Terminology

Research has shown that the medical literature is steadily becoming more complicated and even out of reach for some medical readers. With a rapidly expanding body of increasingly difficult medical literature, it is essential that the fundamentals of epidemiology be understood. A command of the terminology used in epidemiology is crucial to an understanding of the medical literature.

The common epidemiological terms incidence and prevalence have worked their way into the general vernacular, perhaps without an understanding of their meanings. Incidence is the proportion of a group without a condition that will go on to develop the condition in a specific period of time. For example, the incidence of abusive head trauma would be the number of new cases in a year among the population at risk for abusive head trauma. Prevalence is defined as the proportion of a group who has the condition either at one point in time or during a period of time. For example, the prevalence of sexual abuse among women is said to be 25% or higher but most abuse would not have occurred in the past year.

It should be obvious that the method of data collection will strongly influence the determination of the incidence or prevalence of a disease. When physicians rely on their own personal experience to informally assess the prevalence of a disease, their anecdotal experience will be based on the peculiarities of their practices. It follows, then, that clinicians will be influenced to be alert to a given condition based on their perception of the prevalence of the disease in their population. For example, if a clinician tells a colleague, “I don’t screen for domestic violence in my patient population because it doesn’t occur often in my area,” the clinician’s perception of the prevalence of domestic violence (DV) among his patients influences his clinical decision not to screen for DV. This results in circularity of reasoning: there is little risk of DV in my patient population, so I do not need to screen for DV. I do not screen for DV in my patient population, so I do not identify DV. For this reason, clinicians must be especially aware of the hazards of relying on their anecdotal experiences. They should also carefully read the methodology sections of the research they rely on to determine the measured prevalence or incidence of a disease.

Accuracy and reliability in diagnosing conditions is a concern for all clinicians. If a patient’s signs and symptoms are misdiagnosed as a condition attributable to abuse or neglect, there are adverse consequences to the family and the child, such as further unnecessary diagnostic testing or a child protective services investigation of the family. On the other hand, if a child is abused or neglected and their presenting condition is erroneously identified as being attributable to something else, then the child is underdiagnosed and the necessary cascade of events that should take place to assess and protect the child (and potentially other siblings in the home) does not take place, leaving the child in a potentially dangerous environment.

There are four possible outcomes that result when a diagnosis of abuse or neglect is considered: (1) The patient is correctly identified as abused or neglected, a true positive diagnosis; (2) the patient is correctly identified as not abused or neglected, a true negative diagnosis; (3) the patient is not abused or neglected, but is diagnosed as having been so, a false positive diagnosis; and, (4) the patient is abused or neglected and is not identified as having been so, a false negative diagnosis ( Figure 1-1 ). To determine the accuracy of the diagnosis, one must know what the “gold standard” is. In other words, what is the most accurate way of knowing whether a given patient has a given disease at a given point in time? For many diseases, the gold standard is an autopsy. Since most patients will not have an autopsy to confirm a diagnosis, we are forced to rely on other tests to arrive at a diagnosis. We then weigh our results against what the results could be if we were able to perform the gold standard study.

Diagnostic outcomes in medical decision making.

Sensitivity and specificity are measures of diagnostic accuracy. The sensitivity of a test is defined as its ability to accurately identify the true positives. Increasing the sensitivity likely also results in generating more false positives as an unintended consequence. Very sensitive tests, therefore, are most helpful when the results are negative. Ruling out a condition with a sensitive test can assure the clinician (and the family) that the condition in question most likely does not exist. However, there will be patients who will test positive with sensitive tests that do not actually have the condition. These patients will possibly require further intervention or, in the case of suspected child abuse, further investigation. The specificity of a test is defined as its ability to accurately classify true negatives as persons without the disease in question. If a patient is screened with a highly specific test and the results are positive, it can be stated with certainty that the patient does indeed have the condition (i.e., it is unlikely that the positive test results could be attributed to another condition). Specific tests are more useful when the test is positive, as the condition has been ruled in by the positive results.

Balancing the specificity and the sensitivity of a test is challenging. A test with low sensitivity could miss true positives and result in a child left in a potentially abusive situation, whereas tests with low specificity can result in false positives, which may subject a family to intrusive investigations and a child to unnecessary testing. When the testing is being performed for a disease or condition where the consequences for missing the diagnosis are severe, it is important to choose the most sensitive test, sacrificing specificity if necessary.

