The apheresis platelet donation was increased after a nationwide ban on family/replacement donation in China

Based on the availability of the data, we used two independent full samples between October 1, 2012 – September 30, 2019 from Guangzhou Blood Center, the second largest blood center in China and Chengdu Blood Center, the largest blood centers in Western China, which were named as “GZ set” and “CD set” at the individual level, respectively. The GZ and CD sets consisted of 135,851 and 82,129 plateletpheresis donors, respectively. We enrolled the donors with an age of 18 ~ 60 years old, weight ≥ 50 kg for male and ≥ 45 kg for female, systolic blood pressure 90 ~ 140 mmHg and diastolic blood pressure 60 ~ 90 mmHg, pulse 60 ~ 100 beats/min, normal body temperature, platelet counts before donation 150 ~ 450 × 10⁹/L, and no any other health conditions according to the China national standard for the eligible donor selection criteria (GB 18467–2011). All data used in the present study were de-identified.

We chose 5.5 years and 1.5 years before and after April 1, 2018, respectively, to set our investigation time window with a six-month interval as a cross-section, generating a total of 14 repeated cross-sections with 11 and 3 cross-sections before and after the ban, respectively. Based on the GB 18467–2011, a donor may donate up to a total of 24 plateletpheresis collections during a 12-month rolling period. Thus, some plateletpheresis donors may appear more than once across the 14 cross-sections, but not all donors appear in every cross-section. Therefore, there were some different degrees of the correlations of the data across 14 cross-sections, and our datasets presented a pseudo-panel structure [13].

The records for the total amount of the platelets donated by each donor within each cross-section, grouping covariates – gender, birth year, and blood donation history (see detailed information below), as well as the variable –family/replacement plateletpheresis donation (yes vs. no) were extracted from the archived blood donation documents in both blood centers. When a blood donor came to donate blood and assigned his donated blood to a specific patient on the waiting list, his/her donation was marked as FRD. The outcome variable at the cohort (i.e., cell) level within the pseudo-panel datasets (see detailed information below) was defined as average plateletpheresis units per donor in each cell.

The pseudo-panel data approach is actually a solution to transform the individual-level cross-sectional data into the group-level data (i.e., pseudo-panel data) such that the typical longitudinal models can be applied to efficiently and consistently estimate the change of the interested outcome variable over time [13]. This approach has been increasingly applied to public health [14, 15].

According to the methods described in the literature [16, 17], we constructed two pseudo-panel datasets from our GZ and CD data by grouping three time-invariant variables - gender, birth year, and blood donation history. Other potential covariates such as ethnicity, occupation, and education with large proportions of missing data were excluded from the analyses. Briefly, individual platelet donors were first classified based on gender that had two categories – male and female. To balance the size of each cohort (≥100, named as large cohort) and the number of large cohorts [16] within our pseudo-panel datasets, we defined birth year as three categories – “1952–1974”, “1975–1984”, and “1985–2001”. The variable blood donation history was coded as 4 levels – “None” (i.e., the blood donor had never donated blood before the current cross-section, also called no history of donations), “Whole Blood Donation Only” (i.e., the blood donor had donated whole blood only before the current cross-section, abbreviated to “WB”), “Apheresis Platelet Donation Only” (i.e., the blood donor had donated apheresis platelet only before the current cross-section, abbreviated to “PLT”), and “Both WB and PLT Donations” (i.e., the blood donor had donated both whole blood and apheresis platelet before the current cross-section, abbreviated to “Both”). The individuals were then further divided by these two variables, generating 2 × 3 × 4 = 24 cohort groups across the 14 cross-sections, i.e., 24 × 14 = 336 cells for each of two pseudo-panel datasets. The generated each cohort group had common gender, birth year, and blood donation history. To reduce the measurement error [17], we removed the cells with less than 30 individual donors in each dataset to generate two final pseudo-panel datasets, named as “overall Guangzhou pseudo-panel set” (abbreviated to “overall GZ set”) (n = 330 cells) and “overall Chengdu pseudo-panel set” (abbreviated to “overall CD set”) (n = 316 cells), respectively, for further analyses. More detailed information about the summaries of the constructed pseudo-panel datasets is presented in Additional file 7.

