A1258150805cOspB

1
Design of the TELSIS trial Martin
BlandProfessor of Health StatisticsDepartment
of Health SciencesUniversity of York with Dawn
Dowding and Helen Cheyne www-users.york.ac.uk/mb
55/
2
TELSIS The Early Labour Study In Scotland
3

• Background
• Rapid rise in the Caesarean section (CS) rate
  and increased medical interventions in
  labour
• Women admitted while not yet in labour, or in
  the latent phase of labour, are more likely
  to receive some form of intervention than
  those admitted in active labour
• Once diagnosed as being in active labour, a
  woman is expected to progress within strict
  time parameters
• cascade of intervention

4

• Background
• Superficially, diagnosing labour appears
  straightforward but is often problematic in
  practice.
• Between 10 and 30 of women admitted to labour
  wards in the UK are subsequently found not
  to be in labour.
• Systematic review found one study that used
  explicit criteria for diagnosis of labour
  suggests introduction may improve physician
  performance and reduce intervention.

5

• Aims
• to develop explicit criteria for diagnosis of
  labour,
• to compare the use of explicit criteria for
  diagnosis of labour with standard care in
  terms of maternal and neonatal outcomes and
  costs.

6

• Developing the criteria
• Informational cues for labour diagnosis taken
  from clinical literature.
• Used to develop an algorithm to assist with the
  diagnostic process.
• Intervention tested using modelling and
  simulations to identify vulnerabilities or
  weak points.
• Face content validity, inter-rater reliability

• focus groups,
• vignettes,
• questionnaire.

7

• Developing the criteria
• Algorithm had good face and content validity.
• After two stage testing and re-drafting
  inter-rater reliability was good (kappa
  0.86).
• Just as good as midwives using judgement alone
  (kappa 0.83).

8

• Developing the criteria
• Algorithm had good face and content validity.
• After two stage testing and re-drafting
  inter-rater reliability was good (kappa
  0.86).
• Just as good as midwives using judgement alone
  (kappa 0.83).
• Ready for the trial!

9

• TELSIS trial design
• Individual or cluster randomisation?
• The intervention is targeted at midwives with the
  aim of studying the impact on clinical outcomes
  for women.
• Not possible to randomise either individual
  midwives or women without contamination.
• Randomise delivery units.
• Cluster randomised trial.

10

• Feasibility
• There were 20 maternity units with at least 800
  deliveries/year at time of study.
• All were willing to take part.
• Two took part in the feasibility study.

11

• Feasibility
• Test of feasibility of key components of the main
  trial in two delivery units
• consent,
• compliance,
• study materials,
• training needs.

12

• Results of feasibility study
• Consent
• unit 100
• midwives 6/76
• women 89/82
• Compliance 100 / 60
• Midwives reported decision aid acceptable and a
  useful tool for teaching inexperienced midwives .
• Data collection procedures acceptable

13

• Outcomes
• Primary Oxytocin use
• primary marker of slow progress,
• common (30 of primagravidae in Scotland),
• previous study showed a reduction.
• Secondary intervention in labour, maternal and
  neonatal outcomes, NHS resource use cost to
  women.

14

• What sample size do we need
• number of clusters (hospitals).
• number of women within clusters.
• Expected untreated augmentation rate 40.
• Planned difference sought 10 percentage points.

15

• What sample size do we need?
• number of clusters (hospitals).
• number of women within clusters.
• Expected untreated augmentation rate 40.
• Planned difference sought 10 percentage points.
• In an individually randomised trial, to detect a
  difference between 40 and 30 with power 0.90
  would need 476 per group.

16

• What sample size do we need?
• number of clusters (hospitals).
• number of women within clusters.
• Expected untreated augmentation rate 40.
• Planned difference sought 10 percentage points.
• In an individually randomised trial, to detect a
  difference between 40 and 30 with power 0.90
  would need 476 per group.
• For a cluster randomised trial, we must multiply
  this by the design effect.

17

• Design effect
• number of where ?ni2/?ni, ni is the number
  in the i th cluster, and ? intracluster
  correlation coefficient (ICC).

18

• Design effect
• number of where ?ni2/?ni, ni is the number
  in the i th cluster, and ? intracluster
  correlation coefficient (ICC).
• Question what value should we use for ICC?

19

• Design effect
• number of where ?ni2/?ni, ni is the number
  in the i th cluster, and ? intracluster
  correlation coefficient (ICC).
• Question what value should we use for ICC?
• Answer pilot study data.

