A good theory of change or a good logic model, the subjects of the last chapter, can be a bridge to good measurement. This chapter describes key measurement concepts, such as how to identify study constructs and find or create indicators of these constructs.
Glennerster and Takavarasha (2013) provide helpful definitions of the terms most often used in discussions of measurement and data collection. The following table expands on this work.
|Construct||A characteristic, behavior, or phenomenon to be assessed and studied. Often cannot be measured directly.||Depression|
|Outcome||In an impact evaluation, ‘constructs’ will be referred to as outcomes—the intended results of the program. Also referred to as an endpoint in a trial.||Decreased depression|
|Indicator||Observable measures of outcomes or other study constructs.||Depression severity score on a depression scale|
|Instrument||The tools used to measure indicators. Also referred to as a measure.||A depression scale, made up of questions/items about symtoms of depression|
|Variable||The numeric values of the indicators.|
|Respondent||The person (or group) that we measure.|
For example, Patel et al. (2016) designed a randomized controlled trial (RCT) in India to test the efficacy of a lay counsellor-delivered brief psychological treatment for severe depression. The hypothesized outcome was a reduction in severe depression. In a theory of change or logic model, outcomes take on the language of change: increases and decreases.
But the word “outcome” is also used more generally and synonymously with “indicator,” particularly in articles reporting study results. For example, Patel et al. (2016) write:
Primary outcomes were depression symptom severity on the Beck Depression Inventory version II and remission from depression (PHQ-9 [Patient Health Questionnaire] score of <10) at 3 months in the intention-to-treat population, assessed by masked field researchers.
Using the language above, the primary outcome was severe depression and the team measured two indicators of severe depression: (1) a depression symptom severity score on the Beck Depression Inventory version II (BDI-II) and (2) a score of less than 10 on the PHQ-9. They used two instruments to measure depression: the PHQ-9 (pdf) and the BDI-II.
Outside of the impact evaluation literature, the word “outcome” is often replaced with “dependent variable” or “response variable.” Additional constructs of interest might be called “covariates,” “independent variables,” or “exposure variables.”
To simplify matters, the following questions are useful when planning a study:
- What is the key construct under study? What other constructs need to be measured at the same time to fully understand the key construct?
- What are the indicators for these constructs? In other words, how can these constructs be quantified?
- What measurement instrument will enumerate these quantities? What type of data will this instrument yield? (This is the topic of the next chapter.)
Fundamentally, there must be a logical flow from the research problem to the measurement of primary study outcomes/constructs. Figure 7.1 demonstrates this idea using Patel et al. (2016) as an example.
The language of qualitative studies is a bit different. These studies emphasize study constructs, but not indicators or measures. Quantification is not the goal.
7.2 Identify Constructs
Most studies are designed to provide the best evidence possible about one or two primary outcomes linked directly to the main study objective. Secondary outcomes may be registered, investigated, and reported as well, but these analyses may be more exploratory in nature if the study design is not ideal for measuring these additional outcomes.
For instance, Patel et al. (2016) included the following secondary outcomes in addition to depression severity and remission from depression:
Secondary outcomes were disability on the WHO Disability Assessment Schedule II and total days unable to work in the previous month, behavioural activation on the five-item abbreviated Activation Scale based on the Behavioural Activation for Depression Scale-Short Form, suicidal thoughts or attempts in the past 3 months, intimate partner violence (not a prespecified hypothesis), and resource use and costs of illness estimated from the Client Service Receipt Inventory.
These outcomes were labeled secondary because the study was powered on the primary outcomes (a topic of a later chapter):
…we aimed to recruit 500 participants to detect the hypothesised effects (a standardised mean difference of 0·42), with 90% power for the primary continuous outcome of depression severity and 92% power to detect a recovery of 65% in the HAP group for our primary binary outcome of depression remission.
The basic idea is that one study cannot definitively answer every possible research question. There are tradeoffs in terms of the time, money, and resources, so investigators must prioritize among all possible outcomes.
