Sie sind auf Seite 1von 8

2009, NUMBER 12

2ND EDITION

PERFORMANCE MONITORING & EVALUATION

TIPS
DATA QUALITY STANDARDS

ABOUT TIPS
These TIPS provide practical advice and suggestions to USAID managers on issues related to
performance monitoring and evaluation. This publication is a supplemental reference to the
Automated Directive System (ADS) Chapter 203.

WHY IS DATA Figure 1. Data Quality Plays a Central Role in Developing


Effective Performance Management Systems
QUALITY
IMPORTANT?
Results-focused development Cycle:
programming requires Plan: Identify or Refine Key Program Objectives
managers to design and Design: Develop or Refine the Performance Management Plan
Analyze Data
implement programs based
Use Data: Use Findings from Data Analysis to Improve Program Effectiveness
on evidence. Since data play a Data
central role in establishing Quality
effective performance
management systems, it is
essential to ensure good data
quality (see Figure 1).
Without this, decision makers
do not know whether to have
confidence in the data, or
worse, could make decisions  Ensuring that limited  Ensuring that Agency
based on misleading data. development resources are program and budget
Attention to data quality used as effectively as decisions in Washington
assists in: possible and the field are as well

1
informed as practically a decision must be made as to
possible The Five Data Quality whether the Ministry’s data
Standards would allow the Assistance
 Meeting the requirements
Objective team to have
of the Government 1. Validity
confidence when assessing
Performance and Results
2. Reliability program performance or
Act (GPRA)
whether they are so flawed as
 Reporting the impact of 3. Precision to be useless, or perhaps
USAID programs to external misleading, in reporting and
stakeholders, including 4. Integrity
managing for results. The
senior management, OMB, main point is that managers
5. Timeliness
the Congress, and the should not let the ideal drive
public with confidence information should be out the good.
consistent with the
DATA QUALITY requirements of good 1. VALIDITY
management. Determining
STANDARDS appropriate or adequate Validity refers to the extent to
Data quality is one element of thresholds of indicator and which a measure actually
a larger interrelated data quality is not an exact represents what we intend to
performance management science. This task is made measure.1
system. Data quality flows even more difficult by the
complicated and often data- Though simple in principle,
from a well designed and
poor development settings in validity can be difficult to
logical strategic plan where
which USAID operates. assess in practice, particularly
Assistance Objectives (AOs)
when measuring social
and Intermediate Results (IRs)
As with performance phenomena. For example,
are clearly identified. If a
indicators, we sometimes have how can we measure political
result is poorly defined, it is
to consider trade-offs, or power or sustainability? Is the
difficult to identify quality
make informed judgments, poverty gap a good measure
indicators, and further,
when applying the standards of the extent of a country’s
without quality indicators, the
for data quality. This is poverty? However, even valid
resulting data will often have
especially true if, as is often indicators have little value, if
data quality problems.
the case, USAID relies on the data collected do not
One key challenge is to others to provide data for correctly measure the variable
determine what level of data indicators. For example, if our or characteristic encompassed
quality is acceptable (or “good only existing source of data by the indicator. It is quite
enough”) for management for a critical economic growth possible, in other words, to
purposes. It is important to indicator is the Ministry of identify valid indicators but to
understand that we rarely Finance, and we know that the then collect inaccurate,
require the same degree of Ministry’s data collection unrepresentative, or
rigor as needed in research or methods are less than perfect, incomplete data. In such
for laboratory experiments. we may have to weigh the cases, the quality of the
Standards for data quality alternatives of relying on less- indicator is moot. It would be
must be keyed to our than-ideal data, having no equally undesirable to collect
intended use of the data. That data at all, or conducting a
1
potentially costly USAID- This criterion is closely related
is, the level of accuracy,
funded primary data to “directness” criteria for
currency, precision, and
collection effort. In this case, indicators.
reliability of performance

