Session 5: Solutions

Download the PDF · Back to Session 5

Foundations of Clinical Research · instructor copy (do not distribute before the session)

Exercise numbers follow Practicals/s5_exercises.md. All FLUID-ICU numbers come from Course/Data/key_results.md (simulated data).

Exercise 1: Classify the variables

# Variable Answer Explanation
a Age (years) Continuous Measured on a scale. Keep the exact number on the CRF (or collect date of birth).
b Sex Binary Two categories. Coding 1/2 does not make it numerical.
c Glasgow Coma Scale Ordinal Ordered scores (3–15) with unequal steps. Often analysed like a number, which is acceptable if stated, but it is ordinal by nature.
d Source of sepsis Nominal Several categories, no natural order.
e Number of vasopressors Count Whole numbers 0–3. Accept “ordinal” if the participant argues it well; not continuous.
f Cumulative fluid balance (ml/kg) Continuous Can be negative or positive, with any value on the scale. The Paper 1 primary outcome.
g Alive at day 28 Binary Yes/no.
h Time to death, censored at day 28 Time-to-event Two pieces of information: how long, and whether the event happened. Analysed with Kaplan–Meier / Cox.
i Pain none–severe Ordinal Ordered categories.
j Blood group Nominal No order.

Bonus: any continuous variable (age, fluid balance, lactate). You can always group later, but you cannot recover the exact value from a group.

Common errors: GCS called continuous; time-to-death called continuous (ignores censoring); number of vasopressors called continuous.

Exercise 2: Fix the bad CRF items (model answers)

# Bad item Problem Model rewrite
a Fluid given: ____ No unit, no time window, no definition (crystalloid? blood? drug diluents?) “Total IV fluid received from ICU admission to 24 h (ml), including crystalloids, colloids and drug diluents: [ ][ ][ ][ ][ ] ml ☐ not recorded”
b Weight: ____ No unit; measured or estimated? when? “Body weight at ICU admission (kg): [ ][ ][ ].[ ] ☐ measured ☐ estimated ☐ not available”
c Comorbidities: ________ Free text: impossible to analyse, spelling varies Tick list: “Comorbidities present at admission (tick ALL that apply): ☐ diabetes ☐ chronic kidney disease ☐ heart failure ☐ COPD ☐ cirrhosis ☐ cancer ☐ none ☐ unknown”
d Lactate high? Y / N Opinion (“high” by whose cut-off?), loses the value, no time point “Serum lactate at ICU admission (mmol/L): [ ][ ].[ ] ☐ not done” (define “high” later, in the analysis)

Accept any version that passes the four checks (unit, time point, one reading, unknown option). Praise answers that notice that collecting the number is better than collecting a yes/no.

Exercise 3: Choose the right summary

# Variable Shape / type Summary Graph
a Age, bell-shaped Symmetric continuous Mean (SD) Histogram
b ICU length of stay Right-skewed Median (IQR) Histogram or boxplot
c Died in hospital Binary n (%) Bar chart
d CRP Right-skewed Median (IQR) Histogram / boxplot
e Source of sepsis Nominal n (%) per category Bar chart (not 3-D pie)
f SBP by treatment group Symmetric, two groups Mean (SD) per group Side-by-side boxplots

Exercise 4: Judge a bad figure

  1. Each bar is the mean fluid balance per group. The error bar is the standard error (SE), shown in one direction only. Missing: the individual patients, the spread (SD or IQR), the shape (skew), the n per group on the figure, and what the error bar represents.
  2. Any two of: the data are skewed; one extreme patient (160 ml/kg) pulls the usual-care mean up; the groups overlap a lot; the median; the number of patients; negative values.
  3. Dots (one per patient) plus a boxplot per group (see the “dots + box” slide), y-axis labelled with units, n under each group. You gain: shape, spread, outliers and overlap become visible, and the reader can judge the comparison.
  4. The truncated axis makes a 2-percentage-point difference (31% vs 29%) look like a three-fold difference. Fix: start bars at 0 and label the values (or show the difference with its 95% CI).
    1. The author chose the two scales, so the lines can be made to cross or move together wherever wanted; (ii) two trends over time moving together is correlation, not causation (other changes may have happened at the same time). Better: two panels sharing the time axis, and an appropriate design (e.g. interrupted time series, Session 3).

Exercise 5: [FLUID-ICU] Spot the problems in the raw FLUID-ICU data

(See Course/Data/cleaning_log.md for the full log.)

# What is wrong What the script does
a One variable coded many ways (and a leading space) Recode to Female / Male
b Dates in different formats (03/04: March or April?) Read each format explicitly and convert to one date format (YYYY-MM-DD)
c ICU C reported lactate in mg/dL, not mmol/L (about 9 times larger; found by comparing medians by ICU) Divide ICU C values by 9.01 (1 mmol/L = 9.01 mg/dL); confirm with the site laboratory
d 999 is a missing-value code, not a lactate Set to missing (NA)
e Impossible age (outside the allowed range) Set to missing, send a query to the site; keep the patient
f Duplicate record Check the two rows are identical, keep one copy (623 rows become 620 patients)

Bonus: editing by hand is invisible and not reproducible. A script plus a log lets anyone check and repeat every change. The log records the problem, how many records, what was done, and why (plus queries sent).