The basic idea is that one study cannot definitively answer every possible research question. There are tradeoffs in terms of the time, money, and resources, and investigators must prioritize among all possible outcomes.
7.3 Select Good Indicators
To define a study construct in terms of an indicator and to specify its measurement is to operationalize the construct. Indicators should be DREAMY™:
|Relevant||related to the construct|
|Expedient||feasible to obtain|
|Accurate||valid measure of construct|
|Measurable||able to be quantified|
It is important to clearly specify and define all study variables, especially the indicators of primary outcomes. This is a basic requirement that enables a reader to critically appraise the work, and it serves as a building block for future replication attempts.
For instance, the construct of interest in Patel et al. (2016) was severe depression, and the two indicators were (a) depression symptom severity and (b) remission from depression. The authors preregistered the trial and defined these outcomes as follows:
Mean difference in total score measured at 3 months by the Beck’s Depression Inventory (BDI-II), a 21-item questionnaire assessment of depressive symptoms. Each item is scored on a Likert scale of 0 to 3. It measures depression severity based on symptom scores.
Remission, defined as a score of <10 measured at 3 months by the Patient Health Questionnaire (PHQ-9), a nine-item questionnaire for the detection and diagnosis of depression based on DSM-IV criteria. It is scored on a scale of 0 to 3 based on frequency of symptoms.
Indicators should be relevant to the construct of interest. In Patel et al. (2016), scores on the BDI-II and PHQ-9 are clearly measures of depression severity and remission. An example of a nonrelevant indicator would be scores on the Beck Anxiety Inventory, a separate measure of anxiety. While anxiety and depression are often comorbid, anxiety is a distinct construct.
It should be feasible to collect data on the indicator given a specific set of resource constraints. Asking participants to complete a 21-item questionnaire and a 9-item questionnaire (as in Patel et al. (2016)) does not represent a large burden on study staff or participants. However, collecting and analyzing biological samples (e.g., hair, saliva, or blood) might.
Accurate is another word for “valid.” Indicators must be valid measures of study constructs. In other words, do scores on the BDI-II and PHQ-9 measure a concept called depression?
Indicators must be quantifiable. Psychological constructs like depression are often measured using scales like the BDI-II and the PHQ-9. Other constructs require more creativity. For instance, Olken (2005) measured corruption in Indonesia by digging core samples of newly build roads to estimate the amount of materials used in construction and then compared cost estimates against reported expenditures to calculate a measure of corruption (i.e., by determining the missing expenditures).
Whenever possible, it is smart to use standard indicators and follow existing definitions and calculation methods. One way to learn about standards and customs is to follow the current literature and locate articles that measure the same constructs. Familiarity with what is being published and the methods being used is a significant advantage in achieving publication of a research study. Following these methods lends instant credibility in the submission evaluation and peer-review process. For example, to successfully publish the results of an impact evaluation of a microfinance program in an economics journal, other current papers by economists in high-impact journals can provide important examples of research and data collection methods. How do they measure outcomes like income, consumption, and wealth? Generally, data collection methods should follow the approaches and use the instruments that have already been established in a research field unless the purpose of the study is to overcome the limitations of the standard methods.
In addition, the United Nations Sustainable Development Goals (SDG) provides 230 indicators to measure 169 targets for 17 goals. The SDG indicators are available on a website to assist researchers with their study design.
7.4 Constructing Indicators
7.4.1 SINGLE ITEM INDICATORS
Some indicators are measured with responses to a single item (or a short series of items) on a survey. For instance, in Malaria Indicator Surveys, the “proportion of households with at least one ITN” is defined as the “number of households surveyed with at least one ITN” (numerator) divided by the “total number of households surveyed” (denominator).
The numerator for this indicator is obtained from asking the household respondent if there is any mosquito net in the house that can be used while sleeping and from determining whether each net found in a household is a factory-treated net that does not require any treatment (an LLIN) or a net that has been soaked with insecticide within the past 12 months. The denominator is the total number of surveyed households.
