Chronic obstructive pulmonary disease (COPD) is one of the most studied conditions in clinical research, and its trials share a recognisable statistical pattern. Lung function is measured repeatedly and analysed with a mixed model for repeated measures. Exacerbations are counted over time and analysed with a negative binomial model. Health status is measured with a questionnaire and often summarised as a responder analysis.
This article explains the main COPD endpoints and then works through a complete analysis on a dummy dataset: the statistical concepts, the SAS code, the output, and the interpretation.
What makes COPD trials distinctive
- Several endpoints matter at once. Lung function, exacerbations, symptoms and health status each describe a different part of the disease. No single endpoint is enough for a full picture.
- Repeated measurements. Spirometry is performed at every visit, so each patient contributes several correlated observations.
- Patients withdraw. Trials last 12 to 52 weeks in an elderly population with other illnesses. Dropout of 10% to 25% is common, and it is often related to how the patient is doing.
- Exacerbations are rare and uneven. Most patients have none during a trial and a few have several. The counts are more variable than a Poisson distribution allows.
- Background therapy and rescue medication. Patients continue inhaled maintenance treatment and use rescue bronchodilators, both of which affect what the trial can measure.
The main endpoints
| Endpoint | What it measures | Usual analysis | Clinically important difference |
|---|---|---|---|
| Trough FEV1 | Forced expiratory volume in one second, measured before the next dose | Change from baseline, mixed model for repeated measures (MMRM) | About 100 mL |
| Moderate or severe exacerbations | Worsening needing systemic corticosteroids or antibiotics (moderate) or hospitalisation (severe) | Annual rate, negative binomial regression; time to first event, Cox model | No single agreed value; relative reductions of 15% to 25% are typical targets |
| SGRQ total score | St George's Respiratory Questionnaire, health status from 0 (best) to 100 (worst) | Change from baseline, MMRM; responder analysis, logistic regression | Decrease of 4 units |
| Symptom scores | Breathlessness, cough and sputum (for example CAT, TDI, E-RS) | Change from baseline, MMRM | Depends on the instrument |
| Rescue medication use | Puffs per day of short-acting bronchodilator | Change from baseline, MMRM | Not established |
Bronchodilator trials usually have a lung function primary endpoint. Trials of treatments intended to modify the disease course usually need an exacerbation endpoint, which requires more patients and at least 52 weeks of follow-up. Early-phase trials often measure all of these and treat lung function as the main efficacy signal.
Statistical concepts behind the analysis
1. The estimand
ICH E9(R1) asks the trial to define precisely what is being estimated: the population, the variable, the handling of intercurrent events and the summary measure. In COPD the important intercurrent events are stopping study treatment, starting another maintenance therapy, and death. A treatment policy strategy uses all data regardless of these events, so patients must be followed after they stop treatment. A hypothetical strategy estimates the effect as if patients had stayed on treatment. The choice decides which data are collected and which analysis is valid.
2. Change from baseline with baseline as a covariate
The response is the change from baseline in FEV1, and baseline FEV1 is included as a covariate. This adjusts for chance imbalance between arms and removes between-patient variation, which narrows the confidence interval. The treatment estimate is the same whether the response is the change or the post-baseline value, provided baseline is in the model.
3. Mixed model for repeated measures (MMRM)
The MMRM uses every post-baseline visit from every patient in one model. It includes treatment, visit and the treatment-by-visit interaction as fixed effects, so that the treatment effect is estimated separately at each visit. There are no random effects. Instead, the correlation between visits on the same patient is modelled directly through the residual covariance matrix.
4. Unstructured covariance
An unstructured matrix estimates a separate variance for each visit and a separate covariance for each pair of visits. It makes no assumption about the pattern of correlation, which is why regulators prefer it. With three visits it has six parameters. If the model does not converge, the analysis plan should name the simpler structures to try, in order.
5. Missing at random
A patient who withdraws after week 12 still contributes the week 4 and week 12 values. The MMRM uses these, together with the estimated correlation between visits, to inform the week 24 estimate. The result is unbiased if the data are missing at random: the chance of a value being missing depends only on values already observed. This is more reasonable than analysing completers only, but it cannot be proven from the data, so a sensitivity analysis is needed.
