Background
Methods
The prospective study of outcomes in Ankylosing Spondylitis (PSOAS)
Analysis cohort
Statistical approach
quantreg
for the existing optimization algorithms [4, 21]. Specifically, we used the option ‘Powell’ for method and ‘left’ for censoring type (i.e., ctype). It is well known that even if the missing data depend on the observed data, the weighted estimating equations provide unbiased estimation, when the missing data process is modeled with correctly specified probability [10, 12].Simulation studies
mice
[25], imputing only missing values (MI-MCMC 1), as well as imputing both censored and missing values (MI-MCMC 2). MCMC-MI algorithm obtains the posterior distribution of parameters by sampling iteratively from conditional distributions based on Gibbs sampling method.Analysis of PSOAS data
Results
Simulation study results
α
|
α
0
|
α
1
|
α
2
| |||
---|---|---|---|---|---|---|
Method | Bias | 100xRE | Bias | 100xRE | Bias | 100xRE |
10% censored
| ||||||
OMNI | 0.0015 | – | -0.0004 | – | -0.0009 | – |
CC-DL/2 | -0.0353 | 40.885 | 0.2295 | 32.190 | 0.0163 | 49.142 |
MI 1 | 0.0151 | 19.644 | -0.1900 | 15.025 | -0.0531 | 29.805 |
MI 2 | 0.0289 | 54.944 | 0.0044 | 39.458 | -0.0055 | 67.701 |
MI-CQR | 0.0588 | 61.407 | -0.0237 | 49.594 | -0.0133 | 80.753 |
MI-wCQR 1 | 0.0548 | 64.175 | -0.0142 | 53.410 | -0.0109 | 82.931 |
MI-wCQR 2 | 0.0554 | 64.336 | -0.0143 | 53.619 | -0.0110 | 83.657 |
MI-wCQR 3 | 0.0562 | 63.070 | -0.0119 | 53.832 | -0.0112 | 82.383 |
15% censored
| ||||||
OMNI | 0.0015 | – | -0.0004 | – | -0.0009 | – |
CC-DL/2 | 0.0155 | 34.977 | 0.1073 | 20.623 | 0.0005 | 43.807 |
MI 1 | 0.2081 | 12.694 | -0.2884 | 11.962 | -0.0673 | 20.976 |
MI 2 | 0.0318 | 51.749 | 0.0043 | 33.300 | -0.0061 | 65.404 |
MI-CQR | 0.0612 | 60.611 | -0.0265 | 48.581 | -0.0131 | 80.495 |
MI-wCQR 1 | 0.0541 | 62.868 | -0.0195 | 52.850 | -0.0097 | 81.684 |
MI-wCQR 2 | 0.0521 | 61.869 | -0.0194 | 51.751 | -0.0060 | 80.580 |
MI-wCQR 3 | 0.0566 | 61.969 | -0.0155 | 51.419 | -0.0104 | 81.064 |
20% censored
| ||||||
OMNI | 0.0015 | – | -0.0004 | – | -0.0009 | – |
CC-DL/2 | 0.0769 | 25.633 | -0.0500 | 14.061 | -0.0084 | 35.731 |
MI 1 | 0.2753 | 11.360 | -0.3920 | 10.965 | -0.0793 | 15.871 |
MI 2 | 0.0339 | 47.699 | 0.0041 | 28.147 | -0.0067 | 61.568 |
MI-CQR | 0.0636 | 59.024 | -0.0273 | 44.248 | -0.0129 | 75.347 |
MI-wCQR 1 | 0.0542 | 62.054 | -0.0215 | 49.542 | -0.0087 | 79.569 |
MI-wCQR 2 | 0.0485 | 62.323 | -0.0233 | 48.830 | -0.0066 | 78.507 |
MI-wCQR 3 | 0.0618 | 60.503 | -0.0178 | 45.908 | -0.0112 | 75.721 |
30% censored
| ||||||
OMNI | 0.0015 | – | -0.0004 | – | -0.0009 | – |
CC-DL/2 | 0.2094 | 12.909 | -0.4255 | 7.810 | 0.0062 | 23.053 |
MI 1 | 0.4420 | 8.714 | -0.6175 | 10.370 | -0.0973 | 10.664 |
MI 2 | 0.0383 | 42.703 | 0.0019 | 21.804 | -0.0076 | 54.866 |
MI-CQR | 0.0621 | 58.587 | -0.0198 | 39.975 | -0.0111 | 75.797 |