To determine whether a household owns an ITN, survey administrators asked the following sequence of questions.
|119||Does your household have any mosquito nets?|
|120||How many mosquito nets does your household have?|
|121||ASK THE RESPONDENT TO SHOW YOU ALL THE NETS IN THE HOUSEHOLD|
|122||How many months ago did your household get the mosquito net?|
|123||OBSERVE OR ASK BRAND/TYPE OF MOSQUITO NET|
|124||Since you got the net, was it ever soaked or dipped in a liquid to kill or repel mosquitoes?|
|125||How many months ago was the net last soaked or dipped?|
The end result is a binary indicator (yes/no) of whether the household has a bednet that has been dipped in the past 12 months or is factory-treated. In theory, it is possible to ask this in one question—“Does your household have any factory-treated mosquito nets or nets that have been dipped in a liquid to kill or repel mosquitoes in the past 12 months?” But this is a long and complicated question, and it is more effective to break it up into smaller parts.
Sometimes more abstract constructs can be measured with just one survey item. For instance, Konrath et al. (2014) ran 11 studies and found that narcissism can be measured with one question:
To what extent do you agree with this statement: “I am a narcissist.” Response options range from “not very true about me” (1) to “very true of me” (7).37
Most often, however, constructs like narcissism and depression are measured with multiple items that are combined into indexes or scales. The terms index and scale are often used interchangeably, but they are not synonymous. While they share in common the fact that multiple items or observations go into their construction, making them composite measures, the method for and purpose of combining these items or observations are distinct.
Indexes combine items into an overall composite, often without concern for how the individual items relate to each other. For instance, the Dow Jones Industrial Average is a stock-market index that represents a scaled average of stock prices of 30 major U.S. companies such as Walt Disney and McDonald’s. The Dow Jones is a popular indicator of market strength and is constantly monitored during trading hours. Every index has its quirks, and the Dow Jones is no exception. Companies with larger share prices have more influence on the index.
An index popular in the global health field is the DHS wealth index. As a predictor of many health behaviors and outcomes, economic status is a covariate in high demand. Failing to measure economic status in a household survey would be as grave as failing to note a respondent’s gender or age, but measuring economic status is not nearly as easy.38
In an ideal data world, every survey would include accurate information on household income and consumption as measures of household wealth. Income is volatile, however, and consumption is very hard to measure over short periods. Thus, in the late 1990s, researchers proposed creating an index of household assets as a measure of a household’s economic status (Rutstein and Johnson 2004).
Data for the wealth index come from DHS surveys conducted in a particular country. Indicator variables include individual and household assets (e.g., phone, television, car), land ownership, and dwelling characteristics, such as water and sanitation facilities, housing materials (i.e., wall, floor, roof), persons sleeping per room, and cooking facilities. The figure below shows a snapshot of the DHS Household Questionnaire.
A key decision in creating indexes like the wealth index is whether to weight the individual components. Should owning a car be given the same weight as owning a phone? In other words, in constructing an index that measures someone’s wealth, should owning a phone contribute as much to the index as owning a car? Most researchers would probably say no, so the next question is how to assign differential weights to the components. Filmer and Pritchett (2001) first proposed assigning weights via principal component analysis, or PCA.
Principal component analysis is a data reduction technique in which indicators are standardized (i.e., transformed into z-scores) so that they each have a mean of 0 and a variance of 1. If there are 10 items, the total variance is therefore 10, and there are 10 principal components. A principal component (i.e., an eigenvector) is a linear combination of the original indicators; thus, every indicator (e.g., yes/no response to owning a phone) has a loading factor that represents the correlation between the individual indicator and the principal component.
The first principal component always explains the most variance, and in descending order, each component explains a smaller amount of total variance. In constructing the wealth index, the first component measures the concept called “wealth,” so the factor loadings on the first principal component are used to create a score for each household. For example, consider the 2014 Bangladesh DHS survey.
