Attrition in a 30-year follow-up of a perinatal birth risk cohort: factors change with age

Base cohort

The birth cohort originates from Kätilöopisto maternity hospital, Helsinki, Finland, in 1971–1974 and has been prospectively studied up to adulthood. The hospital is one of two major maternity hospitals, the other being the University Central Hospital, in the Helsinki metropolitan area of approximately 900 residents at the time of enrollment. The hospital served as the primary maternity hospital serving an area of approximately 4,000 km². Sixty percent of the mothers lived in Helsinki (mean distance to hospital 8 km), 7% in Espoo (distance 15 km), 15% in Vantaa (distance 14 km) and the remaining 15% in the surrounding municipalities (average distance to hospital 30 km) (Fig. 1). All risk pregnancies known in advance as well as expected birth complications, were directed to the University Central Hospital equipped with a neonatal intensive care unit. Those born in Kätilöopisto therefore represent an unselected sample of deliveries within the area.

There were 22,359 consecutive births, which at the time accounted for about 10% of all births in Finland and about one third of those in the county of Uusimaa. Of those, 2,113 (9.5%) were cesarean sections. Newborns were consecutively screened for predefined perinatal risks: APGAR score lower than 7 at 5 or 15 min (n = 260), birth weight under 2,000 g or less (n = 221) out of whom 46 under 1500 g, significant hyperbilirubinemia (n = 353), severe respiratory difficulties including infections (n = 83), neurological symptoms (n = 114), maternal diabetes (n = 86), infant hypoglycemia (n = 92). Altogether 1,196 (5.3% of the 22,359 births) had at least one risk, 22% of them had more than one risk factor. Additionally, 130 cases were screened, but did not fulfill the inclusion criteria. Of the 1,196 infants, 202 died or were severely disabled and excluded from the study. The inclusion of patients is given in Fig. 2.

Follow-up and other clinical visits

There were several clinical visits after birth, and the children were seen at 6, 12, 18, and 24 months by the physicians involved in the longitudinal study. Additionally, almost all infants were seen by a nurse at child-health centers as part of a public healthcare program, at 6, 10, 18, 24, and 48 months (Vuorenkoski, Mladovsky & Mossialos, 2008). Vaccinations were given on these visits following a schedule adapted from the WHO recommendations, and the vaccination program was continued in school health care. Therefore, the children and parents had frequent face-to-face contacts with health care personnel in addition to study visits. Moreover, clinical visits and psychological and logopedic test sessions took usually several hours, which means that the time parents were in contact with the study personnel was substantial. The parents did not receive financial compensation. The participants of the study had normal access to the public health care system at all times, as the study was not meant to substitute for public health services.

Methods

At birth, detailed data was recorded about maternal risk factors (e.g., smoking, prescribed medication and radiological examinations), family genetic traits, and medical data of delivery.

At five years, a clinical evaluation, age-appropriate psychological tests, logopedic tests, and questionnaires were used. The clinical evaluation included the Neurodevelopmental screen (NDS; Bax & Whitmore, 1973; Michelsson & Donner, 1981). The psychological evaluation included the Frostig Developmental Test of Visual Perception (Frostig, Lefever & Whittlesey, 1966), the Dubowizt screening test (Dubowitz, Leibowitz & Goldberg, 1977), and the Draw a Person Test (Goodenough, 1926). The logopedic evaluation included the Illinois Test of Psycholinguistic Abilities (Kirk, McCarthy & Kirk, 1968), a systematic evaluation of speech and articulation, and a Name writing test. The parents filled out a structured questionnaire regarding the health, development and behavior of the child. In addition, some kindergartens were contacted. The parents were also interviewed.

At nine years a clinical evaluation, age-appropriate psychological tests, logopedic tests, and questionnaires were again used. The clinical evaluation included the Test of Motor Impairment (Stott, Moyes & Henderson, 1972) and the assessment of neurological signs (Lindahl, Michelsson & Donner, 1988; Stokman et al., 1986). The psychological evaluation included the Draw a Person Test (Goodenough, 1926) and the Wechsler Intelligence Scale for Children (Wechsler, 1949). The logopedic evaluation included the Illinois Test of Psycholinguistic Abilities (Kirk, McCarthy & Kirk, 1968), the systematic evaluation of speech and articulation, the Diagnostic Reading and Writing tests for Finnish school children (Ruoppila, Roman & Vasti, 1968; Ruoppila, Roman & Vasti, 1969) and the Name Printing Test (Reimer et al., 1975). The parents filled out a structured questionnaire regarding the health, development and behavior of the child (Supplemental Information S1). Parents were asked about the usefulness of the study and rehabilitation measures initiated based on the visit at 5 years of age. The teachers filled out a structured questionnaire regarding school performance and behavior as well as the personality of the child (Supplemental Information S2). The parents were also interviewed face-to-face.

