Quality Improvement Charts

Shewhart’s 3-sigma rule

Shewhart devised one test to identify special cause variation, the 3-sigma rule, which signals if one or more data points fall outside the 3-sigma limits. The 3-sigma rule is effective in detecting larger (> 2SD) shifts in data.

# Simulate a special cause to y in the form of a large, transient shift of 4 standard deviations.
y[22] <- 4

qic(y, chart = 'i')

However, minor to moderate shifts may go unnoticed by the 3-sigma rule for long periods of time.

# Simulate a special cause to y in the form of a moderate, persistent shift of 2 standard deviations.
y[13:24] <- rnorm(12, mean = 2)

qic(y, chart = 'i')

Therefore, many additional tests have been proposed to increase the sensitivity to non-random variation.

It is important to realise that while it may be tempting to apply more tests in order to increase the sensitivity of the chart, more test will increase the risk of false signals, thereby reducing the specificity of the chart. The choice of which and how many tests to apply is therefore critical.

The Western Electric rules

The best known tests for non-random variation are probably the Western Electric Rules described in the Statistical Quality Control Handbook (Western Electric Company 1956). The WE rules consist of four simple tests that can be applied to control charts by visual inspection and are based on the identification of unusual patterns in the distribution of data points relative to the control and centre lines.

1. One or more points outside of the 3-sigma limits (Shewhart’s original 3-sigma rule).

2. Two out of three successive points beyond a 2-sigma limit (two thirds of the distance between the centre line and the control line).

3. Four out of five successive points beyond a 1-sigma limit.

4. A run of eight successive points on one side of the centre line.

The WE rules have proven their worth during half a century. One thing to notice is that the WE rules are most effective with control charts that have between 20 and 30 data points. With fewer data points, they lose sensitivity (more false negatives), and with more data points they lose specificity (more false positives).

The Anhøj rules

The least known tests for non-random variation are probably the Anhøj rules, proposed and validated by me in two recent publications (Anhøj 2014, Anhøj 2015). The Anhøj rules are the core of qicharts2, and consist of two tests that are based solely on the distribution of data points in relation to the centre line:

• Unusually long runs: A run is one or more consecutive data points on the same side of the centre line. Data points that fall on the centre line neither break nor contribute to the run. The upper 95% prediction limit for longest run is approximately \(log_2(n)+3\) (rounded to the nearest integer), where \(n\) is the number of useful data points. For example, in a run chart with 24 data points a run of more than round(log2(24) + 3) = 8 would suggest a shift in the process.

• Unusually few crossings: A crossing is when two consecutive data points are on opposite sides of the centre line. In a random process, the number of crossings has a binomial distribution, \(b(n-1,0.5)\). The lower 5% prediction limit for number of crossings is found using qbinom(p = 0.05, size = n - 1, prob = 0.5). Thus, in a run chart with 24 useful data points, fewer than qbinom(0.05, 24 - 1, 0.5) = 8 crossings would suggest that the process is shifting.

Critical values for longest runs and number of crossings for 10 – 100 data points are tabulated in Appendix 2.

Apart from being comparable in sensitivity and specificity to WE rules 2-4 with 20-30 data points, the Anhøj rules have some advantages:

• They do not depend on sigma limits and thus are useful as stand-alone rules with run charts.

• They adapt dynamically to the number of available data points, and can be applied to charts with as few as 10 and up to indefinitely many data points without losing sensitivity and specificity.

In the previous control chart, a shift of 2 standard deviations was introduced halfway through the process. While the shift is too small to be picked up by the 3-sigma rule, both Anhøj rules detect the signal. The longest run includes 13 data points (upper limit = 8), and the number of crossings is 4 (lower limit = 8). Note that the shift is also picked up by WE rule 4.

Non-random variation found by the Anhøj rules is visualised by qic() using a red and dashed centre line.

The parameters of the analysis are available using the summary() function on a qic object. See ?summary.qic for details.