20
Pilot study data Number of Number of
Number of Percentage Hospital Months
Primigravid Women Given Women Given
Women Syntocinon Syntocinon
1 2 143 62
43.35 2 6 772 185
23.96 3 2 33
8 24.24 4 3 60
28 46.66 5 6 425
119 28.00 6 2
68 21 30.88 7 2
25 7 28.00 8
12 1563 626 40.05 9
2 85 4 4.70
11 3 230 43
18.69 12 5 482 173
35.89 13 1 35
16 45.71 14 5 275
120 43.63
21
Pilot study data Number of Number of
Number of Percentage Hospital Months
Primigravid Women Given Women Given
Women Syntocinon Syntocinon
1 2 143 62
43.35 2 6 772 185
23.96 3 2 33
8 24.24 4 3 60
28 46.66 5 6 425
119 28.00 6 2
68 21 30.88 7 2
25 7 28.00 8
12 1563 626 40.05 9
2 85 4 4.70
11 3 230 43
18.69 12 5 482 173
35.89 13 1 35
16 45.71 14 5 275
120 43.63 Percentage given syntocinon
varies a lot between hospitals.
22
Pilot study data Pilot study proportion of
syntocinon 34. Deliveries per month mean61,
SD40. To obtain ICC, put the data through
one-way analysis of variance. Outcome variable
1 if syntocinon, 0 if not. Factor
hospital. Use Stata.
23
Pilot study data . loneway usesyn hospital
One-way Analysis of Variance for usesyn
Number
of obs 4196
R-squared 0.0358 Source
SS df MS F Prob
gt F ----------------------------------------------
-------------------- Between hospital 33.550954
12 2.7959128 12.95 0.0000 Within
hospital 903.29557 4183 .21594443 ---------
--------------------------------------------------
------- Total 936.84652 4195
.22332456 Intraclass Asy.
correlation S.E. 95 Conf.
Interval ---------------------------------
--------------- 0.04124 0.02588
0.00000 0.09196 Estimated SD of
hospital effect .096372 Estimated
SD within hospital .4646982
Est. reliability of a hospital mean 0.92276
(evaluated at n277.79)
24
Pilot study data . loneway usesyn hospital
One-way Analysis of Variance for usesyn
Number
of obs 4196
R-squared 0.0358 Source
SS df MS F Prob
gt F ----------------------------------------------
-------------------- Between hospital 33.550954
12 2.7959128 12.95 0.0000 Within
hospital 903.29557 4183 .21594443 ---------
--------------------------------------------------
------- Total 936.84652 4195
.22332456 Intraclass Asy.
correlation S.E. 95 Conf.
Interval ---------------------------------
--------------- 0.04124 0.02588
0.00000 0.09196 Estimated SD of
hospital effect .096372 Estimated
SD within hospital .4646982
Est. reliability of a hospital mean 0.92276
(evaluated at n277.79)
25
Pilot study data ICC 0.041. SD between
hospitals 0.096. SD within hospitals (single
deliveries) 0.465. Pilot study proportion
of syntocinon 34.
26
Pilot study data ICC 0.041. SD between
hospitals 0.096. SD within hospitals (single
deliveries) 0.465. Pilot study proportion
of syntocinon 34. Want to detect a reduction
of 10 percentage points.
27
Pilot study data ICC 0.041. SD between
hospitals 0.096. SD within hospitals (single
deliveries) 0.465. Pilot study proportion
of syntocinon 34. Want to detect a reduction
of 10 percentage points. In an individually
randomised trial, to detect a difference between
34 and 24 with power 0.90 would need 431 per
group.
28
Pilot study data where ?ni2/?ni, ni is
the number in the i th cluster, and ?
intracluster correlation coefficient (ICC). We
have ? 0.041. For the design effect, we need
.
29
Pilot study data Number of Number of
Number of Number of Hospital months
primigravid women per women per
women month, ni month2, ni2 1
2 143 71.50 5112.25
2 6 772 128.67
16555.11 3 2 33
16.50 272.25 4 3 60
20.00 400.00 5 6
425 70.83 5017.36 6
2 68 34.00 1156.00 7
2 25 12.50 156.25
8 12 1563 130.25
16965.06 9 2 85
42.50 1806.25 11 5 482
96.40 9292.96 12 1
35 35.00 1225.00 13
5 275 55.00 3025.00 Total
789.82
66861.28 Mean 65.8
30
Pilot study data Number of Number of
Number of Number of Hospital months
primigravid women per women per
women month, ni month2, ni2 1
2 143 71.50 5112.25
2 6 772 128.67
16555.11 3 2 33
16.50 272.25 4 3 60
20.00 400.00 5 6
425 70.83 5017.36 6
2 68 34.00 1156.00 7
2 25 12.50 156.25
8 12 1563 130.25
16965.06 9 2 85
42.50 1806.25 11 5 482
96.40 9292.96 12 1
35 35.00 1225.00 13
5 275 55.00 3025.00 Total
789.82
66861.28 Mean
65.8 ?ni2/?ni, ni 66861.28/789.82 85
for one month.
31
Pilot study data ? 0.041. 85 for one
month, 2 85 for two months, etc. For
one month, design effect 1 (85 1) 0.041
4.44
32
Pilot study data ? 0.041. 85 for one
month, 2 85 for two months, etc. For
one month, design effect 1 (85 1) 0.041
4.44 If we were able to recruit 12 of the 15
available hospitals, 6 per group, this would give
us 6 65.8 366 patients per group.
33
Pilot study data ? 0.041. 85 for one
month, 2 85 for two months, etc. For
one month, design effect 1 (85 1) 0.041
4.44 If we were able to recruit 12 of the 15
available hospitals, 6 per group, this would give
us 6 65.8 366 patients per group. Equivalent
to a sample of 366/4.44 82.4 in an individually
randomised study.
34
Pilot study data ? 0.041. 85 for one
month, 2 85 for two months, etc. For
one month, design effect 1 (85 1) 0.041
4.44 If we were able to recruit 12 of the 15
available hospitals, 6 per group, this would give
us 6 65.8 366 patients per group. Equivalent
to a sample of 366/4.44 82.4 in an individually
randomised study. We need the equivalent of 431
per group.
35
Pilot study data ? 0.041. 2 85 for
two months. For two months, design effect 1
(2 85 1) 0.041

