Research Design and Methods

blueprint. It's the plan for who you study, what you measure, and how you protect the results from nonsense.

design

the overall plan to answer the question and test hypotheses. you choose a design that can actually produce valid results for that question. if your question is causal ("does Drug X reduce MI?"), some designs are better than others.

research design

internal validity

"did the treatment/exposure actually cause the outcome in this study?"

threats: confounding, selection bias, measurement bias, regression to the mean, maturation, loss to follow-up.

fixes: randomization, allocation concealment, blinding, prespecified outcomes, good adherence/retention, intention-to-treat analysis.

"will these results apply to my patients?"

threats: narrow inclusion/exclusion, unrealistic setting, intensive monitoring, short follow-up.

fixes: representative samples, pragmatic designs, clear description of the population and context.

external validity

when can we say "because of" instead of "with"

causality

select all that apply:

what are Bradford Hill's nine considerations?

1. temporality

2. strength

3. biological gradient

4. consistency

5. specificity

6. plausibility

7. coherence

8. analogy

9. experiment

cause before effect (non-negotiable).

temporality

big effect sizes are more believable.

strength

more exposure → more effect.

biological gradient

seen across studies, places, methods.

consistency

one cause → one effect (rare in biomed).

specificity

fits biology/pharmacology.

plausibility

aligns with existing knowledge.

coherence

similar agents cause similar effects.

analogy

manipulating the cause changes the effect (RCTs or natural).

experiment

you observe exposures/outcomes without assigning treatment.

observational designs

snapshot; prevalence & associations.

cross-sectional

start with outcome, look back at exposures; efficient for rare outcomes; prone to recall/selection biases.

case-control

start with exposure, follow for outcome; better for incidence and temporal order; confounding is the dragon you must slay (matching, stratification, multivariable adjustment, propensity scores, instrumental variables).

cohort

you assign the exposure/treatment.

experimental designs

gold standard for internal validity.Variants: parallel, crossover, factorial, cluster, pragmatic (real-world). features that protect truth: randomization, allocation concealment, blinding, prespecification, adherence monitoring, ITT analysis. use when: you want clean causal inference and can ethically randomize.

randomized controlled trial

tell me what's out there. answers who/what/when/where; no causal claims. examples: prevalence of uncontrolled HTN this year, a case series, a cross-sectional survey.

descriptive studies

do X and Y move together (or does X cause Y)? tests relationships and sometimes causality; can be experimental or observational. examples: "GLP-1 users have lower MACE than DPP-4 users," "new refill program improves PDC."

analytical studies

if the paper says "associated with," that's ______________; if it says "described" or "prevalence," that's _______________.

analytical, descriptive

forward in time. you define variables now and follow patients. strengths: clean measurement, you can enforce definitions, temporality (exposure → outcome) is clear. tradeoffs: time + money, loss to follow-up, Hawthorne effect.

prospective studies

backward in time. you use existing data (charts, claims, registries). strengths: fast, cheap, big samples, great for rare outcomes. tradeoffs: stuck with the data you've got (missingness, misclassification), weaker control of confounding, temporality sometimes messy.

retrospective studies

you assign the treatment.If you randomize, it's an RCT (gold standard for internal validity). if you assign without randomization, it's quasi-experimental.

intervention studies

you change something, but not randomly. designs: before-after, non-equivalent cohorts, interrupted time series, difference-in-differences. use when: randomization isn't feasible (policy rollouts, system-level changes). you still battle confounding, but the time/design structure helps.

quasi-experimental

randomize to intervention vs control and follow for outcomes.

pros: strongest for causality (high internal validity).

cons: tightly controlled settings can limit external validity (generalizability), can be expensive/slow, sometimes unethical.

randomized controlled trials

no randomization; you watch what naturally happens.

cohort (prospective or retrospective), case-control, cross-sectional.

pros: real-world, scalable, captures harms and rare events.

cons: confounding is the dragon—handle with matching, stratification, multivariable models, propensity scores, IVs, etc.

observational

you collect it yourself (interviews, surveys, measurements).

primary data

you reuse what already exists (medical charts, EHR, insurance claims).