Exercise 6: [FLUID-ICU] Read the FLUID-ICU Table 1 (whole class)

All numbers from Course/Data/key_results.md (simulated data).

  1. Negative balance n = 245, positive n = 373, overall 618. Two of the 620 enrolled patients have no computable day-3 fluid balance (an impossible weight or fluid output was set to missing during cleaning), so they cannot be placed in a group.
  2. Expected answers: mean (SD) for age, which is roughly symmetric (64.6 vs 61.4 years); median (IQR) for SOFA, a skewed score (6 (4 to 8) vs 8 (6 to 10)); median (IQR) for lactate, which is right-skewed (2.2 (1.4 to 3.0) vs 2.9 (1.9 to 4.0) mmol/L); n (%) for vasopressor, ventilation, heart failure and sex.
  3. Severity: SOFA 8 vs 6; vasopressor 205/373 (55.0%) vs 61/245 (24.9%); mechanical ventilation 216/373 (57.9%) vs 97/245 (39.6%). (Heart failure goes the other way: 43/373 (11.5%) vs 54/245 (22.0%), because clinicians give less fluid in heart failure.)
  4. Lactate: 91 missing among the 618 (93 of all 620 enrolled, 15.0%).
  5. Model sentences: “Of 620 enrolled patients, 618 had a computable day-3 fluid balance, of whom 373 (60.4%) had a positive balance. Patients with a positive balance were more severely ill at admission (median SOFA 8 vs 6; vasopressor 55.0% vs 24.9%; mechanical ventilation 57.9% vs 39.6%).”
  6. Larger: the positive-balance group was sicker, and sicker patients die more, so the crude association mixes fluid with severity (confounding by indication). Confirmed in Session 6: crude OR 2.49 vs adjusted 1.62.

Exercise 7: [FLUID-ICU] Your manuscript question as a PECO sentence

In class, Q1 is done together; each participant later completes the table for the question they chose.

Q Outcome (type) Exposure (type) Planned comparison / model
Q1 Death by day 28 (binary; also time-to-event) Positive balance by day 3 (binary) Chi-square; logistic regression; Kaplan–Meier + Cox
Q2 AKI, KDIGO ≥ 1 (binary) Cumulative balance, ml/kg (continuous, per 10 ml/kg) Chi-square by group; logistic regression
Q3 Ventilator-free days to day 28 (count 0–28, skewed, piled at 0 and 28) Cumulative balance (continuous) Mann–Whitney; linear regression
Q4 Death by day 28 (binary) Lactate at admission, mmol/L (continuous, 15% missing) Logistic regression; complete case vs multiple imputation
Q5 ICU length of stay in survivors (continuous, right-skewed) Positive balance (binary) Mann–Whitney; linear regression on log scale

Example (Q1): “In adults with sepsis admitted to the ICU, is a positive cumulative fluid balance by day 3, compared with a negative balance, associated with higher 28-day mortality?”

Exercise 8: Proposal builder 5 (what a good answer looks like)

Variables table

Role Variable How measured Type
Primary outcome Death by day 28 Hospital records / phone call at day 28 Binary
Secondary outcome Cumulative fluid balance at day 5 Fluid charts, ml/kg of admission weight Continuous
Intervention Restrictive vs usual fluid strategy Randomised allocation Binary
Baseline / confounders Age, SOFA score, lactate at admission, source of sepsis CRF at ICU admission Continuous, ordinal-like score, continuous, nominal

Sample size: 40% vs 30% mortality, alpha 5% two-sided, power 80% gives 356 per group (spreadsheet sheet 2). Dividing by (1 − 0.10) gives 396 per group, 792 in total.

Analysis plan: describe with mean (SD) / median (IQR) / n (%); compare 28-day mortality with a chi-square test; report the risk ratio and risk difference with 95% CIs; fluid balance compared with Mann–Whitney and reported as median (IQR) per group. (Observational designs: add “adjusted for age, SOFA and lactate using logistic regression”.)

Quick check for the instructor (while modelling, and when reading drafts sent by email):

Element Good enough Needs work
Variables table Outcome, exposure and ≥ 3 confounders, all with a type Confounders that are actually outcomes; no types
Mini-CRF 6–10 items, units, time points, unknown option > 15 items; free text; no units
Data dictionary 3–5 rows with codes and allowed ranges Names with spaces, no codes
Sample size 4 inputs with a stated source “All patients in one year”
Analysis plan Summary + test + effect measure with 95% CI “Data will be analysed with SPSS”

Data management (good enough): EDC with range and logic checks at entry (or paper with double entry of key variables); coded study IDs with the linking log kept separately and locked; audit trail/cleaning log; raw export read-only, dated versions, encrypted 3-2-1 backups; data-sharing statement in the consent form.