The factor loading for water piped into a dwelling (i.e., indoor plumbing) was 0.056 in the PCA run on the 2014 Bangladesh DHS data. To create the index, this loading is converted into a score representing whether the household has or does not have the asset, and these indicator scores are summed for an overall index score for each household. Once every household has an index score, every participant is assigned to of 1 of 5 wealth quintiles reflecting their economic status (relative to the sample). Hereby, the relationship between health outcomes and wealth can be examined.
In an index, indicators “cause” the concept that is being measured. For example, a household’s wealth is determined by the assets it owns (e.g., livestock, floor quality). Conversely, in a scale, the concept “causes” the indicators.
For an example, consider depression. There is no blood test for depression, so depression is a construct or concept that needs a definition. According to the Diagnostic and Statistical Manual for Mental Disorders, currently the DSM-V, the criteria for major depressive disorder are as follows, A-E:
A. Five (or more) of the following symptoms have been present during the same 2-week period and represent a change from previous functioning; at least one of the symptoms is either (1) depressed mood or (2) loss of interest or pleasure.
- Depressed mood most of the day, nearly every day, as indicated by either subjective report (e.g., feels sad, empty, hopeless) or observation made by others (e.g., appears tearful).
- Markedly diminished interest or pleasure in all, or almost all, activities most of the day, nearly every day (as indicated by either subjective account or observation.)
- Significant weight loss when not dieting or weight gain (e.g., a change of more than 5% of body weight in a month), or decrease or increase in appetite nearly every day.
- Insomnia or hypersomnia nearly every day.
- Psychomotor agitation or retardation nearly every day (observable by others, not merely subjective feelings of restlessness or being slowed down).
- Fatigue or loss of energy nearly every day.
- Feelings of worthlessness or excessive or inappropriate guilt (which may be delusional) nearly every day (not merely self-reproach or guilt about being sick).
- Diminished ability to think or concentrate, or indecisiveness, nearly every day (either by subjective account or as observed by others).
- Recurrent thoughts of death (not just fear of dying), recurrent suicidal ideation without a specific plan, or a suicide attempt or a specific plan for committing suicide.
B. The symptoms cause clinically significant distress or impairment in social, occupational, or other important areas of functioning.
C. The episode is not attributable to the physiological effects of a substance or to another medical condition.
D. The occurrence of the major depressive episode is not better explained by schizoaffective disorder, schizophrenia, schizophreniform disorder, delusional disorder, or other specified and unspecified schizophrenia spectrum and other psychotic disorders.
E. There has never been a manic episode or a hypomanic episode.
If someone meets criteria A–E, they are diagnosed with major depressive disorder (MDD). A diagnosis by a trained mental health professional like a psychiatrist is considered the gold standard measure of depression. Gold standards are in short supply in many places, however, and more feasible methods of measuring this concept called depression are needed. A reasonable alternative is to develop a set of questions (i.e., with answers on scale) that can be administered to measure symptom severity. Presumably, if a person scores high enough on this scale, he or she would be considered “depressed”.
In this example, depression is the latent variable that cannot be measured directly. To create an indicator of depression, a combination of manifest variables that are “caused” by the latent variable depression must be defined.
- I do not feel sad
- I feel sad much of the time
- I am sad all the time
- I am so sad or unhappy that I can’t stand it
- I am not discouraged about my future
- I feel more discouraged about my future than I used to be
- I do not expect things to work out for me
- I feel my future is hopeless and will only get worse
Each item is a manifest variable—something measured directly by asking the question. The latent variable depression is measured indirectly by summing the responses to all 21 manifest variables to create the BDI-II scale score.
Determining the factor strucutre of scales
Exploratory factor analysis
Typically, when developing a new scale, researchers start with a large pool of potential items, many more than can ever be used an applied context (where administration time is a relevant constraint). They then use exploratory factor analysis or some other method of data reduction to shrink the pool.
Exploratory factor analysis (EFA) looks a lot like PCA, but they are conceptually and computationally distinct. Whereas PCA results in a linear combination of indicators that maximized total variance, factor analysis maximizes the common or shared variance.
