CMMI High Maturity

High Maturity in CMMI for Development - Part 1

Reference: http://www.connect2hcb.com/tiki-index.php

The high maturity concept in CMMI is centered around four Process Areas (a Process Area is set of practices aligned together for achieving a defined purpose):

1. OPP (Organizational Process Performance)
2. QPM (Quantitative Project Management)
3. OID (Organizational Innovation and Deployment)
4. CAR (Causal Analysis and Resolution)

OPP is about establishing organization-level process performance baselines and predictive/forecasting models

QPM is about planning and managing a project in quantitative terms

OID is about planning and implemeting organization-wide improvement initiatives (in quantitative terms)

CAR is about systematically identifying and addressing root causes behind the problems encountered in performing a process

High Maturity in CMMI for Development - Part 2

Here are links to some excellent articles and reference material on CMMI High Maturity from the SEI website:

• hi-matmis.pdf
• Young08.pdf
• konrad.pdf
• hmworkshop.pdf

Note: Links no longer work. Try search using document names…

High Maturity in CMMI for Development - Part 3

For organizations that intend to achieve CMMI High Maturity, they should be able to demonstrate the following key practices:

- Declaration of process performance objectives in quantitative terms. The objectives must be set-up for at least the "critical" processes and sub-processes. The method to do this effectively is to ascertain or determine the organizations' business goals, identify the processes and sub-processes critical to business success (in terms of achievement of business goals), define parameters and metrics for measuring the process performance and finally assign performance targets or goals for the measurement parameters.

- Clear quantitative understanding of the performance of processes and sub-processes selected as "critical" in terms of their central tendency (mean, etc.) and dispersion (standard deviation, etc.). It may be a good idea to also gain an understanding of both the extent of how symmetric the distribution is (skewness) and also how peaked the distribution is (kurtosis), at least for the "super-critical" processes and sub-processes. This statistical characterization represents the "Process Performance Baselines".

- Process composition based on quantitative understanding of the performance of the available processes and sub-processes. Selection of processes and sub-processes for a specific project should be based on comparing theperformance objectives that must be achieved against what will, with a good likelihood, be achieved by following a selected set of processes and sub-processes

- Usage of prediction models (or "Process Performance Models"), in the form Y = f(x), for projecting or predicting the likely outcome that will be achieved by actually performing a process. The predictedperformance of the selected parameter (dependent Y factor) will be based on actual values of some influencing parameters (independent X factors). When actual data on the X factors becomes available as the project moves forward, the projected or the predicted value of the Y factor should be progressively calibrated and refined.

- Usage of control chart (SPC theory) for monitoring and controlling the process performance on a real-time basis

- Usage of statistical tools such as Hypothesis Testing (t-test, etc.) for making decisions based on solid quantitative understanding of variation

- Planning and execution of improvement projects using rigorous statistical tools and techniques. This can be done using a 4-step approach as listed below:

1. Establish in quantitative terms the current level of performance
2. Determine the statistical significance of the difference between the current and the desired level of performance. If the difference is significant, statistically speaking, plan and implement improvement actions; otherwise, no action is required.
3. In case improvement actions are taken, establish in quantitative terms the achieved level of performance
4. Determine the statistical significance of the difference between the achieved and the desired level of performance. If the difference is significant, statistically speaking, plan and implement further improvement actions; otherwise, stop.

High Maturity in CMMI for Development - Part 4

One of the most difficult (supposedly so) concept in CMMI High Maturity is the development of "valid" and "usable" statistical models for predicting and hence quantitatively managing the outcome of a process - commonly known as Prediction or Predictive Models.

The outcome of a process can be measured in terms of certain variables of interest. For example, the quality of work performed can be measured in terms of the metrics ‘rework effort as a %age of total effort spent on performing the work’ – this metrics can hence be taken as a variable of interest. Knowing the “future” value of this variable right at the beginning of the work can be effectively used for quantitative management (like in this case it could be proactive defect prevention).

Computing the metrics at the end of the work getting completed provides lagging indicators of the variables of interest and as such leaves limited or no scope for effective proactive actions. The ability to predict the “future” value of thevariables of interest gives the power to quantitatively manage the process towards predefined desired goals.

Prediction Model, as the name also suggests, is meant to be used before the associated process has been performed to a significant degree of completion. At the beginning, the variable of interest has an unknown value and hence prediction makes business sense. After the process reaches a significant degree of completion, it may not be cost-effective to do predictions and hence no prediction should be carried out.

Hence, a prediction model should operate in two stages:

Stage 1: Right at the starting point, it should provide an initial predicted value

Stage 2: At intermediate points, it should provide progressively refined predicted value as actual data on certain variables becomes known. The refinement should also consider any non-random events that may occur so that the prediction can be relied upon.

High Maturity in CMMI for Development - Part 5

Statistical models based on multiple (linear) regression can be easily built to serve as prediction models as required by CMMI High Maturity. A prediction model in this case is, simply speaking, nothing but a multiple (linear) regression equation.

A regression equation takes the form Y = f(x1, x2, …, xn), where,

* Y = dependant variable (representing the variable of interest that has to be predicted)

* x1, x2, …, xn = set of independent variables (representing the variables whose value is known and is fed into the model to obtain the predicted value of Y)

The steps to build a prediction model are as follows – the details of the statistical methods to be used for these steps are easily available and hence have not been duplicated:

* Check the data for basic sanity – erroneous data, missing data, etc.

* Determine the outlier values in the data and treat them in an appropriate manner. Outlier values are typically wayward data and may be retained or removed based on certain considerations

* Test the variables Y, x1, x2, …, xn for Normality. Most of the statistical results used commonly assume the underlying distribution to be a Normal Distribution. In case the distribution is Non-normal appropriate transformation is required to be done

* Understand the statistical behavior of the variables. This can be done by computing the measures of central tendency (typically mean) and measures of dispersion (typically standard deviation). A good starting point for this is to draw a histogram which will additionally give insight into the shape of the distribution

* Build the regression equation in the form Y = a + b1x1 + b2x2 + … + bnxn

* Analyze the statistical results (p-value, etc.) for the regression equation for the significance of the overall regression fit (test for R-square) and significance of the regression coefficients (test b1, b2, b3)

* Validate the regression equation for its prediction power. The preferred method is to use the regression equation on a different set of values than the ones that were used to build the model

If both the overall regression fit and the regression coefficients are significant the regression equation can be confidently used as a Prediction Model.

If that is not the case, however, alternative approaches need to be considered and adopted to develop the Prediction Model. Keep watching this space for future posts related to discussion on the alternative approaches that can be considered.