By definition, the specificity and sensitivity of a given test in a given patient require information as to whether the patient truly has or does not have the condition. Because this information is not known when the clinician actually orders the test, it becomes necessary to understand the predictive value of the test for this patient. Two types of probability are used to describe the predictive value of a test: the positive predictive value (PPV) and the negative predictive value (NPV). The positive predictive value of a test is the likelihood that a patient has the condition when the test results indicate that they have the condition. The negative predictive value of a test is the likelihood that patient does not have the condition when the test results indicate that they do not have the condition.

The predictive value of a test is different than the test sensitivity and specificity. Sensitivity and specificity are characteristics of the test; the predictive value is a characteristic of the test and the prevalence of the disease in the patients being tested. Consider the situation where a condition has a very low prevalence in a given population; for example, HIV infections in U.S. children. In such a situation, if a test is highly sensitive and the results are positive, it is still most likely to be a false positive. If a highly sensitive test for HIV is performed in a child following a sexual assault, the pretest probability (the prevalence) is low. Therefore, the predictive value (or the posttest probability) of a positive HIV test remains low in this population. If the results are positive, there is a good chance it will be a false positive and will therefore require further testing.

In another example, consider the negative predictive value of the skeletal survey for rib fractures in the potentially abused infant. If an infant presents with a head injury, multiple bruises and a limb fracture and then receives a skeletal survey, the pretest probability is high that other fractures such as rib fractures will be found. Rib fractures are prevalent in abused infants. If no fractures are identified on the skeletal survey, it is quite possible this is a false negative result, despite x-rays being a sensitive test for rib fractures. For this reason, the negative predictive value of this study to correctly identify rib fractures is low and further testing is necessary to confirm the absence of this type of fracture if other data or history prompts concerns about physical abuse.

Clinicians must understand the prevalence of a condition in their patient population to accurately determine the predictive value of a test. As previously discussed, the prevalence of a condition can be difficult to accurately assess in clinicians’ personal patient populations. Prevalence of abuse can be estimated by starting with the population prevalence and then considering risk factors such as persistent crying, demographics factors such as young maternal age, and the specific details of a clinical situation. By relying upon published epidemiological studies, clinicians can hone their skill at estimating the pretest probability of a given condition and therefore better assess the posttest probability when the results are received.

Still, the medical literature must be viewed through a cautious lens. Often, literature will describe risk factors associated with a given condition. Quantifying the effect of a risk factor on the likelihood of a condition can be done in different statistical fashions. One term is attributable risk (AR). The AR is the incidence of a condition in patients with a specific risk factor minus the incidence of the condition in patients without that risk factor. AR estimates should be reserved for situations in which the factors might be considered causal and not simply as confounders. For example, smoking is a causal risk factor associated with lung cancers but male gender is not thought to be causal and is a marker for increased rates of smoking (a confounder).

The relative risk (RR), also known as the risk ratio, describes the risk (probability) of developing a condition when a risk factor is present versus the risk of developing the same condition without the risk factor. One might ask, “What is the relative risk for an infant being fatally abused if there is an unrelated male caregiver in the home?” To answer this question, one would need to know the probability of an infant being fatally abused with and without an unrelated male caregiver. Relative risks are the statistic most often encountered in cohort studies (i.e., longitudinal studies of a defined group of individuals).

Odds ratios (OR) are distinct from RR, but are sometimes used interchangeably when the disease or outcome being looked at is rare. ORs are used in retrospective studies and case-control studies as measures of effect describing the strength of a relationship. They describe the ratio of the odds of having the exposure if you have the disease compared with the odds of having been exposed if you don’t have the disease or outcome. If an OR is greater than 1, the first group has a higher odds of the condition relative to the second group; if the OR is less than 1, the first group has a lower odds of the condition relative to the second group. It is apparent that the OR would be more appropriate in a case-control study or a retrospective study when you consider the fact that the condition (e.g., some type of abuse) is already present in the sample and the researcher is interested in knowing how often a risk factor is present in the group that is abused versus the group that is not. To use RR, the condition is not yet present, the risk factor is, and the researcher is watching to see which subjects develop the condition.