6. Kenward-Roger degrees of freedom
In a mixed model the standard error and degrees of freedom of a treatment difference are approximate. The Kenward-Roger method adjusts the standard error for the fact that the covariance matrix is estimated and calculates matching degrees of freedom. It is the standard choice for MMRM, particularly in small trials.
7. Count data and overdispersion
A Poisson model assumes that the variance of the count equals its mean. Exacerbation counts break this assumption because patients differ in their underlying risk. The negative binomial model adds a dispersion parameter, k, so that the variance is mean + k × mean2. Ignoring overdispersion makes the confidence interval too narrow and the p-value too small.
8. The offset and the rate ratio
Patients are followed for different lengths of time, so the model analyses a rate, not a raw count. The log of each patient's follow-up time in years is included as an offset, a term with its coefficient fixed at 1. The exponential of the treatment coefficient is then the rate ratio: the annual exacerbation rate on active treatment divided by the rate on placebo.
9. Responder analysis
A responder is a patient whose SGRQ total score improves by at least 4 units, the minimum clinically important difference. The proportion of responders is compared with logistic regression and reported as an odds ratio. Patients with a missing week 24 value are usually counted as non-responders. A responder analysis is easy to interpret but uses less information than the continuous score, so it supports the main analysis and does not replace it.
10. Multiplicity
With several endpoints, the chance of at least one false-positive result rises above 5%. Confirmatory trials control this with a pre-specified testing order: the primary endpoint is tested first, and each secondary endpoint is tested only if all the earlier ones were significant. In an exploratory trial, secondary endpoints are reported with nominal p-values and interpreted with caution.
Study design and dummy data
The example is a simulated Phase II, randomised, double-blind, placebo-controlled, parallel-group trial of an investigational inhaled treatment added to standard maintenance therapy.
| Item | Description |
|---|---|
| Population | Moderate to severe COPD, at least one moderate or severe exacerbation in the previous year |
| Randomisation | 60 patients, 1:1 to active or placebo |
| Treatment period | 24 weeks (168 days), with visits at weeks 4, 12 and 24 |
| Primary endpoint | Change from baseline in trough FEV1 (mL) at week 24 |
| Secondary endpoints | Annual rate of moderate or severe exacerbations; SGRQ response at week 24 |
The dataset has one row per patient. The columns are: patient identifier, treatment (1 = placebo, 2 = active), baseline FEV1, FEV1 at weeks 4, 12 and 24, number of exacerbations in the previous year, days of follow-up, number of exacerbations during the trial, baseline SGRQ and week 24 SGRQ. A dot is a missing value. Eleven patients withdrew before week 24.
data copd;
length usubjid $8 trt01p $8;
input usubjid $ trt01pn base fev4 fev12 fev24
prior days nexac sgrqbl sgrq24;
if trt01pn = 1 then trt01p = 'Placebo';
else trt01p = 'Active';
datalines;
CP-001 1 1350 1550 1540 1620 1 168 0 47.7 50.2
CP-002 2 800 800 1110 1020 2 168 2 40.9 31.8
CP-003 1 1130 1040 1290 . 2 113 0 69.8 .
CP-004 2 1470 1590 1490 1570 2 168 1 59.8 48.4
CP-005 1 1720 2020 2020 2010 1 168 1 37.4 38.8
CP-006 2 1490 1180 1370 1380 3 168 0 50.8 50.3
CP-007 1 1570 1720 1710 1420 2 168 1 51.3 56.0
CP-008 2 920 1130 1050 1110 2 168 0 48.6 51.3
CP-009 1 1340 1360 1520 1240 2 168 1 58.6 63.4
CP-010 2 1090 1050 . . 2 69 0 46.4 .
CP-011 1 1170 1070 1070 1020 1 168 0 22.6 9.0
CP-012 2 1260 1200 1380 1460 1 168 0 45.6 50.7
CP-013 1 1010 960 . . 1 71 0 59.4 .