# Save qic object to o
o <- qic(y, chart = 'i')

# Print summary data frame
summary(o)
#>   facet1 facet2 part n.obs n.useful longest.run longest.run.max n.crossings
#> 1      1      1    1    24       24          13               8           4
#>   n.crossings.min runs.signal      aLCL       CL     aUCL sigma.signal
#> 1               8           1 -2.114884 1.057091 4.229067            0

Many additional tests for special cause variation have been proposed since Shewhart invented the 3-sigma rule, and many tests are currently being used. However, the more tests that are applied to the analysis, the higher the risk of false signals. qicharts2 takes a conservative approach by applying only three tests, the 3-sigma rule and the two Anhøj rules. That means that a signal from qic() most likely represents a significant change in the underlying process.

Run chart of adverse events per 1000 patient days

qic(month, harms, days,
    data     = gtt,
    multiply = 1000,
    title    = 'Patient harm',
    ylab     = 'Adverse events per 1000 patient days',
    xlab     = 'Month')

Since only random variation is present in the run chart, a U control chart for harm rate is appropriate for computing the limits of the common cause variation.

U chart of adverse events per 1000 patient days

qic(month, harms, days,
    data     = gtt,
    chart    = 'u',
    multiply = 1000,
    title    = 'Patient harm (U chart)',
    ylab     = 'Adverse events per 1000 patient days',
    xlab     = 'Month')

Because the sample sizes (number of patient days) are not the same each month, the 3-sigma limits vary. Large samples give narrow limits and vice versa. Varying limits are also seen in P and Xbar charts.

Run and P control charts of percent harmed patients

To plot the percentage of harmed patients (patients with 1 or more harms) we need to add a variable and to aggregate data by month.

suppressPackageStartupMessages(library(dplyr))

gtt_by_month <- gtt %>%
  mutate(harmed = harms > 0) %>% 
  group_by(month) %>% 
  summarise(harms    = sum(harms),
            days     = sum(days),
            n.harmed = sum(harmed),
            n        = n())

head(gtt_by_month)
#> # A tibble: 6 x 5
#>   month      harms  days n.harmed     n
#>   <date>     <dbl> <dbl>    <int> <int>
#> 1 2010-01-01     5   113        3    20
#> 2 2010-02-01     7   132        6    20
#> 3 2010-03-01     5   121        4    20
#> 4 2010-04-01     7   134        5    20
#> 5 2010-05-01     7   116        7    20
#> 6 2010-06-01     5   131        5    20

qic(month, n.harmed,
    n         = n,
    data      = gtt_by_month,
    y.percent = TRUE,
    title     = 'Harmed patients',
    ylab      = NULL,
    xlab      = 'Month')

Again, since the run chart finds random variation, a control chart may be applied. For proportion (or percent) harmed patients we use a P chart.

qic(month, n.harmed, n,
    data      = gtt_by_month,
    chart     = 'p',
    y.percent = TRUE,
    title     = 'Harmed patients (P chart)',
    ylab      = NULL,
    xlab      = 'Month')

Since both the harm rate and the proportion of harmed patients exhibit common cause variation only, we conclude that the risk of adverse events during hospitalisation is predictable: we in the future should expect around 60 adverse events per 1000 patient days, harming around 33% of patients (regardless of their length of stay), although harm rates anywhere between ~0 – 60% would be considered within common cause variation.

Actually, we often do not need to aggregate data in advance. With a little creativity we can make qic do the hard work. The previous plot could have been produced with this code. Note the use of the original gtt dataset and the number of harmed and total number of patients created on the fly.

qic(month, 
    harms > 0,        # TRUE if patient had one or more harms
    days > 0,         # Always TRUE 
    data      = gtt,
    chart     = 'p',
    y.percent = TRUE,
    title     = 'Harmed patients (P chart)',
    ylab      = NULL,
    xlab      = 'Month')

Pareto analysis of adverse event types and severity

The Pareto chart, named after Vilfred Pareto, was invented by Joseph M. Juran as a tool to identify the most important causes of a problem.

