Title: Statistics in Medical Research RCTs and Cohort
1Statistics in Medical ResearchRCTs and Cohort
- Jemila Hamid
- Clinical Epidemiology Biostatistics
- Pathology Molecular Medicine
- McMaster University
- jhamid_at_mcmaster.ca
- July 21, 2011
2Outline
- I. Introduction
- II. Group comparison
- Paper I Edwards et al., submitted
- III. Survival Analysis
- Paper II Krag et al, Lancet oncol, 2010
- Paper III Weaver et al., NEJM, 2011.
- IV. Design Issues sample size
- V. Summary
3I. Introduction
- Study types are classified into two broad
categories - Experimental researcher investigates the
effects of intervention - They are prospective studies and are usually
comparative in nature, longitudinal or cross
sectional, parallel vs crossover designs eg.
clinical trials Krag et al., Lancet oncol - Observational researcher doesnt influence
events - Case-control (retrospective and cross sectional
in general) or cohort (prospective and mostly
longitudinal), Surveys (cross sectional) eg.
epidemiology, diagnostic testing, public health
Weaver et al., NEJM
4Statistical Question?
- Estimation? Estimating prevalence of disease,
treatment effect, risk, hazard, accuracy etc. - Comparative? Comparing treatment effect with a
constant (single sample)? Comparing a new
diagnostic test with a gold standard? Comparing
two or more treatments? Comparing before and
after an intervention? - Association and regression? Relationship between
two variables? Effect of one variable on another?
Effect of multiple variables on an outcome? - Prediction? Predict (classify into) disease
subtype? predict an outcome based on risk
factors?
5- Paper II Krag et al.
- Comparing survival and disease free survival
between two surgical procedures
Sentinel-lymph-node resection and
axillary-lymph-node dissection - Paper III Weaver al al.
- Comparison of disease recurrence and survival
between two groups of patients those with occult
lymph-node metastases and those in whom no occult
metastases was detected - In both Papers, estimation and confidence
intervals are also a part of the statistical
question - eg. estimate overall survival,
disease free survival, hazard ratio etc.
6Outcome Measures?
- Continuous - mean, standard deviation, mean
difference - Normal distribution is often assumed
- Transformations, non-parametric approaches
- Binary risk, odds, relative risk, risk
difference, odds ratio, sensitivity, specificity,
classification accuracy, AUC - Logistic regression
- Binomial distribution
- Count - mean count, proportions, rates
- Poisson regression
- Negative binomial
- Survival hazard, hazard ratio, time to event
- Cox regression
- Weibull, lognormal and generalized Gamma
distributions
7Statistical Analysis
- Descriptive Table 1 of medical articles
- Summarizing and evaluating data using
graphical, tabular - This can be done using boxplots, histograms,
normal probability plots - Gives a good feel of data
- Assess distributions normal? need
transformation? - Outliers? What are we going to do about them?
- Missing values? Why are they missing? What are we
going to do about them? -
8Paper II
9Paper III
10Estimation and Confidence Intervals
- Estimator parameter of interest could be mean,
response rate, proportion etc - Confidence interval (CI) quantifies imprecision
or uncertainty associated with an estimate - the reader can assess whether a result is
estimated precisely or not, definitive or not - presenting CI has been widely promoted in the
literature
11Interpretation of 95 CI
- If we repeated the experiment many many times,
95 of the time the TRUE parameter value would be
in the interval - Before performing the experiment, the probability
that the interval would contain the true
parameter value was 0.95
12Set of 95 C.I.s from samples of size n12 drawn
from a normal distribution with ? 211 and s2
46.