CP-014 2 1550 1860 1840 1660 1 168 1 66.2 59.3
CP-015 1 1470 1420 1560 1500 1 168 3 51.0 43.0
CP-016 2 1700 1700 1600 1640 4 168 1 49.2 43.7
CP-017 1 1340 1360 1430 1330 1 168 0 45.4 43.3
CP-018 2 1080 1070 1320 1180 1 168 0 54.9 46.0
CP-019 1 1310 1350 990 1180 1 168 0 57.6 68.2
CP-020 2 1370 1200 1140 1310 2 168 2 61.8 57.4
CP-021 1 1800 1640 1920 1940 2 168 2 42.0 32.0
CP-022 2 1250 1500 1480 1460 4 168 0 48.1 41.1
CP-023 1 970 1010 . . 2 53 0 50.5 .
CP-024 2 1610 1610 1930 1740 2 168 0 57.7 57.7
CP-025 1 1390 1440 1400 1380 1 168 1 38.7 44.5
CP-026 2 1600 1550 1640 1380 2 168 0 64.0 56.5
CP-027 1 1110 1100 1020 1040 1 168 0 61.9 59.8
CP-028 2 1480 1440 1950 1800 1 168 0 50.6 42.9
CP-029 1 1650 1750 . . 1 65 0 29.0 .
CP-030 2 1180 1330 1490 1250 1 168 0 52.2 54.6
CP-031 1 1690 1530 1740 1520 2 168 0 75.9 78.4
CP-032 2 1480 1370 1360 . 2 102 0 67.9 .
CP-033 1 1100 1130 . . 1 63 0 39.1 .
CP-034 2 1550 1590 1830 1620 1 168 0 75.6 65.7
CP-035 1 970 1040 1040 930 2 168 1 41.5 28.1
CP-036 2 1440 1690 1590 1580 1 168 0 46.9 47.4
CP-037 1 1420 1340 1440 1410 1 168 0 44.7 37.3
CP-038 2 1340 1540 1440 1340 1 168 0 47.6 42.6
CP-039 1 800 860 1030 . 1 126 0 52.9 .
CP-040 2 1460 1570 1550 1720 1 168 0 64.3 59.2
CP-041 1 1660 1600 1420 1710 1 168 1 38.4 34.6
CP-042 2 1750 1750 2050 1770 1 168 0 71.4 57.3
CP-043 1 1110 840 . . 3 76 0 50.5 .
CP-044 2 1970 2110 1820 . 3 126 0 48.6 .
CP-045 1 1080 1010 1150 1010 1 168 0 32.3 29.8
CP-046 2 1270 1410 1290 1390 1 168 0 34.4 23.2
CP-047 1 1300 1190 1360 1380 2 168 0 38.8 40.1
CP-048 2 1590 1730 1860 1790 1 168 1 30.2 30.1
CP-049 1 1630 1670 1590 1760 1 168 0 27.8 17.8
CP-050 2 1490 1380 1250 1560 3 168 0 36.9 52.4
CP-051 1 2000 2210 1910 1970 3 168 0 50.5 54.3
CP-052 2 1070 1130 1160 1070 2 168 0 29.6 11.1
CP-053 1 1040 1350 1180 950 2 168 0 41.1 41.1
CP-054 2 1570 1690 1690 1630 1 168 0 43.9 43.2
CP-055 1 1720 1760 1790 1550 4 168 3 51.8 41.3
CP-056 2 1190 1380 . . 3 48 0 43.6 .
CP-057 1 1360 1300 1190 1170 1 168 0 28.4 35.2
CP-058 2 1320 1690 1760 1570 1 168 1 42.2 32.4
CP-059 1 1330 1600 1530 1460 1 168 0 49.4 45.8
CP-060 2 1700 1610 1600 1710 1 168 0 51.0 36.8
;
run;
SAS code, step by step
Step 1: Build the analysis datasets
The mixed model needs one row per patient per visit. The exacerbation model needs the log of follow-up time in years. The responder flag counts a missing week 24 SGRQ as a non-response.