Types and severity of adverse events in gtt are organised in an untidy wide format. Since paretochart() needs a vector of categorical data, we need to rearrange harm severity and harm category into separate columns.

suppressPackageStartupMessages(library(tidyr))

gtt_ae_types <- gtt %>%
  gather(severity, category, E:I) %>% 
  filter(complete.cases(.))

head(gtt_ae_types)
#>   admission_id admission_dte discharge_dte      month days harms severity
#> 1            2    2009-12-24    2010-01-06 2010-01-01   14     2        E
#> 2           11    2010-01-08    2010-01-18 2010-01-01   11     2        E
#> 3           12    2010-01-11    2010-01-19 2010-01-01    9     1        E
#> 4           29    2010-01-27    2010-02-10 2010-02-01   15     1        E
#> 5           41    2010-02-24    2010-03-01 2010-03-01    6     1        E
#> 6           45    2010-02-22    2010-03-06 2010-03-01   13     1        E
#>           category
#> 1   Pressure ulcer
#> 2        Infection
#> 3   Pressure ulcer
#> 4   Pressure ulcer
#> 5 Gastrointestinal
#> 6 Gastrointestinal

paretochart(gtt_ae_types$category,
            title = 'Pareto chart of harm category')

The bars show the number in each category, and the curve shows the cumulated percentage over categories. Almost 80% of harms come from 3 categories: gastrointestinal, infection, and procedure.

paretochart(gtt_ae_types$severity,
            title = 'Pareto chart of harm severity')

Almost all harms are temporary (E-F).

Run charts for number of infections

We begin by plotting a run chart of the number of CDIs per month.

qic(month, n,
    notes    = notes,
    data     = cdi,
    decimals = 0,
    title    = 'Hospital acquired Clostridium difficile infections',
    ylab     = 'Count',
    xlab     = 'Month')

A downward shift in the number of CDIs is clearly visible to the naked eye and is supported by the runs analysis resulting in the centre line being dashed red. The shift seem to begin around the time of the intervention.

Even though it is not necessary in this case, we can strengthen the analysis by using the median of the before-intervention period to test for non-random variation in the after period.

qic(month, n,
    data        = cdi,
    decimals    = 0,
    freeze      = 24,
    part.labels = c('Baseline', 'Intervention'),
    title       = 'Hospital acquired Clostridium difficile infections',
    ylab        = 'Count',
    xlab        = 'Month')

The median number of CDIs per month in the before period is 19. An unusually long run of 15 data points below the median proves that the CDI rate is reduced.

When a shift in the desired direction is the result of a deliberate change to the process, we may split the chart to compare the numbers before and after the intervention.

qic(month, n,
    data     = cdi,
    decimals = 0,
    part     = 24,
    title    = 'Hospital acquired Clostridium difficile infections',
    ylab     = 'Count',
    xlab     = 'Month')

C chart for number of infections

Since both periods show only random variation, a control chart may be applied to test for larger transient shifts in data and to establish the natural limits of the current process. The correct chart in this case is the C chart for number of infections.

qic(month, n,
    data  = cdi,
    chart = 'c',
    part  = 24,
    title = 'Hospital acquired Clostridium difficile infections (C chart)',
    ylab  = 'Count',
    xlab  = 'Month')

The split C chart shows that the CDI rate has dropped from an average of 19 to 7 per month. Furthermore, the 3-sigma limits tell that the current process is predictable and that we in the future should expect between 0 and 15 CDIs per month.