secondary data

one unusual patient = ______________; a handful of similar patients = ________________. value: early signal. limitation: no comparison group → no causality. pharm example: a single patient develops rhabdo on a weird statin-macrolide combo.

case report, case series

("snapshot", prevalence)

exposure and outcome measured at the same time. good for: how common something is; quick associations. weak for: cause → effect ordering (temporality missing). pharm example: survey of PPI use and B12 levels during one clinic visit.

cross sectional study

(start with outcome)

identify people with the outcome (cases) and similar people without (controls), then look back for exposures. efficient for rare outcomes, gives odds ratios. vulnerable to recall and selection bias. pharm example: among patients with tendon rupture vs without, compare prior fluoroquinolone use.

case-control study

(start with exposure)

group by exposure (exposed vs unexposed) and follow over time to see who develops the outcome. prospective cohort: define groups now, follow forward. strong temporality, cleaner data, but expensive. retrospective cohort: use past records to reconstruct follow-up—cheaper, but variables may be messy. gives risk, incidence, risk ratios. pharm example: compare new users of SGLT2 inhibitors vs DPP-4 inhibitors and track hospitalization for HF.

cohort study

how you get your data and with what instruments. choice depends on the research question, population, feasibility, and cost.

research methodologies

surveys, interviews, clinic measurements. tailored to your question; high control over variables. can be time/$$ heavy.

primary methods

charts, registries, claims. fast(er), cheaper, huge samples possible. variables may be missing or poorly defined; operational. definitions matter (what exactly counts as "med adherence"?).

secondary methods

select all that apply:

what are the decision filters for research methodologies?

1. reliability

2. practicality

3. validity

consistency is queen. you can't be valid if you're not ______________.

reliable

same instrument, same people, different times → are scores stable?

example: re-administer a medication adherence scale 2 weeks later.

test-retest reliability

different raters, same thing → do they agree?

example: two pharmacists independently classify ADE severity from charts.

inter-rater reliability

do items within a multi-item scale hang together? measured by Cronbach's alpha (rule of thumb: ≥0.80 is good for research use; 0.7-0.8 acceptable, >0.9 can mean redundancy). example: the SF-12 quality-of-life scale—items on the physical domain should correlate.

internal consistency

if I measure again, do I get the same answer?

reliability

is it the right answer?

validity

new/unusual signals; no control group; generates hypotheses.

case report/series

snapshot; prevalence; association ≠ causation.

cross-sectional

start with outcome; look back at exposures; odds ratio; efficient for rare outcomes; recall/selection bias.

case-control

start with exposure; follow for outcome; risk & risk ratio; stronger temporality (especially prospective).

cohort

custom & clean but pricey. secondary data: fast & big but messy definitions.

primary data

test-retest, inter-rater, internal consistency (Cronbach's α).

reliability types

basically, "does this thing actually measure what we say it's measuring?" imagine you have a scale that's supposed to weigh your suitcase, but instead it tells you how tall you are. that's not _____________________!

validity

does this test look like it measures what it's supposed to? if a depression questionnaire only asked about your favorite foods, people would be like, "uhh... what?" low _________________.

face validity

are we covering all the right topics? example: a pharmacy exam that only covers antibiotics wouldn't have __________________ for "pharmacology" as a whole.

content validity

now we're asking: does this test actually measure the big-picture idea (the "construct")? for instance, does an anxiety survey really capture "anxiety"—not just stress, or excitement, or shyness?

construct validity

can we show that our measure matches up with some "gold standard" or real-world outcome? (if your new blood pressure cuff gives the same results as the clinic's, that's _________________.)

criterion validity

does this test agree with similar measures? (your anxiety scale should match up with other well-established anxiety scales.)

convergent validity

does this test stay in its lane and NOT overlap with unrelated constructs? (your anxiety scale shouldn't look just like a depression scale.)

discriminant validity

select all that apply:

which of the following are primary data?

1. self-reports

2. observations

3. biological assessments

select all that apply:

which of the following are secondary data?