There are situations where a public health policy makes sense, even if it addresses a risk factor which is relatively weak, if that risk factor affects many people in a community. A risk factor with a small relative risk, but which is widely prevalent may play a larger role in the development of a condition than a risk factor that has a high relative risk but is rare in the population. The statistical measure that describes this impact is the population attributable risk (PAR). Public health initiatives related to lead abatement in urban population centers, for example, are targeting a known population attributable risk. In the field of child abuse research, the work on social capital as a modifiable risk factor is an example of attempting to address a PAR. By increasing the social capital in a neighborhood, the risk of abuse for the children in the neighborhood can be reduced more broadly than by attempting to address the risk factors in an individual home. Therefore, while the relative risk is smaller to the individual children in the neighborhood, the widely prevalent risk of decreased social capital will affect more children placing them at increased risk of being abused. This is the so called “prevention paradox. ”

Epidemiological Studies in Child Abuse

Countless studies address risk factors, prevalence, and incidence of child abuse and neglect in the extant body of medical literature. A few of these studies are widely referenced as sources for our understanding of the epidemiology of child abuse and neglect. The National Child Abuse and Neglect Data System (NCANDS) and the National Incidence Studies (NIS) are two of the most commonly cited studies. These studies both address rates of various forms of child abuse, but gather information in very different ways. Data collected in the NCANDS are compiled by child protection agencies across the United States. With an enormous data set consisting of millions of data points, numerous questions can be addressed but there are limitations that must be considered. Children represented in the NCANDS are only those who came to the attention of local authorities. It is unknown how many others are abused or neglected but aren’t identified in this sample. Although many epidemiological questions can be answered by examining the case details and risk factors of children who are reported for suspicions of abuse or neglect, the sample is inherently biased, and these biases can influence our understanding of the problem.

The NIS is a survey that occurs about every 10 years as a form of active surveillance. The goal of this study is to more accurately identify the number of child abuse and neglect cases that occur in a community but that may not come to the attention of child protection authorities. By using community sentinels (i.e., trained observers to report abuse), cases can be identified at a broader population level. With probability sampling techniques, the NIS attempts to demonstrate the impact of abuse nationally through a more affordable research methodology than would be possible if the entire population were studied.

Direct sampling of the population is another methodology used to assess the rate or prevalence of abuse. Examples of this methodology include the CarolinaSAFE study, the National Gallup Poll, and the Juvenile Victimization Questionnaire. In these types of studies, families are contacted and questioned about specific acts of child maltreatment. In some studies parents are questioned, and in others the children are questioned. Interestingly, despite the obvious concern that a parent would be reluctant to report acts of violence against, or inappropriate contact with, their children, these studies have repeatedly found this not to be the case.

Although the studies discussed above are essential tools for researchers, clinicians, and policy makers, there are numerous inherent challenges that researchers and epidemiologists face when designing studies on child abuse and neglect. All of the studies mentioned have some limitations.

Problems in Conducting Research in Child Abuse and Neglect

Difficulties Identifying Child Abuse and Neglect

Both researchers and clinicians are faced with the issue of accurately identifying cases of abuse or neglect. Just as researchers must be cognizant of biases that adversely affect their studies, clinicians should also be aware of biases that adversely affect their ability to diagnose abuse or neglect accurately. These biases are referred to as cognitive biases or cognitive errors. Errors in the correct diagnosis of abuse can directly lead to inaccuracies in the reported epidemiology of abuse. Take, for example, the NCANDS data. Official reports of abuse are often based on the diagnosis of abuse by a medical professional. If medical professionals are systematically misdiagnosing (either over- or under-diagnosing) abuse, the official reported numbers will not be accurate, leading to a skewed understanding of the prevalence of abuse. Several types of cognitive bias are particularly relevant to the recognition of child maltreatment. For example, selection bias can influence whether a sample accurately reflects the underlying population. When selection bias affects a clinician in a clinical encounter, it can influence who is evaluated for a given condition. If a clinician feels that a family is a “nice family” and is therefore at low risk for abusing their children, injuries in their children will not be evaluated in the same way they would be if the child were from a “bad family.” This selection bias will tend to create a self-fulfilling prophecy. If clinicians only search for abuse in “bad families,” they will find examples in troubled families because they are looking for it. By failing to consider abuse in “nice families,” they will fail to identify abuse and will feel justified in their continued practice of relying on their subjective sense of “good” versus “bad” families.