1395 CI for continuous outcome
14 1595 CI for proportions
16 17Examples from Papers II and III
- Paper II
- 8 year overall survival were 91.8 (95 CI
90.4-93.4) for group 1 and 90.3 (95
CI88.8-91.8) in group 2 - HR 1.05 (95 CI 0.90-1.22)
- Paper III
- Occult metastases were detected in 15.9 of the
patients (95CI 14.7-17.1) - Adjusted hazard ratio HR 1.40 (95 CI
1.05-1.86)
18Hypothesis Testing
19Examples from papers I-III
- Papers I III, overall and disease free survival
were compared among the groups considered eg
paper II, HR1.2,p value 0.12 paper III,
HR1.40, p-value0.03 - Paper I
- Statistical Question - Comparison of clinical
characteristics of B-Cell lymphoma unclassifiable
(BCLU) with that of Burkitt Lymphoma (BL) and
Diffuse Large B-Cell Lymphoma (DLBL) - Several clinical variables were considered some
binary and some categorical eg. Researchers
compared DLBL and BCLU with respect to Gender
(P-value0.51), CNS involved (p-value0.01)
20Analysis of variance (ANOVA)
21Source df SS MS F P- value
Between Groups k-1
Within Groups n-k
Total n-1
22Example for ANOVA
- In paper I, if the researches were to compare all
the three tumor types DLBL, DCLU and BL with
respect to the a continuous clinical
characteristics, Aanova would be appropriate - For the other non-continous characteristics,
one can apply an appropriate transformation
before applying anova - If they dont use anova and opt to use pairwise
comparisons, there will be an issue of multiple
comparison
23Example Anova
- E.g. (Altman, 1991) Twenty two patients
undergoing cardiac bypass surgery were randomized
to one of three ventilation groups - 50 nitrous oxide and 50 oxygen mixture for 24
hours - Same as 1, but received received treatment
during the operation - No nitrous oxide but received 35-50 oxygen for
24 hours - Compare if three groups have the same red
cell folate (RCL) - levels
24Summary of RCL levels Summary of RCL levels Summary of RCL levels
Group Mean Std Dev Freq.
I 316.6 58.7 s1 8 n1
II 256.4 37.1 s2 9 n2
II 278.0 33.8 s3 5 n3
Total 283.2 22 n
25ANOVA table
Source df SS MS F P-value
Between Groups 2 15515.9 7757.9 3.71 0.04
Within Groups 19 39716.1 2090.3
Total 21 55232 2510.5
At the 5 level, there is evidence to suggest
there is a significant difference in RCL levels
among the three groups.
26Regression Methods
- So far we considered estimation, confidence
intervals and comparisons. Some of these can be
framed as a simple regression model - Anova can, for example, be framed as a regression
model where the treatment groups are independent
variables - Estimation is a big part of regression methods
- But, will present regression models in general
and talk about special cases eg. Cox regression
27Regression Methods
- A method for analyzing relationship between two
or more variables - There is a causal direction investigator seeks
to ascertain the causal effect of one variable
upon another - Otherwise, it will be association or correlation
analysis no causal relationship, here one needs
to measure the strength of association between
variables without assuming any causal
relationship
28- Two kinds of variables outcome (dependent
variable) and predictor (independent variable) - Predictors are sometimes called risk factors,
exposure variables, prognostic factors depending
on the nature of data - Simple linear regression
-
- Y a ß1X1 ß2X2 ßpXp
- Depending on the distribution of the outcome
variable, we have different types of regression
Anova (MD), logistic regression (RR and OR), Cox
regression (HR)
29Example
- E.g. (Altman, 1991) Twenty two patients
undergoing cardiac bypass surgery were randomized
to one of three ventilation groups - 50 nitrous oxide and 50 oxygen mixture for 24
hours - Same as 1, but received received treatment
during the operation - No nitrous oxide but received 35-50 oxygen for
24 hours - Compare if three groups have the same red
cell folate (RCL) - levels
30Examples
- Study of biomarkers Kazu et al., work in
progress - Evaluate the diagnostic ability of a panel of
five immunohistochemical markers in
distinguishing between Endometrioid
Adenocarcinoma (EC) and Serous Carcinoma (SC) - Study the relative contribution of each of the
five markers towards predicting the two
histologic types. - Clinical covariates such as age, body mass index
(BMI), stage, and history of hormone replacement
therapy - Multiple logistic regression and ROC analysis was
performed to estimate odds ratio and construct a
predictive model
31Examples
- Paper II
- Cox regression is used to estimate hazard ratio,
model and compare survival between the two
surgical procedures - Here outcome (dependent variable) is survival and
disease free survival, predictor variable is
surgical groups, other covariates are also
included in the model to estimate adjusted HR - Paper III
- Again, Cox regression is used to model and
compare survival between the two group of
patients - Outcome variables overall survival, disease free
survival and distant-disease free interval.