U chart for infection rate

Until now we have assumed that the area of opportunity, or exposure, is constant, i.e., the number of patients or patient days have not changed significantly over time. When the area of opportunity varies, a U chart (U for unequal area of opportunity) is more appropriate.

qic(month, n, days,
    data     = cdi,
    chart    = 'u',
    part     = 24,
    multiply = 10000, 
    title    = 'Hospital acquired Clostridium difficile infections (U chart)',
    ylab     = 'Cases per 10,000 risk days',
    xlab     = 'Month')

In fact, U charts can always be used instead of C charts. When the area of opportunity is constant the two charts will reach the same conclusion. But the raw numbers in a C chart may be easier to communicate than counts per something.

It is very important to recognise that while run and control charts can identify non-random variation, they cannot identify its causes. The analysis only tells that the CDI rate was reduced after an intervention. The causal relationship between the intervention and the shift must be established using other methods.

When should you recompute the limits?

When splitting a chart as in the example above you make a deliberate decision based on your understanding of the process. Some software packages offer automated recalculation of limits whenever a shift is detected. I find this approach very wrong. Limits are the voice of the process and recomputing limits without understanding the context, which computers don’t, is like telling someone to shut up, when they are really trying to tell you something important.

Wheeler suggests that limits may be recomputed when the answers to all four of these question is “yes” (Wheeler 2010, p. 224):

• Do the data display a distinctly different kind of behaviour than in the past?

• Is the reason for this change in behaviour known?

• Is the new process behaviour desirable?

• Is it intended and expected that the new behaviour will continue?

If the answer to any of these question is “no”, you should rather seek to find the assignable cause of the shift and, if need be, eliminate it.

Thus, when a deliberate change has been made to the process and the chart finds a shift attributable to this change, it may be time to recompute the limits.

I and MR charts for age of individual patients

In this and the following cases we skip the obligatory first step of using a run chart.

First, we will make a control chart of the age of the last 100 patients. Since we are plotting individual measurements, we use an I chart.

qic(age,
    data      = tail(cabg, 100), 
    chart     = 'i',
    show.grid = TRUE,
    title     = 'Age of the last 100 patients (I chart)',
    ylab      = 'Years',
    xlab      = 'Patient #')

Two data points (patients number 45 and 70) are below the lower 3-sigma limit indicating that these patients are “unusually” young, i.e. representative of phenomena that are not normally present in the process.

I charts are often displayed along moving range charts that show the ranges of neighbouring data points.

qic(age,
    data      = tail(cabg, 100), 
    chart     = 'mr',
    show.grid = TRUE,
    title     = 'Age of the last 100 patients (MR chart)',
    ylab      = 'Years',
    xlab      = 'Patient #')

The MR chart identifies three unusually large ranges, which are clearly produced by the two unusually young patients also found in the I chart. Note that each data point in an I chart is contained in two moving ranges in the MR chart.

We should seek the cause(s) of these special causes. We may then chose to eliminate the outliers from the computations of the process centre and 3-sigma limits in order to predict the expected age range of future patients.

qic(age,
    data    = tail(cabg, 100), 
    chart   = 'i',
    exclude = c(45, 70),
    title   = 'Age of the last 100 patients (I chart)',
    ylab    = 'Years',
    xlab    = 'Patient #')

Thus, on average we should expect future patients to be 67 years of age, and patients younger than 43 years or older than 91 years would be unusual.

Xbar and S charts for average and standard deviation of patient age

We could also plot the mean age of patients by month. For this we choose an Xbar chart for multiple measurement per subgroup. First, we need to add month as the grouping variable.

cabg <- cabg %>% 
  mutate(month = as.Date(cut(date, 'month')))

Using month as the x axis variable, qic() automatically calculates the average age per month.

qic(month, age,
    data  = cabg,
    chart = 'xbar',
    title = 'Average age (Xbar chart)',
    ylab  = 'Years',
    xlab  = 'Month')

Xbar charts are usually displayed along S charts displaying the within subgroup standard deviation.

qic(month, age,
    data  = cabg,
    chart = 's',
    title = 'Age standard deviation (S chart)',
    ylab  = 'Years',
    xlab  = 'Month')

Since both the Xbar and the S chart display common cause variation, we may conclude that the monthly average and standard deviation of patient age is predictable. Note that the mean of the Xbar chart is close to the mean of the individuals chart but that the 3-sigma limits are narrower.