Confirmation bias can also appear in a clinician’s evaluation for abuse. When abuse is in the differential diagnosis, we ask questions and order tests to strengthen our certainty of the diagnosis. We will tend to incorporate positive findings that support our theory and disregard information that does not fit into our cognitive framework for abuse. Similarly, when the patient with injuries that could be due to abuse is seen through the filter of a “good family,” we will be subject to a measurement bias and will look for alternate explanations for the injury and disregard information that would lead us to believe that a “good family” could have harmed their child.

The concept of anchoring is important in a discussion of cognitive biases. With anchoring, we attach great significance to a piece of information and build from there. This focus can derail a clinical evaluation and result in overdiagnosing or underdiagnosing of abuse. As an example, consider a mother presenting to the ED with her 18-month-old child who is limping and refusing to bear weight on her leg. The mother reports this has been going on for 3 days and has progressed to the point where the child wants to be carried everywhere. The mother is young, unemployed, and has two other children under the age of 5 years at home. She appears tired and disheveled. While waiting for the physician, the nurse overhears the mother yelling at someone on her cell phone and using very explicit language. The physician examining the child is told about the overheard phone call before the examination. When examining the patient, the physician feels the mother is rude and abrupt with him. Given the social information up to this point, the physician believes this patient has a high likelihood of an abusive injury. Anchoring on these facts and the clinician’s subjective assessment of the mother’s appearance and behavior, he begins to question her specifically about discipline, other caregivers, and previous reports to child protective services. As the exchange escalates, the mother becomes extremely frustrated and attempts to explain that the child has been sick recently and has had low grade fevers. Because this information does not fit in with a diagnosis of abuse, the physician does not incorporate it into the overall clinical picture. He interprets her increasing hostility as a sign that she has harmed her child, again adding weight to his cognitive anchor. He now will make clinical and diagnostic decisions based on where he has arrived as a result of his anchor: an evaluation for child abuse. Consider the alternative scenario: a mother comes in with the same child but is neatly dressed and calm. She is young but married and employed. Her husband is at home with her other two small children. When the physician arrives to examine the child, he is struck by how articulately the mother describes the child’s symptoms and clinical course. He begins to develop a diagnostic strategy to determine whether this is toxic synovitis. Both of these cognitive anchors drive the interactions and influence the diagnosis. What if the first child had toxic synovitis, but was incorrectly labeled as abused and the second child was abused and incorrectly labeled as toxic synovitis? Cognitive anchors must be consciously recognized to be sure one does not miss important diagnostic information.

Implicit biases or stereotypes are another example of potential cognitive pitfalls. It is clear that a patient’s race plays a significant role in the type and quality of medical care received. The first three rounds of the NIS have failed to show a difference in the rate of abuse by race. Despite this lack of evidence, researchers have shown that minority children, particularly African-American children, are more likely to be medically evaluated for the possibility of abusive injuries and are more likely to be reported to CPS, regardless of whether their injuries are likely accidental or inflicted. A recent study clearly demonstrated that in children with identical injuries, minority children were significantly more likely to be reported to CPS than their white counterparts. There is evidence that abusive head trauma will be missed more frequently in white children. It is unknown whether minority children have greater errors of “overdiagnosis,” but a detection bias is strongly suggested by the work of Lane et al. Clinicians need to center their judgments about the presence or absence of child abuse on risk factors other than race since race has no known contribution to the a priori risk.

Strength of the Evidence

With the field of literature rapidly expanding, we are developing an increasing understanding of the public health impact of child abuse and neglect. We also are recognizing the need to accurately define and measure child maltreatment to identify how to prevent it. As researchers and clinicians become increasingly aware of cognitive pitfalls, we become better able to address them, strengthening our work along the way. It would be wrong to “throw the baby out with the bath water” and walk away from earlier studies because of the flaws we now recognize. The early literature is the foundation on which to build sound hypotheses that we can then test in a more rigorous fashion. Research, like the clinical diagnostic process, is an iterative process.

Future Directions

As we hone our abilities to effectively measure abuse and neglect, we will need to continue to sample diverse populations, both nationally and internationally, to more fully understand who is at risk and how we can modify that risk. We also need to begin to explore how to teach clinicians to be “better thinkers.” We all rely on cognitive shortcuts without realizing it. Problems arise when we do not assess what we as a field know and apply that to what we as individual clinicians do in our clinical practice.