3.3 Statistical analysis
3.3.3 Data analysis
Activity data (AC)
The primary outcome was the proportion of time spent upright as measured by the AC. For each patient the amount of time spent upright as a proportion of the total recording time was calculated. The total upright time for each patient for each hour of the monitoring period was calculated with the group data for each hour reported as the median time. Additionally, the total upright time (sum of standing and stepping time) and total sedentary time was calculated for every hour of the monitoring period. Time did not have a normal distribution,
therefore group data for each hour were summarised as medians with the 25th and 75th percentiles (interquartile range [IQR]) reported. The HSC PAL software provides the time at which a change in output category (i.e. a transition from sitting to standing) occurs and also the duration of each event (defined as
continuous periods of one activity). This information was used to investigate the accumulation of events throughout the day. Each event was classified into one of the following time intervals which have been used previously (≤ 5 minutes, > 5 to ≤ 10 minutes, > 10 to ≤ 30 minutes, >30 to ≤ 60 minutes, > 60 minutes).130 The amount of time spent in each of the time intervals as a proportion of total time spent upright/sedentary was calculated.
This required decisions to be made regarding the appropriate extraction of data.
In particular, whether or not to include events that crossed imposed start and end times. The HSC software is programmed to include the first event that crossed the start time so this may mean crediting a patient with a five minute duration event, when in fact it was 125 minutes long. Some researchers enforce rules (personal communication, 2011) that if the proportion of the overlapping event is more than 50% out with the monitoring time then it should be excluded.
Whether or not to apply such a rule to these data was assessed by investigating the output for individual patients. Considering this potential underestimation it was decided that these events should be included. An example of an AC output is provided in Appendix 10. The same consideration was given to events that overlapped times when the monitor was removed or reattached. Again, using this case by case assessment approach the end of the event (monitor off) usually
inferred the beginning of a new event such as a transition from sedentary to upright in preparation for washing or a MRI scan, therefore events that
overlapped were included. Subgroup analysis using categories of stroke severity measured on the day of observation (mild stroke: NIHSS ≤ 7; moderate and severe stroke: NIHSS ≥ 8) was conducted to investigate the association between severity and the activity outcome measures. The number of patients with severe stroke (NIHSS >16) was too low (n = 5) to justify separate analysis. The NIHSS score that was extracted from medical notes on admission was used where the NIHSS had not been assessed on the day of monitoring.
Other activity-related data (BMT)
The activPALTM does not detect between lying and sitting so the BMT data were used to provide information of these types of sedentary behaviour. For example, a patient shown to have been consistently sedentary all day may have actually been hoisted from bed to chair. The activPALTM would have missed this
important information. For the BMT data, the total number of observations for each type of motor activity was calculated. As more than one activity may have been observed in one observation period the highest level of activity obtained in each of the observations was used. The motor activity categories were
classified, again, into upright activity or sedentary behaviour. Sedentary behaviour was further classified as in-bed or out-of-bed (Table 3-2).
Table 3-2 Classifications of motor activity categories from BMT
Motor activity Type of behaviour
No active motor (supine or side-lying)
Sedentary (in-bed) Sit support in bed
Sit support out of bed
Sedentary (out-of-bed) Hoist transfer
Sit no support Transfer feet on floor
Upright activity Stand
Walk Stairs
To investigate relationships between upright activity and person present or location data from the BMT and AC were synchronised and combined. Firstly, to
summarise the person present data, the groups were collapsed into eight categories (Table 3-3) and the total number of observations was calculated for each person present and location categories.
Table 3-3 Classifications of person present categories from BMT
Original ‘person present’ category New ‘person present’ category
Alone Alone
Speech and language therapist Speech and language therapist
Family Family
Patient transport Patient transport
Interpreter
Other∗ Other MD team
Other
∗ The ‘other’ category included ‘other MDT staff’ (pharmacists and dieticians), other hospital staff (phlebotomists, smoking cessation representatives, cleaning or catering staff) other patients or talking on mobile phone.
Secondly, appropriate summary estimates for each time point were calculated.
The amount of time spent upright as a proportion of the total recording time in each 10 minute time interval (i.e. 08:00 to 08:10) was calculated for each patient. This was summarised as the mean proportion of time spent upright. For each 10 minute time interval (08:00, 08:10 etc) the total number of nurses and therapists present as a proportion of the total number of observations was calculated. Likewise, the total number of patients observed in each location category as a proportion of the total number of observations was calculated.
This required the assumption that the person present and location remained the same for each 10 minute interval.
Different methods to monitor activity
In order to assess the agreement between AC and BMT two comparable units of measurement were identified and calculated for each method. For the AC data the time spent upright in seconds as a proportion of the total monitoring time for each one minute time interval i.e. 08:00 to 08:01, 08:10 to 08:11 was
calculated. This was summarised as the mean proportion of time spent upright for each time interval and enabled the AC data to be time-matched with the BMT data. For the BMT data the number of times upright as a proportion of the total number of observations i.e. 08:00 to 08:01, 08:10 to 08:11 was calculated.
These proportions were plotted against each other. Linear regression analysis was used to quantify the extent researcher observation can be predicted by accelerometry.
Predictors of upright physical activity levels
Multivariate linear regression was used to assess which baseline characteristics were significantly (p ≤ 0.05) predictive of upright physical activity. A logarithmic transformation was applied to the dependent variable, time spent upright. The variables that were entered into the regression model were as follows: stroke severity (NIHSS baseline), mobility (MSAS score), time in hours from stroke onset, previous stroke and family present (total number of times family were present as a proportion of the total number of observations for each patient).46
Upright physical activity as a predictor of functional outcome
Univariate analysis was undertaken to examine the association between
potentially predictive baseline characteristics and functional outcome at three and six months. The factors identified as predictive of mobility at 30 days (refer to Chapter 2) were used here. These were as follows: age, stroke type (OCSP classification, coded as TACS = 1, no TACS = 0), living alone (coded as alone = 1, not alone = 0), level of disability (mRS ≥ 3 coded as high disability = 1, mRS < 2 coded as 0), level of function (BI < 17, coded as dependent = 1, BI ≥ 18 coded as 0) and stroke severity (NIHSS ≤ 7, coded as mild stroke = 0, NIHSS ≥ 8 coded as moderate and severe stroke). Patient scores from the BI were dichotomised to create a binary outcome (BI ≥ 18 coded as independent = 1 or not independent = BI < 17). The variables that were statistically significant (p < 0.1) on univariate analysis were included in the multivariate model.72 Logistic regression was employed, using backward stepwise regression to drop the least significant variables (p < 0.1), to identify the variables which best predict function.