Should analyses of large, national palliative care data sets with patient reported outcomes (PROs) be restricted to services with high patient participation? A register-based study

The analyses were performed using SAS statistical software version 9.4.

The 40,316 adult Danish cancer patients in DPD who received specialized palliative care and died between 2010 and 2015 had 49,307 admittances to SPC services (some patients were admitted to more than one service). In 24,589 (49.9%) of these patient admittances, based on 22,420 patients, the EORTC QLQ-C15-PAL was answered. These 24,589 admittances were included in this study.

Of the 24,589 SPC admittances, 49.0% represented women and the average age was 68.5 years (Table 1). Three of four (74.0%) of the SPC admittances were in a palliative care team. The number of SPC admittances increased from 2010 to 2014 but decreased from 2014 to 2015 (patients who died later than 2015 were not included). The largest differences between patients responding to the EORTC QLQ-C15-PAL questionnaire (and thus included in the study) and non-respondents was that the study population were less likely to be hospice patients (26.0% vs. 47.9%) and their average survival time was longer (94.0 vs. 64.5 days) compared to non-respondents (Table 1).

The patients were from 44 services, and each service admitted between 2 and 591 patients each year. The overall response rate was 49.9% and varied between year and service from 0.0 to 93.2% (Table 2).

Most patients were categorized in the response rate group ≥60.0% but 42.3% were categorized in a lower group, i.e. were admitted to a service having a response rate below 60.0% in the year of admission (Table 1).

The results of the multiple linear regressions are shown in Fig. 1 and are summarized in Fig. 2. The results were interpreted using the four criteria listed in the Methods section.

Criterion 1. Significant associations with response rate were found for six of the ten symptoms/problem scales (dyspnea, appetite loss, fatigue, nausea, emotional function and physical function).

Criterion 2. Mean scores for the six symptoms/problems were compared between the ≥60.0% response rate group and each of the five lower response rate groups (< 20.0%, 20.0–29.9%, 30.0–39.9%, 40.0–49.9% and 50.0–59.9%, respectively), i.e., in total 30 comparisons. In 13 of the 30 comparisons, a significant difference was found between the ≥60.0% response rate group and the lower response rate group.

Criterion 3. In two of these 13 comparisons, patients from the lower response rate groups (20.0–29.9% and 30.0–39.0%) had better scores (higher mean physical function) compared to the ≥60.0% response rate group, in accordance with the study hypothesis.

The remaining 11 of the 13 significant comparisons contradicted the study hypothesis as the patients from the lower response rate group had worse symptoms/problems. Thus, compared to the ≥60.0% response rate group, the 30.0–59.9% response rate groups had worse appetite loss, the 40.0–49.9% response rate group had worse fatigue, the < 20.0–49.9% response rate groups had worse nausea and the 20.0–59.9% response rate groups had lower emotional function.

Criterion 4. For physical function, the 20.0–29.9% and 30.0–39.0% response rate groups had 1.9 and 1.7 point higher mean physical function, respectively, compared to the ≥60.0% response rate group and these differences were therefore not clinically relevant. For the 11 significant comparisons that were not in accordance with the study hypothesis, only one was possibly clinically relevant (patients from the < 20.0% response rate group reported 5.7 point more nausea compared to patients from the ≥60.0% response rate group).

The linear regression analyses found more significant associations between response rates and scale scores than the logistic regression analyses did, and thus the linear regression analyses did not generally miss significant associations, except for QOL but in that case the logistic regression analysis did not find systematically worse (or better) QOL in the lower response rate groups compared to the highest response rate group (Table 3).

Change in mean scale scores when patients from SPC services with response rates < 60.0% are excluded.

The mean scale scores were almost identical for the whole population and the sub-populations with exclusion of the < 20.0, < 30.0, < 40.0, < 50.0 and < 60.0% response rate groups, respectively. The largest change in mean score after exclusion of the lower response rate groups was 2.0 (in emotional function, range 0–100) (Fig. 3). Thus, removing patients from services with lower response rates from the mean scale score calculations had almost no effect on the mean scores.