1. medical records

2. medical claims

3. research data

questionnaires, interviews

self-reports

watching behaviors

observations

labs, physical measurements

biological assessments

info someone else already collected

secondary data

written down during care, not for your study, but you can use them.

medical charts

for insurance/payment, but can reveal trends.

medical claims

done for policy, big population-level info.

national surveys

collected for a different study but reused for yours (no need to reinvent the wheel).

research data

about asking people directly. want to know how someone feels about a medication? ask!

self-reports

written questions (could be "what's your pain score?" or "what's your biggest fear?"). can be open-ended (write whatever you want) or close-ended (pick from choices).

surveys

face-to-face or phone convo, where the researcher can go off-script, ask follow-ups, and really listen.

interviews

what they do: let people express themselves in their own words. ("tell me about your pain today.")

pros: deep, rich info. sometimes you get stories or reasons you never thought to ask about.

cons: takes time to answer, hard to analyze with stats. If you ask this in a waiting room, you might end up waiting all day!

open-ended questions

what they do: give a menu of answers ("did you have any side effects? yes or no?")

pros: quick, easy to answer, easy to tally up for statistics.

cons: might miss out on unique answers; you're boxed in.

close-ended questions

old school. send that survey by snail mail. good if your folks aren't tech-savvy.

mail surveys

modern vibes. SurveyMonkey, Qualtrics, etc. quick, cheap, and most college courses use this style now.

online or internet surveys

all verbal, no written record for the patient to review.

telephone

they always ask about your prof at the end of the semester, usually online

course evaluations

select all that apply:

what are the types of interviews?

1. structured

2. semi-structured

3. unstructured

think: "scripted talk show". everyone gets the exact same questions, asked in the exact same way. this makes it super easy to compare answers across participants because there's no room for improvisation. if you ever filled out a multiple-choice test with an interviewer—this is the vibe.

structured interviews

think: "guided conversation". there's a set list of questions, but you can riff a little. no strict response categories, so the participant can elaborate and the interviewer can follow up as needed. you get both structure and flexibility—like having talking points but letting the convo flow naturally.

semi-structured interviews

think: "freestyle, let's talk". no script, no predetermined answer categories. you might start with a general question and let the participant take you on a journey. it's like meeting someone at brunch and just vibing wherever the convo goes.

unstructured interviews

researchers collect info by watching participants do stuff—taking notes on their activities, communication style, interactions, how long it takes to complete tasks, etc.

observation

translation: "Big Brother is watching, and you know it." the participant knows they're being watched. this can be good for gaining insight into thought processes—think "skills check-off" in the lab where you know the instructor is right there with a clipboard.

PRO: you can directly see how someone performs or reacts in real time.

CON: people might act "better" or differently ('cause nobody wants to look silly in front of a judge!), also known as the "social desirability bias."

obtrusive observation

translation: "Undercover Boss—but for science." the participant does not know they're being observed (hidden or disguised observer), so their behavior is likely more natural, authentic, and less "put on for the cameras."

PRO: more genuine data—people do what they normally do!

CON: ethical concerns—can feel sneaky or invasive depending on the setting.

unobtrusive observation

what's involved? anything using biophysical, biochemical, or microbiological methods. if you need a machine, a lab tech, or a tube of blood, this is your lane.

examples:

EKG/ECG, glucometer, microscope, x-ray, etc.

requirements:

you need specialized equipment and trained professionals (don't try this at home!).

reliability/validity: usually very reliable if using standardized procedures and equipment.

biological measures

scientific observations: should be unbiased, using best practices.

device development: lots of new tech is making it easier to do these measures, but they must be tested ("validated") before using in real-world clinics.

objective markers: things like blood sugar, cholesterol, heart rhythms—these are clear, numerical, and matter for diagnosis or treatment decisions.

biological measures

physical stuff—think imaging or measuring a physical parameter (x-ray, BP, EKG, MRI).

biophysical assessments

chemical stuff in bodily fluids—blood glucose, urine creatinine, serum drug levels. this means you're analyzing the "ingredients" in blood, pee, etc. usually done with specialized chemical instruments.

biochemical methods

bugs & germs—identifying bacteria/viruses in fluids via cultures, microscopes, etc.

microbiological methods