7.93 If we were able to recruit 12 of the 15
available hospitals, 6 per group, this would give
us 2 6 65.8 732 patients per
group. Equivalent to a sample of 732/7.93 92.3
in an individually randomised study. We need the
equivalent of 431 per group.
36
Design effect and sample size Design effect
increases as the sample size increases, because
the number of clusters is fixed Design
Number in Effective Months effect each
group sample size 1 4.444 366
82.4 2 7.929 732 92.3
3 11.414 1098 96.2 4
14.899 1464 98.3 5 18.384
1830 99.5 6 21.869 2196
100.4
37
Design effect and sample size Design effect
increases as the sample size increases, because
the number of clusters is fixed Design
Number in Effective Months effect each
group sample size 1 4.444 366
82.4 2 7.929 732 92.3
3 11.414 1098 96.2 4
14.899 1464 98.3 5 18.384
1830 99.5 6 21.869 2196
100.4 We need the equivalent of 431 per group.
38
Design effect and sample size Design effect
increases as the sample size increases, because
the number of clusters is fixed Design
Number in Effective Months effect each
group sample size 1 4.444 366
82.4 2 7.929 732 92.3
3 11.414 1098 96.2 4
14.899 1464 98.3 5 18.384
1830 99.5 6 21.869 2196
100.4 We need the equivalent of 431 per
group. There are not enough months to reach this,
ever.
39
Design effect and sample size What can we do?
40
Design effect and sample size What can we do? We
can get five times the number of hospitals!
41
Design effect and sample size What can we do? We
can get five times the number of hospitals! But
there are only 15 available in Scotland and we
plan to use 12 in this design.
42
Design effect and sample size What can we do? We
can get five times the number of hospitals! But
there are only 15 available in Scotland and we
plan to use 12 in this design. Go to England?
43
Design effect and sample size What can we do? We
can get five times the number of hospitals! But
there are only 15 available in Scotland and we
plan to use 12 in this design. Go to
England? But, Sir, let me tell you, the noblest
prospect which a Scotchman ever sees, is the
high road that leads him to England! Dr.
Johnson.
44
Design effect and sample size What can we do? We
can get five times the number of hospitals! But
there are only 15 available in Scotland and we
plan to use 12 in this design. Go to
England? But, Sir, let me tell you, the noblest
prospect which a Scotchman ever sees, is the
high road that leads him to England! Dr.
Johnson. Make Scotland bigger?
45
Design effect and sample size Need to remove the
between-hospital variation.
46
Design effect and sample size Need to remove the
between-hospital variation. Cannot do a
cross-over trial, because midwives cannot unlearn
what they have learned.
47
Design effect and sample size Need to remove the
between-hospital variation. Cannot do a
cross-over trial, because midwives cannot unlearn
what they have learned. Solution use baseline
measurements.
48
Design effect and sample size Need to remove the
between-hospital variation. Cannot do a
cross-over trial, because midwives cannot unlearn
what they have learned. Solution use baseline
measurements. Observe the proportion given
syntocinon for several months before intervention
point and several months after. Proportions
before intervention can be used to remove the
inter-hospital variation.
49
Design effect and sample size Need to remove the
between-hospital variation. Cannot do a
cross-over trial, because midwives cannot unlearn
what they have learned. Solution use baseline
measurements. Observe the proportion given
syntocinon for several months before intervention
point and several months after. Proportions
before intervention can be used to remove the
inter-hospital variation. We could use the
change in syntocinon use after intervention
point in a two-sample t test.
50
Design effect and sample size Need to remove the
between-hospital variation. Cannot do a
cross-over trial, because midwives cannot unlearn
what they have learned. Solution use baseline
measurements. Observe the proportion given
syntocinon for several months before intervention
point and several months after. Proportions