Predictor variable is two groups of patients,
other covariates are also included here
32Other Methods
- Diagnostic testing
- Agreement studies
- Multivariate methods cluster analysis,
discriminant analysis, factor analysis, PCA, CCA - Meta analysis combining data from different
studies - Methods for correlated data - longitudinal and
repeated measures data - Methods for high-dimensional data genomics and
genetics
33II. Group Comparison
- We will talk about Paper I
- We will focus only on the comparative aspect of
the paper - Talk about group comparison, multiple comparison
using same data
34Paper I
- Statistical Question - Comparison of clinical
characteristics of B-Cell lymphoma unclassifiable
(BCLU) and Diffuse Large B-Cell Lymphoma (DLBL) - Several clinical variables were considered some
binary and some categorical eg. Researchers
compared DLBL and BCLU with respect to the
clinical variables - Survival is also considered in this paper but
we will focus on the comparative aspect of the
paper here
35Materials and methods Paper I
- A ten-year retrospective examination of the
clinical characteristics, survival, treatment
response and molecular profile of BCLU (n34)
compared to DLBL (n97) - Variables considered include Gender, age at
diagnosis, International Prognostic Index (IPI),
Eastern Cooperative Oncology Group (EGOC)
performance status, Ann Arbour stage, presence of
B-symptoms, bone marrow (BM) and central nervous
system (CNS) involvement, extranodal and bulky
disease - Chi-square and one way Analysis of Variance were
used respectively for categorical and continuous
data to compare the baseline characteristics
between groups
36Results paper I
Table 2. Clinical characteristics at diagnosis
and treatment regimes.
Variable DLBL n () DLBL n () DLBL n () BCLU n () BCLU n () P-value
Male Gender 45 (46) 18 18 (53) 0.51
B symptoms 37 (41) 12 12 (40) 0.95
Positive BM 21 (26) 5 5 (19) 0.46
CNS involved 4 (5) 6 6 (20) 0.01
Bulky disease 18 (20) 12 12 (40) 0.03
Extranodal disease 35 (39) 14 14 (47) 0.43
IPI Score 0 1 2 3 4 5 Â 6 16 19 18 11 14 Â (7) (19) (23) (21) (13) (17) Â 3 5 8 6 7 0 Â 3 5 8 6 7 0 Â (10) (17) (28) (21) (24) (0) 0.22 Â
EGOC 0 1 2 3 4 Â 31 27 12 15 6 Â (34) (30) (13) (17) (7) Â 11 10 6 1 2 Â 11 10 6 1 2 Â (37) (33) (20) (3) (7) 0.43 Â
Ann Arbour stage 1 2 3 4 Â 18 18 14 41 Â (20) (20) (15) (45) Â 8 5 8 9 Â 8 5 8 9 Â (27) (17) (27) (30) 0.33 Â
Treatment regime BL-like DLBL-like Palliative No treatment  4 74 13 6  (4) (76) (13) (6)  4 24 1 5  4 24 1 5  (12) (71) (3) (15) Â
      Â
37II. Survival Analysis
- Focus on statistical methods used Papers II and
III - We will discuss
- Study type and design
- Materials and Methods
- Statistical Analysis
- Results
38Paper II
39Paper III
40Study type - Paper II
- Randomized controlled phase 3 trial done at 80
centers across Canada and the USA - Women with invasive breast cancer were randomly
assigned to two surgical procedures - Sentinel-lymph-node resection (SLN) plus
axillary-lymph-node dissection (ALND) - SLN alone with ALND only if the SLNs were
positive - Randomization was stratified by age ( 49, 50),
tumor size and surgical plan (lumpectomy,
mastectomy) - Primary outcome was overall survival but other
secondary outcomes were considered
41(No Transcript)
42Study type - Paper III
- Retrospective and observational study from
previously conducted RCT - Paraffin-embedded tissue blocks of sentinel lymph
nodes obtained from patients with pathologically
negative SLNs were centrally evaluated for occult
metastases - Objective is estimate proportion of patients
with occult metastases and compare survival
between group of patients with and without occult
metasetases
43(No Transcript)
44Methods, survival analysis
- In both papers, the primary outcome is overall
survival secondary outcome disease free
survival, regional control - Both papers used the log rank test and Cox
proportional hazard models - Kaplan-Meier corves were used in both papers
- In both papers, HRs and adjusted HRs with 95 CIs
were provided
45Survival analysis
- Survival analysis is used to analyze time to
event data arises in both clinical and cohort
studies - Event Death, disease occurrence, disease
recurrence, recovery, or other experience of
interest - Time The time from the beginning of an
observation period (e.g., surgery) to (a) an
event, or (b) end of the study, or (c) loss of
contact or withdrawal from the study - We almost never observe the event of interest in
all subjects for these patients, we dont know
their survival time
46Survival analysis
- Censoring/censored observation
- When a subject does not have an event during the
observation time, they are described as censored,
meaning that we cannot observe what has happened
to them subsequently. - A censored subject may or may not have an event
after the end of observation time - Such survival times are called censored.
