Physician consultation in young children with recurrent pain—a population-based study

Subjects

We analyzed a subset of the participants of the German Health Interview and Examination Survey for Children and Adolescents (KiGGS). The KiGGS was approved by the Charité/Universitätsmedizin Berlin ethics committee and the Federal Office for the Protection of Data. The KiGGS study gathered nation-wide data on health in children and adolescents in Germany. Between May 2003 and May 2006, data from 10,333 children between 0 and 10 years and 7,308 adolescents between 11 and 17 years were gathered in 167 cities and towns throughout Germany. KiGGS consists of different modules for various domains of well-being including health-related quality of life and pain (Kurth et al., 2008). We analyzed only a subset of the children whose parents reported (1) the child had pain at least once per month in the past three months and (2) data was available on all predictor variables (see below). Because the prevalence for pain varies across gender and age, this study sample is systematically different to the full sample (Ellert, Neuhauser & Roth-Isigkeit, 2007). Since there are no systematic biases that prevent children in pain from participating in KiGGS, we believe that the sample analyzed in the present manuscript is representative for children with recurrent pain.

Measures

As part of the KiGGS study parents answered questions regarding demographic information and different aspects of pain in the preceding three months. Data were based on parental report. Health care utilization was assessed with one item asking whether the child visited a doctor due to his/her pain problem. The five response categories ranged from “never” to “always”. For the present analysis these were dichotomized by collapsing all responses from “once” to “always” so that the analysis compared those who attended a physician at least once to those who never attended a physician. While this reduces the amount of variance this is in line with previous studies that also used this contrast to define predictors (Perquin et al., 2001; Huguet & Miró, 2008).

We used demographic, pain-related and generic variables as predictors of health care utilization in children. Demographic variables were age, gender, SES, and migration background. The SES includes information on the parental education, their occupation as well as the family income (Lange et al., 2007). It was divided into “high”, “middle” and “low”. Migration background was defined as either (1) the child being an immigrant and at least one parent not being born in Germany or (2) both parents being immigrants and being no German citizen (Lange et al., 2007).

We assessed several pain-related predictors for health care utilization; pain-related disability, pain-frequency, and pain-intensity. We calculated a comprehensive pain disability measure by adding up the impairment in different areas of everyday life including school, meeting friends, appetite, sleep, hobbies (Hirschfeld & Zernikow, 2013). Parents were asked to judge their children’s pain-related impairment by indicating how often specific scenarios were met e.g., “child was not able to attend daycare because of their children’s pain” and responded on a scale ranging from “never” (=1) impaired due to pain to “always” (=5). Higher scores are indicative of higher impairment. Average pain intensity was measured with a 100 mm visual analogue scale. Pain frequency in the past three months was described on a scale ranging from “once a month” (=2) to “daily” (=6).

Furthermore, we assessed general predictors for health care utilization: HRQL, and mental health. HRQL was assessed with the KINDL-R (Bullinger et al., 2008). This 24-item questionnaire covers six dimensions of HRQL. For this study we used the scales “psychological well-being” and “physical well-being.” Aspects of well-being were rated on a 5-point Likert-scale. The strengths and difficulties questionnaire (SDQ) (Goodman, 1997) is a screening instrument covering mental health problems. In KiGGs study the parent version of the SDQ was used for all participants. In the present analysis the total difficulties score was used, that was computed as the sum of the four difficulty scales; “Emotional symptoms,” “Conduct problems,” “Hyperactivity/inattention” and “Peer relationship problems.”

Data analysis

Data were analyzed in two levels: the full sample and the pseudosamples that were generated as part of the sequential sampling approach. In the full sample, predictors of health care utilization were identified using multiple logistic regression analysis. In order to make the odds ratios (ORs) comparable across predictors measured on different scales, predictors were scaled by dividing by two standard deviations before being entered into the logistic regression (Gelman, 2008). This allows interpreting ORs for continuous predictors similar to ORs from binary predictors as the added—or reduced—risk of participants with higher scores compared to lower scores.

