I have the equation: $ 0= 2.01106 - 0.00274(34.647+24.24a)-0.02059(45.647+21.122b)+1.37984(2.05-0.206c)-0.01176(10.588+11.963d)+0.00394(118.29-21.097e)-0.03552(92.17+2.855f)$

I have the above equation. The letter's $a$, $b$, $c$, $d$, $e$, and f correspond to z-scores. I would like to solve this equation for the lowest sum of $a+b+c+d+e+f$.

The range of $a$ = -1.42 to 2.69, $b$ = -2.16 to 2.57, $c$ = -2.18 to 2.67, $d$ = -0.88 to 3.29, $e$ = -1.03 to 0.87, $f$ = -4.26 to 2.74.

These variables represent Z-scores. Each z-score can be used to calculate a probability of a Z-score at the level or greater occurring. For example, a Z-score of -1.1, has a p-value of an event occurring at that Z-score or higher of 0.8643. I want to multiply all of the p-values associated with Z-scores that satisfy this equation and then find the set of z-scores that has the maximum value for P(a)*P(b)*P(c)*P(d)*P(e)*P(f). How can this be done?

  • $\begingroup$ Is this the same as your other question? It looks very similar -- rather than posting two questions which are slightly different, we prefer people to edit existing questions to get at what they actually want to know stats.stackexchange.com/questions/228835/… $\endgroup$ – Sycorax Aug 8 '16 at 20:44

Your mention above "I would like to solve this equation for the lowest sum of a+b+c+d+e+f" pertained to your previous question at How to minimize the six z-scores in this linear equation and then calculate a probability? , but is not the correct objective function for the problem you are stating in this question.

Let $\Phi(x)$ denote the cumulative Normal distribution function.

Then your objective is to maximize $$(1-\Phi(a)) (1-\Phi(b)) (1-\Phi(c)) (1-\Phi(d)) (1-\Phi(e)) (1-\Phi(f))$$ Using the fact that $1-\Phi(x) = \Phi(-x)$ and taking the logarithm (which does not affect the optimal values of a,b,c,d,e,f), we arrive at the objective function $$log(\Phi(-a))+log(\Phi(-b))+log(\Phi(-c))+log(\Phi(-d))+log(\Phi(-e))+log(\Phi(-f))$$ which is to be maximized subject to the linear equality constraint on a,b,c,d,e,f and the lower and upper bound constraints on a,b,c,d,e,f.

For convenience, I chose to use CVX http://cvxr.com/cvx/, which uses an approximation log_normcdf(x) (so not exact) to $log(\Phi(x)$. Using other optimization software could eliminate this inaccuracy. CVX provides the globally optimal solution (subject to the use of approximation via log_normcdf(x)).

variables a b c d e f
% constraints follow
-1.42 <= a <= 2.69
-2.16 <= b <= 2.57
-2.18 <= c <= 2.67
-0.88 <= d <= 3.29
-1.03 <= e <= 0.87
-4.26 <= f <= 2.74
0 == 2.01106 - 0.00274*(34.647+24.24*a)-0.02059*(45.647+21.122*b)+1.37984*(2.05-0.206*c)-0.01176*(10.588+11.963*d)+0.00394*(118.29-21.097*e)-0.03552*(92.17+2.855*f)

The optimal [a,b,c,d,e,f] = [ -0.8615, 1.8084, 1.0194, -0.2159, -0.7163, -0.5573]. The approximation in log_normcdf(x) has manifested itself in a discrepancy in the CVX objective function of -6.5780 (actual value) vs. -6.4742 (as calculated using the approximation during the optimization).

Exponentiating back to the original objective function (i.e., before taking log), provides the objective value 1.3906e-03.

By contrast, if we use the values of a,b,c,d,e,f instead optimized to minimize a+b+c+d+e+f subject to the constraints, the objective value per the objective function which optimally produced 1.3906e-03, produces the value 1.3599e-04, which is of course a markedly inferior solution, as should not be a surprise given that it was based on optimizing a different objective.

I leave it as an exercise to the OP to redo the optimization using "exact" values of $\Phi(x)$ or $log(\Phi(x))$.

EDIT: I have now performed the optimization using the "exact" value of the Normal cdf (but could not use CVX to do this). I obtained a slight improvement in the optimal objective value from 1.3906e-03 using the log_normcdf approximation in CVX, to 1.4036e-03 based on using the exact value of the Normal cdf. Corresponding optimal [a,b,c,d,e,f] = [ -0.8676, 1.8278, 0.9082, -0.1279, -0.6724, -0.4829], which is essentially a refinement of the solution I obtained with CVX based on the log_normcdf approximation.

  • $\begingroup$ Looks great! The "a" lower threshold should be negative, but that's probably what you're fixing $\endgroup$ – Evan Aug 8 '16 at 21:29
  • $\begingroup$ Yes, indeed I had originally done the optimization having lost the minus sign on the lower bound of a when copying and pasting.. That has now been corrected, and reflected in the answer. $\endgroup$ – Mark L. Stone Aug 8 '16 at 21:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.