Our aim was to test whether the response rate was associated with scale scores of ten quality of life aspects. By doing this we could test the study hypothesis, i.e. that patients from SPC services with low response rates were reporting better QOL than patients from services with high response rates, indicating possible selection bias in services with low response rates. Such selection bias could happen if the health care professional in low response services had primarily asked the most well patients to answer the EORTC QLQ-C15-PAL questionnaire, leading to an overestimation of the QOL in the low response services which would result in better scores on the ten QOL aspects (lower symptom scores and higher function and overall QOL scores) in these low response services compared to high response services where patients were not selected (or less selected). On the other hand, if patients in low response services were randomly asked to answer the EORTC QLQ-C15-PAL questionnaire the QOL would be representative for the patients in the low response service. This would likely result in similar QOL scores in the low and high response services or perhaps worse scores in the low response services if the explanation to the lower response rate was a sicker patient population. Thus, to accept the study hypothesis, the score of one or more of the quality of life aspects should be significantly better (i.e., lower symptom score, higher function score, higher overall QOL score) in palliative care services with low response rates compared to patients from services with high response rates. However, results from the multiple linear regression analyses did not support the study hypothesis: there were overall associations between response rate and six out of ten QOL scale scores. However, within these six scales, only two out of 30 comparisons were significant and in the expected direction. The magnitude of these comparisons was very small (below 2 on a 0–100 scale) and thus not clinically relevant. Furthermore, mean scale scores in the whole population were almost identical to mean scale scores calculated after exclusion of the < 20.0, < 30.0, < 40.0, < 50.0 and < 60.0% response rate groups, respectively, supporting that the response rate had neglectable impact on the scale scores. This is in accordance with findings from a previous, much smaller study in patients with advanced cancer, where no evidence of clinically relevant difference in the quality of life scores due to non-participation was found [6].

There were some cases where patients from services with low response rates reported significantly worse QOL, and not better QOL as hypothesized. The differences were relatively small (1.8–5.7 on a 100-point scale) and due to the large number of statistical test performed, these findings may be due to chance. It is however not unlikely that non-response is associated with poor health [4, 5] and the slightly worse symptom/problem scores in SPC services with low response rates may reflect a slightly more ill patient population in SPC services with low response rates, i.e. not selection bias.

There are several strengths in this study. First, it is based on a large dataset with national coverage including data from all SPC services in Denmark. Second, the possibility that different SPC services may not have the same composition of patients (due to socio-demographic or other differences across referral areas) was accounted for by including the random effect of specific SPC service.

No services had complete data, which is a limitation for at least two reasons. First, because our conclusion that there is not more selection bias in SPC services with low response rates than in those with ≥60.0% response rate is not a full investigation of the possibility of selection bias in SPC services with low response rates. However, due to the nature of SPC, which is provided to patients with very advanced disease and short survival, it will probably not be possible to achieve much higher response rates than achieved in the best performing services. Second, it is impossible to know whether differences in scale scores between SPC services with high and low response rates are in fact do to selection bias in SPC services with low response rates or whether they are caused by SPC services with low response rates actually having different scale scores, since we do not know the true scale scores. Thus, looking at difference in scale scores between SPC services with low and high response rates is only one way to attempt to decide whether selection bias seems plausible in services with low response rates, and the fact that we found no major impact is reassuring.

Only a few large studies (1000–2000 patients) have looked at symptom prevalence or symptom levels in patients at the start of palliative care [17, 20, 21] and to our knowledge this is the first European study to report data from all services in a whole country. Given that the present study indicates that our full data set can be used for analysis without any major risk of impact of selection bias, future analyses can safely be made at the national level to investigate the symptoms/problems and QOL among patients admitted to SPC.