The Life Cycle of a Trend Steven S Prevette Fluor Hanford Environment Safety and Health ASQ Certified Quality Engineer 509-373-9371 http://www.hanford.gov/safety/vpp/trend.htm The aim of this paper is to provide assistance on establishing a statistical baseline for a work process. Baselines are necessary in order to determine if and when a trend has developed in the work process being monitored. Baselines serve as a predictor of future performance. This paper is written with the assumption that your organization is making use of Statistical Process Control (SPC) to monitor work performance. This paper focuses on one aspect of SPC, determination of the baseline average line for a performance indicator chart. For general information on SPC, please refer to http://www.hanford.gov/safety/vpp/trend.htm Trending or a work process or process element can result in continual improvement in performance of work. The life-cycle-of-a trend is a cyclical effort. The figure below provides an overview of this SPC trending tool. 0. Gestation: Choose a Measure Gather data 4. Maturity: Data stabilizes 1. Establish Baseline Take action to correct, reinforce, or stabilize Establish Expectations, Routine Monitoring 3. Youth: Determine special cause(s) 2. Birth: A trend is detected 0. Gestation The very first step (“stage zero”) in trending a set of data from a work process is to setup the initial control chart from available data. Many organizations have massive amounts of data about existing processes, but have never analyzed it. Having decided upon what process element(s) to measure, the initial data set is gathered. Preferably, there should be at least 25 or more data points collected over time. Even if you do not have 25 points, it Life Cycle of a Trend SS Prevette is still worth proceeding, but a reasonable attempt should be made to retrieve 25 time intervals of data. Data intervals may be based upon time (weekly, monthly, quarterly) or upon numbers of units produced. Choosing a Measure Choosing what to measure should not be a daunting task. If one carefully reviews the stated business objectives, mission statement, vision statement, contractual clauses of the organization, one can find measures. Promises of timeliness and accuracy can lend themselves to development into measurable indicators. Common factors also include safety, financial, achievement of schedules and milestones. This review is then further refined into “operational definitions”. Operational definitions supply sufficient detail such that data gatherers can go forth and collect and collate the data, and have a reasonable expectation that the result is repeatable and measures what was intended. No data set is ever “perfect”, so try to plan ahead with a good, agreed upon definition. Most disagreements over performance results arise from a lack of firm definition of how to retrieve the data in the first place. Gathering Data In many cases one will find that existing reporting systems can be utilized. Many organizations are awash in data, but the data have never been analyzed or put to use. There is an advantage to using existing data, as it has already been paid for, and historic data for establishment of the baseline should be available. An “operational definition” for each measure should be established. This is the agreed upon definition as to what data will be collected and under what criteria. For example, if you desire to trend the amount of corrective maintenance performed, you will need to define what is the unit to be counted (such as number of work packages) and what makes a maintenance item be considered “corrective maintenance”. The operational definition can be based upon existing fields and conventions in the existing reporting system. Analysis of reporting system data also serves to validate the data quality of the reporting system. Automation of data gathering can save time, money, and have improved accuracy and traceability of results. However, do not only focus on the data that are immediately on hand. It may be necessary to establish new data sources to reflect the total performance of the organization. See http://www.hanford.gov/safety/vpp/busobj2.htm for further ideas on choosing measures. Good questions to ask when developing measures include: o What are my customer’s needs and wants? o If I were gone on vacation for a month, what data would I look at upon return? o What are my critical success factors? o What have been causes of historical crises and failures? Data for control charts should be independent from time increment to time increment. For example, inventory in a warehouse at the end of this month is highly dependent upon Life Cycle of a Trend SS Prevette the inventory at the end of last month. Run charts can be plotted of the inventories, but it can be better to analyze the flow rates through the warehouse (incoming and outgoing). When gathering the data, do not “cumulate” the data (such as summing year-to-date). Also, the original time order of the data should be maintained. Choosing Data Reporting Intervals The monthly time increment for data gathering and reporting is very common. But one should consider the following when establishing the reporting interval. What is the longest time you could withstand (from a risk perspective) with an undetected trend? If a trend started the day following the closure of data collection, how long could you continue to operate without substantial risk? One should recognize that there may be other anecdotal signals of a trend that may provide backup and trigger an early analysis cycle. It may be appropriate to shorten the time interval (such as from monthly to