/* Lung function: one row per patient per visit */
data adfev;
set copd;
array fev{3} fev4 fev12 fev24;
array wk{3} _temporary_ (4 12 24);
do i = 1 to 3;
avisitn = wk{i};
aval = fev{i};
if aval ne . then do;
chg = aval - base; /* change from baseline */
output;
end;
end;
keep usubjid trt01p trt01pn base avisitn aval chg;
run;
/* Exacerbations and SGRQ: one row per patient */
data adsl;
set copd;
length respfl $1;
logyr = log(days / 365.25); /* offset: log follow-up in years */
if sgrq24 ne . then sgrqchg = sgrq24 - sgrqbl;
if sgrqchg ne . and sgrqchg le -4 then respfl = 'Y'; /* improved by 4 or more */
else respfl = 'N'; /* includes missing */
run;
Step 2: Descriptive statistics
proc means data=adfev n mean std maxdec=1;
class trt01p avisitn;
var chg;
run;
proc means data=adsl n sum mean maxdec=2;
class trt01p;
var nexac days;
run;
Step 3: Primary analysis with MMRM
The REPEATED statement with TYPE=UN fits the unstructured covariance matrix within each patient. DDFM=KR requests the Kenward-Roger method. The LSMEANS statement produces the adjusted mean for each treatment at each visit and all pairwise differences; the last step keeps only the comparisons made within the same visit.
ods output LSMeans = lsm
Diffs = diffs;
proc mixed data=adfev method=reml;
class usubjid trt01p(ref='Placebo') avisitn;
model chg = base trt01p avisitn trt01p*avisitn
/ ddfm=kr solution;
repeated avisitn / subject=usubjid type=un r;
lsmeans trt01p*avisitn / diff cl;
run;
/* Keep Active minus Placebo within each visit */
proc print data=diffs noobs;
where avisitn = _avisitn;
var avisitn trt01p _trt01p estimate stderr df tvalue probt lower upper;
run;
Step 4: Exacerbation rate with negative binomial regression
The model adjusts for the number of exacerbations in the previous year, which is the strongest predictor of future exacerbations. The EXP option on the LSMEANS statement turns the difference in log rates into a rate ratio, and ILINK returns the adjusted annual rate for each arm.
proc genmod data=adsl;
class trt01p(ref='Placebo');
model nexac = trt01p prior
/ dist=negbin link=log offset=logyr;
lsmeans trt01p / ilink cl diff exp;
run;
Step 5: SGRQ responder analysis with logistic regression
proc freq data=adsl;
tables trt01p*respfl / nocol nopercent;
run;
proc logistic data=adsl;
class trt01p(ref='Placebo') / param=ref;
model respfl(event='Y') = trt01p sgrqbl;
oddsratio trt01p;
run;
Step 6: Sensitivity analysis for missing data
The MMRM assumes that patients who withdrew would have continued like similar patients who stayed. A reference-based multiple imputation tests a less favourable assumption: missing values in both arms are imputed from a model fitted to placebo patients only, so a patient who withdraws from active treatment no longer benefits from it. The code below is a sketch of the approach; the output is not shown.