Our study was actually an observational study using pseudo-panel approach to analysis the data [13‐15]. Thus, we used the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) (Additional file 1) statement to ensure standardization and enhance the quality of the reporting [18].

For the individual-level data, we used two-tailed independent t-tests (α = 0.05) to compare total number of donors, total number of donations, and total apheresis platelet units (U), average plateletpheresis donations per donor, and average apheresis platelet units (U) per donor between before and after the ban (also called “after-before mean difference”). To compare two after-before mean differences between groups within each covariate, we applied two-tailed Z-test (Z = (mean difference₁-mean difference₂)/sqrt (se₁² + se₂²), α = 0.05).

For overall GZ pseudo-panel set, we plotted outcome values (i.e., average apheresis platelet units per donor per cell) from each cohort group that had common gender, birth year, and blood donation history versus time (i.e., 14 cross-sections) as well as overall average outcome values from all 24 cohort groups versus time to visualize whether the trend of the outcome over time is linear or non-linear. Given that generalized linear mixed model and mixed-generalized ordered logit model have been successfully applied to analyze the trend of the outcome variable (proportion or probability) over time within a pseudo-panel dataset [14, 19], for our pseudo-panel dataset with a continuous outcome variable, we applied a linear mixed model for modelling the linear trend, or modelling the non-linear trend using a piecewise linear mixed model with the defined time (i.e., cross-section) breakpoint(s) based on the above-mentioned visualization [20]. We conducted a model selection starting from an unadjusted model (i.e., model 1 – pure time trend model without any covariates) to an adjusted model (i.e., model 2 = model 1 + significant covariates), and then to a final model (i.e., model 3 = model 2 + significant interaction terms). To compare two raw regression coefficients within the same final model, we used one-tailed Z-tests [Z = abs(β₂-β₁)/sqrt (SE₁² + SE₂²), α = 0.05].

The assumptions of normality and homoscedasticity for piecewise linear mixed effects models were examined by visualizing marginal (for fixed effects only) and conditional (for both fixed and random effects) Pearson residual plots.

For five-fold cross-validation of the final model, the data were randomly split into 5 roughly equal-sized subsets using SAS PROC SURVEYSELECT procedure, the final model was fitted to the 4 subsets of the data using SAS PROC MIXED procedure with the STORE statement. The prediction error – Root Mean Square Error (RMSE) of the fitted model to predict the fifth subset using SAS PROC PLM procedure was calculated. This procedure was repeated 5 times such that each subset was used for testing exactly once and an average RMSE of 5-fold cross-validation was calculated by using the formula: SQRT((RMSE₁² + RMSE₂² + … + RMSE₅²)/5) and further compared with that from the model-fitting using the full sample to determine if there was an over-fitting issue.

To test for replication of the trend of the overall outcome mean over time obtained from the overall GZ set, we applied the same methods as described above to the independent pseudo-panel dataset from Chengdu Blood Center, i.e., overall CD set.

To detect the potential effect of family/replacement plateletpheresis donations on the model-based trend of the average apheresis platelet units per donor over time identified from the overall GZ and CD sets, we removed the family/replacement plateletpheresis donations in each of the first 11 cross-sections (all donations after the ban were voluntary) from the overall individual-level data. Then we re-grouped the remaining individuals (81,801 and 48,768 voluntary plateletpheresis donors remaining from the GZ and CD sets, respectively) to generate two nested pseudo-panel subsets, named as “voluntary GZ subset” (n = 300 cells) and “voluntary CD subset” (n = 284 cells), respectively. Finally, we used the same methods as described above to fit the piecewise linear mixed models for both subsets.

All data management and statistical analyses described above were conducted with R (R Development Core Team) and SAS v9.4 (SAS Institute, Cary, North Carolina).