Factor analysis helps explain the structure of the data. For instance, the BDI-II consists of 21 items that are designed to measure the latent construct of depression, but many studies have examined whether these items can be grouped into subfactors—different domains of depression.
Manian et al. (2013) administered the BDI-II to 953 new mothers “from a large East coast metropolitan area” and then conducted EFA on data from half of the sample.39 They looked for 2- to 4-factor solutions and found that a 3-factor model made the most sense empirically (based on data) and theoretically (based on their knowledge of the literature). Their model suggested that the latent variable of depression is composed of three subfactors: cognitive symptoms, affective symptoms, and somatic symptoms.
Confirmatory factor analysis
Manian et al. (2013) then used the holdout data (i.e., data from their sample not used in the EFA) to test the fit of their 3-factor model through confirmatory factor analysis (CFA). It fit! The model is shown below.
For a project using an existing scale in a new population or setting, CFA is a good technique to determine whether the original factor structure generalizes to the context of the project. Often, the original scale is developed in a high-income setting, and the research might suggest that it makes sense to construct an overall scale score (of some latent variable like depression) AND 2 or 3 subscale scores that correspond to subfactors (like cognitive symptoms and affective symptoms). However, in a different cultural context, the construct might not manifest itself along the same dimensions. CFA can be a helpful tool to determine the applicability of the scale in the new context.
Constructing scale scores
When it comes to scales, how to actually construct scale scores is the primary question. Suppose that many items have been proposed to measure a latent construct like depression. The survey with these items has been administered to a few hundred people, and the EFA and CFA analyses have been conducted to determine the factor structure. What happens next?
Several options are available for constructing scale scores. DiStefano, Zhu, and Mindrila (2009) classify them in two ways: refined and nonrefined.
- Non-refined methods are most commonly used because they are simple to compute and easy to compare across samples.
- Sum raw scores. If there are 21 items, each with a possible range of 0 to 3, the scores are simply summed for each item. This method is used in the BDI-II, which has a possible range of scores from 0 to 63.
- Average raw scores. This method uses the same idea as summing, but averaging keeps the possible range consistent with the response scale. For instance, if 21 items with response options ranging from 0 to 3 are averaged, the scale scores will also range from 0 to 3. This method makes intuitive sense when interpreting results.
- Sum standardized scores. With this method, each item is first standardized to have the same mean and standard deviation. This is often a good option when the standard deviations of the items vary quite a bit.
- Refined methods may produce more exact scores because items are weighted empirically (as opposed to equal weighting in nonrefined methods), and relationships between factors are reflected in the scoring. But refined methods are more complex and require the analyst to make a number of decisions along the way that can lead to very different results.
7.4.4 EVALUATING PSYCHOMETRICS
To evaluate scales like the BDI-II, several psychometric properties are often considered that can be grouped generally into two buckets: reliability and validity. The following example highlights the basic difference between these terms.
Imagine a bathroom scale and a person who weighs 195 lbs. If he steps on, then off, then on again, and the scale reads 210 lbs and then 180 lbs, he would realize that he is the owner of an unreliable scale. So he purchases a new scale. He steps on and off the new scale, and it reads 400.12 lbs the first time then 400.15 lbs the second time. It is very reliable (i.e., has good precision), but unfortunately its measurement was very wrong (i.e., poor accuracy, invalid) because he actually weighs 195 lbs.
A reliable instrument is a consistent instrument. It is consistent over repeated use (as in the bathroom scale example) and consistent among its component parts. Several methods are available for assessing the reliability of an instrument. Here are a few common approaches.
Test–retest reliability is sometimes referred to as stability. Participants complete a survey today, then take it again in after a short period of time, maybe a few days or a week. If each person’s score is the exactly the same the second time, the instrument is perfectly reliable. This is highly unlikely, of course, but a high correlation coefficient (conventionally higher than 0.70) indicates that the survey is stable. Beck, Steer, and Brown (1996) assessed the test–retest reliability of the BDI-II by giving the screening to 26 outpatients in Philadelphia at their first and second therapy sessions, approximately 1 week apart. The test–retest correlation of this cohort was 0.93. The trick to measuring test–retest reliability is knowing when to conduct the retest. If the period between tests is too long, scores will change because people change. If the period between the test is too short, people are much more likely to simply repeat their answers from memory.
