Time of day that measurements were taken
BMI 1 Jul 2013 to 30 Jun 2014 and
7.3.3 Baseline characteristics
Some baseline data for both arms was collected prior to 1 July 2014 through the My Health Guardian program run by Healthways, and definitions of these covariates are included in Table 7.5 and Table 7.6 for the Glucose arm, and in Table 7.7 for the BP arm.
Tables in the Appendix show how diagnoses and medications were classified and derived from Healthways data. It should be noted that all participants had some diagnoses data so, for the purposes of the analysis, we assumed this information was complete; that is, we assumed that the absence of a certain diagnosis was not due to missing data but instead, due to that person not having the corresponding condition. It was also assumed that those participants without medication data were, in fact, not taking any. Given the number of participants, however, it is highly likely that errors and omissions exist in the Diagnoses and Medications data.
The level of baseline data missing for Glucose arm participants is shown separately in Table 7.5 to enable an initial assessment. This also indicates the quantity of data we needed to impute using multiple imputation.
7.3 Results and Discussion
Table 7.5 Glucose arm number of participants missing baseline data
Glucose participants missing data Baseline data variable Telemonitoring N = 549 Controls N = 299 N = 848 Total
Age (at 1 July 2014) 0 0 0
Sex 0 0 0
Ethnicity 115 (21%) 88 (29%) 203 (24%)
BMI (last weight recorded July 2013 - June 2014) 183 (33%) 98 (33%) 281 (33%) HbA1c (last recorded Jul 2013 - Jun 2014) 396 (72%) 225 (75%) 621 (73%) Diagnoses (onset before Jul 2014)
(Diabetes type, Hypertension, Hyperlipidaemia, CVD,
Arthritis, Back pain, Walking pain, Eye problem) 7 (1%)* 5 (2%)* 12 (1%)*
Medications (began taking before Jul 2014)
(Insulin/Analogue; Diabetes drugs; Pain relief drug) 36 (7%)† 28 (9%)† 64 (8%)†
Employment status (before July 2014) 255 (46%) 155 (52%) 410 (48%) Moderate exercise (before July 2014) 278 (51%) 144 (48%) 422 (50%) Smoking status (before July 2014) 270 (49%) 141 (47%) 411 (48%) Risk level (last recorded July 2013 - June 2014) 0 0 0 * No diagnoses recorded with onset before July 2014; † No medications recorded with start date
before July 2014
Table 7.6 presents the demographic and observed baseline data of participants with at least one home blood glucose measurement from July to November 2015. We restricted
participants in this table to those with outcome data in the analysis time window because it is the outcomes of these participants that formed the basis for missing data imputation when multiple imputation was used. Around half of the participants were missing important baseline information, such as whether they engaged in moderate exercise or were current or past smokers, while around three quarters did not have baseline HbA1c results.
This table also highlights a potential problem with an intention-to-treat analysis for the Glucose arm with outcome data available for 88% of the Control group but only 49% of the Telemonitoring group. As with the BP arm, the difference derives from the fact that Control
group participants enrolled during the analysis time window, so those who measured at least once but later dropped out, did not have missing outcome data. Telemonitoring group participants, on the other hand, enrolled between 4 and 12 months before the analysis time window started, so many had either dropped out or stopped measuring by 1 July 2015. And some of the participants might have dropped out for reasons predictive of their blood glucose level, such as poor motivation to measure because they had not been taking their medication and did not want to see anticipated unfavourable glucose readings. If this was the case, then with more participants dropping out from the Telemonitoring group, the participants from each group with available data would no longer have been exchangeable due to selection bias from dropout, and the intent-to-treat estimates would be biased. And the missing data mechanism would likely have been MNAR because variable such as motivation were not measured.
Some differences are suggested by the range of p-values in Table 7.6, though with the number of tests conducted, some of the low p-values may be due to chance. Nevertheless, overall the differences suggest that the Telemonitoring group participants providing data were, on average, slightly older and not as healthy. This can be seen with the variables Age, Diabetes type, previous diagnosis of Hypertension, Hyperlipidemia, Cardiovascular disease or Arthritis, the prescription of a Pain relief drug, and the ‘hospitalisation’ risk level. To reduce potential confounding from the imbalances in the available data, which may carry over into the imputed data, these variables were incorporated into the regression models constructed in Chapter 8.