Table 1 and Additional file 2 showed a similar demographics of plateletpheresis donors across 14 successive cross-sections between overall GZ and CD sets. Our overall GZ set consisted of 69.3–80.3% of plateletpheresis donors who were male across all 14 cross-sections, 5.0–6.5% and 8.8–9.7% ≥46 years old across the first 11 cross-sections (i.e., before the ban) and the 12th – 14th cross-sections (i.e., after the ban), respectively, as well as 45.1–51.8% of donors who had a blood donation history (14.4–14.5%: WB only, 18.6–21.9%: PLT only, and 12.0–15.5%: Both) across the cross-sections before the ban and 74.8–76.8% with a blood donation history (17.5–20.9%: WB only, 27.9–28.5%: PLT only, and 25.4–31.5%: Both) across the cross-sections after the ban (Table 1). The overall GZ set also contained about 15.4–41.9% of FRDs across the first 11 cross-sections (for the 12th – 14th cross-sections, no donations were FRD) (Table 1). In the overall CD set, about 63.3–75.5% were male across all 14 cross-sections; 10.0–13.9% and 16.3–17.9% were ≥ 46 years old across the first 11 cross-sections and the 12th – 14thcross-sections, respectively; 30.1–40.6% and 63.9–72.4% had a blood donation history across the cross-sections before and after the ban, respectively; and 3.9–52.6% were family/replacement donors across the first 11 cross-sections (Additional file 2).

In either GZ or CD dataset, both average total number of donations and average total apheresis platelet units were significantly increased after the ban compared to those before the ban; in contrast, the change of average total number of donors from before to after the ban was the opposite (t = 2.42–5.56, p = 0.0325–0.0001) (Table 2, Additional file 3 and Additional file 4). These results indicated that there is a phenomenon of “fewer donors, more blood” after the FRD ban in plateletpheresis donation practice.

Additional file 5 demonstrated that both average apheresis platelet units per donor and average total number of plateletpheresis donations per donor after the ban were significantly higher than those before the ban for all covariates including gender, age, and blood donation history in both datasets (all p-values< 0.05). Further comparisons revealed that male donors or donors with plateletpheresis donation history had significantly larger increase (i.e., after-before mean difference) of both average apheresis platelet units per donor and average total number of plateletpheresis donations per donor compared to their peer groups (i.e., female donors or donors with other donation history) (all p-values< 0.05).

In addition, in terms of average apheresis platelet units per donor among family/replacement donors, voluntary donors before the ban, and voluntary donors after the ban, the family/replacement donors contributed the least, voluntary donors before the ban the second, and voluntary donors after the ban the most (ANOVA with post-hoc Tukey tests, all p-values< 0.0001) (Additional file 6).

Additional file 7 summarized the characteristics of two independent pseudo-panel datasets (overall GZ set: n = 330 cells; overall CD set: n = 316 cells) with the number of plateletpheresis donors per cell ≥30. Figure 1a and b showed 24 individual trajectories of the average apheresis platelet units per donor representing 24 cohort groups with each who shared common gender, birth year, and blood donation history over 14 cross-sections in the overall GZ and CD pseudo-panel sets, respectively. Figure 1c and d visualized an overall mean profile of the outcome over time in the GZ and CD sets, respectively, indicating that there was a breakpoint at the 11th cross-section (just right before the ban) with a roughly horizontal line to the left of the breakpoint and significantly positive linear trend to the right.

Thus, to the overall GZ set, we fitted a two-piecewise linear mixed-effects model in which we specified intercept as the random term with an unstructured covariance-structure. The covariate birth year was excluded due to its non-significance. As shown in Table 3 and Fig. 2a, the final model presented a two-piecewise linear trend of average apheresis platelet units per donor over time (i.e., cross-section) with a horizontal line to the left of the breakpoint (β_timeBefore11 = 0.0111, p = 0.0976) and a significantly positive linear trend to the right (β_timeAfter11 = 0.0404, p < 0.0001). This result suggests that the average apheresis platelet units per donor were maintained at lower level and did not change with time before the ban, but started increasing linearly with time after the ban.