P charts for proportion of patients who were readmitted or died within 30 after surgery

To plot readmissions and mortality we first need to summarise these counts by month.

cabg_by_month <- cabg %>% 
  group_by(month) %>% 
  summarise(deaths       = sum(death),
            readmissions = sum(readmission),
            n            = n())

head(cabg_by_month)
#> # A tibble: 6 x 4
#>   month      deaths readmissions     n
#>   <date>      <int>        <int> <int>
#> 1 2011-07-01      1           14    52
#> 2 2011-08-01      3           12    64
#> 3 2011-09-01      2           15    70
#> 4 2011-10-01      1            8    60
#> 5 2011-11-01      1           16    67
#> 6 2011-12-01      3           11    69

Then we use P charts for proportion of readmissions.

qic(month, readmissions, n,
    data      = cabg_by_month,
    chart     = 'p',
    y.percent = TRUE,
    title     = 'Readmissions within 30 days (P chart)',
    ylab      = NULL,
    xlab      = 'Month')

On average, 22% of patients are readmitted within 30 days after surgery, and if nothing changes we would expect this to continue.

What about mortality?

qic(month, deaths, n,
    data      = cabg_by_month,
    chart     = 'p',
    y.percent = TRUE,
    title     = '30 days mortality (P chart)',
    ylab      = NULL,
    xlab      = 'Month')

Thankfully, post-surgery mortality is a rare event, and the P chart suggests common cause variation. However, it seems curious that only one patient died during the summer of 2012. To investigate this further, G or T charts for rare events are useful.

T and G charts for rare events

Rare events are often better analysed as time or opportunities between events. First, we need to calculate the number of surgeries and days between fatalities.

fatalities <- cabg %>% 
  mutate(x = row_number()) %>% 
  filter(death) %>% 
  mutate(dd = x - lag(x),
         dt = date - lag(date))

head(fatalities)
#>         date    age gender los death readmission      month   x dd      dt
#> 1 2011-07-21 91.227 Female   5  TRUE       FALSE 2011-07-01  37 NA NA days
#> 2 2011-08-06 71.773   Male   2  TRUE       FALSE 2011-08-01  60 23 16 days
#> 3 2011-08-26 73.942   Male   1  TRUE       FALSE 2011-08-01  99 39 20 days
#> 4 2011-08-31 80.293   Male   3  TRUE       FALSE 2011-08-01 114 15  5 days
#> 5 2011-09-13 75.359   Male  12  TRUE       FALSE 2011-09-01 148 34 13 days
#> 6 2011-09-14 72.019 Female  19  TRUE       FALSE 2011-09-01 149  1  1 days

Then we use a G chart for opportunities (surgeries) between events (deaths).

qic(dd,
    data  = fatalities,
    chart = 'g',
    title = 'Surgeries between deaths (G chart)',
    ylab  = 'Count',
    xlab  = 'Death #')

If we do not have information on the opportunities (surgeries with survival), we may instead use a T chart for time between events.

qic(dt,
    data  = fatalities,
    chart = 't',
    title = 'Days between deaths (T chart)',
    ylab  = 'Days',
    xlab  = 'Death #')

The conclusions from the two charts are the same: Deaths number 24 and 25 are unusually far apart. We should look for an assignable cause to explain this as it may help us get ideas to reduce postoperative mortality in the future.