47Survival analysis
48Survival analysis.
- Median survival- time point at which 50 of the
population survives - Mean survival the average survival time (not
commonly used) calculated as the number if
years survived by all patients divide by the
number of deaths - 5 year survival proportion of patients that
survive 5 years ( it can be 1 year, 2 years, 5
years or 10 years depending on the nature of the
event)
Time to event
49Paper II
50Paper III
51Paper III
52Paper I
53Comparing survival curves
- One use KM curves is to compare survival between
two or more groups - We can visually compare median survival, mean
survival, 5 year survival etc. - We need more objective way of comparing long
rank test
Remission time for acute myelogenous leukemia.
Group 1 received maintenance chemo, group 2,
didnt
54Log rank test
Alive Dead Total
Group 1 a1 d1 n1
Group 2 a2 d2 n2
Total a d n
Â
55Cox proportional hazard model
56Results from paper II
- Comparison of survival between group 1 (SLR
ALND) and group 2 (SLN alone) resulted in
unadjusted HR 1.20 (95CI 0.86-1.50,
p-value0.12) - 8 year KM estimates for overall survival were
91.8 (95CI 90.4-93.3) for group 1 and 90.3
(95CI88.8-91.8) for group 2 - Comparison for disease free survival resulted in
HR of 1.05 (95CI0.90-1.22,p-value0.54) - 8 year KM estimates for disease free survival
were 82.4 (95CI 80.5-84.4) for group 1 and
81.5 (95CI79.6-83.4) for group 2
57Results from Paper III
- Occult metastases were detected in 15.9 (95CI
14.7 -17.1) of the patients - A significant difference in overall survival
(p-value0.03), disease free survival
(p-value0.02) and distant disease free survival
(p-value0.04) between the two groups of patients - The 5 year KM survival estimates for group 1 (in
whom occult metastases was detected were 94.6,
86.4, 89.7 for overall, disease free and
distant disease free survival, respectively. - For group 2, the estimates are 95.8, 89.2, and
92.5, respectively
58(No Transcript)
59IV. Design Issues sample size
- Determining the sample size for a study is a
crucial component of study design - Small sample size leads to imprecise estimates
and the study will be under powered to detect
differences, in particular when the effect size
is very small - Using too many subjects may result in
statistically significant conclusions and clear
future study directions however, the same
answer could have been obtained with fewer
subjects studies are over powered and lead to
wasted resources - Some treatments are also invasive
60- Important to choose primary outcome
- There are important parameters needed for sample
size calculations desired power (1-type II
error), level of the test (Type one error or a),
variance (or standard deviation) and effect size
(the minimum difference to be detected) - Eg. comparing two means 80 power, 0.05 level,
range of sd, range of MD
- N102 (50 for each group is needed to achieve 80
power when 3.5 and sd3 - Only N9 (3 for each group is needed to achieve
80 power when 1 and sd1
61- Multiple logistic regression we need to chose
the primary predictor variable - Provide parameter values based on this primary
predictor - Consider the correlation between these predictor
with other predictors or covariates - When correlation is high, we need larger sample
size
62Summary
- Statistical question, outcome variable, primary
and secondary outcomes - Estimation, CI, hypothesis testing
- Multiple comparison
- Survival analysis and linear regression
(comparison between two or more groups) - Design issues sample size