A sequential sampling approach was used to assess the variability of the parameter estimates. The sequential sampling approach tries to illustrate how the ORs of the predictors change when the sample size is gradually increased and identifies a sample size from which on these are stable. Specifically, this entails adding the participants one by one to the dataset and computing the logistic regression with each addition. The resulting sequence of ORs can be plotted against the sample-size showing the trajectory of the ORs. Some ORs for the individual predictors may either be relatively stable across the number of participants—e.g., the OR for pain-intensity in Fig. 2—while others may show some changes depending on the number of participants that were included in the analysis—e.g., the OR for migration in Fig. 2. Of special importance was the point of stability (POS), i.e., the number of participants that had to be included until the significance of this specific effect did not change any more or stabilized (Schönbrodt & Perugini, 2013). Based on the trajectory one can see that at low sample sizes adding participants affects the estimated OR much more than at larger sample sizes. Hence, adding more participants to the analysis before the POS may change not only the magnitude but also whether an effect is significant or not. After the POS, the effect remains stable, i.e., it remains either insignificant or significant. In the example below, the OR for migration was significantly smaller than one when only the first 100 participants were included and significantly larger than one when more than 500 participants were included. The POS for this effect in this specific sampling order was around 500 as from this sample size on adding participants did not change the sign and significance of the effect. Because trajectories and its’ corresponding POS are specific to the particular order in which participants were added to the analysis, we replicated this analysis for 1.000 random orders of participants. For each of the 1.000 random orders, the trajectory and POS were calculated. This resulted in 1,000 slightly different POS for each OR. From the distribution of these POS we calculated the POS_crit as the 80th percentile of the POS. This indicates the sample size from which on 80% of the ORs stabilized. Inspection of the POS for 1.000 sequences and POS_crit provides an index for how many participants would have to be sampled before the solution stabilizes irrespective of the specific order in which the participants were sampled (Schönbrodt & Perugini, 2013). Data analysis was performed in R. Analysis scripts; the file “Public Use File KiGGS 2003-2006” may be requested at: http://www.rki.de/EN/Content/Health_Monitoring/Public_Use_Files/application/application_node.html will be made available upon request.

Patients

Data from 2,149 children (3–10 years; 7.41 years ± 2.25; 54% female) were analyzed. Table 1 gives an overview of the demographic and pain-related characteristics of the analyzed sample.

Predictors of health care utilization in the full sample

Overall, we found that 1,144 (53%) of the children with recurrent pain consulted a physician due to their recurrent pain. Pain-related disability was the strongest predictor of whether or not a child would visit a physician due to recurrent pain (OR = 5.37 ; 95% CI [4.15–7.02]), children with high disability, due to pain were five times more likely to visit a physician (Table 2). As can be seen in Fig. 1, five other variables were also significant predictors but had much smaller ORs; pain-intensity (OR = 2.05; 95% CI [1.66–2.54]), pain-frequency (OR = 1.67; 95% CI [1.37–2.03]), migrant (OR = 1.64; 95% CI [1.12–2.41]), sex (OR = 1.22 ; 95% CI [1.01–1.47]), age (OR =.66; 95% CI [0.54–0.79]). Overall, young girls with high pain-related disability, -intensity, –frequency, and migration background had higher probabilities to visit a physician. Socioeconomic status, health-related quality of life and mental health problems were not systematically related to physician visits.

Variability of predictor estimates

Figure 3 shows a trajectory of the estimates for one specific sampling order, i.e., the ORs from the logistic regression based on different subsamples of patients. In this sampling order the estimates of the effects change depending on how many participants are added to the analysis. The ORs for most effects converge to 1, i.e., a null-effect, in the first 250 participants. Critically the ORs may even reverse, e.g., the effect of migration background is smaller than 1 up until about 450 included participants, from which point on the effect becomes larger than 1.

Inspecting the POS for this specific order (Table 2) revealed that the effects of HRQL and SES did not change from the smallest sample size studied, i.e., the POS was 50. However, even the huge effect of pain-related disability was only stable from 117 participants onwards, and the effect of pain frequency needed 531 participants to stabilize. Of note three variables (Migration, Sex, and SDQ) needed over 1,000 participants until their effect could be reliably detected in this specific sampling order.

Inspection of the POS_crit from the 1,000 different random sequences (Fig. 3) demonstrates that the findings concerning the POS described above are not specific for the order in which the participants were sampled. The POS_crit indicate that most effects stabilized only after several hundred participants were included in the analysis (Table 2). Of note, while the effect of HRQL-Phy. seemed to stabilize with few participants in the specific order studied above, this effect stabilizes relatively late when alternative orders are considered. Across the thousand different orders, the effect of pain-related disability stabilized with the fewest number of participants while the effect of sex, which was borderline significant in the full sample, stabilized extremely late.

Limitations

Several limitations need to be kept in mind when interpreting the results of the present study. Most of these are due to the fact that the data were from a large population-based survey that did not assess pain-related variables with the level of detail that is possible when a smaller group of patients are assessed at individual treatment centers (Hirschfeld & Zernikow, 2013). This is most relevant with regard to an assessment of the etiology of pain and the measure of pain-related disability. Based on the present data it is not possible to distinguish between primary and secondary pain problems, i.e., some children within this sample may report pain because of recurrent acute infections. At the same time we do not believe that this explains differences to earlier studies because only some excluded patients with known etiology (Perquin et al., 2001) while others do not differentiate (Toliver-Sokol et al., 2011). The presumably low stability of the results concerning pain-related disability cannot be explained by lack of reliability alone because the more comprehensive measures for HRQL and mental health problems did not yield more reliable effects. Nevertheless, future population-based studies should expand their focus on pain as a central variable, and utilize measures that balance the need for more detailed information on pain-related health care utilization and brevity.