2
good data for an invalid Indicator: Number of population. Consider the
indicator. houses in the target following:
community with tin roofs
There are a number of ways to Result: Increased
organize or present concepts This example does not appear revenues in targeted
related to data validity. In the to have a high degree of face municipalities.
USAID context, we focus on validity as a measure of
three key dimensions of increased income, because it Indicator: Number of
validity that are most often is not immediately clear how municipalities where tax
relevant to development tin roofs are related to revenues have increased
programming, including: face increased income. The by 5%.
validity, attribution, and indicator above is a proxy In this case, assume that
measurement error. indicator for increased increased revenues are
income. Proxy indicators measured among all
FACE VALIDITY measure results indirectly, and municipalities nationwide,
Face validity means that an their validity hinges on the while the program only
outsider or an expert in the assumptions made to relate focuses on a targeted group
field would agree that the the indicator to the result. If of municipalities. This means
data is a true measure of the we assume that 1) household that the data would not be a
result. For data to have high income data are too costly to valid measure of performance
face validity, the data must be obtain and 2) research shows because the overall result is
true representations of the that when the poor have not reasonably attributable to
indicator, and the indicator increased income, they are program activities.
must be a valid measure of likely to spend it on tin roofs,
the result. For example: then this indicator could be an MEASUREMENT ERROR
appropriate proxy for
Result: Increased Measurement error results
increased income.
household income in a primarily from the poor
target district ATTRIBUTION design or management of
data collection processes.
Indicator: Value of Attribution focuses on the Examples include leading
median household income extent to which a change in questions, unrepresentative
in the target district the data is related to USAID sampling, or inadequate
interventions. The concept of training of data collectors.
In this case, the indicator has a attribution is discussed in Even if data have high face
high degree of face validity detail as a criterion for validity, they still might be an
when compared to the result. indicator selection, but inaccurate measure of our
That is, an external observer is reemerges when assessing result due to bias or error in
likely to agree that the data validity. Attribution means the measurement process.
measure the intended that changes in the data can
objective. On the other hand, be plausibly associated with Judgments about acceptable
consider the following USAID interventions. For measurement error should
example: example, an indicator that reflect technical assessments
measures changes at the about what level of reductions
Result: Increased national level is not usually in measurement error are
household income in a appropriate for a program possible and practical. This
target district targeting a few areas or a can be assessed on the basis
particular segment of the of cost as well as management
judgments about what level of

3
accuracy is needed for sample to infer the status of chosen sample. Rather, there
decisions. the population as a whole. If are established techniques for
we are interested in describing determining acceptable levels
Some degree of measurement the characteristics of a of non-response or for
error is inevitable, particularly country’s primary schools, for substituting new respondents.
when dealing with social and example, we would not need
economic changes, but the to examine every school in the If a sample is necessary, it is
level of measurement error country. Depending on our important for managers to
associated with all focus, a sample of a hundred consider the sample size and
performance data collected or schools might be enough. method relative to the data
used by operating units However, when the sample needs. While data validity
should not be so large as to 1) used to collect data are not should always be a concern,
call into question either the representative of the there may be situations where
direction or degree of change population as a whole, accuracy is a particular
reflected by the data or 2) significant bias can be priority. In these cases, it may
overwhelm the amount of introduced into the data. For be useful to consult a
anticipated change in an example, if we only use data sampling expert to ensure the
indicator (making it from 100 schools in the capital data are representative.
impossible for managers to area of the country, our data
determine whether progress. Non-Sampling Error
will not likely be
reflected in the data is a result representative of all primary Non-sampling error includes
of actual change or of schools in the country. poor design of the data
measurement error). The two collection instrument, poorly
main sources of measurement Drawing a sample that will trained or partisan
error are sampling and non- allow managers to confidently enumerators, or the use of
sampling error. generalize data/findings to questions (often related to
the population requires that sensitive subjects) that elicit
Sampling Error (or two basic criteria are met: 1) incomplete or untruthful
representativeness) that all units of a population answers from respondents.
(e.g., households, schools, Consider the earlier example:
Data are said to be
enterprises) have an equal
representative if they
chance of being selected for Result: Increased
accurately reflect the
the sample and 2) that the household income in a
population they are intended
sample is of adequate size. target district
to describe. The
The sample size necessary to
representativeness of data is a Indicator: Value of
ensure that resulting data are
function of the process used median household
representative to any specified
to select a sample of the income in the target
degree can vary substantially,
population from which data district
depending on the unit of
will be collected.
analysis, the size of the While these data appear to
It is often not possible, or population, the variance of the have high face validity, there is
even desirable, to collect data characteristics being tracked, the potential for significant
from every individual, and the number of
measurement error through
household, or community characteristics that we need to
reporting bias. If households
involved in a program due to analyze. Moreover, during are asked about their income,
resource or practical data collection it is rarely they might be tempted to
constraints. In these cases, possible to obtain data for under-report income to
data are collected from a every member of an initially demonstrate the need for