/* One row per patient, changes as columns (monotone missing pattern) */
data wide;
set copd;
chg4 = fev4 - base;
chg12 = fev12 - base;
chg24 = fev24 - base;
run;
/* Impute missing changes from the placebo arm */
proc mi data=wide nimpute=100 seed=20261 out=mi_ref;
class trt01pn;
monotone reg(/details);
mnar model(chg4 chg12 chg24 / modelobs=(trt01pn='1'));
var base chg4 chg12 chg24;
run;
/* Analyse week 24 in each imputed dataset */
ods output Diffs = mi_diffs;
proc mixed data=mi_ref;
by _imputation_;
class trt01p(ref='Placebo');
model chg24 = base trt01p;
lsmeans trt01p / diff cl;
run;
/* Combine the results with Rubin's rules */
proc mianalyze data=mi_diffs;
modeleffects estimate;
stderr stderr;
run;
Output
Descriptive statistics for trough FEV1 (mL)
| Statistic | Placebo (N = 30) | Active (N = 30) |
|---|---|---|
| Baseline, mean (SD) | 1351.3 (292.9) | 1401.3 (257.4) |
| Change at week 4: n, mean (SD) | 30, 22.7 (138.6) | 30, 60.3 (151.2) |
| Change at week 12: n, mean (SD) | 25, 45.6 (149.0) | 28, 117.1 (196.2) |
| Change at week 24: n, mean (SD) | 23, -11.7 (135.4) | 26, 92.3 (124.7) |
MMRM: estimated unstructured covariance matrix
| Visit | Week 4 | Week 12 | Week 24 |
|---|---|---|---|
| Week 4 | 21342 | 11255 | 7537 |
| Week 12 | 11255 | 30497 | 12096 |
| Week 24 | 7537 | 12096 | 16997 |
MMRM: least-squares means of change from baseline at week 24 (mL)
| Treatment | LS mean | Standard error | DF | 95% confidence limits |
|---|---|---|---|---|
| Placebo | -12.0 | 26.6 | 50.2 | -65.5 to 41.5 |
| Active | 85.8 | 25.2 | 49.3 | 35.1 to 136.5 |
MMRM: treatment difference, Active minus Placebo (mL)
| Visit | Estimate | Standard error | DF | t value | Pr > |t| | 95% confidence limits |
|---|---|---|---|---|---|---|
| Week 4 | 39.0 | 37.8 | 57.5 | 1.03 | 0.3067 | -36.7 to 114.8 |
| Week 12 | 78.8 | 47.6 | 52.0 | 1.65 | 0.1043 | -16.8 to 174.4 |
| Week 24 (primary) | 97.8 | 36.7 | 49.7 | 2.67 | 0.0103 | 24.1 to 171.5 |
Exacerbations: summary
| Statistic | Placebo (N = 30) | Active (N = 30) |
|---|---|---|
| Patients with at least one exacerbation, n (%) | 9 (30.0%) | 7 (23.3%) |
| Total exacerbations | 14 | 9 |
| Total follow-up (patient-years) | 12.13 | 12.90 |
| Unadjusted rate per patient-year | 1.15 | 0.70 |
Negative binomial model: parameter estimates
| Parameter | Estimate | Standard error | Wald chi-square | Pr > ChiSq |
|---|---|---|---|---|
| Intercept | -0.5324 | 0.5173 | 1.06 | 0.3034 |
| Treatment: Active | -0.5545 | 0.4755 | 1.36 | 0.2435 |
| Exacerbations in previous year | 0.3865 | 0.2351 | 2.70 | 0.1003 |
| Dispersion | 0.5035 | 0.6952 |
Negative binomial model: adjusted annual rates and rate ratio
| Quantity | Estimate | 95% confidence limits | p-value |
|---|---|---|---|
| Placebo, rate per year | 1.11 | 0.61 to 2.02 | |
| Active, rate per year | 0.64 | 0.31 to 1.32 | |
| Rate ratio, Active / Placebo | 0.57 | 0.23 to 1.46 | 0.2435 |
SGRQ total score and responder analysis
| Statistic | Placebo (N = 30) | Active (N = 30) |
|---|---|---|
| Baseline, mean (SD) | 46.2 (12.3) | 51.0 (11.6) |
| Change at week 24: n, mean (SD) | 23, -1.9 (6.8) | 26, -5.1 (7.2) |
| Responders (improved by 4 units or more), n (%) | 7 (23.3%) | 17 (56.7%) |
| Odds ratio, Active vs Placebo (95% CI) | 4.27 (1.37 to 13.28) | |
| p-value | 0.0121 | |
Interpretation
- The primary endpoint is met. At week 24 the adjusted mean change in trough FEV1 was 85.8 mL on active treatment and -12.0 mL on placebo. The difference is 97.8 mL (95% CI 24.1 to 171.5, p = 0.0103).
- The size of the effect is uncertain. The estimate is close to the 100 mL usually regarded as clinically important, but the confidence interval runs from 24 mL to 172 mL. The trial shows that an effect exists; it does not show precisely how large it is.
