Presented by Tom Smith, Swiss Tropical and Public Health Institute, as part of a symposium organized by MEASURE Evaluation and MEASURE DHS at the 6th MIM Pan-African Malaria Conference.
Methodological Challenges in Evaluating Malaria Control Program Impact: How do we ever find out what worked?
1. Dept. Epidemiology & Public Health
Methodological challenges in
evaluating malaria control program
impact: how do we ever find out what
worked?
Thomas Smith, Melissa Penny, Nakul Chitnis
2. Issues in monitoring and evaluation
According to the Global Fund to Fight AIDS, Tuberculosis and Malaria (GFATM)
“As a general guideline, it is expected that 5 to 10 per cent of the national
program budget be allocated for Monitoring and Evaluation”*.
Many malaria control programs spend far less than this in M & E
36/45 countries have insufficient data to assess trends (Noor, Monday; WMR 2012)
Sufficient data or
Sufficient representative data?
Alignment of data collection and
health systems response?
Which components are working well?
Which components need improving?
Which components are poor value?
*GFATM Guidelines for budgeting in global fund grants (2012)
3. Malaria has decreased in many places.....
Fajara, Gambia
Malaria admissions and deaths 1999-2007
Ceesay Lancet 2008
4. What factors contribute?
We have seen substantial reductions in malaria in many endemic
countries
What are the contributions to this of:
Improved drugs (ACT)?
Improved delivery systems?
LLIN scale-up?
Improvements in housing?
Urbanisation?
Climate (drought in Eastern Africa)?
Evolutionary changes in parasites or vectors?
5. Limitations of analysis of routine data
In real situations multiple
interventions occur at the same
time. Can trends be attributed
to interventions?
5
6. Limitations of analysis of routine data
Improvements in child survival, Tanzania
Between 1999 and 2004:
Important improvements in Tanzania's health
system included:
doubled public expenditure on health;
decentralisation and sector-wide basket
funding
Increased coverage of key child-survival
interventions:
integrated management of childhood illness
insecticide-treated nets
vitamin A supplementation
Immunisation
exclusive breastfeeding.
Non-malaria interventions can also be relevant and
will impact malaria mortality because of disease
interactions
Masanja et al, Lancet 371, 2008
7. Limitations of plausibility designs
Climate change
Economic development
Vector-proof Housing
Urbanisation
Source reduction
Improved case
management
With a complex
system, plausibility can be a
misleading criterion
IRS
Malaria
infection
LLINs
Access to care
Larviciding
Vaccines (?)
Morbidity
Mortality
ACT treatment
Enhancement
Reduction
7
8. Use of field trial data to estimate impact
Field trials tell us the efficacy of an intervention:
-
Coverage data are often available from surveys (e.g. DHS)
-
Can be combined with efficacy data from field trials
-
Example: Cochrane review of ITN impact: LIST model (Eisele et al)
Lengeler, Cochrane Review, 2009
9. Methodological and Policy Limitations of Quantifying the Saving of Lives: A Case Study of the
Global Fund’s Approach
9
McCoy et al, Plos Med 2013
11. Use of static models for estimating impact
In general: meta-analyses of clinical trials do not capture:
Variations between settings in the impact of interventions
Long term temporal dynamics
Effects of loss of immunity
Effects of reduction of infectious reservoir
Interactions between different interventions
12. Complex systems are
characterised by non-linearities
that make the knock-on effects of
changes unpredictable
Limitations of static models
Climate change
Economic development
Vector-proof Housing
Urbanisation
Source reduction
Improved case
management
IRS
Malaria
infection
LLINs
Access to care
Larviciding
Vaccines (?)
Morbidity
Mortality
ACT treatment
Enhancement
Reduction
12
13. Simulation of impact of LLIN campaign
Different models
suggest different
effects on
subsequent clinical
incidence
Briet et al, in preparation
13
14. Statistics from World Malaria Report, 2011
Intervention scale-up
Scale up of ACTs has occurred at the same time as
scale-up of LLINs. Which accounts for the reduction
in burden?
Disease burden
15. Use of static models for estimating impact
Example:
WHO estimates of malaria mortality burden are adjusted
for LLIN coverage, using the figure of 17% mortality
reduction
WHO estimates of malaria mortality burden are not
adjusted for ACT scale-up
Implication:
Evaluations of the impact of ITN scale-up based on
WHO burden statistics will concur with the value 17%
Evaluations of the impact of ACT scale-up based on
WHO burden statistics will attribute minimal impact
16. Simulation modeling
Simulation modelling can be used to support decisions about which
interventions are likely to be most cost effective in given settings, and to
predict what a program should be achieving.
Models used to project disease burden need to consider transmission
dynamics
Such modelling should be seen as a complement to, rather than a
substitute, for the capture of data on coverage or access from the field.
Circular reasoning should be avoided: don’t use a model to estimate disease
burden that assumes specific intervention impacts, and then use the outputs
of this to estimate intervention impacts.
17. Secondary and incidental explanations
Changes in other diseases that interact with malaria
Climate change
Evolution of resistance/insensitivity
Drug resistance
Insecticide resistance
(Vaccine insensitivity)
Evolution of behavioural resistance
Evolution of life-histories (Ferguson et al, 2012)
=> In general these are highly unpredictable
17
18. CDC light trap catches of Anopheles & monthly rainfall patterns
(Tanga region, Tanzania)
Masaika 1998-2001
Kirare, 2004-2009
“Decline(s) in the density of malaria mosquito vectors …. during both study periods
despite the absence of organized vector control”.
Meyrowitsch et al, Malar J. 2011
19. Conclusions so far
Collection of representative data from programs is key for
assessing malariological trends
Attribution of effects to specific interventions is problematical even if
comprehensive data are collected
Mathematical models can help attribute effects to specific
interventions but models should not be used uncritically:
Circular logic should be avoided: don’t use coverage data to simultaneously
estimate trends in burden, and in intervention impact
Models of intervention impact need to allow for temporal dynamics
Some causes of malariological trends may be inherently unpredictable and
hard to model
20. Trial design for empirical estimation of impact
Individually-randomised RCTs are feasible only for estimating drug efficacy.
For testing intervention combinations in the real world, RCTs are infeasible:
Instead:
Plausibility designs?
Before-and-after studies?
One-against-one trials?
Intervention
time
Control
time
Intervention
zone
Control
zone
Comparisons of small numbers of villages/districts?
Intervention
zones
Control
zones
21. Factors that can account for variation between
zones
Climate change
Economic development
Vector-proof Housing
Urbanisation
Source reduction
Improved case
management
IRS
Malaria
infection
LLINs
Access to care
Larviciding
Vaccines (?)
Morbidity
Mortality
ACT treatment
Enhancement
Reduction
21
23. Stepwise introduction of interventions
The inclusion of elements of randomisation in the order of introduction to different
Trial of odour-based traps critical for inferring causality from such data, and the
geographical areas, is on Rusinga Island, Kenya
(population 25,000) should be stressed to program managers.
importance of this
24. Conclusions
Collection of representative data from programs is key for
assessing malariological trends
Attribution of effects to specific interventions is problematical even if
comprehensive data are collected
Mathematical models can help attribute effects to specific
interventions but models should not be used uncritically:
Circular logic should be avoided: don’t use coverage data to simultaneously
estimate trends in burden, and in intervention impact
Models of intervention impact need to allow for temporal dynamics
Direct estimation of the impact of complex intervention programs is
challenging but possible:
Large numbers of clusters need to be evaluated
Stepwise roll-out should be considered
Randomisation of the order of roll-out is critical for generating interpretable
data and is possible even in programme settings.