before intervention can be used to remove the
inter-hospital variation. We could use the
change in syntocinon use after intervention
point in a two-sample t test. It is better to
do a covariance analysis on baseline.
51
Design effect and sample size Need to remove the
between-hospital variation. Cannot do a
cross-over trial, because midwives cannot unlearn
what they have learned. Solution use baseline
measurements. Observe the proportion given
syntocinon for several months before intervention
point and several months after. Proportions
before intervention can be used to remove the
inter-hospital variation. We could use the
change in syntocinon use after intervention
point in a two-sample t test. It is better to
do a covariance analysis on baseline. Efficient
to use the same number of months before and after.
52
Design effect and sample size Need to remove the
between-hospital variation. Solution use
covariance on baseline measurements. To
estimate the effect on sample size calculations,
we need the correlation between before and after
proportions.
53
Design effect and sample size Need to remove the
between-hospital variation. Solution use
covariance on baseline measurements. To
estimate the effect on sample size calculations,
we need the correlation between before and after
proportions. We can estimate this from the pilot
data.
54
Design effect and sample size Need to remove the
between-hospital variation. Solution use
covariance on baseline measurements. To
estimate the effect on sample size calculations,
we need the correlation between before and after
proportions. We can estimate this from the pilot
data. Done by simulation, the lazy way.
55
Design effect and sample size Need to remove the
between-hospital variation. Solution use
covariance on baseline measurements. To
estimate the effect on sample size calculations,
we need the correlation between before and after
proportions. We can estimate this from the pilot
data. Done by simulation, the lazy way. If we
observed 4 months before and 4 months after
intervention, we estimate that the correlation
would be 0.86 and the standard deviation of the
proportions would be 0.10.
56
Design effect and sample size Using the Stata
sampsi command . sampsi .34 .24 , alpha(0.05)
n1(6) n2(6) sd1(0.1) method(ancova) r01(0.86)
pre(1) Estimated power for two samples with
repeated measures Assumptions
alpha 0.0500 (two-sided)
m1 .34
m2
.24 sd1
.1 sd2
.1 sample size
n1 6
n2 6
n2/n1 1.00 number of follow-up
measurements 1 number of
baseline measurements 1 correlation
between baseline follow-up 0.860  
57
Design effect and sample size Using the Stata
sampsi command   Method ANCOVA relative
efficiency 3.840 adjustment to sd
0.510 adjusted sd1 0.051  
Estimated power power 0.924
58
Design effect and sample size Using the Stata
sampsi command   Method ANCOVA relative
efficiency 3.840 adjustment to sd
0.510 adjusted sd1 0.051  
Estimated power power
0.924 Hence two groups of 6 hospitals with 4
months observations before and 4 months after
would give power greater than 0.9 to detect a
difference of 10 percentage points.
59
Design effect and sample size Using the Stata
sampsi command   Method ANCOVA relative
efficiency 3.840 adjustment to sd
0.510 adjusted sd1 0.051  
Estimated power power
0.924 Hence two groups of 6 hospitals with 4
months observations before and 4 months after
would give power greater than 0.9 to detect a
difference of 10 percentage points. This ignores
weighting for number of deliveries and so might
be too liberal. If we could get more hospitals
we should, if we can observe for longer we
should, if we can observe for longer in small
units we should. 
60
Funding Study funded by the Scottish Executive
Health Department.
61
Funding Study funded by the Scottish Executive
Health Department. Recruitment 14 hospitals
recruited. Allocated to two groups of 7, using
minimisation on presence of midwife birth
unit. Data collection has now begun.
62
Design of the TELSIS trial Martin
BlandProfessor of Health StatisticsDepartment
of Health SciencesUniversity of York with Dawn
Dowding and Helen Cheyne www-users.york.ac.uk/mb
55/