For the BP arm, comparison of the baseline characteristics between the intervention and control groups, detailed in Table 7.7, suggests some differences also existed. For Analysis 7, we listed the baseline characteristics separately given the restricted participant inclusion. But although we used this restriction in an attempt to make the comparison groups more exchangeable, in the end, if we judge by comparing the range of p-values between analyses 5 and 6 and analysis 7, this goal does not appear to have been achieved.
7.3 Results and Discussion
Table 7.6 Glucose arm baseline characteristics before multiple imputation
For participants with ≥1 home blood glucose measurement from 1 Jul to 30 Nov 2015; Some variable categories are not shown, with the full details in the Appendix.
Baseline characteristics Telemonitoring N = 271 (49% of 549) Controls N = 263 (88% of 299) P-value Sex, Male 169 (62% of 271) † 171 (65% of 299) † 0.530 Age, mean (SD) 68.8 (9.2) 65.7 (11.1) 0.001
Ethnicity,Caucasian(Missing: 22%)* 202 (87%) 165 (88%) 0.267 HbA1c, mean (SD) (DCCT %) (Missing: 73%) 6.7 (1.2) 6.8 (1.2) 0.944
BMI, mean (SD) (Missing: 31%) 30.5 (5.6) 30.4 (5.4) 0.838
Diabetes Type 2 248 (92%) 237 (90.5%) 0.074
Hypertension 157 (58%) 57 (22%) < .0001
Hyperlipidemia 80 (30%) 56 (22%) 0.037
Cardiovascular disease 145 (54%) 107 (41%) 0.003
Arthritis (any type) 131 (48%) 100 (38%) 0.018
Back pain‡ 55 (20%) 58 (22%) 0.672
Walking pain‡ 48 (18%) 36 (14%) 0.235
Eye problem‡ 34 (13%) 27 (10%) 0.418
Insulin or Analogue 45 (17%) 41 (16%) 0.814
Pain relief drug 155 (57%) 122 (46%) 0.015
Number of Type 2 diabetes drugs
0 drugs prescribed 71 (26%) 92 (35%) 0.263
1 drugs prescribed 127 (47%) 113 (43%)
2 – 4 drugs prescribed 73 (27%) 58 (22%) Employment status(Missing: 81%)
Full-time, Part-time or Self-employed 8 (17%) 12 (23%) 0.734
No employment 15 (31%) 13 (25%)
Retired 25 (52%) 28 (53%)
Moderate exercise(Missing: 88%) 9 (22%) 4 (16%) 0.752 Smoking status(Missing: 45%)
Never smoker 88 (58%) 89 (61%) 0.860 Past smoker 56 (37%) 50 (34%) Current smoker 7 (5%) 6 (4%) Risk level Extreme Risk 11 (4%) 11 (4%) 0.011 High Risk 63 (23%) 49 (19%) Medium Risk 17 (6%) 10 (4%) Low Risk 100 (37%) 77 (29%) Self-Care 80 (30%) 116 (44%)
Table 7.7 BP arm baseline characteristics before multiple imputation
Analyses 5 & 6 Analysis 7
Baseline characteristics N = 1,429 TM N = 1,259 Controls P N = 773 TM Controls N = 617 P
Sex, Male 727 (51%) 661 (52%) 0.40 426 (55%) 370 (60%) 0.07 Age, mean (SD) 70.6 (9.9) 69.1 (9.5) <.0001 70.6 (9.1) 69.4 (9.0) 0.01 Ethnicity,Caucasian(m%)* 1,036 (73%) (21%) 809 (64%) (29%) 0.58 577 (75%) (19%) 424 (69%) (24%) 0.23 BMI, mean (SD) (m%) 29.4 (6.3) (38%) 29.3 (5.3) (38%) 0.74 29.2 (5.9) (37%) 28.8 (4.5) (35%) 0.38 Diabetes type 2 139 (10%) 145 (12%) 0.009 70 (9%) 46 (7%) 0.04 Systolic BP, mean (SD) (m%) 132.6 (13.7) (46%) 132.2 (13.2) (48%) 0.57 132.3 (13.4) (42%) 132.4 (13.2) (42%) 0.88 Diastolic BP, mean (SD) (m%) 75.1 (9.4) (48%) 76.0 (8.7) (49%) 0.08 75.0 (8.9) (43%) 76.2 (8.8) (44%) 0.06 Cholesterol, mean (SD) (m%) 4.5 (1.6) (92%) 4.5 (1.3) (93%) 0.80 4.4 (1.4) (90%) 4.4 (1.2) (91%) 0.92 Hyperlipidemia 504 (35%) 373 (30%) 0.002 283 (37%) 199 (32%) 0.09 Cardiovascular disease 616 (43%) 543 (43%) 0.99 359 (46%) 279 (45%) 0.65