- The effect builds over time. The difference grows from 39 mL at week 4 to 79 mL at week 12 and 98 mL at week 24. Only the week 24 comparison is significant. This pattern is useful when choosing the duration of the next trial.
- The model handled dropout. Seven placebo and four active patients had no week 24 value. They still contributed their earlier visits, which is why the degrees of freedom (49.7) are higher than a completers-only analysis of 49 patients would give. The model-based means are close to the observed means, which suggests that dropout did not distort the result much under the missing-at-random assumption.
- The exacerbation result is inconclusive. The rate ratio of 0.57 suggests 43% fewer exacerbations on active treatment, but the confidence interval of 0.23 to 1.46 includes 1 and the p-value is 0.24. With 23 events in total, the trial cannot distinguish a large benefit from no benefit. This is expected: the trial was not sized for this endpoint.
- Overdispersion is present but poorly estimated. The dispersion estimate of 0.50 is above zero, which supports the negative binomial model over Poisson. Its standard error is large because there are so few events.
- The SGRQ result supports the primary endpoint. 56.7% of active patients and 23.3% of placebo patients improved by 4 units or more (odds ratio 4.27, 95% CI 1.37 to 13.28, p = 0.0121). The interval is wide, and the p-value is nominal because the trial did not control for multiple endpoints.
- Overall. This is a positive Phase II result on lung function with supportive health-status data. It justifies a larger trial; it is not confirmatory evidence, and the exacerbation effect remains to be shown.
Practical points
- Size the trial for the endpoint that matters. To detect a 100 mL difference in FEV1 with a standard deviation of 200 mL, 80% power and a two-sided 5% significance level, about 64 patients per arm are needed before allowing for dropout. An exacerbation endpoint usually needs several hundred patients per arm followed for 52 weeks.
- Standardise spirometry. Use the same equipment, the same time of day and the same withholding rules for bronchodilators at every visit, with central review of the curves. Poor-quality spirometry adds variability that no model can remove.
- Define an exacerbation precisely. State the severity criteria, and the minimum gap that separates two events (often 7 or 14 days), so that one prolonged episode is not counted twice.
- Follow patients after they stop treatment. A treatment policy estimand needs data collected after discontinuation. Without it, the analysis has to rely on untestable assumptions.
- Pre-specify the covariance structure and its fallback. Name the unstructured matrix first and list the alternatives to be used if the model does not converge.
- Adjust for the right covariates. Baseline value, prior exacerbation history, smoking status and region are common. Include the stratification factors used at randomisation.
- Count exposure correctly. The follow-up time for the exacerbation rate should stop when the patient leaves the study, and some analysis plans also exclude the days during an exacerbation, when a new event cannot start.
- Plan a time-to-first-event analysis. A Kaplan-Meier plot and Cox model for time to first exacerbation complement the rate analysis and are less affected by a few patients with many events.
References
- US FDA. Chronic Obstructive Pulmonary Disease: Developing Drugs for Treatment, Draft Guidance for Industry. 2016.
- European Medicines Agency. Guideline on Clinical Investigation of Medicinal Products in the Treatment of Chronic Obstructive Pulmonary Disease (COPD). 2012.
- ICH E9(R1). Addendum on Estimands and Sensitivity Analysis in Clinical Trials. 2019.
- Global Initiative for Chronic Obstructive Lung Disease (GOLD). Global Strategy for the Diagnosis, Management and Prevention of COPD.
- Jones PW. St George's Respiratory Questionnaire: MCID. COPD. 2005;2(1):75-79.
- Keene ON, Jones MRK, Lane PW, Anderson J. Analysis of exacerbation rates in asthma and chronic obstructive pulmonary disease: example from the TRISTAN study. Pharmaceutical Statistics. 2007;6(2):89-97.
- Kenward MG, Roger JH. Small sample inference for fixed effects from restricted maximum likelihood. Biometrics. 1997;53(3):983-997.
The dataset in this article is simulated and is provided for training purposes only. It does not describe any real product or trial.