When responses to items in an instrument are consistent, the items may be tested for interitem reliability. (If the responses are not consisent, they may not be measuring the same underlying construct.) 1. One approach to finding unreliable items in an instrument is calculating item-total correlations. This process is quite simple: correlate responses on each item with the total scale score. Generally, item-total correlations exceeding 0.30 are sufficient. Beck, Steer, and Brown (1996) reported that item-total correlations for the 21 BDI-II items ranged from 0.39 to 0.70 in the outpatient sample. 2. Another approach is Cronbach’s alpha, a measure of internal consistency. Beck, Steer, and Brown (1996) reported a coefficient alpha value of 0.92 for the outpatient sample. To understand Cronbach’s alpha, remember that instruments are imperfect, even the BDI-II. Every person’s observed score (e.g., their total score on the BDI-II) is actually a function of their ‘true’ score (which is unknowable) plus or minus some amount of measurement error. Cronbach’s alpha provides an estimate of how much variance in an individual score is measurement error. When Cronbach’s alpha is calculated in a program like R or Stata, the program does the equivalent of splitting the dataset into two halves over and over again then calculating the correlation between the total scores for the first half with total scores for the second half. Cronbach’s alpha is the average of all possible correlation coefficients. Cronbach’s alpha has several important characteristics:
- Alpha can range from 0 to 1.
- 0.70 is a rough guide for the low-end of acceptable.
- A value of 1 indicates complete redundancy, suggesting that the items are too similar!
- Alpha is sensitive to the number of items, so a high alpha might just reflect that there are a lot of items included in the scale.
- Alpha is not a property of the test but rather a characteristic of the test when used in a particular sample.
- Alpha should not be used when a scale might tap different latent constructs—only use alpha when the scale is unidimensional.
- Dunn, Baguley, and Brunsden (2014) and others reviewed the limitations of alpha and have suggested coefficient omega as an alternative.
Another type of reliability indicates whether two observers are consistent in observational ratings. Imagine that two observers watch a video of a parent and child interacting and ‘code’ the parent’s behaviors using a depression rating system being tested. If the observers agree to a high degree in their video ratings, the observers are reliable. Here are several items to note about interrater reliability:
- Percentage of agreement is one method of evaluating two raters when the rating is binary, but it does not account for agreement that can happen by chance.
- Cohen’s kappa coefficient does account for agreement by chance; generally, a value greater than 0.40 is desirable.
- Weighted kappa works well when the rating scale is ordinal (e.g., good < better < best), and good vs best represents more disagreement than better vs best, which must be considered in the data analysis.
- Intraclass correlation is a good option when 2 or more raters are used.
The validity of the research instrument foretells the validity of the study data and therefore the study conclusions. It is the basis of credibility. Therefore, validity must be very carefully considered in the study design. Does the measurement instrument measure what needs to be measured with this instrument? For example, is the BDI-II an accurate instrument with which to measure this concept called ‘depression’? If not, any project that uses this instrument will not have a valid measure of depression. There are several types of validity to determine whether an instrument is valid.
Face validity is the weakest form of validity. An instrument is said to have face validity if it appears to measure the construct (i.e., “on its face”). For instance, if a depression instrument asks about depression directly, then it probably has face validity as a measure of depression. This is a weak standard, however. An instrument that appears to be effective can perform very poorly in practice, and an instrument that approaches the research question more indirectly might perform very well.
Depression is a hypothetical construct. For a novel depression instrument to have construct validity, it must be more strongly related to other instruments that also measure depression (i.e., convergent validity) and less strongly (or not at all) related to other instruments that measure something other than depression (i.e., discriminant validity). For instance, Beck, Steer, and Brown (1996) reported that the BDI-II was more positively correlated with the Hamilton Psychiatric Rating Scale for depression (0.71; convergent) than the Hamilton Rating Scale for anxiety (0.47, discriminant). If an instrument has both convergent and discriminant validity, it likely measures the construct as described.