At 16 years of age, questionnaires were used for the whole cohort. A clinical evaluation and psychological tests, not analyzed in this paper, were conducted for a subsample of children. The clinical evaluation included the Test of Motor Impairment (Stott, Moyes & Henderson, 1972) and the assessment of neurological disorder (Lindahl, Michelsson & Donner, 1988; Stokman et al., 1986). The psychological evaluation included subtests from the Wechsler Adult Intelligence Scale (Wechsler, 1955). The questionnaires included early versions of the Child Behavior Check List for parents and the Youth Self-Report (Achenbach & Edelbrock, 1983; Achenbach & Edelbrock, 1987). The parents also filled out a structured evaluation of the child’s behavior and personality (Supplemental Information S3, S4, S5).

At 30 years, a questionnaire was used to explore educational and occupational outcome, health, and life satisfaction (Supplemental Information S6). It included also the ADHD Current Symptoms Scale as the ADHD Childhood Symptoms Scale with each of the 18 DSM-IV (American Psychiatric Association, 1994) diagnostic symptom criteria scored from zero to three depending on the severity of symptoms (Barkley & Murphy, 1998).

The family’s socioeconomic status (SES) was recorded at all levels. The social class variable was defined as the best score on a 5-item scale based on father’s occupation, and if unavailable, mother’s occupation was used. The psychosocial distress score was formed to include poor housing conditions, divorces, relocations, alcohol abuse, unemployment, family conflicts, domestic violence, imprisonment, severe diseases, or mental problems in the family.

Statistics

Participation frequencies were calculated at birth, 5, 9, 16 and 30 years. The children, who were not seen on the 5-year visit or any other visit thereafter, were included only in the analysis of 5-year results. First, we tested for significant group differences with the Mann–Whitney-U test with participation at different age levels as the grouping variable. Ordinal and categorical variables were cross-tabulated, the Fisher’s exact test or the Pearson Chi-square test were used for significance testing and Cramér’s V for association. Regression methods were not used, because in our follow-up cohort, neither the dependent nor the predictor variables belong to the same pool of variance as the person who decides of participation changes in course of the follow-up. Bonferroni correction was used to correct the p values for the required alpha level of 0.05 due to multiple significance testing. Statistica 12 software was used for the calculations (Statsoft, 2013).

To identify clinically meaningful patterns and combinations of explanatory variables, we performed a classification tree analysis using the random forest algorithm (Breiman, 2001a). In a situation with a very large numbers of predictors which are likely to interact in an unknown fashion, classification tree analysis (e.g., Rokach and Maimon, 2007) provides a flexible and robust way to find the most important predictors of attrition. The present data was collected at fixed time points. Therefore, classification tree analysis was used instead of survival analysis, which is often used to analyze study designs where time is measured as a continuous variable. The basic idea of classification tree analysis is simple. Prediction of a response or class Y from inputs X₁, X₂, …, X_p is done by growing a binary tree. At each internal node in the tree, one applies a test to one of the inputs X_i. Depending on the outcome of the test, one goes to either the left or the right sub-branch of the tree. Eventually one comes to a leaf node, where prediction is made. This prediction aggregates or averages all the training data points that reach that leaf. In contrast to this, predictors like linear or polynomial regression are global models, where a single predictive formula is supposed to hold over the entire data space. As noted previously, when the data has lots of features which interact in complicated, and possible nonlinear ways, assembling a single global model can be very difficult. Detailed presentation of classification tree algorithms can be found e.g., in Breiman et al. (1984) and in Kohavi & Quinlan (2002).

Data mining techniques are useful in exploratory studies (Mendez et al., 2008) when data cannot be fit into models using goodness-of-fit methods and statistically significantly differences are difficult to interpret. As compared to such classical data mining situations the sample size in the present study is small. Therefore the random forest algorithm (Breiman, 2001a) was used, because it prevents overfitting the model which is problem in classical classification tree analysis (Breiman, 2001b). Model accuracy was optimized by testing predictive accuracy, which was in the range of 0.71–0.97 at different age levels. The Gini index (Gini, 1909) was used as a measure for variable importance. Gini index is the weighted mean of the improvement that each variable causes for the splitting criterion. Maximal classification accuracy and the most important variables using Gini importance score are reported (Sandri & Zuccolotto, 2006). Random forest classification was done using the Salford system Predictive Modeler software (Salford Systems, 2013). In all analysis Bootstrap aggregating was used for training algorithm (Breiman, 1996). Detailed nontechnical description of random forest algorithm can be found in Touw et al. (2013).

