Transforming a variable when original variable does not have explantory power Sometimes in multivariate linear regression, there will be one explanatory variable that does not contribute much in way of explanatory power. Then, we will perform a tranform on that variable, i.e take the log, scale it by some factor etc. and suddenly this variable now has explanatory power and is deemed fit to be included in the model.
Is this a good practice? i.e Simply transform a variable to make it look like it has explanatory power? Does experience in the domain guide what needs tobe done with the variable...i.e exclude it or include it after a transformation?
 A: This depends on if you are a purist, a pragmatist, or a data dredger.  
The purist will only ever try one transformation, and that is based only on the science and previous knowledge and will be chosen before any data is collected (or looked at).  This is the safest approach, but can also loose out on gaining insight from the data.
The data dredger will try every possible transformation until they find one that gives them the answer that they want.  This can result in impressive results, but also ends in over fitting models and unreproducible results.  
The pragmatist (the largest and most productive group) recognizes that there are things that we can learn from the data, but also is wary of over fitting the data.  They will look at a few possible transforms that are justified by the science, or try splines or other smooths with a reasonable smoothness constraint.  They will also do things like cross-validation to make sure that their experimenting has not resulted in over fitting and reign things back if it has.  They will then also be honest about what they did (explaining which transforms were experimented with in the discussion) and look for validation in a follow-up study.
I recommend the pragmatic approach.
A: It's a legitimate practice you are describing. As a matter of fact, having all variables linear is a simplification out of necessity. Usually, we don't know what is a true relationship between variables, so we model them linearly.
If you happen to know the relationships, often non-linear, then definitely go for it. 
Here's an example, where the true process is $y_t=sin(\beta t)+\varepsilon_t$. First I model it with $y_t=\beta' t+\varepsilon_t$, then model with a proper specification. In the first case the coefficient comes not significant, in the second case it's significant.
MATLAB Code:
rng(0);
x = (1:100)'/100*pi;
% data generating process
y = sin(x) + randn(100,1)/2;

% fit linear x
fit = fitlm(x,y)

% fit sin(x)
fit2 = fitlm(sin(x),y)

plot([y fit.Fitted fit2.Fitted])
legend({'actual' ,'fitted x','fitted sin x'})

OUTPUT:
fit = 


Linear regression model:
    y ~ 1 + x1

Estimated Coefficients:
                   Estimate       SE        tStat       pValue  
                   ________    ________    _______    __________

    (Intercept)     0.89621     0.12771     7.0177    2.9539e-10
    x1             -0.12487    0.069885    -1.7868      0.077068


Number of observations: 100, Error degrees of freedom: 98
Root Mean Squared Error: 0.634
R-squared: 0.0315,  Adjusted R-Squared 0.0217
F-statistic vs. constant model: 3.19, p-value = 0.0771

fit2 = 


Linear regression model:
    y ~ 1 + x1

Estimated Coefficients:
                   Estimate      SE       tStat       pValue  
                   ________    _______    ______    __________

    (Intercept)    0.13799     0.13389    1.0306       0.30527
    x1             0.87991     0.18936    4.6468    1.0507e-05


Number of observations: 100, Error degrees of freedom: 98
Root Mean Squared Error: 0.583
R-squared: 0.181,  Adjusted R-Squared 0.172
F-statistic vs. constant model: 21.6, p-value = 1.05e-05






A: This is absolutely good practice! If the true relationship between your response and the explanatory variables is logarithmic then your model ought to reflect this. Of course, we never know for sure the nature of the true relationship so we do things like plot the data or fit the model in different ways (with and without a log transformation) and assess fit.