weekly) in order to gain a sufficient volume of data points (25 minimum). Increasing the amount of reporting (increasing the frequency of data gathering and reporting) has several costs. One must consider the cost to gather, analyze, and publish the data. Also, one may discover that by the time a report is generated, one or more reporting time intervals have elapsed if the reporting interval is too short. Many organizations use a time interval basis for reporting. However, if one is measuring characteristics of items produced, it may make more sense to report in increments of output produced. For example, one may choose to plot a new datum after each fifth item produced. This is the basis for xbar-R control charts. If one is counting items or events, adjust the time interval for the analysis in order to keep the average rate above 1.0 per time interval. As an alternative, low rate trending (see http://www.hanford.gov/safety/vpp/lowrate.htm) could be used. 1. Establish the Baseline The baseline on a control chart consists of the average (center) line, a three-standard deviation Upper Control Limit (UCL) and a three-standard deviation Lower Control Limit (LCL). A procedure for choosing the proper form of control chart, and calculating the baseline and control limits is available at http://www.hanford.gov/safety/vpp/spc.htm. An example control chart is portrayed below. Life Cycle of a Trend SS Prevette UCL 16 14 Average Line Performance Data 12 10 8 6 4 LCL 2 29 27 25 23 21 19 17 15 13 11 9 7 5 3 1 0 In general, it is best to use all of the data you have. At a minimum, one should keep at least 25 points on hand prior to discontinuing use of older data. If you have less than 25, use what you have. The goal is to establish one or more baseline time intervals with no trends within the interval. Experience has indicated a rule of thumb called the “MW Rule” for how many data points are needed to establish a “good” baseline. For purposes of this paper, “goodness” reflects the capability of the baseline to serve in detecting future trends. That is, once a baseline is established, and later a trend is detected, it is not likely that the trend was due to there being too few data points in the baseline, causing the baseline to have been inaccurate. The “MW” Rule A rule of thumb is that one needs at least three changes of direction in the data within a baseline (“MW” Rule) Life Cycle of a Trend SS Prevette “MW” Rule 2 1 3 Empirical experience is that if the data have made three changes of direction (like a letter “M” or “W”), then there are sufficient data to generate a baseline. It is also desirable for the resulting average line to pass between the “corners” or vertices. In the example above, the average line will have corner “2” above and corners “1” and “3” below. The Initial Baseline Try to use the first 25 points for the baseline. If 25 points were not available (either due to lack of data, or trends within the data), use the “MW” rule. It can be worth leaving some recent data out of the baseline, for testing the suitability of the baseline. Also, in some reporting systems the current data are incomplete or subject to change. With less than 25 points the resulting control limits can be quite “soft” and should be viewed with some skepticism until proven to be valid. Achieving a run of 25 points, with no statistically significant changes against the average and control limits would serve to “validate” the control limits. Note that it is not necessary to recalculate the average and control limits once you reach 25 (or any other number). One only recalculates the baseline following a statistically significant change. Plot the data, and add the initial average and control limits as appropriate to the chart type. Next, apply the following detection rules. An suggested set of rules to use are: One point outside the control limits Two out of Three points two standard deviations above/below average Four out of Five points one standard deviation above/below average Seven points in a row all above/below average Ten out of Eleven points in a row all above/below average Seven points in a row all increasing/decreasing. Life Cycle of a Trend SS Prevette Several other sets of rules exist (for example, Dr. Wheeler uses 8 in a row instead of 7). Also, some may choose to implement the easier rules first (such as points outside the control limits and seven in a row) prior to implementing the more difficult rules. Whatever the choice, one should determine the set of rules they will use and stick to it. When one of these rules is triggered”, it can be considered a “trend”. In most requirements referring to “trends” or “trending”, a “trend” is something that has changed, or something that is “new” that should reviewed, causal factors determined, and actions taken. The statistical criteria above provide a method to separate a “trend” from ongoing random noise in the data. If there are not statistically significant changes detected in the initial baseline, the baseline should be accepted, and utilized for reporting. However, in many cases there will be trends detected. These trends represent shifts and the question “What happened” should be asked. The process should be investigated and the causes of the trends determined. Appropriate actions should be taken to bring the process “under control”. However, even if the current process has trends within it, it is necessary to determine some baseline to work with such that one can detect further trends. If the initial attempt to establish average and control limits results in trends being identified, the following strategies can be utilized to establish a valid