4
additional assistance (or over- acceptable level of error not a reliable instrument for
report to demonstrate would be. Unfortunately, measuring fever. In other
success). A similar type of there is no simple standard words, if a data collection
reporting bias may occur that can be applied across all process is unreliable due to
when data is collected in of the data collected for changes in the data collection
groups or with observers, as USAID’s varied programs and instrument, different
respondents may modify their results. As performance implementation across data
responses to match group or management plans (PMPs) are collectors, or poor question
observer norms. This can be a developed, teams should: choice, it will be difficult for
particular source of bias when managers to determine if
collecting data on vulnerable  Identify the existing or changes in data over the life
groups. Likewise, survey or potential sources of error of the project reflect true
interview questions and for each indicator and changes or random error in
sequencing should be document this in the PMP. the data collection process.
developed in a way that  Assess how this error Consider the following
minimizes the potential for compares with the examples:
the leading of respondents to magnitude of expected
predetermined responses. In Indicator: Percent
change. If the anticipated
order to minimize non- increase in income
change is less than the
sampling measurement error, among target
measurement error, then
managers should carefully beneficiaries.
the data are not valid.
plan and vet the data The first year, the project
 Decide whether alternative
collection process with a reports increased total
data sources (or indicators)
careful eye towards potential income, including income as a
need to be explored as
sources of bias. result of off-farm resources.
better alternatives or to
complement the data to The second year a new
Minimizing Measurement
improve data validity. manager is responsible for
Error data collection, and only farm
Keep in mind that USAID is 2. RELIABILITY based income is reported.
primarily concerned with The third year, questions arise
learning, with reasonable Data should reflect stable and as to how “farm based
confidence, that anticipated consistent data collection income” is defined. In this
improvements have occurred, processes and analysis case, the reliability of the data
not with reducing error below methods over time. comes into question because
some arbitrary level. 2 Since it managers are not sure
is impossible to completely Reliability is important so that whether changes in the data
eliminate measurement error, changes in data can be are due to real change or
and reducing error tends to recognized as true changes changes in definitions. The
become increasingly rather than reflections of poor following is another example:
expensive or difficult, it is or changed data collection
important to consider what an methods. For example, if we Indicator: Increased
use a thermometer to volume of agricultural
2
For additional information, refer measure a child’s temperature commodities sold by
to Common Problems/Issues with repeatedly and the results farmers.
Using Secondary Data in the CDIE vary from 95 to 105 degrees,
Resource Book on Strategic A scale is used to measure
even though we know the
Planning and Performance volume of agricultural
child’s temperature hasn’t
Monitoring, April 1997. commodities sold in the
changed, the thermometer is

5
What’s the Difference be perfectly tolerable; for part of the data source to
Between Validity and other indicators, small manipulate the data for
Reliability? amounts of change will be political or personal reasons.
important and even moderate
Validity refers to the Transcription Error
extent to which a levels of measurement error
measure actually will be unacceptable. Transcription error refers to
represents what we simple data entry errors made
intend to measure. Example: The number of
politically active non- when transcribing data from
Reliability refers to the
governmental organizations one document (electronic or
stability of the
measurement process. (NGOs) is 900. Preliminary paper) or database to another.
That is, assuming there is data shows that after a few Transcription error is
no real change in the
years this had grown to avoidable, and Missions
variable being measured, should seek to eliminate any
would the same 30,000 NGOs. In this case, a
10 percent measurement error such error when producing
measurement process
(+/- 3,000 NGOs) would be internal or external reports
provide the same result if
the process were essentially irrelevant. and other documents. When
repeated over and over? Similarly, it is not important to the data presented in a
know precisely whether there document produced by an
market. The scale is jostled operating unit are different
around in the back of the are 29,999 or 30,001 NGOs. A
less precise level of detail is from the data (for the same
truck. As a result, it is no indicator and time frame)
longer properly calibrated at still sufficient to be confident
in the magnitude of change. presented in the original
each stop. Because of this, source simply because of data
the scale yields unreliable Consider an alternative
scenario. If the second data entry or copying mistakes, a
data, and it is difficult for transcription error has
managers to determine point is 1,000, a 10 percent
measurement error (+/- 100) occurred. Such differences
whether changes in the data (unless due to rounding) can
truly reflect changes in would be completely
unacceptable because it be easily avoided by careful
volume sold. cross-checking of data against
would represent all of the
apparent change in the data. the original source. Rounding
3. PRECISION may result in a slight
difference from the source
Precise data have a sufficient 4. INTEGRITY data but may be readily
level of detail to present a fair justified when the underlying
picture of performance and Integrity focuses on whether
there is improper manipulation data do not support such
enable management decision- specificity, or when the use of
making. of data.
the data does not benefit
The level of precision or detail Data that are collected, materially from the originally
reflected in the data should be analyzed and reported should reported level of detail. (For
smaller (or finer) than the have established mechanisms example, when making cost or
margin of error, or the tool of in place to reduce budget projections, we
measurement is considered manipulation. There are typically round numbers.
too imprecise. For some generally two types of issues When we make payments to
indicators, for which the that affect data integrity. The vendors, we do not round the
magnitude of expected first is transcription error. The amount paid in the
change is large, even relatively second, and somewhat more accounting ledger. Different
large measurement errors may complex issue, is whether purposes can accept different
there is any incentive on the levels of specificity.)