Faceting readmission rates by gender

To demonstrate the use of faceted charts, we summarise readmissions by month and gender.

cabg_by_month_gender <- cabg %>% 
  group_by(month, gender) %>% 
  summarise(readmissions = sum(readmission),
            n            = n())
#> `summarise()` has grouped output by 'month'. You can override using the `.groups` argument.

head(cabg_by_month_gender)
#> # A tibble: 6 x 4
#> # Groups:   month [3]
#>   month      gender readmissions     n
#>   <date>     <fct>         <int> <int>
#> 1 2011-07-01 Female            2     9
#> 2 2011-07-01 Male             12    43
#> 3 2011-08-01 Female            0     9
#> 4 2011-08-01 Male             12    55
#> 5 2011-09-01 Female            3    18
#> 6 2011-09-01 Male             12    52

Then we use the facets argument to make a control chart for each gender while keeping the x and y axes the same.

qic(month, readmissions, n,
    data      = cabg_by_month_gender,
    facets    = ~ gender, 
    chart     = 'p',
    y.percent = TRUE,
    title     = 'Readmissions within 30 days (P chart)',
    ylab      = '',
    xlab      = 'Month')

This chart suggests that the readmission rate for females is increasing. This is clearly undesirable and we should investigate this to find and eliminate the special cause.

U chart of the total number of infections per 10,000 risk days

qic(month, n, days,
    data     = hospital_infections,
    chart    = 'u',
    multiply = 10000,
    title    = 'Hospital acquired infections in the Capital Region of Denmark',
    ylab     = 'Cases per 10,000 risk days',
    xlab     = 'Month')

The infection rate appears predictable. However, plotting three types of infection together is like mixing apples and oranges.

Faceted U chart by type of infection

qic(month, n, days,
    data     = hospital_infections,
    facets   = ~ infection,
    chart    = 'u',
    multiply = 10000,
    title    = 'Hospital acquired infections in the Capital Region of Denmark',
    ylab     = 'Cases per 10,000 risk days',
    xlab     = 'Month')

Now each infection has its own facet. We see that all three infections appear to be predictable and that UTI is the most common infection.

But since these data come from six very different hospitals, we should also facet by hospital.

Two-way faceted U chart by infection and hospital

qic(month, n, days,
    data     = hospital_infections,
    facets   = infection ~ hospital,
    chart    = 'u',
    multiply = 10000,
    scales   = 'free_y',
    x.angle  = 45,
    title    = 'Hospital acquired infections in the Capital Region of Denmark',
    ylab     = 'Cases per 10,000 risk days',
    xlab     = 'Month')

This plot makes it easy to display the differences between both hospitals and types of infection. Non-random variation is present in UTI rates of two hospitals (BOH, NOH).

Note the use of individual y axes to better visualise variation over time and the rotated x axis labels to avoid overlapping labels.

Funnels plot of infection rates by hospital

Processes containing only common cause variation over time may be aggregated and compared across other units than time. When data have no natural time order, i.e., when the x variable is categorical, the data points are not connected by lines and runs analysis is not performed. To produce a funnel plot of infections by hospital, we filter data from a recent stable period and reorder the x variable by size.

hai_2016 <- hospital_infections %>% 
  filter(month    >= '2016-07-01') %>% 
  mutate(hospital = reorder(hospital, days))

qic(hospital, n, days,
    data            = hai_2016,
    facets          = ~ infection,
    ncol            = 1,
    scales          = 'free_y',
    chart           = 'u',
    multiply        = 10000,
    decimals        = 1,
    show.labels     = TRUE,
    title           = 'Hospital acquired bacteremia 2016 (U funnel chart)',
    ylab            = 'Cases per 10,000 risk days',
    xlab            = 'Hospital')

One hospital, NOH, has unusually few bacteremias and urinary tract infections compared to the other hospitals. Another hospital, BFH, has unusually many Clostridium difficile infections. The other hospitals are indistinguishable from each other with respect to the hospital infection rates.

What can we learn from NOH and BFH? It is important to stress that data presented in funnel plots should be used for learning rather than for judgement. While it may be interesting for some in need of a “case” to point fingers at the good, the bad, and the ugly, the full power of SPC will only be released when data are used for learning and improvement.