baseline. Cleaning Up the Baseline If the initial baseline has a trend within it, recalculate the average and control limits through the following strategies: Drop data off of the beginning Drop data off of the end Drop individual datum point(s) out Split into two or more baselines Note that the complete data set should still be plotted on the chart, we are only making a choice as to what data to include in the average and control limit calculations. If zeroes are causing problems: Leave runs of 7 or more zeroes out (they will always be below average) Lengthen the data interval (i.e. from monthly to quarterly) Try “Low Rate Trending” (http://www.hanford.gov/safety/vpp/lowrate.htm) Chart Examples: Life Cycle of a Trend SS Prevette 1. Three data points removed from the beginning of the data set, leaving a stable baseline from the fourth point onwards: 2. Three data points removed from the end of the data. In this case, there is a current trend that needs to be acted upon, and there are not enough data to determine a new baseline (there is only one change in direction in the three recent points). 3. Two data points within the data have been removed, and the remaining data used to construct the baseline. The special causes of these two points should be investigated to determine if they could recur. If they can recur, one would need to be careful with the validity of predicting future performance using this baseline (or at least put a clarifier on the prediction that as long as these special causes do not recur, this is the prediction). Life Cycle of a Trend SS Prevette 4. The data have been split into two baselines. A permanent decrease in the level appears to have occurred. The current baseline does have three changes in direction, so it is believed that this is a valid prediction of future performance. The cause for the decrease should be determined. 5. Below is an example of “low rate trending” from Dr. Wheeler’s Understanding Variation, the Key to Managing Chaos. The first chart is based upon counting events per Life Cycle of a Trend SS Prevette month, and does not show a trend. The second chart is based upon trending the time between events, and does show an increasing trend. Sep-05 May-05 Jan-05 Sep-04 May-04 Jan-04 Sep-03 May-03 Jan-03 Sep-02 May-02 Jan-02 May-01 Jan-01 Sep-01 Number of Events per Month 5 4 3 2 1 0 4 Rate per Year, between Events 3 2 1 0 The x-chart Standard Deviation The sample standard deviation formula (“stdev” in Excel) can be used for determining the standard deviation of x-charts. The traditional method uses the moving range of the data, which most statistical texts prefer. With either method, one does need to check for outliers in the data, which can cause the control limits to be too far from the average and fail to detect the trend. An example follows. After removing the datum point that is close to the UCL, the recalculated UCL shows that the point is an outlier and justifies removing it from the baseline. As always, the special cause for the outlier should be determined. With Outlier in the Average and Stdev Life Cycle of a Trend SS Prevette With Outlier Removed Establish Expectations, perform routine monitoring Once the baseline is established, management should determine if the baseline is “acceptable”. In manufacturing, this is referred to as determining the “process capability”. Since the baseline is the prediction of future performance, can the organization continue to prosper (or continue to have crises, as the case may be) if current performance continues? If current performance is indeed “acceptable”, the management expectation that should be established is to maintain current performance. If not, then the expectation that should be established is that improvement is needed. Further, the expectation (or goal) is that a significant improving trend needs to occur. In order to improve performance when the process is stable, one will need to review the common causes of the results, found in the work process elements, across the entire baseline time interval. This should provide insight as to how to improve the process. For organizations that pledge to “continual improvement” in principle, there should be a manageable number of indicators of the work process elements identified as “needing improvement”. As improvement is achieved, it may be appropriate to shift the improvement focus to other work process element indicators. Deciding those indicators that are acceptable as is and those in need of improvement is a rudimentary form of risk management. High-risk areas should be targeted for improvement. Low-risk areas that can be accepted “as is” are simply monitored and action is only taken if a non-improving trend occurs. During routine monitoring and reporting, the data are gathered using the operational definitions previously established, at the time interval previously established. During each reporting cycle, the new data are plotted and reviewed against the trending rules to see if a trend has developed. This trend could be due to an unforeseen change, or due to actions taken to improve the process. The charts are published routinely. Note that there is nothing special about the end of the calendar year or fiscal year. The data continue to be plotted by their reporting interval. In some cases where management focuses upon fiscal or calendar year, a text box with the value of the current and past fiscal year average may be useful, at least until management completely shifts to reading the control chart only. Life Cycle of a Trend SS Prevette The baselines are not shifted unless a significant trend per the trending rules develops. A baseline is “innocent until proven guilty” and is only proven guilty using your established trend rules. Some organizations desire to have “executive summaries” and/or color coding of results. The following phrases and color codes could be utilized for standardized, consistent wording: Data are stable, at a good baseline value (“Green”) Data are stable, improvement is needed (“Yellow”) There is an improving trend (“Green”) There is a non-improving trend (“Red”) Birth of a Trend Your agreed upon set of trend detection rules are utilized in order to detect the birth of the trend. To review, the set of rules to consider are: One point outside the control limits Two out of Three points two standard deviations above/below average Four out of Five points one standard deviation above/below average Seven points in a row all above/below average Ten out of Eleven points in a row all above/below average Seven points in a row all increasing/decreasing. The example charts below demonstrate a trend which has been discovered following establishment of the initial baseline. Initial Stable Process Average = 9.5 (Oct 98 - Oct 00) 20 15 25 points in a stable baseline 10 Life Cycle of a Trend SS Prevette Aug-01 Jun-01 Apr-01 Feb-01 Dec-00 Oct-00 Aug-00 Jun-00 Apr-00 Feb-00 Dec-99 Oct-99 Aug-99 Jun-99 Apr-99 Feb-99 Dec-98 0 Oct-98 5 Initial Indication of Trend Average = 9.5 (Oct 98 - Oct 00) 2 of 3 at 2 standard deviations above average, circled 20 15 10 Aug-01 Jun-01 Apr-01 Feb-01 Dec-00 Oct-00 Aug-00 Jun-00 Apr-00 Feb-00 Dec-99 Oct-99 Aug-99 Jun-99 Apr-99 Feb-99 Dec-98 0 Oct-98 5 Youth Any trends should be reported through the routine performance report. The chart itself should be annotated. A general practice is to circle those data points that make up the trend. Special notifications to responsible management may be necessary to flag that the trend exists. Once the trend has been detected, the special cause(s) of the trend should be determined. Note that the cause might turn out to be the result of a previous improvement action implemented by management. “Root cause” analyses techniques may be useful. Comparison of the common causes in force during the stable time interval (Oct 98 to Nov 00 in this case) to the special causes in force during the trend (Dec 00 and Jan 01) may provide insight. Ishikawa (fishbone) diagrams can also be useful for analyzing cause and effect. Note that ONLY looking at the data from the trend may result in a conclusion that was not the reason for the trend. For example, while investigating bearing failures one company discovered that there was a contaminate in the bearing grease of the failed bearings. A great campaign was instituted to eliminate the contamination. There was no effect on bearing failures. An analysis of “good” bearings found the same contamination existed. Take Action After reviewing the special causes, management should determine if action is needed. If no action is needed, the trend is monitored in order to determine when and at what value it stabilizes. In organizations with “corrective action management” (such as required by CFR 830.122), non-improving trends should be reviewed to see if they are “adverse”. Many corrective action management policies require adverse trends to be documented and the corrective actions documented. Improving trends should be reinforced in order to maximize the shift in the improving direction. Life Cycle of a Trend SS Prevette Lessons learned from all trends should be reviewed for application in other processes. Maturity As time progresses, and actions are taken, the data will eventually stabilize at a new level. The charts below follow on the discussion of the previous data example. The next chart shows more data received for the trend (which has increased to seven points in a row above average), but has not yet stabilized. Waiting for 3 changes of Direction More Data Gathered Average = 9.5 (Oct 98 - Oct 00) > 7 months in a row above average 20 15 10 Aug-01 Jun-01 Apr-01 Feb-01 Dec-00 Oct-00 Aug-00 Jun-00 Apr-00 Feb-00 Dec-99 Oct-99 Aug-99 Jun-99 Apr-99 Feb-99 Dec-98 0 Oct-98 5 Cycle back to step 1 – Establish Baseline In the last chart below, three changes of direction of the new data have occurred, satisfying the “MW” rule. Therefore, new baseline average and control limits are established. The circle around the trend is removed, as the shift in the average line now accounts for the trend. It should be noted that in some cases the data will return to the old baseline, and it may not be necessary to calculate a new baseline. This is a judgment call on the part of the analyst. Eventually, it will become obvious as to whether a new baseline is needed, as further trends would be detected by the trending rules. Life Cycle of a Trend SS Prevette Average = 9.5 (Oct 98 - Oct 00) New Baseline Established Average = 13 (Dec 00 - Aug 01) 20 15 10 Aug-01 Jun-01 Apr-01 Feb-01 Dec-00 Oct-00 Aug-00 Jun-00 Apr-00 Feb-00 Dec-99 Oct-99 Aug-99 Jun-99 Apr-99 Feb-99 Dec-98 0 Oct-98 5 With the new baseline established, one continues on with the life-of-a-trend cycle. This “life cycle”, if executed in a routine, consistent manner, will be of great assistance in generating continual improvement for the practicing organization. -----------------------------------------------------------------------------------------------------References: Quality Control and Industrial Statistics, Acheson Duncan The New Economics and Out of the Crisis, W. Edwards Deming Peer Reviews: Peer Review comments from Rich DeRoeck, Alpha Industries; Phil Monroe, DEMCOM Consulting; and Gary Reid, Department of Energy Rocky Flats were incorporated. Life Cycle of a Trend SS Prevette