Ethics

Ethical review has been conducted over the course of the longitudinal study, and the latest approval was obtained from the Ethical Review Board of the Helsinki and Uusimaa hospital district in May 2013 (number 147/13/3/00/2013).

Attrition

There were 130 infants who were seen neither at the 5-year visit nor on any later visit. The initial loss during the first 5 years therefore was 13%. However, 92 of them had been seen at an early clinical visit or child health centers during the first 40 months of life, thus only 38 were completely LTFU after birth. At 5 years, 839 children were assessed (84% of those alive and not severely disabled, n = 994), at the age of 9 years 756 (76%), including 25 children not assessed at 5 years. At the age of 16 years 560 responded (65% of those who participated at either 5 or 9 years, n = 864) and at the age of 30 years 469 responded (54%). The attrition rate thus increased from 16% at 5 years to 46 at 30 years. There was data on 78 (8%) children on one age level only, on 152 (15%) children on two levels, on 288 (29%) children on three levels, and on 343 (35%) on all four age levels of the study. Participation of subjects on the different age levels is given in Figs. 3 and 4. At the age of 9 years 54% of the parents regarded participation in the study useful, 26% were undetermined, 13% did not answer that question, and 7% did not find participation useful.

Group comparisons of responders and non-responders

Between the groups of children who were not brought to clinical evaluations at all (n = 130) and those who did not participate at five years, but participated later (n = 25), there was a statistically significant difference only in the father’s social class (mean responders 2.2 ± 0.9 and 2.6 ± 1.0 in non-responders). The social class was the only variable that was significantly different between responders and non-responders at all other age levels too, a lower socioeconomic class being associated with non-responding.

A complete list of 63 variables tested for group differences is presented in Supplemental Information S7. Notably, the type of perinatal risk had no apparent effect on the number of visits at any age level (Kruskal–Wallis test H = 8.10, p = 0.32). Also, maternal age, gestational age, birth weight, or Apgar scores or the distance from home to the hospital had no statistically significant differences on any age level.

Classification analysis

Classification accuracy of classifying cases into responders and non-responders, and five of the most important variables with their relative importance are given in Table 1. A complete list of variable importance is available in Supplemental Information S8. The variables that got a high importance score in the random forest model were grouped into three categories: 1. a category characterizing the mother, infant and delivery (Birth related), 2. a category characterizing the neurodevelopmental factors and body mass index (Development related), and 3. a category characterizing the socioeconomic status, perceived security of the child and living conditions of the family (SES related).

For assessing the response of not participating at all, the average importance of the variables selected by the random forest model was 0.50 (range 1.0–0.16; 21 variables) in the Birth related category and 0.39 (range 0.55–0.17; six variables) in the SES related category. At five years the average importance was 0.39 (range 1.0–0.13; 24 variables) in the Birth related category, and 0.27 (range 0.41–0.08; seven variables) in the SES related category. At nine years the average importance was 0.28 (range 0.52–0.09; 20 variables) in the Birth related category, 0.39 (range 0.85–0.07; 13 variables) in the SES related category, and 0.81 (range 1.00–0.71; five variables) in the Development related category. At 16 years of age the average importance of variables in the Birth related category was 0.19 (range 0.44–0.03; 11 variables), 0.14 (range 0.56–0. 03; 14 variables) in the SES related category and 0.36 (range 1.00–0.23; 12 variables) in the Development related category. At 30 years of age the corresponding figures were 0.31 (range 0.87–0.10; 11 variables) for in Birth related category, 0.25 (range 0.79–0.06; 17 variables) in the SES related category and 0.50 (range 1.00–0.39; 13 variables) in the Development related category (Fig. 5).

Within the Birth related category, at all age levels, maternal age, gestational age, birth weight, and Apgar scores were in the upper quartile of the variable importance scores.

Within the SES related category, up to the age of 16, the most important variables the model chose for the classification into responders and non-responders reflected the family’s occupation, housing, domestic disputes, and distance to hospital. At 30 years, the most important variables reflected more the individual’s characteristics, e.g., subjective feeling of security, plans for future, parent’s opinions regarding the child and to a lesser degree the socioeconomic features of the family in which the individual was raised.

Within the Development related category, the Neurodevelopmental screen was the most important variable for classifying into responders and non-responders at 16 and 30 years and it was the fifth most important variable at nine years.