratification or multivariable models; a mediator on the causal path is not adjusted away automatically. Examiners are calibrated and their agreement reported as Cohen’s kappa — but reliability is not validity: an examiner who always misses early lesions is consistent but wrong.
Ethics approval is part of the method. Under the Bulgarian Health Act, medical research is carried out only on persons who have given written informed consent after written information on its nature, significance, scope and risks (Art. 199(1)); consent may be withdrawn at any time (Art. 199(3)); the participant’s interests prevail over scientific and financial ones (Art. 197(5)); and research starts only after a positive opinion of a local ethics committee (Art. 203(1)) — at MU-Plovdiv, the Research Ethics Committee. Health data are a special category under the GDPR, and the Declaration of Helsinki requires registration before the first participant. The Vipeholm study in Sweden (1945–1953), in which residents of an institution for people with intellectual disabilities were fed sticky sweets to study caries, shaped prevention — but the participants never consented.
Analysis plan. Each objective gets its variables and its analysis — prevalence with a 95% CI, comparisons and multivariable models that account for the cluster design — and the primary outcome and analysis are fixed in advance, because changing them after seeing the data inflates false-positive findings.
The worked example
The class ends with a complete plan-programme: a cross-sectional, school-based survey of caries experience among 15-year-olds in Plovdiv, with random schools and classes, two calibrated examiners, a questionnaire piloted on 30 pupils, consent from pupil and parent, referral letters for pupils with untreated caries, pseudonymised data and a report written to STROBE. Its stated limitation matters: current behaviour is measured together with cumulative caries history, so the survey can show associations, not causes.
Before opening the student tasks, try to answer in three sentences: which design fits a rare disease with a long latency, and why; what the difference is between the unit of observation and the technical unit; and why a larger sample cannot correct a biased one.