To determine content validity, researchers consider whether the components of an instrument (e.g., each question in a questionnaire) are relevant to the measurement of the larger construct. For example, a question about difficulty sleeping is relevant to the measurement of depression because insomnia is a common symptom of depression. Conversely, a question about compulsive behaviors is probably not relevant because compulsive behaviors are not typical symptoms of the syndrome. Content validity can also be used to account for missing dimensions of a construct. If the BDI-II lacked a question about A5, “psychomotor agitation,” its content validity may be considered incomplete.
7.5 Indicators Throughout the Causal Chain
Indicators that define inputs, activities, and outputs in a logic model can be classified as process indicators. Process indicators capture how well a program is implemented. In short, the “M” (monitoring) in M&E.
Good researchers care about collecting good process and monitoring data to develop a better understanding why programs do or do not work. For example, program costs must be accurately tracked to estimate cost-effectiveness. Or it may be important to determine whether the intervention was delivered according to the plan. Researchers often rely on program partners to deliver the intervention under investigation, so it is important to closely track fidelity to the treatment or intervention plan and compliance with study protocols. Patel et al. (2016) has provided a few examples of process indicators that are often important to intervention researchers.
Inputs are the resources needed to implement the program. The most basic input of all is money; therefore, one indicator is program cost.
Impact evaluations produce estimates of the effectiveness of a program or intervention. Does the program “work”? For some public health and behavioral health researchers and professionals, evidence of impact is enough because they are narrowly focused on developing and testing novel interventions. But policymakers who are thinking about delivering programs at scale with limited public funding want to know whether the intervention is cost-effective, rather than effective.41
A cost-effectiveness analysis requires close tracking of the costs of all program inputs. Patel et al. (2016) indicate that the HAP program costs $66 per person, or $181 per remission from depression at 3 months.
Treatment fidelity is a measure of how closely the actual implementation of a treatment or program reflects the intended design. The consequence of low treatment fidelity is usually an attenuation (i.e., shrinking) of treatment effects, which is a threat to internal validity. If the study shows no effect but treatment fidelity is low, the null result may not be valid. Implementation failure rather than theory or program failure could be to blame. Low fidelity is also a threat to external validity because it is not possible to truly replicate the study.
Patel et al. (2016) measured fidelity in several ways, including external ratings of a randomly selected 10% of all intervention sessions. An expert not involved in the program listened to recorded sessions and compared session content against the HAP manual.
Treatment compliance is a measure of the extent to which people (or units) were treated or not treated according to their study assignment. Sometimes people assigned to the treatment group do not take the treatment, or they complete only part of the planned intervention. It is also possible for members of the control or comparison group to be treated accidentally. Both are examples of broken randomization. Noncompliance to randomization on the treatment side is called one-sided noncompliance. When some members of the control group are also noncompliant with randomization, the break in randomness is called two-sided noncompliance.
Patel et al. (2016) randomly assigned 495 eligible adults to the HAP plus enhanced usual care group (n=247) or the enhanced usual care condition alone group (n=248). No one in the EUC-only group was treated with HAP, but 31% of the HAP group had an unplanned discharge and did not complete the treatment. Analysis strategies for one- and two-sided noncompliance are discussed in a later chapter.
Additional Resources on Indicators
|Malaria||Roll Back Malaria (2013). Household Survey Indicators for Malaria.|
|Measure Evaluation (2016). Monitoring and Evaluation of Malaria Programs.|
|HIV/AIDS||WHO (2015). Consolidated Strategic Information Guidelines for HIV in the Health Sector.|
|TB||WHO (2015). A Guide to Monitoring and Evaluation for Collaborative TB/HIV Activities: 2015 Revision.|
|Family Planning||FP2020 (2015). Measurement Annex.|
Glennerster, R., and K. Takavarasha. 2013. Running Randomized Evaluations: A Practical Guide. Princeton University Press. http://amzn.to/1eQqpvr.