In the model, independent covariates measured whether baseline outcome differed by group. Table 3 demonstrated that on average at baseline, male donated 1.337 more apheresis platelet units per donor than female (β_{male vs female} = 1.3370, p = 0.0044), and the donated apheresis platelet units per donor was significantly higher in the donors who had platelet donation history than that in their peers who had no blood donation history (β_{PLT vs None} = 2.6243,p = 0.0003; β_{Both vs None} = 3.9440, p < 0.0001) whereas donations by the donors with whole blood donation history only were not significant (β_{WB vs None} = 0.1790, p = 0.7630). Due to the non-significance of the slope for the time_Before11 term, we only considered the interactions between the time_After11 and covariates in the model. After the model selection, a two-way interaction term gender×blood donation history and a three-way interaction term time_After11 × gender×blood donation history were excluded due to their non-significance. The regression coefficient of the interaction term measured whether the outcome change differed by covariate-specific groups. As shown in Table 3, on average after the ban, the outcome change was significant in males than that in females (β_{time × male vs time×female} = 0.0550, p < 0.0001) and in donors with blood donation history than that in their peers without donation history (β_{time × WB vs time×none} = 0.0331, p = 0.0018; β_{time × PLT vs time×none} = 0.0698, p < 0.0001; and β_{time × Both vs time×none} = 0.0444,p< 0.0001). These results suggest that the contributions of gender and blood donation history to the model had heterogeneity for both the baseline outcome value and the outcome change after the ban.

The Pearson residual plots, either marginally for the consideration of fixed effects only or conditionally for the consideration of both fixed and random effects, for the final model indicated that the model’s assumptions – normality and heteroscedasticity were not significantly violated (Additional file 8). Five-fold cross-validation demonstrated that the over-fitting percentage of the final model for overall GZ set accounted for only 0.28% (Additional file 9), suggesting that there was no significant over-fitting issue for the final model.

Next, we used an independent overall CD set to test the replication of the final model obtained from the overall GZ set. As shown in Table 3 and Fig. 2b, all parameters’ estimates in the final model for the overall CD set were similar to those for the overall GZ set, except for that the outcome change for the donors with whole blood donation history only was not significant and that the outcome values across all time-points appeared to be systematically reduced in the overall CD set, compared to those in the overall GZ set. The latter result was consistent with that from the individual-level data, i.e., the overall average apheresis platelet units per donor in Guangzhou (overall mean = 3.3 U, SD = 1.0) was higher than that in Chengdu (overall mean = 2.6 U, SD = 1.0) with a marginally non-significance (mean difference = 0.7 U, 95% CI:-0.04–1.5, p = 0.0644) (data not shown). The assumptions of normality and heteroscedasticity were not significantly violated (Additional file 8), and the over-fitting percentage of the final model for overall CD set was 2.52% (Additional file 9). These results suggest that the two-piecewise linear trend of the average apheresis platelet units per donor over time observed in overall GZ set can be replicated in an independent dataset – overall CD set with some minor variations.

After removing the FRD across the first 11 cross-sections followed by re-grouping the data (Additional file 10), our modeling results demonstrated that although the average apheresis platelet units per donor still significantly increased over time after the ban (β_timeAfter11 = 0.0588 and 0.0583 for voluntary GZ and CD subsets, respectively, all p-values < 0.0001), a positive linear trend of the outcome values over time before the ban also became significant (β_timeAfter11 = 0.0400 and 0.0499 for voluntary GZ and CD subsets, respectively, all p-values < 0.0001) (Additional file 11, Fig. 2c and d). These results suggest that the FRD is a critical factor to influence the trend of the outcome values over time before the ban. The heterogeneity for group-specific baseline outcome and outcome change after the ban was still similar to that from the overall GZ and CD sets with some variations. Neither significant violations of the model’s assumptions (Additional file 12) nor over-fitting issue (over-fitting percentage was 4.64% for voluntary GZ subset and 0.65% for voluntary CD subset) (Additional file 13) were found for the final models.