6
Technology can help to managers suspect that this importance of the data, a
reduce transcription error. may be a problem, 2) if study is commissioned to
Systems can be designed so managers believe that examine data quality and
that the data source can enter numbers appear to be more specifically whether
data directly into a database— unusually favorable, or 3) if there is any tendency for the
reducing the need to send in a the data are of high value and data to be inflated. The study
paper report that is then managers want to ensure the finds that there is a very slight
entered into the system. integrity of the data. tendency to inflate the data,
However, this requires access but it is within an acceptable
to computers and reliable There are a number of ways in range.
internet services. Additionally, which managers can address
manipulation. First, simply
databases can be developed
understand the data collection
5. TIMELINESS
with internal consistency or
range checks to minimize process. A well organized and Data should be available and
transcription errors. structured process is less likely up to date enough to meet
to be subject to manipulation management needs.
The use of preliminary or because each step in the
partial data should not be process is clearly documented There are two key aspects of
confused with transcription and handled in a standard timeliness. First, data must be
error. There are times, where way. Second, be aware of available frequently enough
it makes sense to use partial potential issues. If managers to influence management
data (clearly identified as have reason to believe that decision making. For
preliminary or partial) to data are manipulated, then performance indicators for
inform management decisions they should further explore which annual data collection is
or to report on performance the issues. Managers can do not practical, operating units
because these are the best this by periodically spot will collect data regularly, but
data currently available. When checking or verifying the data. at longer time intervals.
preliminary or partial data are This establishes a principle
updated by the original that the quality of the data is Second, data should be
source, USAID should quickly important and helps to current or, in other words,
follow suit, and note that it determine whether sufficiently up to date to be
has done so. Any discrepancy manipulation is indeed a useful in decision-making. As
between preliminary data problem. If there is a general guideline, data
included in a dated USAID substantial concern about this should lag no more than three
document and data that were issue, managers might years. Certainly, decision-
subsequently updated in an conduct a Data Quality making should be informed
original source does not Assessment (DQA) for the AO, by the most current data that
constitute transcription error. IR, or specific data in question. are practically available.
Frequently, though, data
Manipulation Example: A project assists obtained from a secondary
the Ministry of Water to source, and at times even
A somewhat more complex reduce water loss for USAID-funded primary data
issue is whether data is agricultural use. The Ministry collection, will reflect
manipulated. Manipulation reports key statistics on water substantial time lags between
should be considered 1) if loss to the project. These initial data collection and final
there may be incentive on the statistics are critical for the analysis and publication. Many
part of those that report data Ministry, the project and of these time lags are
to skew the data to benefit USAID to understand program unavoidable, even if
the project or program and performance. Because of the considerable additional

7
resources were to be year time lag for these data. year. Moreover, to the extent
expended. Sometimes that USAID relies on
preliminary estimates may be While it is optimal to collect secondary data sources and
obtainable, but they should be and report data based on the partners for data collection,
clearly flagged as such and U.S. Government fiscal year, we may not be able to dictate
replaced as soon as possible there are often a number of exact timing
as the final data become practical challenges in doing
available from the source. so. We recognize that data ASSESSING DATA
may come from preceding QUALITY
The following example calendar or fiscal years.
demonstrates issues related to Approaches and steps for how
Moreover, data often measure
timeliness: to assess data quality are
results for the specific point in
discussed in more detail in
Result: Primary school time that the data were
TIPS 18: Conducting Data
attrition in a targeted collected, not from September
Quality Assessments. USAID
region reduced. to September, or December to
policy requires managers to
December.
understand the strengths and
Indicator: Rate of Often the realities of the weaknesses of the data they
student attrition at recipient country context will use on an on-going basis. In
targeted schools. dictate the appropriate timing addition, a Data Quality
In August 2009, the Ministry of the data collection effort, Assessment (DQA) must be
of Education published full rather than the U.S. fiscal year. conducted at least once every
enrollment analysis for the For example, if agricultural 3 years for those data
2007 school year. yields are at their peak in July, reported to Washington (ADS
then data collection efforts to 203.3.5.2).
In this case, currency is a measure yields should be
problem because there is a 2 conducted in July of each

For more information:


TIPS publications are available online at [insert website]

Acknowledgements:
Our thanks to those whose experience and insights helped shape this publication including Gerry Britan
and Subhi Mehdi of USAID’s Office of Management Policy, Budget and Performance (MPBP). This
publication was updated by Michelle Adams-Matson of Management Systems International (MSI).

Comments regarding this publication can be directed to:


Gerald Britan, Ph.D.
Tel: (202) 712-1158
gbritan@usaid.gov

Contracted under RAN-M-00-04-00049-A-FY0S-84


Integrated Managing for Results II

Das könnte Ihnen auch gefallen