do not regress if any Y is zero

Hi, I am using regress to perform linear regression on a large nc file (containing lat, lon, time, and temperature). So for each lat, lon, temperature is regressed over 30 years (one measurement each year).

In some years, some temperatures are 0 K, in those cases I want the loop to not regress, so the final mean slope does not include these.

Here is the important part of my code:

I am not getting an error but also not getting any difference in whether i include y(y==0) = NaN;, so I think there may be a better way. I had understood that regress ignores these autmatically, but that also appears to not be the case.

% initialise time parameters
time_begin  = [1981, 1, 1, 0,0,0];
time_end    = [2010,12,31,23,0,0];
years       = (time_begin(1):time_end(1))';
nyears      = length(years);
% create storage and regress
TXx = randi(100, 288, 192, 30);
lat = rand(192, 1);
lon = rand(288, 1);
time = rand(30,1);
M = numel(lon);
N = numel(lat);
slope = zeros(M, N);
intercept = zeros(M, N);
T = numel(time);
x = ([ones(T, 1) years]);
% Regress each lat/lon location
for i = 1 : M
    for j = 1 : N
        % Get all time instances of each lat/lon location
        y = squeeze(TXx(i, j, :));
        y(y==0) = NaN;
        % Create regression problem and solve
        c = regress(y, x);
        intercept(i, j) = c(1);
        slope(i, j) = c(2);      
    end
end

Answer 1

[Matt J]

How about

c = regress(y(y~=0), x(y~=0));

[aine gormley]

Hi, thanks for your comment. I think then though then matlab just excludes data if it is zero and regresses the rest. But is there a way to exclude an entire timeline if one of the timesteps is zero?

For example, at a certain lat/lon, in 1999 - the temperature is 0, is there a way to exclude that entire lat/lon from the regression?

[aine gormley]

Just wondering perhaps your solution works - but now i get the error:

Index exceeds the number of array elements (1).

Error in ensmean2d (line 43)

slope(i, j) = c(2);

but i am unsure if i can change that?

[Matt J]

You need to decide how you want things handled when c is a scalar. I assume this happens when only one of the y's is non-zero.

[aine gormley]

can a slope (i.e. the slopes within which some y's are zero) be replaced by NaN at that lat/lon?

because i then plot the slopes on the grid so if I compress the slopes/have the non-zeros a removed it wont plot.

[Matt J]

If only one y is non-zero there is no way to derive a slope. You cannot fit a line when there is fewer than 2 data points.

[aine gormley]

Hi, sorry - I meant if only 1 y is zero (so the other 29 are not zero), do you know if there a way to exclude this regression in such cases?

[Matt J]

So, tell me again why can't you set the results to NaN?