CREATE Biostatistics Core THRio Statistical Considerations

1
CREATE Biostatistics CoreTHRio Statistical
Considerations

• Analysis Plan

2
Design Review
Intervention Train clinic staff to implement
PPD testing procedures among those without
prior TB or INH give INH prophylaxis to those
testing positive.
3
Phased Clinic Entry Into Intervention Status
1/2 3/4 5/6
7/8 9/10 29
Clinic entry to intervention period
Control period
Follow-up period
Intervention period
1 3
5 7 9 29
36 42 Month
4
FirstNeed to Consider Analytic Approach

• Study will take place over 2.5 years, and there
  may be a strong temporal trend in TB incidence
• Perfectly control for calendar time by comparing,
  ON EACH DAY, TB incidence in clinics that are
  still in control status to incidence in clinics
  that are in intervention status
• Assume Poisson process with time-varying
  intensity

where
is the person-days of exposure in the ith clinic
on the tth day,
represents the effect of the tth day
is the log rate ratio comparing those in the
intervention status ( 1)
to those in control status ( 0)
5
Analytic Approach (continued)

• Condition on each days risk set form partial
  likelihood, comparing covariates of incident
  cases to those of the other patients eliminates
• Use clinic-level bootstrap, or robust variance
  estimator, to account for within-clinic
  correlation over time

6
But

• Big delay between intervention in a clinic
• and intervention in an individual in the
  clinic
• (as per Pachecos K-M graphs)
• Sonew primary analysis
• One covariate is fit, which tracks intervention
  status on a given day for a given patient, it
  is the proportion of patients in that patients
  clinic who have had a clinic visit since
  initiation of the intervention in that clinic.

7
Interpretation of Main Analysis

• The interpretation of the coefficient of this
  covariate is that it represents a log hazard
  ratio comparing a clinic whose entire patient
  population has had a visit during intervention
  phase to clinics in control phase. No
  distinction is made between an intervention phase
  clinic with no patients who have made a visit in
  that phase, and clinics in control phase.

8
Other Analyses

• A. Same as Primary, except with the endpoint of
  the complement of TB-free survival (i.e. time to
  earliest of TB or death). This will rely on
  merging in the mortality data base. The idea is
  to make sure to capture those who leave a clinic,
  get TB without it being noted, and then die.
• B. Original primary calculation use of a
  covariate that is 0 if the patients clinic on a
  given day is in control phase, 1 if it is in
  intervention phase. This will have reduced power
  compared to the primary calculation, due to the
  large lag in a clinic entering intervention
  status and the potential receipt of the
  intervention by individual patients.
•  

9
Other Analyses (continued)

• These (C,D,E,F) will be conducted among the
  subgroup of patients who are eligible for the
  intervention, i.e. who have not had prior TB or
  INH prophylaxis.
•  
• C. Use of a covariate that is 0 for a patient
  who has not yet made a visit to his or her clinic
  during the clinics intervention phase, 1 on the
  day of the patients first visit to the clinic
  during its intervention phase. This may have
  more power than the primary calculation, but may
  have some bias due to a potential correlation
  between an individuals frequency of attendance
  and risk of TB (which could be the case if a
  seldom-attender is taking fewer of the prescribed
  ARVs).
• Among clinics in intervention phase only Use of
  a covariate that is 0 for a person who has not
  had a TST, and 1 as of the day of TST reading
  (after 1 September 2005). This measures the
  value of a TST it is not a randomized
  comparison, but is a better measure than in
  control status clinics where TST tends to be
  given to those thought to be more susceptible to
  TB. This measures the impact of initiating the
  intervention at the individual level.

10
Other Analyses (continued)

• E. INH effectiveness use of a covariate that is
  0 unless a patient has started INH prophylaxis,
  at which point it becomes 1.
• F. Intervention effects on processestime-to-even
  t analyses, accounting for within-clinic
  correlation, that
• Compare time from first visit when eligible to
  first TST between clinics on intervention and
  control status.
• Compare time from positive TST to initiation of
  INH prophylaxis between clinics on intervention
  and control status.

11
Tertiary Analyses

• Further elaboration of the foregoing analyses
  will incorporate relevant patient-level
  covariates
• Time-varying Age, CD4, HIV viral load,
  HAART, time on HAART, PPD result
• Fixed Gender
• We will have a great deal of data on
  opportunistic infections. They may be used as
  potential confounders in the foregoing analyses.
  It may also be of interest to look at the
  association between HAART, CD4 and OIs in this
  population. A major use of these data will be
  for descriptive purposes.
•  
• Other analyses will be specific to those who are
  adherent to INH prophylaxis (gt80 meds, or 180
  days). For example, we can estimate rates among
  the intervention status person-years, comparing
  adherent to not, with a timeline starting 12
  months after initiation of IPT for each patient.

12
Handling Correlation

• Currently, plan to form daily risk sets, do
  conditional logistic regression, with a dummy
  variable for whether each of the 29 clinics is in
  intervention status on that day (same as Cox
  model to TB)
• Correlation can be handled with a sandwich
  covariance estimator or, by bootstrapping entire
  clinic histories
• Q sandwich not a great idea when have lots of
  obs per cluster and few clusters but what if
  those lots of obs only have a few events? Perhaps
  10-20 TB events per clinic.