Arthritis (any type) 712 (50%) 562 (45%) 0.007 393 (51%) 295 (48%) 0.26
Back pain 342 (24%) 257 (20%) 0.03 196 (25%) 132 (21%) 0.08
Walking pain 166 (12%) 147 (12%) 0.96 91 (12%) 88 (14%) 0.17
Eye problem 159 (11%) 107 (9%) 0.02 89 (12%) 55 (9%) 0.11
Insulin or Analogue 229 (16%) 164 (13%) 0.03 113 (15%) 85 (14%) 0.66
Pain relief drug 801 (56%) 580 (46%) <.0001 447 (58%) 318 (52%) 0.02
Employment status(m%) (46%) (53%) (45%) (45%) Full-time 69 (5%) 57 (5%) 0.34 33 (4%) 36 (6%) 0.03 Part-time 47 (3%) 50 (4%) 25 (3%) 32 (5%) Self-employed 43 (3%) 26 (2%) 25 (3%) 16 (3%) No employment 409 (29%) 295 (23%) 241 (31%) 156 (25%) Retired 211 (15%) 168 (13%) 104 (13%) 99 (16%) Moderate exercise(m%) 388 (27%) (51%) 345 (27%) (57%) 0.004 223 (29%) (50%) 211 (34%) (50%) 0.004 Smoking status(m%) (56%) (62%) (55%) (56%) Never smoker 380 (27%) 299 (24%) 0.82 218 (28%) 163 (26%) 0.33 Past smoker 231 (16%) 178 (14%) 122 (16%) 108 (18%) Current smoker 12 (0.8%) 7 (0.6%) 6 (0.8%) 2 (0.3%) Risk level(m%) (6%) (7%) (6%) (6%) Extreme Risk 68 (5%) 35 (3%) <.0001 35 (5%) 11 (2%) 0.009 High Risk 284 (20%) 178 (14%) 140 (18%) 87 (14%) Medium Risk 102 (7%) 107 (9%) 59 (8%) 54 (9%) Low Risk 496 (35%) 467 (37%) 281 (37%) 243 (39%) Self-Care 393 (28%) 378 (30%) 210 (27%) 184 (30%) * missing %
7.3 Results and Discussion
7.3.4 Causal diagrams
Following an assessment of missing data and determination of definitions to be used for the interventions, outcomes and covariates, the causal diagrams in Figure 7.9 and Figure 7.10 were constructed.
Figure 7.9 Causal diagram for the Glucose arm blood glucose outcome group comparisons
Figure 7.9 shows a directed acyclic graph (DAG) that includes both measured and
unmeasured variables. Though it is initially complex to look at, this feature appeared to be an advantage when trying to convey the level of complexity to stakeholders. Nevertheless, a
simpler version was also constructed, shown in the Appendix, and used as an example of a causal diagram in a Three Minute Thesis presentation.
Features to note are:
• variables that are conditioned on are surrounded by a box • the intervention and outcome are coloured blue
• unmeasured variables are coloured red
• the green coloured “Glucometer is used” variable is conditioned on because there is missing outcome data, and this produces selection (collider) bias
In Figure 7.10, a simplified first step type of causal diagram is shown that was thought might have been an easier starting point for researchers not experienced in creating DAGs. Such a strategy might also appeal to some who do have such experience but find using the