Your categorical variable has two levels, so there is no actual difference between dummy-coding vs. simply entering the variable into the analysis. That is, to dummy code you would create one new variable with two values... but your original variable is already one variable with two values. Dummy-coding is important for variables with more than two possible values.

I don't think that the values matter 1/2 is a linear transformation of 0/1 so it won't change any correlations.  But you should use a pair of contiguous integers (e.g., use 0/1, 1/2, etc.; do not use, say, M=1/F=10) so the regression coefficient will tell you the mean difference in the predicted value across sexes.

The use of a dummy variable in this case for two categories (M/F) will allow you to kick in an effect for the category coded (1) relative the category coded (0). It does not matter what one is 1 and 0 so long as you remember how the direction is built. For example if you have and dummy variable (X2)....

y = a + b1 X1 + b2 X2

where X2 = 0 if male, and 1 if female, then clearly the regression for men is

y = a + b1 X1

and for women

y = a + b1 X1 + b2

so, "b2" here is going to give you some info regarding women as compared to men.

If your data are coded 1/2 or some other scheme, you can reclassify -- it is important that ONE of the two classes be valued zero/

One advantage to dummy coding is that the intercept is more easily interpreted. The intercept is the expected value of the outcome when all the independent variables are equal to zero. If you use a dummy code (0 = male 1 = female) the intercept reflects the expected value for male respondents (with zero's on all the other variables) and the coefficient reflects how much larger (or smaller) the expected value will be for women. This is one reason why centering or standardizing is so helpful, it makes the intercept very useful. 

As several have written it is useful to have the coding so that one of the dummies is zero - but other codings are possible and sometimes useful.

This site gives general background  on additional coding systems for categorical variables in regression analysis

 http://www.ats.ucla.edu/stat/sas/webbooks/reg/chapter5/sasreg5.htm

and here dummy coding is contrasted with effect coding

http://www.ats.ucla.edu/stat/mult_pkg/faq/general/dummy.htm

http://www.ats.ucla.edu/stat/mult_pkg/faq/general/effect.htm

With effect coding (and balanced data) the constant is equal to the grand mean of all of the observations. 

You need to know what coding is being used when the strings Male and Female are being turned into numeric values by the software you are using. For example in doing an ANOVA type analysis the effect  coding is often used. 

There are even times when you want to keep BOTH 0/1 and 1/0 dummies in the model but have a model with no estimated intercept (thereby avoiding exact multicollinearity) - this gives the mean for each sex separately and not as the usual difference in the mean - see my answers here

https://www.researchgate.net/post/Why_does_SPSS_exclude_certain_variables_from_a_regression

It's better to use 0-1 encoding because, in this way, the coefficient related to the dummy tells you how much your dependent variable changes (on average) when you have male (if 1 is male and 0 is female) in comparison with females (and is true the contrary if 1 is female and 0 is male) If you use some other values (like 1-2 encoding, 1-20, 20-30 and so on) the interpretation of the coefficent is not so easy.

Its depend upon  reference  group, its better to code 0 -female and 1- male to have better interpretation of coefficient. 

All great advice. I suggest two points. First, a dummy variable coded 0-1 assesses the relationship between the dependent variable and the absence vs. presence of a characteristic, whereas a dummy variable coded 1-2 assesses the relationship between the dependent variable and the presence (dummy=1) vs. "doubled presence" (dummy=2) of a characteristic. To extend the answer by O'Kelly (above) if you have sex coded 1=male and 2=female in the equation

y = [coefficient for sex]*[value for sex variable] + [intercept]

then for males the equation becomes

y = [coefficient for sex]*1 + [intercept]

while for females it becomes

y = [coefficient for sex]*2 + [intercept]

So now instead of modeling the relationship between the dependent variable and being female, you now have modeled the relationship between the dependent variable and being "double-male". The coefficient won't change, as they still have the same distribution about their respective mean, but the intercept will have a different value, because the intercept is calculated from the other variables in the model. I highly recommending reading the conversation suggested by Jones (above) on why SPSS automatically excludes some variables from a regression.

The intercept is all too frequently overlooked when comparing unstandardized regression coefficients, often preferred over standardized coefficients because their meaning is much more easily translated into into arguments about real-world relationships. In other words, an unstandardized coefficient of 2,000 means that a one-unit increase in your independent variable is associated with a 2,000-unit increase in your dependent variable, controlling for the other variables in the model. For example, If you are regressing income with gender (coded 0=m,1=f) and your gender coefficient is 10,000 and your intercept is 50,000 then females make $60,000 while males only make $50,000, meaning that females make about 20% more than males. With this same data set, coding gender as (1=m,2=f) means that females make $70,000 while males make $60,000, meaning that females make only about 16.7% more than males. This means that, dependent upon how you code gender (or any other dummy variable for that matter,) there could be a considerable difference in the relationship between your dependent variable and your independent variable(s).

Second, dependent upon your interest in the influence of sex, I would suggest running separate models for females only and then for males only. In many cases, females are treated by the world much differently than males, resulting in substantially different coefficients across a number of experiences. Using sex only as a control variable will mask much of those different experiences, but comparing the coefficients of estimating models separately for females only and then males only will reveal much more about how sex influences your dependent variable.

Is it advisable to use a dummy variable for sex (male, female) in my regression analysis? - ResearchGate. Available from: https://www.researchgate.net/post/Is_it_advisable_to_use_a_dummy_variable_for_sex_male_female_in_my_regression_analysis [accessed Oct 4, 2016].

I think 1 & 2 coding for females and males is ok. I dont find any extra benefit for using gender as dummy variable in the same analysis.

Dummy variable represent to your creative thinking in one domain to other limitation elements. it's determined to your scale items  and research rationals. 

The problem with a stratified analysis - ie separate analysis for men and women - is that you lose statistical power and the SEs of the coefficients in the two models  are not comparable.  It depends on how many predictors I have and how big is the data set but I would generally use the 0/1 binary coding and then use interactions with this dummy with other predictors to test for differential relations. I agree that men/women divide can operate through other predictors.

see this introductory chapter

https://www.researchgate.net/publication/256801601_An_introduction_to_statistical_modelling 

I am sorry but I do not get the point of 'doubling the presence' - but do agree that there it should not be 'one size fits all' . As I said in an earlier post I often use two dummies and no intercept and this formulation allows me to examine differential heterogeneity. See the section headed 

Complex between-individual heterogeneity in this paper

https://www.researchgate.net/publication/23538539_Modelling_complexity_Analysing_between-individual_and_between-place_variation_-_A_multilevel_tutorial

Article Modelling Complexity: Analysing Between-Individual and Betwe...

Chapter An introduction to statistical modelling

Yes! Sex is dichotomous but you need to be careful and consistent with determining which gender takes the value of 1 or 0.

Easier to code these as 0/1 than 1/2........

Set up a dummy variable as 0,1.

In fact, gender is not dichotomous. How do you code male, female, diverse etc

@katharina Kühne - if I understand your query, having more categories would simple require further dummy coding. For example, male/female/non-binary... etc. Would be 0/1/2... Etc. Where '0' is your 'reference group', which every other gender would be compared to in your output data.

0,1,2...does the job well.

I have found that many articles in top journals have gender coding as 1/2. Why is this?