MI-wCQR 1 | 0.0466 | 61.367 | -0.0177 | 44.947 | -0.0042 | 77.979 |
MI-wCQR 2 | 0.0327 | 62.075 | -0.0262 | 43.389 | 0.0018 | 79.091 |
MI-wCQR 3 | 0.0302 | 61.427 | -0.0072 | 40.062 | 0.0015 | 75.897 |
α
|
α
0
|
α
1
|
α
2
| |||
---|---|---|---|---|---|---|
Method | Bias | 100xRE | Bias | 100xRE | Bias | 100xRE |
Scenario 2: MVN, Unstructured covariance
| ||||||
OMNI | 0.0017 | – | -0.0006 | – | -0.0013 | – |
CC-DL/2 | 0.0527 | 16.527 | -0.5413 | 7.541 | 0.1161 | 28.516 |
MI 1 | 0.1809 | 8.011 | -0.3166 | 11.718 | 0.0028 | 18.080 |
MI 2 | 0.0183 | 42.410 | -0.0014 | 21.517 | -0.0035 | 52.201 |
MI-CQR | 0.0433 | 63.613 | -0.0050 | 36.069 | -0.0089 | 69.551 |
MI-wCQR 1 | 0.0319 | 71.852 | -0.0099 | 42.007 | -0.0031 | 78.523 |
MI-wCQR 2 | 0.0204 | 71.104 | -0.0299 | 40.525 | 0.0031 | 77.691 |
MI-wCQR 3 | 0.0172 | 63.936 | -0.0168 | 37.197 | 0.0030 | 70.009 |
Scenario 3: MVE, Exchangeable covariance
| ||||||
OMNI | 0.0028 | – | 0.0000 | – | -0.0014 | – |
CC-DL/2 | 0.2094 | 12.616 | -0.4212 | 7.660 | 0.0057 | 22.869 |
MI 1 | 0.4430 | 3.682 | -0.6098 | 3.640 | -0.0981 | 10.591 |
MI 2 | 0.0423 | 32.878 | 0.0008 | 11.074 | -0.0088 | 44.840 |
MI-CQR | 0.0667 | 58.224 | -0.0183 | 36.201 | -0.0128 | 71.372 |
MI-wCQR 1 | 0.0501 | 62.644 | -0.0163 | 41.087 | -0.0055 | 77.240 |
MI-wCQR 2 | 0.0379 | 62.214 | -0.0256 | 40.690 | 0.0000 | 77.128 |
MI-wCQR 3 | 0.0273 | 59.215 | -0.0041 | 37.352 | 0.0018 | 71.902 |
Scenario 4: MVE, Heteroscedastic covariance
| ||||||
OMNI | 0.0028 | – | 0.0000 | – | -0.0014 | – |
CC-DL/2 | 0.1739 | 14.389 | -0.4081 | 8.210 | 0.0171 | 23.356 |
MI 1 | 0.1892 | 3.537 | -0.3084 | 1.531 | 0.0002 | 10.432 |
MI 2 | 0.0410 | 32.650 | 0.0011 | 11.025 | -0.0083 | 44.126 |
MI-CQR | 0.0649 | 58.870 | -0.0141 | 36.739 | -0.0125 | 71.618 |
MI-wCQR 1 | 0.0492 | 63.492 | -0.0121 | 41.733 | -0.0054 | 78.362 |
MI-wCQR 2 | 0.0340 | 63.222 | -0.0223 | 40.015 | 0.0009 | 78.325 |
MI-wCQR 3 | 0.0273 | 59.195 | -0.0041 | 37.005 | 0.0018 | 71.929 |
Results of applying the proposed methods to PSOAS data
General characteristics of patients in PSOAS
Analysis results
log(CRP) | ||
---|---|---|
Method | adj. RR (95% CI) | p-value |
CC-DL/2 | 1.001 (0.98, 1.02) | 0.9867 |
MI-MCMC 2 | 1.006 (0.99, 1.03) | 0.5586 |
MI-CQR | 1.010 (0.995, 1.02) | 0.1839 |
MI-wCQR 2 | 1.018 (1.004, 1.03) | 0.0095 |
Discussion and conclusion
quantreg
in R that fits quantile regression, we are currently developing R package for users to easily implement the proposed MI approaches. Censored quantile regression has been extended to data censored at both lower and upper thresholds [4], therefore our method can be also directly extended to doubly censored biomarker data. There is a growing interest in developing MI methods that impute missing data across multiple medications while accounting for the correlations among them, which can be also extended by our proposed method.