Patel, Vikram, Benedict Weobong, Helen A Weiss, Arpita Anand, Bhargav Bhat, Basavraj Katti, Sona Dimidjian, et al. 2016. “The Healthy Activity Program (Hap), a Lay Counsellor-Delivered Brief Psychological Treatment for Severe Depression, in Primary Care in India: A Randomised Controlled Trial.” The Lancet.
Olken, Benjamin A. 2005. “Monitoring Corruption: Evidence from a Field Experiment in Indonesia.” National Bureau of Economic Research.
WHO. 2015a. “Global Reference List of 100 Core Health Indicators.” World Health Organization. http://apps.who.int/iris/bitstream/10665/173589/1/WHO_HIS_HSI_2015.3_eng.pdf?ua=1.
Konrath, Sara, Brian P Meier, and Brad J Bushman. 2014. “Development and Validation of the Single Item Narcissism Scale (Sins).” PLoS One 9 (8):e103469.
Rutstein, S. O., and K. Johnson. 2004. “The Dhs Wealth Index.” DHS Comparative Reports 6. ORC Macro. https://dhsprogram.com/pubs/pdf/CR6/CR6.pdf.
Filmer, Deon, and Lant H Pritchett. 2001. “Estimating Wealth Effects Without Expenditure Data—or Tears: An Application to Educational Enrollments in States of India.” Demography 38 (1):115–32.
Beck, A.T., Steer R.A., and G.K. Brown. 1996. “Manual for the Beck Depression Inventory-Ii.” San Antonio, TX: Psychological Corporation.
Manian, Nanmathi, Elizabeth Schmidt, Marc H Bornstein, and Pedro Martinez. 2013. “Factor Structure and Clinical Utility of Bdi-Ii Factor Scores in Postpartum Women.” Journal of Affective Disorders 149 (1):259–68.
DiStefano, Christine, Min Zhu, and Diana Mindrila. 2009. “Understanding and Using Factor Scores: Considerations for the Applied Researcher.” Practical Assessment, Research & Evaluation 14 (20):1–11.
Dunn, Thomas J, Thom Baguley, and Vivienne Brunsden. 2014. “From Alpha to Omega: A Practical Solution to the Pervasive Problem of Internal Consistency Estimation.” British Journal of Psychology 105 (3):399–412.
Kim, Maria H, Alick C Mazenga, Akash Devandra, Saeed Ahmed, Peter N Kazembe, Xiaoying Yu, Chi Nguyen, and Carla Sharp. 2014. “Prevalence of Depression and Validation of the Beck Depression Inventory-Ii and the Children’s Depression Inventory-Short Amongst Hiv-Positive Adolescents in Malawi.” Journal of the International AIDS Society 17 (1).
Skeptical? From the authors: “We recognize that some readers may be skeptical about whether simply asking people if they are narcissistic is an appropriate measure of narcissism, given that narcissism is associated with a host of defensive processes. Are people really aware of their own levels of narcissism? We would argue that, based on the evidence from the current studies, people who are willing to admit that they are relatively more narcissistic than others, actually are.”↩
According to (2004), the wealth index is more appropriately understood as a measure of economic status rather than socioeconomic status because it does not include type of occupation and level of education.↩
This hold-out approach enables researchers to develop the model using some existing data and cross validating (i.e., testing) the model on new data not used to build the model. This step is important because the best EFA model, if it is too closely fit to the data, it will not be replicable. In other words, if another study was run to collect new data, the model may not be a good fit to the new data.↩
Decreasing false positives (higher specificity) could be favored to avoid labeling nondepressed patients as depressed and to avoid using resources for unnecessary additional evaluations. Doing so would mean missing more true positives, but it may be a defensible tradeoff.↩
As discussed in Chapter 1, the gap between developing evidence for effective programs and actually implementing them at scale is an example of a “T4” translational research bottleneck.↩